Base_JSON/prefixI/json/iwcs/2021.iwcs-1.6.json

{
    "paper_id": "2021",
    "header": {
        "generated_with": "S2ORC 1.0.0",
        "date_generated": "2023-01-19T07:24:33.547787Z"
    },
    "title": "Predicate Representations and Polysemy in VerbNet Semantic Parsing",
    "authors": [
        {
            "first": "James",
            "middle": [],
            "last": "Gung",
            "suffix": "",
            "affiliation": {},
            "email": "james.gung@colorado.edu"
        },
        {
            "first": "Martha",
            "middle": [],
            "last": "Palmer",
            "suffix": "",
            "affiliation": {},
            "email": "martha.palmer@colorado.edu"
        }
    ],
    "year": "",
    "venue": null,
    "identifiers": {},
    "abstract": "Despite recent advances in semantic role labeling propelled by pre-trained text encoders like BERT, performance lags behind when applied to predicates observed infrequently during training or to sentences in new domains. In this work, we investigate how semantic role labeling performance on low-frequency predicates and out-of-domain data can be improved by using VerbNet, a verb lexicon that groups verbs into hierarchical classes based on shared syntactic and semantic behavior and defines semantic representations describing relations between arguments. We find that Verb-Net classes provide an effective level of abstraction, improving generalization on lowfrequency predicates by allowing them to learn from the training examples of other predicates belonging to the same class. We also find that joint training of VerbNet role labeling and predicate disambiguation of VerbNet classes for polysemous verbs leads to improvements in both tasks, naturally supporting the extraction of VerbNet's semantic representations.",
    "pdf_parse": {
        "paper_id": "2021",
        "_pdf_hash": "",
        "abstract": [
            {
                "text": "Despite recent advances in semantic role labeling propelled by pre-trained text encoders like BERT, performance lags behind when applied to predicates observed infrequently during training or to sentences in new domains. In this work, we investigate how semantic role labeling performance on low-frequency predicates and out-of-domain data can be improved by using VerbNet, a verb lexicon that groups verbs into hierarchical classes based on shared syntactic and semantic behavior and defines semantic representations describing relations between arguments. We find that Verb-Net classes provide an effective level of abstraction, improving generalization on lowfrequency predicates by allowing them to learn from the training examples of other predicates belonging to the same class. We also find that joint training of VerbNet role labeling and predicate disambiguation of VerbNet classes for polysemous verbs leads to improvements in both tasks, naturally supporting the extraction of VerbNet's semantic representations.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Abstract",
                "sec_num": null
            }
        ],
        "body_text": [
            {
                "text": "Semantic role labeling (SRL) is a form of shallow semantic parsing that involves the extraction of predicate arguments and their assignment to consistent roles with respect to the predicate, facilitating the labeling of e.g. who did what to whom (Gildea and Jurafsky, 2000) . SRL systems have been broadly applied to applications such as question answering (Berant et al., 2014; Wang et al., 2015) , machine translation (Liu and Gildea, 2010; Bazrafshan and Gildea, 2013) , dialog systems (Tur and Hakkani-T\u00fcr, 2005; Chen et al., 2013) , metaphor detection (Stowe et al., 2019) , and clinical information extraction (Gung, 2013; MacAvaney et al., 2017) . Recent approaches to SRL have achieved * Work done prior to joining Amazon. large gains in performance through the use of pretrained text encoders like ELMo and BERT (Peters et al., 2018; Devlin et al., 2019) . Despite these advances, performance on low-frequency predicates and out-of-domain data remains low relative to in-domain performance on higher frequency predicates. The assignment of role labels to a predicate's arguments is dependent upon the predicate's sense. PropBank divides each predicate into one or more rolesets, which are coarsegrained sense distinctions that each provide a set of core numbered arguments (A0-A5) and their corresponding definitions. VerbNet (VN) groups verbs into hierarchical classes, each class defining a set of valid syntactic frames that define a direct correspondence between thematic roles and syntactic realizations, e.g. Agent REL Patient (e.g. John broke the vase) or Patient REL (e.g. The vase broke) for break-45.1 (Schuler, 2005) .",
                "cite_spans": [
                    {
                        "start": 246,
                        "end": 273,
                        "text": "(Gildea and Jurafsky, 2000)",
                        "ref_id": "BIBREF20"
                    },
                    {
                        "start": 357,
                        "end": 378,
                        "text": "(Berant et al., 2014;",
                        "ref_id": "BIBREF3"
                    },
                    {
                        "start": 379,
                        "end": 397,
                        "text": "Wang et al., 2015)",
                        "ref_id": "BIBREF50"
                    },
                    {
                        "start": 420,
                        "end": 442,
                        "text": "(Liu and Gildea, 2010;",
                        "ref_id": "BIBREF32"
                    },
                    {
                        "start": 443,
                        "end": 471,
                        "text": "Bazrafshan and Gildea, 2013)",
                        "ref_id": "BIBREF2"
                    },
                    {
                        "start": 489,
                        "end": 516,
                        "text": "(Tur and Hakkani-T\u00fcr, 2005;",
                        "ref_id": "BIBREF48"
                    },
                    {
                        "start": 517,
                        "end": 535,
                        "text": "Chen et al., 2013)",
                        "ref_id": "BIBREF9"
                    },
                    {
                        "start": 557,
                        "end": 577,
                        "text": "(Stowe et al., 2019)",
                        "ref_id": "BIBREF45"
                    },
                    {
                        "start": 616,
                        "end": 628,
                        "text": "(Gung, 2013;",
                        "ref_id": "BIBREF21"
                    },
                    {
                        "start": 629,
                        "end": 652,
                        "text": "MacAvaney et al., 2017)",
                        "ref_id": "BIBREF33"
                    },
                    {
                        "start": 821,
                        "end": 842,
                        "text": "(Peters et al., 2018;",
                        "ref_id": "BIBREF41"
                    },
                    {
                        "start": 843,
                        "end": 863,
                        "text": "Devlin et al., 2019)",
                        "ref_id": "BIBREF14"
                    },
                    {
                        "start": 1621,
                        "end": 1636,
                        "text": "(Schuler, 2005)",
                        "ref_id": "BIBREF43"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "Recent PropBank (PB) semantic role labeling models have largely eschewed explicit predicate disambiguation in favor of direct prediction of semantic roles in end-to-end trainable models (Zhou and Xu, 2015; He et al., 2017; Shi and Lin, 2019) . This is possible for several reasons: First, Prop-Bank's core roles and modifiers are shared across all predicates, allowing a single classifier to be trained over tokens or spans. Second, although definitions of PB roles are specific to the different senses of each predicate, efforts are made when creating rolesets to ensure that A0 and A1 exhibit properties of Dowty's prototypical Agent and prototypical Patient respectively (1991) . Finally, PB rolesets are defined based on VN class membership, with predicates in the same classes thus being assigned relatively consistent role definitions (Bonial et al., 2010) .",
                "cite_spans": [
                    {
                        "start": 186,
                        "end": 205,
                        "text": "(Zhou and Xu, 2015;",
                        "ref_id": "BIBREF57"
                    },
                    {
                        "start": 206,
                        "end": 222,
                        "text": "He et al., 2017;",
                        "ref_id": "BIBREF24"
                    },
                    {
                        "start": 223,
                        "end": 241,
                        "text": "Shi and Lin, 2019)",
                        "ref_id": "BIBREF44"
                    },
                    {
                        "start": 609,
                        "end": 680,
                        "text": "Dowty's prototypical Agent and prototypical Patient respectively (1991)",
                        "ref_id": null
                    },
                    {
                        "start": 841,
                        "end": 862,
                        "text": "(Bonial et al., 2010)",
                        "ref_id": "BIBREF4"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "Unlike PropBank, VerbNet's thematic roles are shared across predicates and classes with consistent definitions. However, VN roles are more dependent on the identity of the predicate (Zapirain et al., 2008; Merlo and Van Der Plas, 2009) . Examples of PropBank and VerbNet roles illustrating this are given in Table 1 . Consequently, VN role labeling models may benefit more from predicate features than PropBank. Furthermore, while it is possible to identify PB or VN roles without classifying predicate senses, linking the resulting roles to their definitions or to the syntactic frames and associated semantic primitives in VN does require explicit predicate disambiguation (Brown et al., 2019) . Therefore, predicate disambiguation is often an essential step when applying SRL systems to real-world problems.",
                "cite_spans": [
                    {
                        "start": 182,
                        "end": 205,
                        "text": "(Zapirain et al., 2008;",
                        "ref_id": "BIBREF56"
                    },
                    {
                        "start": 206,
                        "end": 235,
                        "text": "Merlo and Van Der Plas, 2009)",
                        "ref_id": "BIBREF34"
                    },
                    {
                        "start": 675,
                        "end": 695,
                        "text": "(Brown et al., 2019)",
                        "ref_id": "BIBREF5"
                    }
                ],
                "ref_spans": [
                    {
                        "start": 308,
                        "end": 315,
                        "text": "Table 1",
                        "ref_id": "TABREF1"
                    }
                ],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "In this work, we evaluate alternative approaches for incorporating VerbNet classes in English Verb-Net and PropBank role labeling. We propose a joint model for SRL and VN predicate disambiguation (VN classification), finding that joint training leads to improvements in VN classification and role labeling for out-of-domain predicates. We also evaluate VN classes as predicate-specific features. Using gold classes, we observe significant improvements in both PB and VN SRL. We also observe improvements in VN role labeling when using predicted classes and features that incorporate all valid classes for each predicate 1 .",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "VerbNet VerbNet is a broad-coverage lexicon that groups verbs into hierarchical classes based on shared syntactic and semantic behavior (Schuler, 2005) . Each VN class is assigned a set of thematic roles that, unlike PB numbered arguments, maintain consistent meanings across different verbs and classes. VN classes provide an enumeration of syntactic frames applicable to each member verb, describing how the thematic roles of a VN class may be realized in a sentence. Every syntactic frame entails a set of low-level semantic representations (primitives) that describe relations between thematic role arguments as well as changes throughout the course of the event (Brown et al., 2018) . The close relationship between syntactic realizations and semantic representations facilitates straightforward extraction of VN semantic predicates given identification of a VN class and corresponding thematic roles. VN primitives have been applied to problems such as machine comprehension and question generation (Dhole and Manning, 2020) .",
                "cite_spans": [
                    {
                        "start": 136,
                        "end": 151,
                        "text": "(Schuler, 2005)",
                        "ref_id": "BIBREF43"
                    },
                    {
                        "start": 667,
                        "end": 687,
                        "text": "(Brown et al., 2018)",
                        "ref_id": "BIBREF6"
                    },
                    {
                        "start": 1005,
                        "end": 1030,
                        "text": "(Dhole and Manning, 2020)",
                        "ref_id": "BIBREF15"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Background and Related Work",
                "sec_num": "2"
            },
            {
                "text": "Comparing VerbNet with PropBank Yi et al. (2007) use VN role groupings to improve label consistency across verbs by reducing the overloading of PropBank's numbered arguments like A2. Comparing SRL models trained on PB and VN, Zapirain et al. (2008) find that their VerbNet model performs worse on infrequent predicates than their PB model, and suggest that VN is more reliant on the identity of the predicate than PB based on experiments removing predicate-specific features from their models. They suggest that the high consistency of A0 and A1 enables PB to generalize better without relying on predicate-specific information. Merlo and Van Der Plas (2009) provide an information-theoretic perspective on the comparison of PropBank and VerbNet, demonstrating how the identity of the predicate is more important to VN SRL than for PB by comparing the conditional entropy of roles given verbs as well as the mutual information of roles and verbs. In multilingual BERT probing studies comparing several SRL formalisms, Kuznetsov and Gurevych (2020) find that layer utilization for predicates differs between PB and VN. PB emphasizes the same layers used for syntactic tasks, while VN uses layers associated with tasks used more prevalently in lexical tasks. These findings reinforce the importance of predicate representations to VerbNet.",
                "cite_spans": [
                    {
                        "start": 42,
                        "end": 48,
                        "text": "(2007)",
                        "ref_id": null
                    },
                    {
                        "start": 222,
                        "end": 248,
                        "text": "VN, Zapirain et al. (2008)",
                        "ref_id": null
                    },
                    {
                        "start": 629,
                        "end": 658,
                        "text": "Merlo and Van Der Plas (2009)",
                        "ref_id": "BIBREF34"
                    },
                    {
                        "start": 1018,
                        "end": 1047,
                        "text": "Kuznetsov and Gurevych (2020)",
                        "ref_id": "BIBREF30"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Background and Related Work",
                "sec_num": "2"
            },
            {
                "text": "Previous work has investigated the interplay between predicate sense disambiguation and SRL. Dang and Palmer (2005) improve verb sense disambiguation (VSD) using features based on semantic role labels. Moreda and Palomar (2006) find that explicit verb senses improve PB SRL for verb-specific roles like A2 and A3, but hurt on adjuncts. Yi (2007) find that using gold standard PB roleset IDs as features in an SRL model improves performance only on highly polysemous verbs. Dahlmeier et al. (2009) propose a joint probabilistic model for preposition disambiguation and SRL, finding an improvement over independent models.",
                "cite_spans": [
                    {
                        "start": 93,
                        "end": 115,
                        "text": "Dang and Palmer (2005)",
                        "ref_id": "BIBREF12"
                    },
                    {
                        "start": 202,
                        "end": 227,
                        "text": "Moreda and Palomar (2006)",
                        "ref_id": "BIBREF35"
                    },
                    {
                        "start": 473,
                        "end": 496,
                        "text": "Dahlmeier et al. (2009)",
                        "ref_id": "BIBREF11"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "SRL and Predicate Disambiguation",
                "sec_num": null
            },
            {
                "text": "Predicate disambiguation plays a critical role in FrameNet (Baker et al., 1998) parsing, in part because FrameNet's role inventory is more than an order of magnitude larger than that of PB and VN. This richer, more granular role inventory lends advantages to approaches that constrain role identification to the set of valid roles for the predicted frame Hermann et al., 2014) , or that jointly encode argument and role representations given identified frames (FitzGerald et al., 2015) .",
                "cite_spans": [
                    {
                        "start": 59,
                        "end": 79,
                        "text": "(Baker et al., 1998)",
                        "ref_id": "BIBREF1"
                    },
                    {
                        "start": 355,
                        "end": 376,
                        "text": "Hermann et al., 2014)",
                        "ref_id": "BIBREF25"
                    },
                    {
                        "start": 460,
                        "end": 485,
                        "text": "(FitzGerald et al., 2015)",
                        "ref_id": "BIBREF18"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "SRL and Predicate Disambiguation",
                "sec_num": null
            },
            {
                "text": "LM Pre-training and SRL Language model (LM) pre-training has become ubiquitous in natural language processing tasks, with LM encoders like ELMo propelling forward the state of the art in SRL (Peters et al., 2018) . We are interested in whether a strong baseline model using a LM encoder such as BERT can be further improved by incorporating external knowledge from lexical resources like VN.",
                "cite_spans": [
                    {
                        "start": 191,
                        "end": 212,
                        "text": "(Peters et al., 2018)",
                        "ref_id": "BIBREF41"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "SRL and Predicate Disambiguation",
                "sec_num": null
            },
            {
                "text": "BERT (Devlin et al., 2019 ) is a Transformer encoder (Vaswani et al., 2017) jointly trained using two objectives: a masked language modeling objective to predict the identity of randomly-masked tokens in the input, as well as a next sentence prediction task (NSP) intended to encourage the model to encode the relationship between sentence pairs (henceforth referred to as Sent. A and Sent. B). Sentences are tokenized using WordPiece (Wu et al., 2016) . As a Transformer encoder, BERT applies multiple layers of a multi-headed self-attention mechanism to progressively build contextual tokenlevel representations. In our experiments, we use encodings from the final layer.",
                "cite_spans": [
                    {
                        "start": 5,
                        "end": 25,
                        "text": "(Devlin et al., 2019",
                        "ref_id": "BIBREF14"
                    },
                    {
                        "start": 53,
                        "end": 75,
                        "text": "(Vaswani et al., 2017)",
                        "ref_id": "BIBREF49"
                    },
                    {
                        "start": 435,
                        "end": 452,
                        "text": "(Wu et al., 2016)",
                        "ref_id": "BIBREF52"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "SRL and Predicate Disambiguation",
                "sec_num": null
            },
            {
                "text": "Our baseline SRL model closely follows Shi and Lin (2019) . We thus approach SRL as a sequence tagging task, predicting per-word, IOB-encoded (In, Out, Begin) role labels independently for each predicate in a sentence. A predicate-aware encod-ing of a sentence is produced using the target predicate as the Sent. B input to BERT. For example, the sentence I tried opening it is processed as:",
                "cite_spans": [
                    {
                        "start": 39,
                        "end": 57,
                        "text": "Shi and Lin (2019)",
                        "ref_id": "BIBREF44"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Semantic Role Labeling with BERT",
                "sec_num": "3"
            },
            {
                "text": "CLS I tried opening it SEP opening SEP",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Semantic Role Labeling with BERT",
                "sec_num": "3"
            },
            {
                "text": "for the verb open. This enables BERT to incorporate the identity of the predicate in the encoding of each word while clearly delineating it from tokens in the original sentence.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Semantic Role Labeling with BERT",
                "sec_num": "3"
            },
            {
                "text": "To simplify notation, we'll treat LM(a, b) \u2208 R Ta\u00d7D LM as shorthand for the final layer BERT encoding for a pair of sentences a = w 1 , ..., w Ta and b = w 1 , ..., w T b with T a and T b words respectively, where D LM gives BERT's hidden size. This is produced by applying WordPiece tokenization (WP) to each word in each sentence and concatenating the resulting sequences of token IDs with standard BERT-specific IDs:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Semantic Role Labeling with BERT",
                "sec_num": "3"
            },
            {
                "text": "w = CLS, WP(a), SEP, WP(b), SEP",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Semantic Role Labeling with BERT",
                "sec_num": "3"
            },
            {
                "text": "The resulting sequence of tokens w is encoded using BERT. We use the final layer outputs, taking vectors only for the first WordPiece token for each original word in Sent. A (a), filtering out vectors corresponding to Sent. B (b), SEP or CLS. The resulting matrix consists of a vector per word in Sent. A, avoiding any discrepancies between IOBencoded word-level output labels and WordPiece tokens used as inputs.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Semantic Role Labeling with BERT",
                "sec_num": "3"
            },
            {
                "text": "Following previous work (Zhou and Xu, 2015; He et al., 2017) , we use a marker feature as an indicator for the specific location of the predicate within the sentence. For a sentence, w 1 , ..., w T , with a predicate given by index p \u2208 1...T , we compute a predicate-aware, contextualized embedding x pt of each word as",
                "cite_spans": [
                    {
                        "start": 24,
                        "end": 43,
                        "text": "(Zhou and Xu, 2015;",
                        "ref_id": "BIBREF57"
                    },
                    {
                        "start": 44,
                        "end": 60,
                        "text": "He et al., 2017)",
                        "ref_id": "BIBREF24"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Semantic Role Labeling with BERT",
                "sec_num": "3"
            },
            {
                "text": "x pt = LM(w 1...T , w p ) (t) ; W (mark) (t=p) (1) with W (mark) \u2208 R 2\u00d7D mark and x pt \u2208 R D LM +D mark ,",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Semantic Role Labeling with BERT",
                "sec_num": "3"
            },
            {
                "text": "where D mark provides the size of the predicate marker embedding.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Semantic Role Labeling with BERT",
                "sec_num": "3"
            },
            {
                "text": "The predicate's positional information from the marker is integrated using a bidirectional LSTM (Hochreiter and Schmidhuber, 1997) , concatenating the hidden states for the forward and backward LSTMs at each timestep (omitting the p from x pt for brevity):",
                "cite_spans": [
                    {
                        "start": 96,
                        "end": 130,
                        "text": "(Hochreiter and Schmidhuber, 1997)",
                        "ref_id": "BIBREF26"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Semantic Role Labeling with BERT",
                "sec_num": "3"
            },
            {
                "text": "h (f w) t = LSTM (f w) (x 1...T ) (t) h (bw) t = LSTM (bw) (x T...1 ) (T \u2212t) h (f b) t = h (f w) t ; h (bw) t (2)",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Semantic Role Labeling with BERT",
                "sec_num": "3"
            },
            {
                "text": "The BiLSTM output at each timestep t is concatenated with that of the predicate's timestep and passed through a sequentially-applied linear transformation followed by a leaky ReLu (\u03b1 = 0.1):",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Semantic Role Labeling with BERT",
                "sec_num": "3"
            },
            {
                "text": "x (mlp) pt = \u03c3 W (mlp) h (f b) t ; h (f b) p + b (mlp) (3)",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Semantic Role Labeling with BERT",
                "sec_num": "3"
            },
            {
                "text": "We apply a final linear projection from x (mlp) pt to IOB-encoded role labels:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Semantic Role Labeling with BERT",
                "sec_num": "3"
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "s (srl) pt = W (srl) x (mlp) pt + b (srl)",
                        "eq_num": "(4)"
                    }
                ],
                "section": "Semantic Role Labeling with BERT",
                "sec_num": "3"
            },
            {
                "text": "where s",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Semantic Role Labeling with BERT",
                "sec_num": "3"
            },
            {
                "text": "(srl) pt \u2208 R K provides the unnormalized scores for each of K possible role labels, with the probability of predicting a label for a given token t and predicate p given by:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Semantic Role Labeling with BERT",
                "sec_num": "3"
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "P (y (srl) pt |w 1...T , w p ) = softmax(s (srl) pt )",
                        "eq_num": "(5)"
                    }
                ],
                "section": "Semantic Role Labeling with BERT",
                "sec_num": "3"
            },
            {
                "text": "Like He et al. 2017, we apply constrained Viterbi decoding to restrict inferred label sequences to produce valid IOB sequences.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Semantic Role Labeling with BERT",
                "sec_num": "3"
            },
            {
                "text": "Verbs belonging to the same VN class share syntactic and semantic properties and the same set of thematic roles and syntactic frames. Replacing a predicate in a sentence with a different verb from the same class typically produces a syntactically coherent sentence and does not impact the proposition's thematic role labels. VN classes may thus provide an effective level of abstraction for predicates in SRL. We hypothesize that using VN classes as predicate-specific features may help reduce sparsity issues for low-frequency and out-of-vocabulary (OOV) verbs. Intuitively, training examples for each member verb within a class contribute to the estimation of parameters associated with all other members of the same class, enabling the fine-tuning of predicate-level features even for OOV predicates. For example, a verb like traipse may rarely or never occur during training, but may belong to a class which appears hundreds of times in the form of more common verbs like run or rush. We investigate whether by sharing parameter updates across VN members, we can further improve generalization on infrequent verbs.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "VerbNet Classes as Predicate Features",
                "sec_num": "4"
            },
            {
                "text": "Methodology Intuitively, BERT's NSP pretraining task encourages some level of focus on Sent. B tokens from attention heads when processing tokens in Sent. A. The predicate feature presented by Shi and Lin (2019) and applied in our baseline model uses the predicate token as the Sent. B input to BERT and thus allows the encodings of tokens in a sentence to be conditioned directly on the predicate.",
                "cite_spans": [
                    {
                        "start": 193,
                        "end": 211,
                        "text": "Shi and Lin (2019)",
                        "ref_id": "BIBREF44"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "VerbNet Classes as Predicate Features",
                "sec_num": "4"
            },
            {
                "text": "We propose to include tokens corresponding to the predicate's VN class as additional features as part of Sent. B. To realize this, we concatenate the corresponding VN class ID to Sent. B along with the predicate, updating the inputs given in Equation 1:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "VerbNet Classes as Predicate Features",
                "sec_num": "4"
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "LM(w 1...T , w p w s )",
                        "eq_num": "(6)"
                    }
                ],
                "section": "VerbNet Classes as Predicate Features",
                "sec_num": "4"
            },
            {
                "text": "where w s is a token corresponding to the VN class of the predicate w p 2 .",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "VerbNet Classes as Predicate Features",
                "sec_num": "4"
            },
            {
                "text": "VerbNet Classification VN classes can be predicted automatically using a word sense disambiguation system. We propose a simple model for VerbNet classification: fine tune a pre-trained BERT encoder by applying a feedforward multilayer perceptron (MLP) classifier over all VN classes to the BERT encoding associated with the first WordPiece of the target predicate. We again condition BERT on the target predicate by including it as a feature (w p ) in Sent. B:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "VerbNet Classes as Predicate Features",
                "sec_num": "4"
            },
            {
                "text": "x p = LM(w 1...T , w p ) (p) x (mlp) p = \u03c3 W (mlp) x p + b (mlp) s (vncls) p = W (vncls) x (mlp) p + b (vncls) (7)",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "VerbNet Classes as Predicate Features",
                "sec_num": "4"
            },
            {
                "text": "where W (vncls) \u2208 R D mlp \u00d7V projects over all V VN classes for all predicates. The probability for predicting a VN class y (vncls) p for a given predicate and sentence is given by:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "VerbNet Classes as Predicate Features",
                "sec_num": "4"
            },
            {
                "text": "P (y (vncls) p |w 1...T , w p ) = softmax(s (vncls) p ) (8)",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "VerbNet Classes as Predicate Features",
                "sec_num": "4"
            },
            {
                "text": "This single classifier formulation is possible for lexicons like VN and FrameNet in which predicates share senses from a global sense inventory. While individual predicates have a specific set of valid senses, their senses are shared from the global lexicon. Kawahara and Palmer (2014) demonstrate that a single classifier approach to VN classification achieves competitive performance when using shared semantic features. Intuitively, by training the classifier across multiple verbs, the model parameters specific to each sense receive more updates, with infrequent verb-class pairs also benefiting from the examples of other verbs within the same class. At inference time, we constrain sense predictions to predicate-sense combinations observed in the training data, selecting the highestscoring valid sense given the predicate. We evaluate models using both predicted and gold (ground truth) classes for w s as PREDICTED CLASS and GOLD CLASS respectively.",
                "cite_spans": [
                    {
                        "start": 259,
                        "end": 285,
                        "text": "Kawahara and Palmer (2014)",
                        "ref_id": "BIBREF28"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "VerbNet Classes as Predicate Features",
                "sec_num": "4"
            },
            {
                "text": "Like SRL, VerbNet classification accuracy declines in the long tail of low frequency senses and predicates. For this reason, incorrect sense predictions may negate the benefits of VN class features on precisely the instances for which they might be expected to be beneficial: OOV or rare predicates.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "VerbNet Classes without Disambiguation",
                "sec_num": null
            },
            {
                "text": "To avoid this problem while still retaining the benefits of parameter sharing for low frequency predicates with higher-frequency predicates belonging to the same VN class, we propose including the set of all possible classes for a given predicate as Sent. B features. To incorporate multiple senses, we simply concatenate them sequentially to Sent. B:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "VerbNet Classes without Disambiguation",
                "sec_num": null
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "LM(w 1...T , w p w s 1...k )",
                        "eq_num": "(9)"
                    }
                ],
                "section": "VerbNet Classes without Disambiguation",
                "sec_num": null
            },
            {
                "text": "This allows the BERT encoder to attend over all possible VerbNet classes for a given predicate and sentence, without making a discrete decision about which class is correct. The extent and way in which the model incorporates the Sent. B tokens associated with the available classes is learned during training. The inputs to this model, later referred to as ALL CLASSES are identical to PREDICTED CLASS and GOLD CLASS models for monosemous predicates.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "VerbNet Classes without Disambiguation",
                "sec_num": null
            },
            {
                "text": "Features that are useful for SRL may also be useful in predicting the sense of a predicate. For example, surface-level syntactic awareness that the argument of a predicate is a clause instead of a noun phrase may change the expected sense of a verb (bring-11.3 vs. characterize-29.2):",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Joint VerbNet Classification and SRL",
                "sec_num": "5"
            },
            {
                "text": "Bob took Mary to the doctor. John took Mary to be a doctor.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Joint VerbNet Classification and SRL",
                "sec_num": "5"
            },
            {
                "text": "The semantic classes of arguments are also often important in determining the sense of a given predicate (dub-29.3.2 vs. get-13.5.1):",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Joint VerbNet Classification and SRL",
                "sec_num": "5"
            },
            {
                "text": "John called Mary a name. John called Mary a car.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Joint VerbNet Classification and SRL",
                "sec_num": "5"
            },
            {
                "text": "This dependency between SRL and predicate sense disambiguation together with the prevalence of shared features between the two tasks makes them a good candidate for multi-task learning (Caruana, 1998) .",
                "cite_spans": [
                    {
                        "start": 185,
                        "end": 200,
                        "text": "(Caruana, 1998)",
                        "ref_id": "BIBREF8"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Joint VerbNet Classification and SRL",
                "sec_num": "5"
            },
            {
                "text": "Multi-task Model Much of recent work in multitask learning for SRL has focused on syntactic tasks such as syntactic parsing as auxiliary objectives (Strubell et al., 2018; Swayamdipta et al., 2018; Xia et al., 2019; Zhou et al., 2020) . We first investigate an MTL approach that predicts semantic role labels and predicate senses independently given a shared BERT encoder. We extend our baseline SRL model, adding an additional head that is trained to predict the target predicate's sense, as described in Equation 8. The negative log likelihood of a single training instance with predicate p and token sequence x = w 1...T with T tokens is then given by:",
                "cite_spans": [
                    {
                        "start": 148,
                        "end": 171,
                        "text": "(Strubell et al., 2018;",
                        "ref_id": "BIBREF46"
                    },
                    {
                        "start": 172,
                        "end": 197,
                        "text": "Swayamdipta et al., 2018;",
                        "ref_id": "BIBREF47"
                    },
                    {
                        "start": 198,
                        "end": 215,
                        "text": "Xia et al., 2019;",
                        "ref_id": "BIBREF53"
                    },
                    {
                        "start": 216,
                        "end": 234,
                        "text": "Zhou et al., 2020)",
                        "ref_id": "BIBREF58"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Joint VerbNet Classification and SRL",
                "sec_num": "5"
            },
            {
                "text": "\u2212 T t=1 log P (y (srl) pt |x, p) + \u03bb vncls log P (y (vncls) p |x, p) (10)",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Joint VerbNet Classification and SRL",
                "sec_num": "5"
            },
            {
                "text": "with \u03bb vncls weighting the contribution of VerbNet class prediction to the overall objective. For brevity, we henceforth refer to this model as SRL + VSD.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Joint VerbNet Classification and SRL",
                "sec_num": "5"
            },
            {
                "text": "We also investigate conditioning role labeling directly on predicted predicate senses. We implement this by concatenating a weighted label embedding of the target predicate's predicted class to each of the SRL head's input vectors, x ",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Joint VerbNet Classification and SRL",
                "sec_num": "5"
            },
            {
                "text": "y (vncls) p = K k=1 P (y (vncls) p = k|x, p)W (vncls) (k) (11) with W (vncls) \u2208 R K\u00d7D vncls and y (srl) p \u2208 R D vncls .",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Joint VerbNet Classification and SRL",
                "sec_num": "5"
            },
            {
                "text": "The input to the SRL head is then given by:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Joint VerbNet Classification and SRL",
                "sec_num": "5"
            },
            {
                "text": "x (srl) pt = LM(w 1...T , w p ) (t) ; W (mark) (t=p) ; y (vncls) p (12)",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Joint VerbNet Classification and SRL",
                "sec_num": "5"
            },
            {
                "text": "VerbNet class embeddings are initialized using the average of word embeddings corresponding to members of each class. During training, we use embeddings of predicted labels to avoid a discrepancy between the inputs to the SRL head between training and inference, when the gold labels are no longer available. In preliminary experiments, we used gold labels, similar to teacher forcing as described in Williams and Zipser (1989) , but found that performance degraded when applied to predicted labels. We refer to the model described in this section as SRL | VSD.",
                "cite_spans": [
                    {
                        "start": 401,
                        "end": 427,
                        "text": "Williams and Zipser (1989)",
                        "ref_id": "BIBREF51"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Joint VerbNet Classification and SRL",
                "sec_num": "5"
            },
            {
                "text": "All models are implemented using Tensorflow 1.13 (Abadi et al., 2016) and are trained on a single NVIDIA GTX 1080 Ti GPU. We use the 110M parameter cased BERT-Base model available in Tensorflow Hub 3 , with D LM = 768. To align with Shi and Lin (2019) , D mark is set to 10, and LSTM and MLP hidden state sizes are set to 768 and 300 respectively. Dropout rates of 0.1 are applied to BERT outputs as well as after ReLu transforms in MLPs. Recurrent dropout (Gal and Ghahramani, 2016) with a rate of 0.1 is applied in LSTMs on hidden states and outputs. To initialize VerbNet class embeddings, we use 100-dimensional GloVe embeddings (Pennington et al., 2014) averaged over member verbs (D vncls = 100). \u03bb vncls is set to 0.5 after a preliminary search over {0.1, 0.5, 1.0}.",
                "cite_spans": [
                    {
                        "start": 49,
                        "end": 69,
                        "text": "(Abadi et al., 2016)",
                        "ref_id": "BIBREF0"
                    },
                    {
                        "start": 233,
                        "end": 251,
                        "text": "Shi and Lin (2019)",
                        "ref_id": "BIBREF44"
                    },
                    {
                        "start": 457,
                        "end": 483,
                        "text": "(Gal and Ghahramani, 2016)",
                        "ref_id": "BIBREF19"
                    },
                    {
                        "start": 633,
                        "end": 658,
                        "text": "(Pennington et al., 2014)",
                        "ref_id": "BIBREF40"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Experiments",
                "sec_num": "6"
            },
            {
                "text": "We follow the fine-tuning methodology described in Devlin et al. (2019) , using Adam (Kingma and Ba, 2014) with a batch size of 16. The learning rate is warmed up linearly from 0 to 5e-5 for 10% of training, then decayed linearly to 0 for the rest of training. Models are trained for up to 8 epochs. The best-performing checkpoint on the development set, evaluated at every half epoch, is selected for evaluation.",
                "cite_spans": [
                    {
                        "start": 51,
                        "end": 71,
                        "text": "Devlin et al. (2019)",
                        "ref_id": "BIBREF14"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Experiments",
                "sec_num": "6"
            },
            {
                "text": "Unless otherwise mentioned, we train and evaluate all models with at least 7 independent random initializations, and present mean scores in our comparisons. To establish statistical significance, we apply a test for Almost Stochastic Dominance (Dror et al., 2019) between test score distributions, using \u03b1 = 0.05. Numbers in bold indicate highest average performance within a given evaluative setting, with a single star indicating statistical significance of almost stochastic dominance over our baseline Datasets We use English PropBank datasets from CoNLL-2005 (Carreras and M\u00e0rquez, 2005) and the CoNLL-2012 split (Pradhan et al., 2013) for OntoNotes (Hovy et al., 2006) in order to situate our baseline mode among recent work in PB SRL. We compare against models of similar size (120M parameters) with pre-identified predicates. The SemLink corpus (Palmer, 2009) is currently the only dataset that contains explicit VerbNet thematic role annotations with VN sense annotations. SemLink contains mappings between VN, PB and FrameNet, with annotations performed over a subset of the CoNLL-2005 PB WSJ annotations and Brown corpus out-of-domain test set (Carreras and M\u00e0rquez, 2005) . Using SemLink thus allows us to evaluate performance for both PB and VN roles on the same source text. Following Zapirain et al. (2008), we restrict evaluation to propositions with PB core arguments fully mapped to VN thematic roles. This accounts for 56% of the original corpus. We include PB modifier roles in addition to VN thematic roles.",
                "cite_spans": [
                    {
                        "start": 244,
                        "end": 263,
                        "text": "(Dror et al., 2019)",
                        "ref_id": "BIBREF17"
                    },
                    {
                        "start": 553,
                        "end": 577,
                        "text": "CoNLL-2005 (Carreras and",
                        "ref_id": null
                    },
                    {
                        "start": 578,
                        "end": 592,
                        "text": "M\u00e0rquez, 2005)",
                        "ref_id": "BIBREF7"
                    },
                    {
                        "start": 618,
                        "end": 640,
                        "text": "(Pradhan et al., 2013)",
                        "ref_id": "BIBREF42"
                    },
                    {
                        "start": 655,
                        "end": 674,
                        "text": "(Hovy et al., 2006)",
                        "ref_id": "BIBREF27"
                    },
                    {
                        "start": 853,
                        "end": 867,
                        "text": "(Palmer, 2009)",
                        "ref_id": "BIBREF37"
                    },
                    {
                        "start": 1155,
                        "end": 1183,
                        "text": "(Carreras and M\u00e0rquez, 2005)",
                        "ref_id": "BIBREF7"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Experiments",
                "sec_num": "6"
            },
            {
                "text": "Baseline Comparisons Our baseline SRL model achieves comparable performance to Shi and Lin (2019) on both CoNLL-2012 and CoNLL-2005 and thus has performance similar to state-of-the-art models of the same size.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Experiments",
                "sec_num": "6"
            },
            {
                "text": "To compare our VerbNet classification models against prior work, we train and evaluate a publicly available state-of-the-art VN classification system directly on the SemLink corpus. We use Clear-WSD 4 , which is a sense disambiguation library tailored for verb sense disambiguation based on linear models over features constructed from an ensemble of word representations applied over syntactic relations (Palmer et al., 2017) .",
                "cite_spans": [
                    {
                        "start": 405,
                        "end": 426,
                        "text": "(Palmer et al., 2017)",
                        "ref_id": "BIBREF39"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Experiments",
                "sec_num": "6"
            },
            {
                "text": "The results of our experiments are shown in Table 3 . First, we find that incorporating gold VerbNet classes (GOLD CLASS) significantly improves VerbNet SRL, providing a 15% relative error reduction on out-of-domain data (80.1 to 83.0), and 6% reduction on in-domain data (87.4 to 88.2). In PB SRL, gold classes are also beneficial, but to a lesser degree. ALL CLASSES and PRE-DICTED CLASS models improve both in-domain and out-of-domain VN SRL.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 44,
                        "end": 51,
                        "text": "Table 3",
                        "ref_id": "TABREF5"
                    }
                ],
                "eq_spans": [],
                "section": "VerbNet Models",
                "sec_num": null
            },
            {
                "text": "Predicting both VN classes and semantic roles from a single encoder reduces the total computational resources required to make predictions from separate models, providing a practical benefit. Additionally, we are interested in determining whether our multi-task models lead to improvements in generalization. Our multi-task model SRL + VSD, which does not condition thematic role prediction on predicted senses, does not have a significant effect on VN SRL performance. However, we do find that conditioning SRL on VN class predictions in a multi-task model (SRL | VSD) leads to a significant improvement in performance on the out-of-domain Brown test set for VN SRL. No significant change is observed on the in-domain WSJ test set, or when the model is applied to PB SRL.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "VerbNet Models",
                "sec_num": null
            },
            {
                "text": "We also evaluate the impact of multi-task learning on predicate disambiguation (VN classification). First, we find that even our baseline model is competitive with the highly-specialized approach for verb sense disambiguation provided in Clear-WSD (Table 4) . Comparing our joint VN SRL models with a single task baseline for VN classification, we observe a significant improvement on WSJ test data when incorporating multi-task supervision from SRL. This approach is related to earlier use of SRL features for verb sense disambiguation reported in Dang and Palmer (2005) , and the positive result is consistent with their findings.",
                "cite_spans": [
                    {
                        "start": 549,
                        "end": 571,
                        "text": "Dang and Palmer (2005)",
                        "ref_id": "BIBREF12"
                    }
                ],
                "ref_spans": [
                    {
                        "start": 248,
                        "end": 257,
                        "text": "(Table 4)",
                        "ref_id": "TABREF6"
                    }
                ],
                "eq_spans": [],
                "section": "VerbNet Models",
                "sec_num": null
            },
            {
                "text": "Monosemous vs. Polysemous Predicates To understand the impact of VerbNet class features, we break down our evaluation by polysemous and 4 https://github.com/clearwsd/clearwsd monosemous verbs in Table 5 . First, we observe that incorporating VN classes improves F 1 scores for monosemous verbs in both models. This is expected, as monosemous verbs are typically lower frequency, with low-frequency and OOV verbs benefiting the most from parameter sharing with other verbs belonging to the same VN classes. We also observe a significant improvement on polysemous verbs in the WSJ (in-domain) test set when including VN features. However, polysemous verbs in the Brown (out-of-domain) test set only benefit from using explicitly predicted classes, but not when using all valid classes for each predicate.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 195,
                        "end": 202,
                        "text": "Table 5",
                        "ref_id": "TABREF8"
                    }
                ],
                "eq_spans": [],
                "section": "Analysis",
                "sec_num": "7"
            },
            {
                "text": "Why does ALL CLASSES improve performance on out-of-domain data for monosemous verbs, but not polysemous verbs? Intuitively, the per-verb distributions of VerbNet classes may change considerably between two domains. Using a correctlypredicted class may help mitigate errors on verbs for which one class was dominant during training, but a different class or set of classes are observed during testing in the new domain. This benefit would not be observed with ALL CLASSES as for a given verb, the same classes used as model inputs during training would be used as inputs on out-ofdomain data. However, VN classes receive fewer updates during training when using only predicted classes. Thus, verbs appearing in classes that never or rarely appeared during training will not benefit from PREDICTED CLASS features. ALL CLASSES may mitigate this issue, since even if a specific class does not appear in the training data, it still can receive updates from examples of polysemous member verbs that belong to other classes (and improved performance over PREDICTED CLASS on monosemous verbs on the out-of-domain Brown test set supports this). As future work, a promising direction may therefore be to combine PREDICTED CLASS and ALL CLASSES features.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Analysis",
                "sec_num": "7"
            },
            {
                "text": "Out-of-Vocabulary Predicates How well do models incorporating VerbNet features generalize on out-of-vocabulary and rare predicates? We split an evaluation on the WSJ development set into 5 bins by training set predicate frequency (shown in Figure 1 ). Comparing development F 1 scores for ALL CLASSES and PREDICTED CLASS models against our baseline model, we note that VN classes improve SRL performance most for predicates appearing 0-50 times in the training data, which account for 24.4% of instances in the development set. Focusing on low-frequency predicates, we further divide our evaluation of predicates occurring fewer than 50 times in the training data into 6 bins, one of which is reserved for OOV predicates (Figure 2) . From this analysis, we find that VN classes are most impactful on predicates appearing fewer than 10 times in the training data, with a large improvement over the baseline on OOV predicates when applying predicted classes.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 240,
                        "end": 248,
                        "text": "Figure 1",
                        "ref_id": "FIGREF1"
                    },
                    {
                        "start": 721,
                        "end": 731,
                        "text": "(Figure 2)",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Analysis",
                "sec_num": "7"
            },
            {
                "text": "We investigate VerbNet classes as an effective level of abstraction for predicates when performing semantic role labeling. We find that incorporating features based on gold VerbNet classes improves both VerbNet and PropBank SRL, but when predicted classes are used, this effect is only observed for VerbNet. An improvement is also observed without explicit prediction of classes by including a list of all VerbNet classes the target predicate belongs to as features. Breaking down our evaluation into polysemous and monosemous predicates, we find that predicted classes help more on outof-domain polysemous predicates, while using all valid VerbNet classes helps more on out-of-domain low-frequency predicates. In multi-task learning experiments motivated by the interdependence of VN classification and SRL, we find that joint training improves both tasks when conditioning role labeling on predicted predicates, facilitating VN semantic parsing. In future work, we will investigate alternative approaches incorporating the structure of VerbNet into the parsing of VerbNet semantic representations. Finally, we hope to expand our evaluations to larger, more diverse datasets to further investigate domain transfer. Predicted Class All Classes Baseline Figure 2 : Evaluation by training set predicate frequency similar to Figure 1 , but focused on lowfrequency predicates. Most improvements are for predicates appearing fewer than 10 times in the training data.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 1253,
                        "end": 1261,
                        "text": "Figure 2",
                        "ref_id": null
                    },
                    {
                        "start": 1322,
                        "end": 1330,
                        "text": "Figure 1",
                        "ref_id": "FIGREF1"
                    }
                ],
                "eq_spans": [],
                "section": "Conclusions and Future Work",
                "sec_num": "8"
            },
            {
                "text": "Our code is available at https://github.com/ jgung/verbnet-parsing-iwcs-2021.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "In preliminary experiments, we found that directly modifying Sent. A drastically reduces the performance of the model and slows convergence.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            }
        ],
        "back_matter": [
            {
                "text": "We gratefully acknowledge the support of C3 (Cognitively Coherent Human-Computer Communication, subcontracts from UIUC and SIFT), DARPA AIDA Award FA8750-18-2-0016 (RAM-FIS), and DTRA HDTRA1-16-1-0002/Project 1553695 (eTASC -Empirical Evidence for a Theoretical Approach to Semantic Components). Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of any government agency. This work was partially supported by research credits from Google Cloud. Finally, we thank the anonymous IWCS reviewers for their insightful comments and suggestions.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Acknowledgments",
                "sec_num": null
            }
        ],
        "bib_entries": {
            "BIBREF0": {
                "ref_id": "b0",
                "title": "Tensorflow: A system for large-scale machine learning",
                "authors": [
                    {
                        "first": "Mart\u00edn",
                        "middle": [],
                        "last": "Abadi",
                        "suffix": ""
                    },
                    {
                        "first": "Paul",
                        "middle": [],
                        "last": "Barham",
                        "suffix": ""
                    },
                    {
                        "first": "Jianmin",
                        "middle": [],
                        "last": "Chen",
                        "suffix": ""
                    },
                    {
                        "first": "Zhifeng",
                        "middle": [],
                        "last": "Chen",
                        "suffix": ""
                    },
                    {
                        "first": "Andy",
                        "middle": [],
                        "last": "Davis",
                        "suffix": ""
                    },
                    {
                        "first": "Jeffrey",
                        "middle": [],
                        "last": "Dean",
                        "suffix": ""
                    },
                    {
                        "first": "Matthieu",
                        "middle": [],
                        "last": "Devin",
                        "suffix": ""
                    },
                    {
                        "first": "Sanjay",
                        "middle": [],
                        "last": "Ghemawat",
                        "suffix": ""
                    },
                    {
                        "first": "Geoffrey",
                        "middle": [],
                        "last": "Irving",
                        "suffix": ""
                    },
                    {
                        "first": "Michael",
                        "middle": [],
                        "last": "Isard",
                        "suffix": ""
                    }
                ],
                "year": 2016,
                "venue": "OSDI",
                "volume": "",
                "issue": "",
                "pages": "265--283",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Mart\u00edn Abadi, Paul Barham, Jianmin Chen, Zhifeng Chen, Andy Davis, Jeffrey Dean, Matthieu Devin, Sanjay Ghemawat, Geoffrey Irving, Michael Isard, et al. 2016. Tensorflow: A system for large-scale machine learning. In OSDI, pages 265-283.",
                "links": null
            },
            "BIBREF1": {
                "ref_id": "b1",
                "title": "The Berkeley FrameNet project",
                "authors": [
                    {
                        "first": "Collin",
                        "middle": [
                            "F"
                        ],
                        "last": "Baker",
                        "suffix": ""
                    },
                    {
                        "first": "Charles",
                        "middle": [
                            "J"
                        ],
                        "last": "Fillmore",
                        "suffix": ""
                    },
                    {
                        "first": "John",
                        "middle": [
                            "B"
                        ],
                        "last": "Lowe",
                        "suffix": ""
                    }
                ],
                "year": 1998,
                "venue": "36th Annual Meeting of the Association for Computational Linguistics and 17th International Conference on Computational Linguistics",
                "volume": "1",
                "issue": "",
                "pages": "86--90",
                "other_ids": {
                    "DOI": [
                        "10.3115/980845.980860"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Collin F. Baker, Charles J. Fillmore, and John B. Lowe. 1998. The Berkeley FrameNet project. In 36th An- nual Meeting of the Association for Computational Linguistics and 17th International Conference on Computational Linguistics, Volume 1, pages 86-90, Montreal, Quebec, Canada. Association for Compu- tational Linguistics.",
                "links": null
            },
            "BIBREF2": {
                "ref_id": "b2",
                "title": "Semantic roles for string to tree machine translation",
                "authors": [
                    {
                        "first": "Marzieh",
                        "middle": [],
                        "last": "Bazrafshan",
                        "suffix": ""
                    },
                    {
                        "first": "Daniel",
                        "middle": [],
                        "last": "Gildea",
                        "suffix": ""
                    }
                ],
                "year": 2013,
                "venue": "Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics",
                "volume": "2",
                "issue": "",
                "pages": "419--423",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Marzieh Bazrafshan and Daniel Gildea. 2013. Seman- tic roles for string to tree machine translation. In Proceedings of the 51st Annual Meeting of the As- sociation for Computational Linguistics (Volume 2: Short Papers), pages 419-423, Sofia, Bulgaria. As- sociation for Computational Linguistics.",
                "links": null
            },
            "BIBREF3": {
                "ref_id": "b3",
                "title": "Modeling biological processes for reading comprehension",
                "authors": [
                    {
                        "first": "Jonathan",
                        "middle": [],
                        "last": "Berant",
                        "suffix": ""
                    },
                    {
                        "first": "Vivek",
                        "middle": [],
                        "last": "Srikumar",
                        "suffix": ""
                    },
                    {
                        "first": "Pei-Chun",
                        "middle": [],
                        "last": "Chen",
                        "suffix": ""
                    },
                    {
                        "first": "Abby",
                        "middle": [],
                        "last": "Vander Linden",
                        "suffix": ""
                    },
                    {
                        "first": "Brittany",
                        "middle": [],
                        "last": "Harding",
                        "suffix": ""
                    },
                    {
                        "first": "Brad",
                        "middle": [],
                        "last": "Huang",
                        "suffix": ""
                    },
                    {
                        "first": "Peter",
                        "middle": [],
                        "last": "Clark",
                        "suffix": ""
                    },
                    {
                        "first": "Christopher",
                        "middle": [
                            "D"
                        ],
                        "last": "Manning",
                        "suffix": ""
                    }
                ],
                "year": 2014,
                "venue": "Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP)",
                "volume": "",
                "issue": "",
                "pages": "1499--1510",
                "other_ids": {
                    "DOI": [
                        "10.3115/v1/D14-1159"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Jonathan Berant, Vivek Srikumar, Pei-Chun Chen, Abby Vander Linden, Brittany Harding, Brad Huang, Peter Clark, and Christopher D. Manning. 2014. Modeling biological processes for reading compre- hension. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1499-1510, Doha, Qatar. Associa- tion for Computational Linguistics.",
                "links": null
            },
            "BIBREF4": {
                "ref_id": "b4",
                "title": "PropBank Annotation Guidelines",
                "authors": [
                    {
                        "first": "Claire",
                        "middle": [],
                        "last": "Bonial",
                        "suffix": ""
                    },
                    {
                        "first": "Olga",
                        "middle": [],
                        "last": "Babko-Malaya",
                        "suffix": ""
                    },
                    {
                        "first": "D",
                        "middle": [],
                        "last": "Jinho",
                        "suffix": ""
                    },
                    {
                        "first": "Jena",
                        "middle": [],
                        "last": "Choi",
                        "suffix": ""
                    },
                    {
                        "first": "Martha",
                        "middle": [],
                        "last": "Hwang",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Palmer",
                        "suffix": ""
                    }
                ],
                "year": 2010,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Claire Bonial, Olga Babko-Malaya, Jinho D Choi, Jena Hwang, and Martha Palmer. 2010. PropBank Anno- tation Guidelines. Technical report, Center for Com- putational Language and Education Research Insti- tute of Cognitive Science University of Colorado at Boulder.",
                "links": null
            },
            "BIBREF5": {
                "ref_id": "b5",
                "title": "VerbNet representations: Subevent semantics for transfer verbs",
                "authors": [
                    {
                        "first": "Julia",
                        "middle": [],
                        "last": "Susan Windisch Brown",
                        "suffix": ""
                    },
                    {
                        "first": "James",
                        "middle": [],
                        "last": "Bonn",
                        "suffix": ""
                    },
                    {
                        "first": "Annie",
                        "middle": [],
                        "last": "Gung",
                        "suffix": ""
                    },
                    {
                        "first": "James",
                        "middle": [],
                        "last": "Zaenen",
                        "suffix": ""
                    },
                    {
                        "first": "Martha",
                        "middle": [],
                        "last": "Pustejovsky",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Palmer",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "Proceedings of the First International Workshop on Designing Meaning Representations",
                "volume": "",
                "issue": "",
                "pages": "154--163",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/W19-3318"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Susan Windisch Brown, Julia Bonn, James Gung, An- nie Zaenen, James Pustejovsky, and Martha Palmer. 2019. VerbNet representations: Subevent semantics for transfer verbs. In Proceedings of the First Inter- national Workshop on Designing Meaning Represen- tations, pages 154-163, Florence, Italy. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF6": {
                "ref_id": "b6",
                "title": "Integrating Generative Lexicon event structures into VerbNet",
                "authors": [
                    {
                        "first": "James",
                        "middle": [],
                        "last": "Susan Windisch Brown",
                        "suffix": ""
                    },
                    {
                        "first": "Annie",
                        "middle": [],
                        "last": "Pustejovsky",
                        "suffix": ""
                    },
                    {
                        "first": "Martha",
                        "middle": [],
                        "last": "Zaenen",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Palmer",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018)",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Susan Windisch Brown, James Pustejovsky, Annie Za- enen, and Martha Palmer. 2018. Integrating Gen- erative Lexicon event structures into VerbNet. In Proceedings of the Eleventh International Confer- ence on Language Resources and Evaluation (LREC 2018), Miyazaki, Japan. European Language Re- sources Association (ELRA).",
                "links": null
            },
            "BIBREF7": {
                "ref_id": "b7",
                "title": "Introduction to the CoNLL-2005 shared task: Semantic role labeling",
                "authors": [
                    {
                        "first": "Xavier",
                        "middle": [],
                        "last": "Carreras",
                        "suffix": ""
                    },
                    {
                        "first": "Llu\u00eds",
                        "middle": [],
                        "last": "M\u00e0rquez",
                        "suffix": ""
                    }
                ],
                "year": 2005,
                "venue": "Proceedings of the Ninth Conference on Computational Natural Language Learning (CoNLL-2005)",
                "volume": "",
                "issue": "",
                "pages": "152--164",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Xavier Carreras and Llu\u00eds M\u00e0rquez. 2005. Introduc- tion to the CoNLL-2005 shared task: Semantic role labeling. In Proceedings of the Ninth Confer- ence on Computational Natural Language Learning (CoNLL-2005), pages 152-164, Ann Arbor, Michi- gan. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF8": {
                "ref_id": "b8",
                "title": "Multitask learning",
                "authors": [
                    {
                        "first": "Rich",
                        "middle": [],
                        "last": "Caruana",
                        "suffix": ""
                    }
                ],
                "year": 1998,
                "venue": "Learning to Learn",
                "volume": "",
                "issue": "",
                "pages": "95--133",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Rich Caruana. 1998. Multitask learning. In Learning to Learn, pages 95-133. Springer.",
                "links": null
            },
            "BIBREF9": {
                "ref_id": "b9",
                "title": "Unsupervised Induction and Filling of Semantic Slots for Spoken Dialogue Systems Using Frame-Semantic Parsing",
                "authors": [
                    {
                        "first": "Yun-Nung",
                        "middle": [],
                        "last": "Chen",
                        "suffix": ""
                    },
                    {
                        "first": "William",
                        "middle": [
                            "Yang"
                        ],
                        "last": "Wang",
                        "suffix": ""
                    },
                    {
                        "first": "Alexander",
                        "middle": [
                            "I"
                        ],
                        "last": "Rudnicky",
                        "suffix": ""
                    }
                ],
                "year": 2013,
                "venue": "ASRU",
                "volume": "",
                "issue": "",
                "pages": "120--125",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Yun-nung Chen, William Yang Wang, and Alexander I Rudnicky. 2013. Unsupervised Induction and Fill- ing of Semantic Slots for Spoken Dialogue Systems Using Frame-Semantic Parsing. ASRU, pages 120- 125.",
                "links": null
            },
            "BIBREF10": {
                "ref_id": "b10",
                "title": "What happened? leveraging verbnet to predict the effects of actions in procedural text",
                "authors": [
                    {
                        "first": "Peter",
                        "middle": [],
                        "last": "Clark",
                        "suffix": ""
                    },
                    {
                        "first": "Bhavana",
                        "middle": [],
                        "last": "Dalvi",
                        "suffix": ""
                    },
                    {
                        "first": "Niket",
                        "middle": [],
                        "last": "Tandon",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:1804.05435"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Peter Clark, Bhavana Dalvi, and Niket Tandon. 2018. What happened? leveraging verbnet to predict the effects of actions in procedural text. arXiv:1804.05435.",
                "links": null
            },
            "BIBREF11": {
                "ref_id": "b11",
                "title": "Joint learning of preposition senses and semantic roles of prepositional phrases",
                "authors": [
                    {
                        "first": "Daniel",
                        "middle": [],
                        "last": "Dahlmeier",
                        "suffix": ""
                    },
                    {
                        "first": "Tanja",
                        "middle": [],
                        "last": "Hwee Tou Ng",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Schultz",
                        "suffix": ""
                    }
                ],
                "year": 2009,
                "venue": "Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing",
                "volume": "",
                "issue": "",
                "pages": "450--458",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Daniel Dahlmeier, Hwee Tou Ng, and Tanja Schultz. 2009. Joint learning of preposition senses and se- mantic roles of prepositional phrases. In Proceed- ings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 450-458, Singapore. Association for Computational Linguis- tics.",
                "links": null
            },
            "BIBREF12": {
                "ref_id": "b12",
                "title": "The role of semantic roles in disambiguating verb senses",
                "authors": [
                    {
                        "first": "Trang",
                        "middle": [],
                        "last": "Hoa",
                        "suffix": ""
                    },
                    {
                        "first": "Martha",
                        "middle": [],
                        "last": "Dang",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Palmer",
                        "suffix": ""
                    }
                ],
                "year": 2005,
                "venue": "Proceedings of the 43rd Annual Meeting of the Association for Computational Linguistics (ACL'05)",
                "volume": "",
                "issue": "",
                "pages": "42--49",
                "other_ids": {
                    "DOI": [
                        "10.3115/1219840.1219846"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Hoa Trang Dang and Martha Palmer. 2005. The role of semantic roles in disambiguating verb senses. In Proceedings of the 43rd Annual Meeting of the As- sociation for Computational Linguistics (ACL'05), pages 42-49, Ann Arbor, Michigan. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF13": {
                "ref_id": "b13",
                "title": "Frame-semantic parsing",
                "authors": [
                    {
                        "first": "Dipanjan",
                        "middle": [],
                        "last": "Das",
                        "suffix": ""
                    },
                    {
                        "first": "Desai",
                        "middle": [],
                        "last": "Chen",
                        "suffix": ""
                    },
                    {
                        "first": "F",
                        "middle": [
                            "T"
                        ],
                        "last": "Andr\u00e9",
                        "suffix": ""
                    },
                    {
                        "first": "Nathan",
                        "middle": [],
                        "last": "Martins",
                        "suffix": ""
                    },
                    {
                        "first": "Noah",
                        "middle": [
                            "A"
                        ],
                        "last": "Schneider",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Smith",
                        "suffix": ""
                    }
                ],
                "year": 2014,
                "venue": "Computational Linguistics",
                "volume": "40",
                "issue": "1",
                "pages": "9--56",
                "other_ids": {
                    "DOI": [
                        "10.1162/COLI_a_00163"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Dipanjan Das, Desai Chen, Andr\u00e9 F. T. Martins, Nathan Schneider, and Noah A. Smith. 2014. Frame-semantic parsing. Computational Linguis- tics, 40(1):9-56.",
                "links": null
            },
            "BIBREF14": {
                "ref_id": "b14",
                "title": "BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding",
                "authors": [
                    {
                        "first": "Jacob",
                        "middle": [],
                        "last": "Devlin",
                        "suffix": ""
                    },
                    {
                        "first": "Ming-Wei",
                        "middle": [],
                        "last": "Chang",
                        "suffix": ""
                    },
                    {
                        "first": "Kenton",
                        "middle": [],
                        "last": "Lee",
                        "suffix": ""
                    },
                    {
                        "first": "Kristina",
                        "middle": [],
                        "last": "Toutanova",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "NAACL",
                "volume": "",
                "issue": "",
                "pages": "4171--4186",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of Deep Bidirectional Transformers for Language Un- derstanding. In NAACL, pages 4171-4186.",
                "links": null
            },
            "BIBREF15": {
                "ref_id": "b15",
                "title": "Syn-QG: Syntactic and shallow semantic rules for question generation",
                "authors": [
                    {
                        "first": "Kaustubh",
                        "middle": [],
                        "last": "Dhole",
                        "suffix": ""
                    },
                    {
                        "first": "D",
                        "middle": [],
                        "last": "Christopher",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Manning",
                        "suffix": ""
                    }
                ],
                "year": 2020,
                "venue": "Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics",
                "volume": "",
                "issue": "",
                "pages": "752--765",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/2020.acl-main.69"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Kaustubh Dhole and Christopher D. Manning. 2020. Syn-QG: Syntactic and shallow semantic rules for question generation. In Proceedings of the 58th An- nual Meeting of the Association for Computational Linguistics, pages 752-765, Online. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF16": {
                "ref_id": "b16",
                "title": "Thematic Proto-Roles and Argument Selection",
                "authors": [
                    {
                        "first": "David",
                        "middle": [],
                        "last": "Dowty",
                        "suffix": ""
                    }
                ],
                "year": 1991,
                "venue": "Language",
                "volume": "67",
                "issue": "3",
                "pages": "547--619",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "David Dowty. 1991. Thematic Proto-Roles and Argu- ment Selection. Language, 67(3):547-619.",
                "links": null
            },
            "BIBREF17": {
                "ref_id": "b17",
                "title": "Deep dominance -how to properly compare deep neural models",
                "authors": [
                    {
                        "first": "Rotem",
                        "middle": [],
                        "last": "Dror",
                        "suffix": ""
                    },
                    {
                        "first": "Segev",
                        "middle": [],
                        "last": "Shlomov",
                        "suffix": ""
                    },
                    {
                        "first": "Roi",
                        "middle": [],
                        "last": "Reichart",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics",
                "volume": "",
                "issue": "",
                "pages": "2773--2785",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/P19-1266"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Rotem Dror, Segev Shlomov, and Roi Reichart. 2019. Deep dominance -how to properly compare deep neural models. In Proceedings of the 57th Annual Meeting of the Association for Computational Lin- guistics, pages 2773-2785, Florence, Italy. Associa- tion for Computational Linguistics.",
                "links": null
            },
            "BIBREF18": {
                "ref_id": "b18",
                "title": "Semantic role labeling with neural network factors",
                "authors": [
                    {
                        "first": "Nicholas",
                        "middle": [],
                        "last": "Fitzgerald",
                        "suffix": ""
                    },
                    {
                        "first": "Oscar",
                        "middle": [],
                        "last": "T\u00e4ckstr\u00f6m",
                        "suffix": ""
                    },
                    {
                        "first": "Kuzman",
                        "middle": [],
                        "last": "Ganchev",
                        "suffix": ""
                    },
                    {
                        "first": "Dipanjan",
                        "middle": [],
                        "last": "Das",
                        "suffix": ""
                    }
                ],
                "year": 2015,
                "venue": "Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing",
                "volume": "",
                "issue": "",
                "pages": "960--970",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/D15-1112"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Nicholas FitzGerald, Oscar T\u00e4ckstr\u00f6m, Kuzman Ganchev, and Dipanjan Das. 2015. Semantic role la- beling with neural network factors. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 960-970, Lis- bon, Portugal. Association for Computational Lin- guistics.",
                "links": null
            },
            "BIBREF19": {
                "ref_id": "b19",
                "title": "A theoretically grounded application of dropout in recurrent neural networks",
                "authors": [
                    {
                        "first": "Yarin",
                        "middle": [],
                        "last": "Gal",
                        "suffix": ""
                    },
                    {
                        "first": "Zoubin",
                        "middle": [],
                        "last": "Ghahramani",
                        "suffix": ""
                    }
                ],
                "year": 2016,
                "venue": "Advances in Neural Information Processing Systems",
                "volume": "29",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Yarin Gal and Zoubin Ghahramani. 2016. A theoret- ically grounded application of dropout in recurrent neural networks. In Advances in Neural Information Processing Systems, volume 29. Curran Associates, Inc.",
                "links": null
            },
            "BIBREF20": {
                "ref_id": "b20",
                "title": "Automatic labeling of semantic roles",
                "authors": [
                    {
                        "first": "Daniel",
                        "middle": [],
                        "last": "Gildea",
                        "suffix": ""
                    },
                    {
                        "first": "Daniel",
                        "middle": [],
                        "last": "Jurafsky",
                        "suffix": ""
                    }
                ],
                "year": 2000,
                "venue": "Proceedings of the 38th Annual Meeting of the Association for Computational Linguistics",
                "volume": "",
                "issue": "",
                "pages": "512--520",
                "other_ids": {
                    "DOI": [
                        "10.3115/1075218.1075283"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Daniel Gildea and Daniel Jurafsky. 2000. Automatic labeling of semantic roles. In Proceedings of the 38th Annual Meeting of the Association for Com- putational Linguistics, pages 512-520, Hong Kong. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF21": {
                "ref_id": "b21",
                "title": "Using Relations for Identification and Normalization of Disorders: Team CLEAR in the ShARe/CLEF 2013 eHealth Evaluation Lab",
                "authors": [
                    {
                        "first": "James",
                        "middle": [],
                        "last": "Gung",
                        "suffix": ""
                    }
                ],
                "year": 2013,
                "venue": "CLEF (Working Notes)",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "James Gung. 2013. Using Relations for Identification and Normalization of Disorders: Team CLEAR in the ShARe/CLEF 2013 eHealth Evaluation Lab. In CLEF (Working Notes).",
                "links": null
            },
            "BIBREF22": {
                "ref_id": "b22",
                "title": "A joint many-task model: Growing a neural network for multiple NLP tasks",
                "authors": [
                    {
                        "first": "Kazuma",
                        "middle": [],
                        "last": "Hashimoto",
                        "suffix": ""
                    },
                    {
                        "first": "Caiming",
                        "middle": [],
                        "last": "Xiong",
                        "suffix": ""
                    },
                    {
                        "first": "Yoshimasa",
                        "middle": [],
                        "last": "Tsuruoka",
                        "suffix": ""
                    },
                    {
                        "first": "Richard",
                        "middle": [],
                        "last": "Socher",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing",
                "volume": "",
                "issue": "",
                "pages": "1923--1933",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/D17-1206"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Kazuma Hashimoto, Caiming Xiong, Yoshimasa Tsu- ruoka, and Richard Socher. 2017. A joint many-task model: Growing a neural network for multiple NLP tasks. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 1923-1933, Copenhagen, Denmark. Associa- tion for Computational Linguistics.",
                "links": null
            },
            "BIBREF23": {
                "ref_id": "b23",
                "title": "Jointly predicting predicates and arguments in neural semantic role labeling",
                "authors": [
                    {
                        "first": "Luheng",
                        "middle": [],
                        "last": "He",
                        "suffix": ""
                    },
                    {
                        "first": "Kenton",
                        "middle": [],
                        "last": "Lee",
                        "suffix": ""
                    },
                    {
                        "first": "Omer",
                        "middle": [],
                        "last": "Levy",
                        "suffix": ""
                    },
                    {
                        "first": "Luke",
                        "middle": [],
                        "last": "Zettlemoyer",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics",
                "volume": "2",
                "issue": "",
                "pages": "364--369",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/P18-2058"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Luheng He, Kenton Lee, Omer Levy, and Luke Zettle- moyer. 2018. Jointly predicting predicates and argu- ments in neural semantic role labeling. In Proceed- ings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Pa- pers), pages 364-369, Melbourne, Australia. Asso- ciation for Computational Linguistics.",
                "links": null
            },
            "BIBREF24": {
                "ref_id": "b24",
                "title": "Deep semantic role labeling: What works and what's next",
                "authors": [
                    {
                        "first": "Luheng",
                        "middle": [],
                        "last": "He",
                        "suffix": ""
                    },
                    {
                        "first": "Kenton",
                        "middle": [],
                        "last": "Lee",
                        "suffix": ""
                    },
                    {
                        "first": "Mike",
                        "middle": [],
                        "last": "Lewis",
                        "suffix": ""
                    },
                    {
                        "first": "Luke",
                        "middle": [],
                        "last": "Zettlemoyer",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics",
                "volume": "1",
                "issue": "",
                "pages": "473--483",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/P17-1044"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Luheng He, Kenton Lee, Mike Lewis, and Luke Zettle- moyer. 2017. Deep semantic role labeling: What works and what's next. In Proceedings of the 55th Annual Meeting of the Association for Computa- tional Linguistics (Volume 1: Long Papers), pages 473-483, Vancouver, Canada. Association for Com- putational Linguistics.",
                "links": null
            },
            "BIBREF25": {
                "ref_id": "b25",
                "title": "Semantic frame identification with distributed word representations",
                "authors": [
                    {
                        "first": "Karl",
                        "middle": [],
                        "last": "Moritz Hermann",
                        "suffix": ""
                    },
                    {
                        "first": "Dipanjan",
                        "middle": [],
                        "last": "Das",
                        "suffix": ""
                    },
                    {
                        "first": "Jason",
                        "middle": [],
                        "last": "Weston",
                        "suffix": ""
                    },
                    {
                        "first": "Kuzman",
                        "middle": [],
                        "last": "Ganchev",
                        "suffix": ""
                    }
                ],
                "year": 2014,
                "venue": "Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics",
                "volume": "1",
                "issue": "",
                "pages": "1448--1458",
                "other_ids": {
                    "DOI": [
                        "10.3115/v1/P14-1136"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Karl Moritz Hermann, Dipanjan Das, Jason Weston, and Kuzman Ganchev. 2014. Semantic frame iden- tification with distributed word representations. In Proceedings of the 52nd Annual Meeting of the As- sociation for Computational Linguistics (Volume 1: Long Papers), pages 1448-1458, Baltimore, Mary- land. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF26": {
                "ref_id": "b26",
                "title": "Long Short-Term Memory",
                "authors": [
                    {
                        "first": "Sepp",
                        "middle": [],
                        "last": "Hochreiter",
                        "suffix": ""
                    },
                    {
                        "first": "J\u00fcrgen",
                        "middle": [],
                        "last": "Schmidhuber",
                        "suffix": ""
                    }
                ],
                "year": 1997,
                "venue": "Neural Computation",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Sepp Hochreiter and J\u00fcrgen Schmidhuber. 1997. Long Short-Term Memory. Neural Computation.",
                "links": null
            },
            "BIBREF27": {
                "ref_id": "b27",
                "title": "OntoNotes: The 90% solution",
                "authors": [
                    {
                        "first": "Eduard",
                        "middle": [],
                        "last": "Hovy",
                        "suffix": ""
                    },
                    {
                        "first": "Mitchell",
                        "middle": [],
                        "last": "Marcus",
                        "suffix": ""
                    },
                    {
                        "first": "Martha",
                        "middle": [],
                        "last": "Palmer",
                        "suffix": ""
                    },
                    {
                        "first": "Lance",
                        "middle": [],
                        "last": "Ramshaw",
                        "suffix": ""
                    },
                    {
                        "first": "Ralph",
                        "middle": [],
                        "last": "Weischedel",
                        "suffix": ""
                    }
                ],
                "year": 2006,
                "venue": "Proceedings of the Human Language Technology Conference of the NAACL, Companion Volume: Short Papers",
                "volume": "",
                "issue": "",
                "pages": "57--60",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Eduard Hovy, Mitchell Marcus, Martha Palmer, Lance Ramshaw, and Ralph Weischedel. 2006. OntoNotes: The 90% solution. In Proceedings of the Human Language Technology Conference of the NAACL, Companion Volume: Short Papers, pages 57-60, New York City, USA. Association for Computa- tional Linguistics.",
                "links": null
            },
            "BIBREF28": {
                "ref_id": "b28",
                "title": "Single classifier approach for verb sense disambiguation based on generalized features",
                "authors": [
                    {
                        "first": "Daisuke",
                        "middle": [],
                        "last": "Kawahara",
                        "suffix": ""
                    },
                    {
                        "first": "Martha",
                        "middle": [],
                        "last": "Palmer",
                        "suffix": ""
                    }
                ],
                "year": 2014,
                "venue": "Proceedings of the Ninth International Conference on Language Resources and Evaluation (LREC'14)",
                "volume": "",
                "issue": "",
                "pages": "4210--4213",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Daisuke Kawahara and Martha Palmer. 2014. Sin- gle classifier approach for verb sense disambigua- tion based on generalized features. In Proceedings of the Ninth International Conference on Language Resources and Evaluation (LREC'14), pages 4210- 4213, Reykjavik, Iceland. European Language Re- sources Association (ELRA).",
                "links": null
            },
            "BIBREF29": {
                "ref_id": "b29",
                "title": "Adam: A Method for Stochastic Optimization",
                "authors": [
                    {
                        "first": "P",
                        "middle": [],
                        "last": "Diederik",
                        "suffix": ""
                    },
                    {
                        "first": "Jimmy",
                        "middle": [],
                        "last": "Kingma",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Ba",
                        "suffix": ""
                    }
                ],
                "year": 2014,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:1412.6980"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Diederik P Kingma and Jimmy Ba. 2014. Adam: A Method for Stochastic Optimization. arXiv:1412.6980.",
                "links": null
            },
            "BIBREF30": {
                "ref_id": "b30",
                "title": "A matter of framing: The impact of linguistic formalism on probing results",
                "authors": [
                    {
                        "first": "Ilia",
                        "middle": [],
                        "last": "Kuznetsov",
                        "suffix": ""
                    },
                    {
                        "first": "Iryna",
                        "middle": [],
                        "last": "Gurevych",
                        "suffix": ""
                    }
                ],
                "year": 2020,
                "venue": "Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)",
                "volume": "",
                "issue": "",
                "pages": "171--182",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/2020.emnlp-main.13"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Ilia Kuznetsov and Iryna Gurevych. 2020. A matter of framing: The impact of linguistic formalism on prob- ing results. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Process- ing (EMNLP), pages 171-182, Online. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF31": {
                "ref_id": "b31",
                "title": "Dependency or Span, End-to-End Uniform Semantic Role Labeling",
                "authors": [
                    {
                        "first": "Zuchao",
                        "middle": [],
                        "last": "Li",
                        "suffix": ""
                    },
                    {
                        "first": "Shexia",
                        "middle": [],
                        "last": "He",
                        "suffix": ""
                    },
                    {
                        "first": "Hai",
                        "middle": [],
                        "last": "Zhao",
                        "suffix": ""
                    },
                    {
                        "first": "Yiqing",
                        "middle": [],
                        "last": "Zhang",
                        "suffix": ""
                    },
                    {
                        "first": "Zhuosheng",
                        "middle": [],
                        "last": "Zhang",
                        "suffix": ""
                    },
                    {
                        "first": "Xi",
                        "middle": [],
                        "last": "Zhou",
                        "suffix": ""
                    },
                    {
                        "first": "Xiang",
                        "middle": [],
                        "last": "Zhou",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "",
                "volume": "33",
                "issue": "",
                "pages": "6730--6737",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Zuchao Li, Shexia He, Hai Zhao, Yiqing Zhang, Zhu- osheng Zhang, Xi Zhou, and Xiang Zhou. 2019. De- pendency or Span, End-to-End Uniform Semantic Role Labeling. AAAI, 33:6730-6737.",
                "links": null
            },
            "BIBREF32": {
                "ref_id": "b32",
                "title": "Semantic role features for machine translation",
                "authors": [
                    {
                        "first": "Ding",
                        "middle": [],
                        "last": "Liu",
                        "suffix": ""
                    },
                    {
                        "first": "Daniel",
                        "middle": [],
                        "last": "Gildea",
                        "suffix": ""
                    }
                ],
                "year": 2010,
                "venue": "Proceedings of the 23rd International Conference on Computational Linguistics",
                "volume": "",
                "issue": "",
                "pages": "716--724",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Ding Liu and Daniel Gildea. 2010. Semantic role features for machine translation. In Proceedings of the 23rd International Conference on Computa- tional Linguistics (Coling 2010), pages 716-724, Beijing, China. Coling 2010 Organizing Committee.",
                "links": null
            },
            "BIBREF33": {
                "ref_id": "b33",
                "title": "GUIR at SemEval-2017 task 12: A framework for cross-domain clinical temporal information extraction",
                "authors": [
                    {
                        "first": "Sean",
                        "middle": [],
                        "last": "Macavaney",
                        "suffix": ""
                    },
                    {
                        "first": "Arman",
                        "middle": [],
                        "last": "Cohan",
                        "suffix": ""
                    },
                    {
                        "first": "Nazli",
                        "middle": [],
                        "last": "Goharian",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "Proceedings of the 11th International Workshop on Semantic Evaluation (SemEval-2017)",
                "volume": "",
                "issue": "",
                "pages": "1024--1029",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/S17-2180"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Sean MacAvaney, Arman Cohan, and Nazli Goharian. 2017. GUIR at SemEval-2017 task 12: A frame- work for cross-domain clinical temporal information extraction. In Proceedings of the 11th International Workshop on Semantic Evaluation (SemEval-2017), pages 1024-1029, Vancouver, Canada. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF34": {
                "ref_id": "b34",
                "title": "Abstraction and generalisation in semantic role labels: PropBank, VerbNet or both?",
                "authors": [
                    {
                        "first": "Paola",
                        "middle": [],
                        "last": "Merlo",
                        "suffix": ""
                    },
                    {
                        "first": "Lonneke",
                        "middle": [],
                        "last": "Van Der",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Plas",
                        "suffix": ""
                    }
                ],
                "year": 2009,
                "venue": "Proceedings of the Joint Conference of the 47th Annual Meeting of the ACL and the 4th International Joint Conference on Natural Language Processing of the AFNLP",
                "volume": "",
                "issue": "",
                "pages": "288--296",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Paola Merlo and Lonneke Van Der Plas. 2009. Ab- straction and generalisation in semantic role labels: PropBank, VerbNet or both? In Proceedings of the Joint Conference of the 47th Annual Meeting of the ACL and the 4th International Joint Conference on Natural Language Processing of the AFNLP, pages 288-296, Suntec, Singapore. Association for Com- putational Linguistics.",
                "links": null
            },
            "BIBREF35": {
                "ref_id": "b35",
                "title": "The Role of Verb Sense Disambiguation in Semantic Role Labeling",
                "authors": [
                    {
                        "first": "Paloma",
                        "middle": [],
                        "last": "Moreda",
                        "suffix": ""
                    },
                    {
                        "first": "Manuel",
                        "middle": [],
                        "last": "Palomar",
                        "suffix": ""
                    }
                ],
                "year": 2006,
                "venue": "Lecture Notes in Computer Science Advances in Natural Language Processing",
                "volume": "",
                "issue": "",
                "pages": "684--695",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Paloma Moreda and Manuel Palomar. 2006. The Role of Verb Sense Disambiguation in Semantic Role Labeling. Lecture Notes in Computer Science Ad- vances in Natural Language Processing, pages 684- 695.",
                "links": null
            },
            "BIBREF36": {
                "ref_id": "b36",
                "title": "A span selection model for semantic role labeling",
                "authors": [
                    {
                        "first": "Hiroki",
                        "middle": [],
                        "last": "Ouchi",
                        "suffix": ""
                    },
                    {
                        "first": "Hiroyuki",
                        "middle": [],
                        "last": "Shindo",
                        "suffix": ""
                    },
                    {
                        "first": "Yuji",
                        "middle": [],
                        "last": "Matsumoto",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing",
                "volume": "",
                "issue": "",
                "pages": "1630--1642",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/D18-1191"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Hiroki Ouchi, Hiroyuki Shindo, and Yuji Matsumoto. 2018. A span selection model for semantic role la- beling. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 1630-1642, Brussels, Belgium. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF37": {
                "ref_id": "b37",
                "title": "SemLink: Linking PropBank, VerbNet and FrameNet",
                "authors": [
                    {
                        "first": "Martha",
                        "middle": [],
                        "last": "Palmer",
                        "suffix": ""
                    }
                ],
                "year": 2009,
                "venue": "Proceedings of the Generative Lexicon Conference",
                "volume": "",
                "issue": "",
                "pages": "9--15",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Martha Palmer. 2009. SemLink: Linking PropBank, VerbNet and FrameNet. In Proceedings of the Gen- erative Lexicon Conference, pages 9-15. Pisa Italy.",
                "links": null
            },
            "BIBREF38": {
                "ref_id": "b38",
                "title": "The Proposition Bank: An annotated corpus of semantic roles",
                "authors": [
                    {
                        "first": "Martha",
                        "middle": [],
                        "last": "Palmer",
                        "suffix": ""
                    },
                    {
                        "first": "Daniel",
                        "middle": [],
                        "last": "Gildea",
                        "suffix": ""
                    },
                    {
                        "first": "Paul",
                        "middle": [],
                        "last": "Kingsbury",
                        "suffix": ""
                    }
                ],
                "year": 2005,
                "venue": "Computational Linguistics",
                "volume": "31",
                "issue": "1",
                "pages": "71--106",
                "other_ids": {
                    "DOI": [
                        "10.1162/0891201053630264"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Martha Palmer, Daniel Gildea, and Paul Kingsbury. 2005. The Proposition Bank: An annotated cor- pus of semantic roles. Computational Linguistics, 31(1):71-106.",
                "links": null
            },
            "BIBREF39": {
                "ref_id": "b39",
                "title": "The Pitfalls of Shortcuts: Tales from the Word Sense Tagging Trenches. Springer series Text, Speech and Language Technology",
                "authors": [
                    {
                        "first": "Martha",
                        "middle": [],
                        "last": "Palmer",
                        "suffix": ""
                    },
                    {
                        "first": "James",
                        "middle": [],
                        "last": "Gung",
                        "suffix": ""
                    },
                    {
                        "first": "Claire",
                        "middle": [],
                        "last": "Bonial",
                        "suffix": ""
                    },
                    {
                        "first": "Jinho",
                        "middle": [],
                        "last": "Choi",
                        "suffix": ""
                    },
                    {
                        "first": "Orin",
                        "middle": [],
                        "last": "Hargraves",
                        "suffix": ""
                    },
                    {
                        "first": "Derek",
                        "middle": [],
                        "last": "Palmer",
                        "suffix": ""
                    },
                    {
                        "first": "Kevin",
                        "middle": [],
                        "last": "Stowe",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "Essays in Lexical Semantics and Computational Lexicography",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Martha Palmer, James Gung, Claire Bonial, Jinho Choi, Orin Hargraves, Derek Palmer, and Kevin Stowe. 2017. The Pitfalls of Shortcuts: Tales from the Word Sense Tagging Trenches. Springer series Text, Speech and Language Technology, Essays in Lexi- cal Semantics and Computational Lexicography -In Honor of Adam Kilgarriff.",
                "links": null
            },
            "BIBREF40": {
                "ref_id": "b40",
                "title": "GloVe: Global vectors for word representation",
                "authors": [
                    {
                        "first": "Jeffrey",
                        "middle": [],
                        "last": "Pennington",
                        "suffix": ""
                    },
                    {
                        "first": "Richard",
                        "middle": [],
                        "last": "Socher",
                        "suffix": ""
                    },
                    {
                        "first": "Christopher",
                        "middle": [],
                        "last": "Manning",
                        "suffix": ""
                    }
                ],
                "year": 2014,
                "venue": "Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP)",
                "volume": "",
                "issue": "",
                "pages": "1532--1543",
                "other_ids": {
                    "DOI": [
                        "10.3115/v1/D14-1162"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Jeffrey Pennington, Richard Socher, and Christopher Manning. 2014. GloVe: Global vectors for word representation. In Proceedings of the 2014 Confer- ence on Empirical Methods in Natural Language Processing (EMNLP), pages 1532-1543, Doha, Qatar. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF41": {
                "ref_id": "b41",
                "title": "Deep contextualized word representations",
                "authors": [
                    {
                        "first": "Matthew",
                        "middle": [],
                        "last": "Peters",
                        "suffix": ""
                    },
                    {
                        "first": "Mark",
                        "middle": [],
                        "last": "Neumann",
                        "suffix": ""
                    },
                    {
                        "first": "Mohit",
                        "middle": [],
                        "last": "Iyyer",
                        "suffix": ""
                    },
                    {
                        "first": "Matt",
                        "middle": [],
                        "last": "Gardner",
                        "suffix": ""
                    },
                    {
                        "first": "Christopher",
                        "middle": [],
                        "last": "Clark",
                        "suffix": ""
                    },
                    {
                        "first": "Kenton",
                        "middle": [],
                        "last": "Lee",
                        "suffix": ""
                    },
                    {
                        "first": "Luke",
                        "middle": [],
                        "last": "Zettlemoyer",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies",
                "volume": "1",
                "issue": "",
                "pages": "2227--2237",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/N18-1202"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Matthew Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and Luke Zettlemoyer. 2018. Deep contextualized word rep- resentations. In Proceedings of the 2018 Confer- ence of the North American Chapter of the Associ- ation for Computational Linguistics: Human Lan- guage Technologies, Volume 1 (Long Papers), pages 2227-2237, New Orleans, Louisiana. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF42": {
                "ref_id": "b42",
                "title": "Towards robust linguistic analysis using OntoNotes",
                "authors": [
                    {
                        "first": "Alessandro",
                        "middle": [],
                        "last": "Sameer Pradhan",
                        "suffix": ""
                    },
                    {
                        "first": "Nianwen",
                        "middle": [],
                        "last": "Moschitti",
                        "suffix": ""
                    },
                    {
                        "first": "Hwee Tou",
                        "middle": [],
                        "last": "Xue",
                        "suffix": ""
                    },
                    {
                        "first": "Anders",
                        "middle": [],
                        "last": "Ng",
                        "suffix": ""
                    },
                    {
                        "first": "Olga",
                        "middle": [],
                        "last": "Bj\u00f6rkelund",
                        "suffix": ""
                    },
                    {
                        "first": "Yuchen",
                        "middle": [],
                        "last": "Uryupina",
                        "suffix": ""
                    },
                    {
                        "first": "Zhi",
                        "middle": [],
                        "last": "Zhang",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Zhong",
                        "suffix": ""
                    }
                ],
                "year": 2013,
                "venue": "Proceedings of the Seventeenth Conference on Computational Natural Language Learning",
                "volume": "",
                "issue": "",
                "pages": "143--152",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Sameer Pradhan, Alessandro Moschitti, Nianwen Xue, Hwee Tou Ng, Anders Bj\u00f6rkelund, Olga Uryupina, Yuchen Zhang, and Zhi Zhong. 2013. Towards ro- bust linguistic analysis using OntoNotes. In Pro- ceedings of the Seventeenth Conference on Computa- tional Natural Language Learning, pages 143-152, Sofia, Bulgaria. Association for Computational Lin- guistics.",
                "links": null
            },
            "BIBREF43": {
                "ref_id": "b43",
                "title": "VerbNet: A Broad-Coverage",
                "authors": [
                    {
                        "first": "Karin Kipper",
                        "middle": [],
                        "last": "Schuler",
                        "suffix": ""
                    }
                ],
                "year": 2005,
                "venue": "Comprehensive Verb Lexicon. Dissertation Abstracts International, B: Sciences and Engineering",
                "volume": "66",
                "issue": "6",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Karin Kipper Schuler. 2005. VerbNet: A Broad- Coverage, Comprehensive Verb Lexicon. Disserta- tion Abstracts International, B: Sciences and Engi- neering, 66(6).",
                "links": null
            },
            "BIBREF44": {
                "ref_id": "b44",
                "title": "Simple BERT Models for Relation Extraction and Semantic Role Labeling",
                "authors": [
                    {
                        "first": "Peng",
                        "middle": [],
                        "last": "Shi",
                        "suffix": ""
                    },
                    {
                        "first": "Jimmy",
                        "middle": [],
                        "last": "Lin",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:1904.05255"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Peng Shi and Jimmy Lin. 2019. Simple BERT Models for Relation Extraction and Semantic Role Labeling. arXiv:1904.05255.",
                "links": null
            },
            "BIBREF45": {
                "ref_id": "b45",
                "title": "Linguistic analysis improves neural metaphor detection",
                "authors": [
                    {
                        "first": "Kevin",
                        "middle": [],
                        "last": "Stowe",
                        "suffix": ""
                    },
                    {
                        "first": "Sarah",
                        "middle": [],
                        "last": "Moeller",
                        "suffix": ""
                    },
                    {
                        "first": "Laura",
                        "middle": [],
                        "last": "Michaelis",
                        "suffix": ""
                    },
                    {
                        "first": "Martha",
                        "middle": [],
                        "last": "Palmer",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "Proceedings of the 23rd Conference on Computational Natural Language Learning (CoNLL)",
                "volume": "",
                "issue": "",
                "pages": "362--371",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/K19-1034"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Kevin Stowe, Sarah Moeller, Laura Michaelis, and Martha Palmer. 2019. Linguistic analysis improves neural metaphor detection. In Proceedings of the 23rd Conference on Computational Natural Lan- guage Learning (CoNLL), pages 362-371, Hong Kong, China. Association for Computational Lin- guistics.",
                "links": null
            },
            "BIBREF46": {
                "ref_id": "b46",
                "title": "Linguistically-informed self-attention for semantic role labeling",
                "authors": [
                    {
                        "first": "Emma",
                        "middle": [],
                        "last": "Strubell",
                        "suffix": ""
                    },
                    {
                        "first": "Patrick",
                        "middle": [],
                        "last": "Verga",
                        "suffix": ""
                    },
                    {
                        "first": "Daniel",
                        "middle": [],
                        "last": "Andor",
                        "suffix": ""
                    },
                    {
                        "first": "David",
                        "middle": [],
                        "last": "Weiss",
                        "suffix": ""
                    },
                    {
                        "first": "Andrew",
                        "middle": [],
                        "last": "Mccallum",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing",
                "volume": "",
                "issue": "",
                "pages": "5027--5038",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/D18-1548"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Emma Strubell, Patrick Verga, Daniel Andor, David Weiss, and Andrew McCallum. 2018. Linguistically-informed self-attention for semantic role labeling. In Proceedings of the 2018 Confer- ence on Empirical Methods in Natural Language Processing, pages 5027-5038, Brussels, Belgium. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF47": {
                "ref_id": "b47",
                "title": "Syntactic scaffolds for semantic structures",
                "authors": [
                    {
                        "first": "Swabha",
                        "middle": [],
                        "last": "Swayamdipta",
                        "suffix": ""
                    },
                    {
                        "first": "Sam",
                        "middle": [],
                        "last": "Thomson",
                        "suffix": ""
                    },
                    {
                        "first": "Kenton",
                        "middle": [],
                        "last": "Lee",
                        "suffix": ""
                    },
                    {
                        "first": "Luke",
                        "middle": [],
                        "last": "Zettlemoyer",
                        "suffix": ""
                    },
                    {
                        "first": "Chris",
                        "middle": [],
                        "last": "Dyer",
                        "suffix": ""
                    },
                    {
                        "first": "Noah",
                        "middle": [
                            "A"
                        ],
                        "last": "Smith",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing",
                "volume": "",
                "issue": "",
                "pages": "3772--3782",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/D18-1412"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Swabha Swayamdipta, Sam Thomson, Kenton Lee, Luke Zettlemoyer, Chris Dyer, and Noah A. Smith. 2018. Syntactic scaffolds for semantic structures. In Proceedings of the 2018 Conference on Em- pirical Methods in Natural Language Processing, pages 3772-3782, Brussels, Belgium. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF48": {
                "ref_id": "b48",
                "title": "Semi-Supervised Learning for Spoken Language Understanding Using Semantic Role Labeling. Language",
                "authors": [
                    {
                        "first": "Gokhan",
                        "middle": [],
                        "last": "Tur",
                        "suffix": ""
                    },
                    {
                        "first": "Dilek",
                        "middle": [],
                        "last": "Hakkani-T\u00fcr",
                        "suffix": ""
                    }
                ],
                "year": 2005,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "232--237",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Gokhan Tur and Dilek Hakkani-T\u00fcr. 2005. Semi- Supervised Learning for Spoken Language Under- standing Using Semantic Role Labeling. Language, pages 232-237.",
                "links": null
            },
            "BIBREF49": {
                "ref_id": "b49",
                "title": "Attention is all you need",
                "authors": [
                    {
                        "first": "Ashish",
                        "middle": [],
                        "last": "Vaswani",
                        "suffix": ""
                    },
                    {
                        "first": "Noam",
                        "middle": [],
                        "last": "Shazeer",
                        "suffix": ""
                    },
                    {
                        "first": "Niki",
                        "middle": [],
                        "last": "Parmar",
                        "suffix": ""
                    },
                    {
                        "first": "Jakob",
                        "middle": [],
                        "last": "Uszkoreit",
                        "suffix": ""
                    },
                    {
                        "first": "Llion",
                        "middle": [],
                        "last": "Jones",
                        "suffix": ""
                    },
                    {
                        "first": "Aidan",
                        "middle": [
                            "N"
                        ],
                        "last": "Gomez",
                        "suffix": ""
                    },
                    {
                        "first": "Illia",
                        "middle": [],
                        "last": "Kaiser",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Polosukhin",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "Advances in Neural Information Processing Systems",
                "volume": "30",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, \u0141 ukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in Neural Information Pro- cessing Systems, volume 30. Curran Associates, Inc.",
                "links": null
            },
            "BIBREF50": {
                "ref_id": "b50",
                "title": "Machine comprehension with syntax, frames, and semantics",
                "authors": [
                    {
                        "first": "Hai",
                        "middle": [],
                        "last": "Wang",
                        "suffix": ""
                    },
                    {
                        "first": "Mohit",
                        "middle": [],
                        "last": "Bansal",
                        "suffix": ""
                    },
                    {
                        "first": "Kevin",
                        "middle": [],
                        "last": "Gimpel",
                        "suffix": ""
                    },
                    {
                        "first": "David",
                        "middle": [],
                        "last": "Mcallester",
                        "suffix": ""
                    }
                ],
                "year": 2015,
                "venue": "Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing",
                "volume": "2",
                "issue": "",
                "pages": "700--706",
                "other_ids": {
                    "DOI": [
                        "10.3115/v1/P15-2115"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Hai Wang, Mohit Bansal, Kevin Gimpel, and David McAllester. 2015. Machine comprehension with syntax, frames, and semantics. In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Volume 2: Short Papers), pages 700-706, Beijing, China. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF51": {
                "ref_id": "b51",
                "title": "A Learning Algorithm for Continually Running Fully Recurrent Neural Networks",
                "authors": [
                    {
                        "first": "Ronald",
                        "middle": [
                            "J"
                        ],
                        "last": "Williams",
                        "suffix": ""
                    },
                    {
                        "first": "David",
                        "middle": [],
                        "last": "Zipser",
                        "suffix": ""
                    }
                ],
                "year": 1989,
                "venue": "Neural Computation",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Ronald J. Williams and David Zipser. 1989. A Learn- ing Algorithm for Continually Running Fully Recur- rent Neural Networks. Neural Computation.",
                "links": null
            },
            "BIBREF52": {
                "ref_id": "b52",
                "title": "Google's neural machine translation system: Bridging the gap between human and machine translation",
                "authors": [
                    {
                        "first": "Yonghui",
                        "middle": [],
                        "last": "Wu",
                        "suffix": ""
                    },
                    {
                        "first": "Mike",
                        "middle": [],
                        "last": "Schuster",
                        "suffix": ""
                    },
                    {
                        "first": "Zhifeng",
                        "middle": [],
                        "last": "Chen",
                        "suffix": ""
                    },
                    {
                        "first": "V",
                        "middle": [],
                        "last": "Quoc",
                        "suffix": ""
                    },
                    {
                        "first": "Mohammad",
                        "middle": [],
                        "last": "Le",
                        "suffix": ""
                    },
                    {
                        "first": "Wolfgang",
                        "middle": [],
                        "last": "Norouzi",
                        "suffix": ""
                    },
                    {
                        "first": "Maxim",
                        "middle": [],
                        "last": "Macherey",
                        "suffix": ""
                    },
                    {
                        "first": "Yuan",
                        "middle": [],
                        "last": "Krikun",
                        "suffix": ""
                    },
                    {
                        "first": "Qin",
                        "middle": [],
                        "last": "Cao",
                        "suffix": ""
                    },
                    {
                        "first": "Klaus",
                        "middle": [],
                        "last": "Gao",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Macherey",
                        "suffix": ""
                    }
                ],
                "year": 2016,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:1609.08144"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Yonghui Wu, Mike Schuster, Zhifeng Chen, Quoc V Le, Mohammad Norouzi, Wolfgang Macherey, Maxim Krikun, Yuan Cao, Qin Gao, Klaus Macherey, et al. 2016. Google's neural machine translation system: Bridging the gap between human and machine translation. arXiv:1609.08144.",
                "links": null
            },
            "BIBREF53": {
                "ref_id": "b53",
                "title": "A syntax-aware multi-task learning framework for Chinese semantic role labeling",
                "authors": [
                    {
                        "first": "Qingrong",
                        "middle": [],
                        "last": "Xia",
                        "suffix": ""
                    },
                    {
                        "first": "Zhenghua",
                        "middle": [],
                        "last": "Li",
                        "suffix": ""
                    },
                    {
                        "first": "Min",
                        "middle": [],
                        "last": "Zhang",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP)",
                "volume": "",
                "issue": "",
                "pages": "5382--5392",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/D19-1541"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Qingrong Xia, Zhenghua Li, and Min Zhang. 2019. A syntax-aware multi-task learning framework for Chinese semantic role labeling. In Proceedings of the 2019 Conference on Empirical Methods in Nat- ural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 5382-5392, Hong Kong, China. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF54": {
                "ref_id": "b54",
                "title": "Robust Semantic Role Labeling Using Parsing Variations and Semantic Classes",
                "authors": [
                    {
                        "first": "Yi",
                        "middle": [],
                        "last": "Szu-Ting",
                        "suffix": ""
                    }
                ],
                "year": 2007,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Szu-ting Yi. 2007. Robust Semantic Role Labeling Us- ing Parsing Variations and Semantic Classes. Ph.D. thesis, University of Pennsylvania.",
                "links": null
            },
            "BIBREF55": {
                "ref_id": "b55",
                "title": "Can semantic roles generalize across genres?",
                "authors": [
                    {
                        "first": "Edward",
                        "middle": [],
                        "last": "Szu-Ting Yi",
                        "suffix": ""
                    },
                    {
                        "first": "Martha",
                        "middle": [],
                        "last": "Loper",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Palmer",
                        "suffix": ""
                    }
                ],
                "year": 2007,
                "venue": "Human Language Technologies 2007: The Conference of the North American Chapter of the Association for Computational Linguistics; Proceedings of the Main Conference",
                "volume": "",
                "issue": "",
                "pages": "548--555",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Szu-ting Yi, Edward Loper, and Martha Palmer. 2007. Can semantic roles generalize across genres? In Hu- man Language Technologies 2007: The Conference of the North American Chapter of the Association for Computational Linguistics; Proceedings of the Main Conference, pages 548-555, Rochester, New York. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF56": {
                "ref_id": "b56",
                "title": "Robustness and generalization of role sets: PropBank vs. VerbNet",
                "authors": [
                    {
                        "first": "Eneko",
                        "middle": [],
                        "last": "Be\u00f1at Zapirain",
                        "suffix": ""
                    },
                    {
                        "first": "Llu\u00eds",
                        "middle": [],
                        "last": "Agirre",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "M\u00e0rquez",
                        "suffix": ""
                    }
                ],
                "year": 2008,
                "venue": "Proceedings of ACL-08: HLT",
                "volume": "",
                "issue": "",
                "pages": "550--558",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Be\u00f1at Zapirain, Eneko Agirre, and Llu\u00eds M\u00e0rquez. 2008. Robustness and generalization of role sets: PropBank vs. VerbNet. In Proceedings of ACL-08: HLT, pages 550-558, Columbus, Ohio. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF57": {
                "ref_id": "b57",
                "title": "End-to-end learning of semantic role labeling using recurrent neural networks",
                "authors": [
                    {
                        "first": "Jie",
                        "middle": [],
                        "last": "Zhou",
                        "suffix": ""
                    },
                    {
                        "first": "Wei",
                        "middle": [],
                        "last": "Xu",
                        "suffix": ""
                    }
                ],
                "year": 2015,
                "venue": "Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing",
                "volume": "1",
                "issue": "",
                "pages": "1127--1137",
                "other_ids": {
                    "DOI": [
                        "10.3115/v1/P15-1109"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Jie Zhou and Wei Xu. 2015. End-to-end learning of se- mantic role labeling using recurrent neural networks. In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Lan- guage Processing (Volume 1: Long Papers), pages 1127-1137, Beijing, China. Association for Compu- tational Linguistics.",
                "links": null
            },
            "BIBREF58": {
                "ref_id": "b58",
                "title": "LIMIT-BERT : Linguistics informed multi-task BERT",
                "authors": [
                    {
                        "first": "Junru",
                        "middle": [],
                        "last": "Zhou",
                        "suffix": ""
                    },
                    {
                        "first": "Zhuosheng",
                        "middle": [],
                        "last": "Zhang",
                        "suffix": ""
                    },
                    {
                        "first": "Hai",
                        "middle": [],
                        "last": "Zhao",
                        "suffix": ""
                    },
                    {
                        "first": "Shuailiang",
                        "middle": [],
                        "last": "Zhang",
                        "suffix": ""
                    }
                ],
                "year": 2020,
                "venue": "Findings of the Association for Computational Linguistics: EMNLP 2020",
                "volume": "",
                "issue": "",
                "pages": "4450--4461",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/2020.findings-emnlp.399"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Junru Zhou, Zhuosheng Zhang, Hai Zhao, and Shuail- iang Zhang. 2020. LIMIT-BERT : Linguistics in- formed multi-task BERT. In Findings of the Associ- ation for Computational Linguistics: EMNLP 2020, pages 4450-4461, Online. Association for Computa- tional Linguistics.",
                "links": null
            }
        },
        "ref_entries": {
            "FIGREF0": {
                "num": null,
                "text": "compute the weighted label embedding of a given VN class y (vncls) p we follow Hashimoto et al. (2017):",
                "type_str": "figure",
                "uris": null
            },
            "FIGREF1": {
                "num": null,
                "text": "Evaluation by training set predicate frequency on the SemLink development data comparing the impact of VerbNet features.",
                "type_str": "figure",
                "uris": null
            },
            "TABREF1": {
                "text": "Comparison of PropBank (PB) and VerbNet (VN) roles for predicates console and walk. VerbNet's thematic role assignments (e.g. Stimulus vs. Agent and Experiencer vs. Theme) are more dependent on the predicate than PropBank's numbered arguments.",
                "type_str": "table",
                "html": null,
                "num": null,
                "content": "<table/>"
            },
            "TABREF3": {
                "text": "",
                "type_str": "table",
                "html": null,
                "num": null,
                "content": "<table><tr><td>: Comparison of baseline SRL system on</td></tr><tr><td>CoNLL-2005 and CoNLL-2012 against models apply-</td></tr><tr><td>ing pre-trained encoders of comparable size (F 1 ).</td></tr><tr><td>models for each experiment, and two stars indicat-</td></tr><tr><td>ing stochastic dominance ( = 0). For example,</td></tr><tr><td>a value in a table of 88.2 \u00b10.2 indicates that a</td></tr><tr><td>model has a mean test score (e.g. F 1 or accuracy)</td></tr><tr><td>of 88.2, with a standard deviation of 0.2, and is</td></tr><tr><td>stochastically dominant over the baseline.</td></tr></table>"
            },
            "TABREF4": {
                "text": "\u00b10.0 82.8 \u00b10.2 88.2 \u00b10.2 83.0 \u00b10.9",
                "type_str": "table",
                "html": null,
                "num": null,
                "content": "<table><tr><td/><td>PropBank</td><td/><td>VerbNet</td><td/></tr><tr><td>System</td><td>WSJ</td><td>Brown</td><td>WSJ</td><td>Brown</td></tr><tr><td colspan=\"2\">Zapirain et al. (2008) 78.9\u00b10.9</td><td/><td>77.0\u00b10.9</td><td>62.9\u00b11.0</td></tr><tr><td>Baseline</td><td>88.5\u00b10.1</td><td>82.4\u00b10.5</td><td>87.4\u00b10.2</td><td>80.1\u00b10.4</td></tr><tr><td>SRL + VSD</td><td>88.2\u00b10.2</td><td colspan=\"2\">82.8 \u00b10.6 87.3\u00b10.1</td><td>80.0\u00b10.7</td></tr><tr><td>SRL | VSD</td><td>88.3\u00b10.2</td><td>82.2\u00b10.4</td><td>87.4\u00b10.2</td><td>80.6 \u00b10.4</td></tr><tr><td>PREDICTED CLASS</td><td>88.3\u00b10.1</td><td>81.2\u00b10.6</td><td colspan=\"2\">87.6 \u00b10.1 80.9 \u00b10.6</td></tr><tr><td>ALL CLASSES</td><td colspan=\"2\">88.6 \u00b10.3 82.3\u00b10.5</td><td colspan=\"2\">87.6 \u00b10.2 81.1 \u00b10.6</td></tr><tr><td>GOLD CLASS</td><td>88.7</td><td/><td/><td/></tr></table>"
            },
            "TABREF5": {
                "text": "F 1 scores of models incorporating different predicate representations and sense distinctions on VerbNet and PropBank SRL on SemLink. SRL + VSD and SRL | VSD are multitask models for SRL and VerbNet classification, with the latter using predicted classes as features for SRL. ALL CLASSES, PREDICTED CLASS, and GOLD CLASS are SRL models using VerbNet class features (the list of all VerbNet classes the predicate belongs to, predicted VerbNet classes, and gold VerbNet classes respectively). VSD 97.7 \u00b10.1 91.3\u00b10.4 SRL | VSD 97.6 \u00b10.1 91.3\u00b10",
                "type_str": "table",
                "html": null,
                "num": null,
                "content": "<table><tr><td>System</td><td>WSJ</td><td>Brown</td></tr><tr><td colspan=\"2\">ClearWSD 97.0\u00b10</td><td>89.3\u00b10</td></tr><tr><td>Baseline</td><td>97.3\u00b10.1</td><td>90.7\u00b10</td></tr><tr><td>SRL +</td><td/><td/></tr></table>"
            },
            "TABREF6": {
                "text": "VerbNet classification (sense disambiguation) accuracy on SemLink.",
                "type_str": "table",
                "html": null,
                "num": null,
                "content": "<table/>"
            },
            "TABREF7": {
                "text": "+0.0) 88.2\u00b10.3 (+0.0) 81.8\u00b10.8 (+0.0) 85.9\u00b10.3 (+0.0) 77.7\u00b11.3 ALL CLASSES (+0.4) 88.6 \u00b10.2 (\u22120.2) 81.6\u00b10.8 (+0.2) 86.1 \u00b10.4 (+2.6) 80.3 \u00b10.8 PREDICTED CLASS (+0.3) 88.5 \u00b10.2 (+0.5) 82.3 \u00b10.8 (+0.2) 86.1 \u00b10.2 (+0.9) 78.6 \u00b11.3",
                "type_str": "table",
                "html": null,
                "num": null,
                "content": "<table><tr><td/><td>Polysemous</td><td/><td>Monosemous</td><td/></tr><tr><td>System</td><td>WSJ</td><td>Brown</td><td>WSJ</td><td>Brown</td></tr><tr><td>Baseline</td><td>(</td><td/><td/><td/></tr></table>"
            },
            "TABREF8": {
                "text": "Evaluation of contribution of VerbNet features on polysemous vs. monosemous predicates for VerbNet SRL averaged over all models. Average change over the baseline performance is given in parentheses.",
                "type_str": "table",
                "html": null,
                "num": null,
                "content": "<table><tr><td/><td>90</td></tr><tr><td/><td>80</td></tr><tr><td>Dev Set F 1</td><td>70</td></tr><tr><td/><td>60</td></tr><tr><td/><td>50</td></tr><tr><td/><td>0</td><td>1-10 11-20 21-30 31-40 41-50</td></tr><tr><td/><td colspan=\"2\">Predicate Frequency in Training Data</td></tr></table>"
            }
        }
    }
}