| { | |
| "paper_id": "I05-1007", | |
| "header": { | |
| "generated_with": "S2ORC 1.0.0", | |
| "date_generated": "2023-01-19T07:26:02.500743Z" | |
| }, | |
| "title": "Parsing the Penn Chinese Treebank with Semantic Knowledge", | |
| "authors": [ | |
| { | |
| "first": "Deyi", | |
| "middle": [], | |
| "last": "Xiong", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "Chinese Academy of Sciences", | |
| "location": { | |
| "postBox": "P.O. Box 2704", | |
| "postCode": "100080", | |
| "settlement": "Beijing", | |
| "country": "China" | |
| } | |
| }, | |
| "email": "dyxiong@ict.ac.cn" | |
| }, | |
| { | |
| "first": "Shuanglong", | |
| "middle": [], | |
| "last": "Li", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "Chinese Academy of Sciences", | |
| "location": { | |
| "postBox": "P.O. Box 2704", | |
| "postCode": "100080", | |
| "settlement": "Beijing", | |
| "country": "China" | |
| } | |
| }, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Qun", | |
| "middle": [], | |
| "last": "Liu", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "Chinese Academy of Sciences", | |
| "location": { | |
| "postBox": "P.O. Box 2704", | |
| "postCode": "100080", | |
| "settlement": "Beijing", | |
| "country": "China" | |
| } | |
| }, | |
| "email": "liuqun@ict.ac.cn" | |
| }, | |
| { | |
| "first": "Shouxun", | |
| "middle": [], | |
| "last": "Lin", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "Chinese Academy of Sciences", | |
| "location": { | |
| "postBox": "P.O. Box 2704", | |
| "postCode": "100080", | |
| "settlement": "Beijing", | |
| "country": "China" | |
| } | |
| }, | |
| "email": "sxlin@ict.ac.cn" | |
| }, | |
| { | |
| "first": "Yueliang", | |
| "middle": [], | |
| "last": "Qian", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "Chinese Academy of Sciences", | |
| "location": { | |
| "postBox": "P.O. Box 2704", | |
| "postCode": "100080", | |
| "settlement": "Beijing", | |
| "country": "China" | |
| } | |
| }, | |
| "email": "" | |
| } | |
| ], | |
| "year": "", | |
| "venue": null, | |
| "identifiers": {}, | |
| "abstract": "We build a class-based selection preference sub-model to incorporate external semantic knowledge from two Chinese electronic semantic dictionaries. This sub-model is combined with modifier-head generation sub-model. After being optimized on the held out data by the EM algorithm, our improved parser achieves 79.4% (F1 measure), as well as a 4.4% relative decrease in error rate on the Penn Chinese Treebank (CTB). Further analysis of performance improvement indicates that semantic knowledge is helpful for nominal compounds, coordination, and N V tagging disambiguation, as well as alleviating the sparseness of information available in treebank.", | |
| "pdf_parse": { | |
| "paper_id": "I05-1007", | |
| "_pdf_hash": "", | |
| "abstract": [ | |
| { | |
| "text": "We build a class-based selection preference sub-model to incorporate external semantic knowledge from two Chinese electronic semantic dictionaries. This sub-model is combined with modifier-head generation sub-model. After being optimized on the held out data by the EM algorithm, our improved parser achieves 79.4% (F1 measure), as well as a 4.4% relative decrease in error rate on the Penn Chinese Treebank (CTB). Further analysis of performance improvement indicates that semantic knowledge is helpful for nominal compounds, coordination, and N V tagging disambiguation, as well as alleviating the sparseness of information available in treebank.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Abstract", | |
| "sec_num": null | |
| } | |
| ], | |
| "body_text": [ | |
| { | |
| "text": "In the recent development of full parsing technology, semantic knowledge is seldom used, though it is known to be useful for resolving syntactic ambiguities. The reasons for this may be twofold. The first one is that it can be very difficult to add additional features which are not available in treebanks to generative models like Collins (see [1] ), which are very popular for full parsing. For smaller tasks, like prepositional phrase attachment disambiguation, semantic knowledge can be incorporated flexibly using different learning algorithms (see [2, 3, 4, 5] ). For full parsing with generative models, however, incorporating semantic knowledge may involve great changes of model structures. The second reason is that semantic knowledge from external dictionaries seems to be noisy, ambiguous and not available in explicit forms, compared with the information from treebanks. Given these two reasons, it seems to be difficult to combine the two different information sources-treebank and semantic knowledge-into one integrated statistical parsing model.", | |
| "cite_spans": [ | |
| { | |
| "start": 345, | |
| "end": 348, | |
| "text": "[1]", | |
| "ref_id": "BIBREF0" | |
| }, | |
| { | |
| "start": 554, | |
| "end": 557, | |
| "text": "[2,", | |
| "ref_id": "BIBREF1" | |
| }, | |
| { | |
| "start": 558, | |
| "end": 560, | |
| "text": "3,", | |
| "ref_id": "BIBREF2" | |
| }, | |
| { | |
| "start": 561, | |
| "end": 563, | |
| "text": "4,", | |
| "ref_id": "BIBREF3" | |
| }, | |
| { | |
| "start": 564, | |
| "end": 566, | |
| "text": "5]", | |
| "ref_id": "BIBREF4" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "One feasible way to solve this problem is to keep the original parsing model unchanged and build an additional sub-model to incorporate semantic knowledge from external dictionaries. The modularity afforded by this approach makes it easier to expand or update semantic knowledge sources with the treebank unchanged or vice versa. Further, the combination of the semantic sub-model and the original parsing model can be optimized automatically.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "In this paper, we build a class-based selection preference sub-model, which is embedded in our lexicalized parsing model, to incorporate external semantic knowledge. We use two Chinese electronic dictionaries and their combination as our semantic information sources. Several experiments are carried out on the Penn Chinese Treebank to test our hypotheses. The results indicate that a significant improvement in performance is achieved when semantic knowledge is incorporated into parsing model. Further improvement analysis is made. We confirm that semantic knowledge is indeed useful for nominal compounds and coordination ambiguity resolution. And surprisingly, semantic knowledge is also helpful to correct Chinese N V mistagging errors mentioned by Levy and Manning (see [12] ). Yet another great benefit to incorporating semantic knowledge is to alleviate the sparseness of information available in treebank.", | |
| "cite_spans": [ | |
| { | |
| "start": 776, | |
| "end": 780, | |
| "text": "[12]", | |
| "ref_id": "BIBREF11" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "Our baseline parsing model is similar to the history-based, generative and lexicalized Model 1 of Collins (see [1] ). In this model, the right hand side of lexicalized rules is decomposed into smaller linguistic objects as follows:", | |
| "cite_spans": [ | |
| { | |
| "start": 111, | |
| "end": 114, | |
| "text": "[1]", | |
| "ref_id": "BIBREF0" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "The Baseline Parser", | |
| "sec_num": "2" | |
| }, | |
| { | |
| "text": "P (h) \u2192 #L n (l n )...L 1 (l 1 )H(h)R 1 (r 1 )...R m (r m )# .", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "The Baseline Parser", | |
| "sec_num": "2" | |
| }, | |
| { | |
| "text": "The uppercase letters are delexicalized nonterminals, while the lowercase letters are lexical items, e.g. head word and head tag (part-of-speech tag of the head word), corresponding to delexicalized nonterminals. H(h) is the head constituent of the rule from which the head lexical item h is derived according to some head percolation rules. 1 The special termination symbol # indicates that there is no more symbols to the left/right. Accordingly, the rule probability is factored into three distributions. The first distribution is the probability of generating the syntactic label of the head constituent of a parent node with label P , head word Hhw and head tag Hht: P r H (H|P, Hht, Hhw) .", | |
| "cite_spans": [ | |
| { | |
| "start": 342, | |
| "end": 343, | |
| "text": "1", | |
| "ref_id": "BIBREF0" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "The Baseline Parser", | |
| "sec_num": "2" | |
| }, | |
| { | |
| "text": "Then each left/right modifier of head constituent is generated in two steps: first its syntactic label M i and corresponding head tag M i ht are chosen given context features from the parent (P ), head constituent (H, Hht, Hhw), previously generated modifier (M i\u22121 , M i\u22121 ht) and other context information like the direction (dir) and distance 2 (dis) to the head constituent:", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "The Baseline Parser", | |
| "sec_num": "2" | |
| }, | |
| { | |
| "text": "P r M (M i , M i ht|HC M ) .", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "The Baseline Parser", | |
| "sec_num": "2" | |
| }, | |
| { | |
| "text": "where the history context HC M is defined as the joint event of", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "The Baseline Parser", | |
| "sec_num": "2" | |
| }, | |
| { | |
| "text": "P, H, Hht, Hhw, M i\u22121 , M i\u22121 ht, dir, dis .", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "The Baseline Parser", | |
| "sec_num": "2" | |
| }, | |
| { | |
| "text": "Then the new modifier's head word M i hw is also generated with the probability:", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "The Baseline Parser", | |
| "sec_num": "2" | |
| }, | |
| { | |
| "text": "P r Mw (M i hw|HC Mw ) .", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "The Baseline Parser", | |
| "sec_num": "2" | |
| }, | |
| { | |
| "text": "where the history context HC Mw is defined as the joint event of", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "The Baseline Parser", | |
| "sec_num": "2" | |
| }, | |
| { | |
| "text": "P, H, Hht, Hhw, M i\u22121 , M i\u22121 ht, dir, dis, M i , M i ht .", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "The Baseline Parser", | |
| "sec_num": "2" | |
| }, | |
| { | |
| "text": "All the three distributions are smoothed through Witten-Bell interpolation just like Collins (see [1] ). For the distribution P r M , we build back-off structures with six levels, which are different from Collins' since we find our back-off structures work better than the three-level back-off structures of Collins. For the distribution P r Mw , the parsing model backs off to the history context with head word Hhw removed, then to the modifier head tag M i ht, just like Collins. Gildea (see [9] ) and Bikel (see [10] ) both observed that the effect of bilexical dependencies is greatly impaired due to the sparseness of bilexical statistics. Bikel even found that the parser only received an estimate that made use of bilexical statistics a mere 1.49% of the time. However, according to the wisdom of the parsing community, lexical bigrams, the word pairs (M i hw, Hhw) are very informative with semantic constraints. Along this line, in this paper, we build an additional class-based selectional preference sub-model, which is described in section 3, to make good use of this semantic information through selectional restrictions between head and modifier words.", | |
| "cite_spans": [ | |
| { | |
| "start": 98, | |
| "end": 101, | |
| "text": "[1]", | |
| "ref_id": "BIBREF0" | |
| }, | |
| { | |
| "start": 495, | |
| "end": 498, | |
| "text": "[9]", | |
| "ref_id": "BIBREF8" | |
| }, | |
| { | |
| "start": 516, | |
| "end": 520, | |
| "text": "[10]", | |
| "ref_id": "BIBREF9" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "The Baseline Parser", | |
| "sec_num": "2" | |
| }, | |
| { | |
| "text": "Our parser takes segmented but untagged sentences as input. The probability of unknown words, P r(uword|tag), is estimated based on the first character of the word and if the first characters are unseen, the probability is estimated by absolute discounting.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "The Baseline Parser", | |
| "sec_num": "2" | |
| }, | |
| { | |
| "text": "We do some linguistically motivated re-annotations for the baseline parser. The first one is marking non-recursive noun phrases from other common noun phrases without introducing any extra unary levels (see [1, 8] ). We find this basic NP re-annotation very helpful for the performance. We think it is because of the annotation style of the Upenn Chinese Treebank (CTB). According to Xue et al. (see [11] ), noun-noun compounds formed by an uninterrupted sequence of words POS-tagged as NNs are always left flat because of difficulties in determining which modifies which. The second re-annotation is marking basic VPs, which we think is beneficial for reducing multilevel VP adjunction ambiguities (see [12] ).", | |
| "cite_spans": [ | |
| { | |
| "start": 207, | |
| "end": 210, | |
| "text": "[1,", | |
| "ref_id": "BIBREF0" | |
| }, | |
| { | |
| "start": 211, | |
| "end": 213, | |
| "text": "8]", | |
| "ref_id": "BIBREF7" | |
| }, | |
| { | |
| "start": 400, | |
| "end": 404, | |
| "text": "[11]", | |
| "ref_id": "BIBREF10" | |
| }, | |
| { | |
| "start": 704, | |
| "end": 708, | |
| "text": "[12]", | |
| "ref_id": "BIBREF11" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "The Baseline Parser", | |
| "sec_num": "2" | |
| }, | |
| { | |
| "text": "To speed up parsing, we use the beam thresholding techniques in Xiong et al. (see [13] ). In all cases, the thresholding for completed edges is set at ct = 9 and incomplete edges at it = 7. The performance of the baseline parser is 78.5% in terms of F1 measure of labeled parse constituents on the same CTB training and test sets with Bikel et al. (see [14] )", | |
| "cite_spans": [ | |
| { | |
| "start": 82, | |
| "end": 86, | |
| "text": "[13]", | |
| "ref_id": "BIBREF12" | |
| }, | |
| { | |
| "start": 353, | |
| "end": 357, | |
| "text": "[14]", | |
| "ref_id": "BIBREF13" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "The Baseline Parser", | |
| "sec_num": "2" | |
| }, | |
| { | |
| "text": "In this section, we describe how to incorporate semantic knowledge from external semantic dictionaries into parsing model to improve the performance. Firstly, we extract semantic categories through two Chinese electronic semantic dictionaries and some heuristic rules. Then we build a selection preference sub-model based on extracted semantic categories. In section 3.3, we present our experiments and results in detail. And finally, we compare parses from baseline parser with those from the new parser incorporated with semantic knowledge. We empirically confirm that semantic knowledge is helpful for nominal compound, coordination and POS tagging ambiguity resolution. Additionally, we also find that semantic knowledge can greatly alleviate problems caused by data sparseness.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Incorporating Semantic Knowledge", | |
| "sec_num": "3" | |
| }, | |
| { | |
| "text": "Semantic knowledge is not presented in treebanks and therefore has to be extracted from external knowledge sources. We have two Chinese electronic semantic dictionaries, both are good knowledge sources for us to extract semantic categories. One is the HowNet dictionary 3 , which covers 67,440 words defined by 2112 different sememes. The other is the \"TongYiCi CiLin\" expanded version (henceforth CiLin) 4 , which represents 77,343 words in a dendrogram.", | |
| "cite_spans": [ | |
| { | |
| "start": 405, | |
| "end": 406, | |
| "text": "4", | |
| "ref_id": "BIBREF3" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Extracting Semantic Categories", | |
| "sec_num": "3.1" | |
| }, | |
| { | |
| "text": "Each sememe defined by the HowNet is regarded as a semantic category. And through the hypernym-hyponym relation between different categories, we can extract semantic categories at various granularity levels. Since words may have different senses, and therefore different definitions in HowNet, we just use the first definition of words in HowNet. At the first level HN1, we extract the first definitions and use them as semantic categories of words. Through the hypernym ladders, we can get HN2, HN3, by replacing categories at lower level with their hypernyms at higher level. Table 1 shows information about words and extracted categories at different levels.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 578, | |
| "end": 585, | |
| "text": "Table 1", | |
| "ref_id": "TABREF0" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "HowNet (HN):", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "CL is a branching diagram, where each node represents a semantic category. There are three levels in total, and from the top down, 12 categories in the first level (CL1), 97 categories in the second level (CL2), 1400 categories in the third level (CL3). We extract semantic categories at level CL1, CL2 and CL3.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "CiLin (CL):", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "HowNet+CiLin: Since the two dictionaries have different ontologies and representations of semantic categories, we establish a strategy to combine them: HowNet is used as a primary dictionary, and CiLin as a secondary dictionary. If a word is not found in HowNet but found in Cilin, we will look up other words from its synset defined by CiLin in HowNet. If HowNet query succeeds, the corresponding semantic category in HowNet will be assigned to this word. According to our experimental results, we choose HN2 as the primary semantic category set and combine it with CL1, CL2 and CL3.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "CiLin (CL):", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "Heuristic Rules (HR): Numbers and time expressions are recognized using simple heuristic rules. For a better recognition, one can define accurate regular expressions. However, we just collect suffixes and feature characters to match strings. For example, Chinese numbers are strings whose characters all come from a predefined set. These two classes are merged into HowNet and labelled by semantic categories from HowNet.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "CiLin (CL):", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "In our experiments, we combine HN2, CL1/2/3, and HR as our external sources. In these combinations {HN2+CL1/2/3+HR}, all semantic classes come from the primary semantic category set HN2, therefore we get the same class coverage that we obtain from the single source HN2 but a bigger word coverage. The number of covered words of these combinations in {train, test, both} is {7911, 1672, 1372} respectively.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "CiLin (CL):", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "There are several ways to incorporate semantic knowledge into parsing model. Bikel (see [15] ) suggested a way to capture semantic preferences by employing bilexical-class statistics, in other words, dependencies among head-modifier word classes. Bikel did not carry it out and therefore greater details are not available. However, the key point, we think, is to use classes extracted from semantic dictionary, instead of words, to model semantic dependencies between head and modifier. Accordingly, we build a similar bilexical-class sub-model as follows:", | |
| "cite_spans": [ | |
| { | |
| "start": 88, | |
| "end": 92, | |
| "text": "[15]", | |
| "ref_id": "BIBREF14" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Building Class-Based Selection Preference Sub-model", | |
| "sec_num": "3.2" | |
| }, | |
| { | |
| "text": "P r class (C Mihw |C Hhw , Hht, M i ht, dir) .", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Building Class-Based Selection Preference Sub-model", | |
| "sec_num": "3.2" | |
| }, | |
| { | |
| "text": "where C Mihw and C Hhw represent semantic categories of words M i hw and Hhw, respectively. This model is combined with sub-model P r Mw to form a mixture model P mix :", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Building Class-Based Selection Preference Sub-model", | |
| "sec_num": "3.2" | |
| }, | |
| { | |
| "text": "EQUATION", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [ | |
| { | |
| "start": 0, | |
| "end": 8, | |
| "text": "EQUATION", | |
| "ref_id": "EQREF", | |
| "raw_str": "P r mix = \u03bbP r Mw + (1 \u2212 \u03bb)P r class .", | |
| "eq_num": "(1)" | |
| } | |
| ], | |
| "section": "Building Class-Based Selection Preference Sub-model", | |
| "sec_num": "3.2" | |
| }, | |
| { | |
| "text": "\u03bb is hand-optimized, and an improvement of about 0.5% in terms of F1 measure is gained. However, even a very slight change in the value of \u03bb, e.g. 0.001, will have a great effect on the performance. Besides, it seems that the connection between entropy, i.e. the total negative logarithm of the inside probability of trees, and F1 measure, is lost, while this relation is observed in many experiments. Therefore, automatic optimization algorithms, like EM, can not work in this mixture model. The reason, we guess, is that biclass dependencies among head-modifier word classes seem too coarse-grained to capture semantic preferences between head and modifier. In most cases, a head word has a strong semantic constraints on the concept \u03ba of mw, one of its modifier words, but that doesn't mean other words in the same class with the head word has the same semantic preferences on the concept \u03ba. For example, the verb eat impose a selection restriction on its object modifier 5 : it has to be solid food. On the other hand, the verb drink specifies its object modifier to be liquid beverage. At the level HN2, verb eat and drink have the same semantic category metabolize. However, they impose different selection preferences on their PATIENT roles.", | |
| "cite_spans": [ | |
| { | |
| "start": 975, | |
| "end": 976, | |
| "text": "5", | |
| "ref_id": "BIBREF4" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Building Class-Based Selection Preference Sub-model", | |
| "sec_num": "3.2" | |
| }, | |
| { | |
| "text": "To sum up, bilexical dependencies are too fine-grained when being used to capture semantic preferences and therefore lead to serious data sparseness. Biclass dependencies, which result in an unstable performance improvement, on the other hand, seem to be too coarse-grained for semantic preferences. We build a class-based selection preference model:", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Building Class-Based Selection Preference Sub-model", | |
| "sec_num": "3.2" | |
| }, | |
| { | |
| "text": "P r sel (C Mihw |Hhw, P ) .", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Building Class-Based Selection Preference Sub-model", | |
| "sec_num": "3.2" | |
| }, | |
| { | |
| "text": "This model is similar to Resnik (see [2] ). We use the parent node label P to represent the grammatical relation between head and modifier. Besides, in this model, only modifier word is replaced with its semantic category. The dependencies between head word and modifier word class seem to be just right for capturing these semantic preferences.", | |
| "cite_spans": [ | |
| { | |
| "start": 37, | |
| "end": 40, | |
| "text": "[2]", | |
| "ref_id": "BIBREF1" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Building Class-Based Selection Preference Sub-model", | |
| "sec_num": "3.2" | |
| }, | |
| { | |
| "text": "The final mixture model is the combination of the class-based selection preference sub-model P r sel and modifier-head generation sub-model P r Mw :", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Building Class-Based Selection Preference Sub-model", | |
| "sec_num": "3.2" | |
| }, | |
| { | |
| "text": "EQUATION", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [ | |
| { | |
| "start": 0, | |
| "end": 8, | |
| "text": "EQUATION", | |
| "ref_id": "EQREF", | |
| "raw_str": "P r mix = \u03bbP r Mw + (1 \u2212 \u03bb)P r sel .", | |
| "eq_num": "(2)" | |
| } | |
| ], | |
| "section": "Building Class-Based Selection Preference Sub-model", | |
| "sec_num": "3.2" | |
| }, | |
| { | |
| "text": "Since the connection between entropy and F1 measure is observed again, EM algorithm is used to optimize \u03bb. Just like Levy (see [12] ), we set aside articles 1-25 in CTB as held out data for EM algorithm and use articles 26-270 as training data during \u03bb optimization.", | |
| "cite_spans": [ | |
| { | |
| "start": 127, | |
| "end": 131, | |
| "text": "[12]", | |
| "ref_id": "BIBREF11" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Building Class-Based Selection Preference Sub-model", | |
| "sec_num": "3.2" | |
| }, | |
| { | |
| "text": "We have designed several experiments to check the power of our class-based selection preference model with different semantic data sources. In all experiments, we first use the EM algorithm to optimize the parameter \u03bb. As mentioned above, during parameter optimization, articles 1-25 are used as held out data and articles 26-270 are used as training data. Then we test our mixture model with optimized parameter \u03bb using the training data of articles 1-270 and test data of articles 271-300 of length at most 40 words. Firstly, we carry out experiments on HowNet and CiLin, separately. Experimental results are presented in Table 2 . As can be seen, CiLin has a greater coverage of words than that of HowNet, however, it works worse than HowNet. And at the level CL1, coarse-grained classes even yield degraded results. It's difficult to explain this, but the main reason may be that HowNet has a fine-grained and substantial ontology while CiLin is designed only as a synset container.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 624, | |
| "end": 631, | |
| "text": "Table 2", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Experimental Results", | |
| "sec_num": "3.3" | |
| }, | |
| { | |
| "text": "Since HowNet has a better semantic representation and CiLin better coverage, we want to combine them. The combination is described in Sect. 3.1, where HN2 is used as the primary semantic category set. Words found by CiLin and heuristic rules are labelled by semantic categories from HN2. Results are shown in Table 3 . Although external sources HN2+CL1/2/3+HR have the identical word coverage and yield exactly the same number of classes, the different word-class distributions in them lead to the different results.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 309, | |
| "end": 316, | |
| "text": "Table 3", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Experimental Results", | |
| "sec_num": "3.3" | |
| }, | |
| { | |
| "text": "Due to the combination of HN2, CL2 and HR, we see that our new parser with external semantic knowledge outperforms the baseline parser by 0.9% in F1 measure. Given we are already at the 78% level of accuracy, an improvement of 0.9% is well worth obtaining and confirms the importance of semantic dependencies on parsing. Further, we do the significance test using Bikel's significance tester 6 which is modified to output p-value for F1. The significance level for F-score is at most (43376 + 1)/(1048576 + 1) = 0.041. A second 1048576 iteration produces the similar result. Therefore the improvement is statistically significant.", | |
| "cite_spans": [ | |
| { | |
| "start": 392, | |
| "end": 393, | |
| "text": "6", | |
| "ref_id": "BIBREF5" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Experimental Results", | |
| "sec_num": "3.3" | |
| }, | |
| { | |
| "text": "We manually analyze parsing errors of the baseline parser (BP ) as well as performance improvement of the new parser (IP ) with semantic knowledge from the combination of HN2, CL2 and HR. Improvement analysis can provide an additional valuable perspective: how semantic knowledge helps to resolve some ambiguities. We compare BP and IP on the test data parse by parse. There are 299 sentences of length at most 40 words among the total 348 test sentences. The two parsers BP and IP found different parses for 102 sentences, among which IP yields better parse trees for 47 sentences according to the gold standard trees. We have concentrated on these 47 sentences and compared parse trees found by IP with those found by BP . Frequencies of major types of parsing improvement is presented in Table 4 . Levy and Manning (see [12] )(henceforth L&M) observed the top three parsing error types: NP-NP modification, Coordination and N V mistagging, which are also common in our baseline parser. As can be seen, our improved parser can address these types of ambiguities to some extent through semantic knowledge.", | |
| "cite_spans": [ | |
| { | |
| "start": 823, | |
| "end": 827, | |
| "text": "[12]", | |
| "ref_id": "BIBREF11" | |
| } | |
| ], | |
| "ref_spans": [ | |
| { | |
| "start": 791, | |
| "end": 798, | |
| "text": "Table 4", | |
| "ref_id": "TABREF2" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Performance Improvement Analysis", | |
| "sec_num": "3.4" | |
| }, | |
| { | |
| "text": "Nominal Compounds (NCs) Disambiguation: Nominal compounds are notorious \"every way ambiguous\" constructions. 7 The different semantic interpretations have different dependency structures. According to L&M, this ambiguity will be addressed by the dependency model when word frequencies are large enough to be reliable. However, even for the treebank central to a certain topic, many very plausible dependencies occur only once. 8 A good technique for resolving this conflict is to generalize the dependencies from word pairs to wordclass pairs. Such generalized dependencies, as noted in section 3.2, can capture semantic preferences, as well as alleviate the data sparseness associated with standard bilexical statistics. In our class-based selection preference model, if the frequency of pair [C Mhw , Hhw] 9 is large enough, the parser can interpret nominal compounds correctly, that is, it can tell which modify which. NCs are always parsed as flatter structures by our baseline parser, just like the tree a. in Figure 1 . This is partly because of the annotation style of CTB, where there is no NP-internal structure. For these NCs without internal analysis, we re-annotated them as basic NPs with label NPB, as mentioned in section 2. This re-annotation really helps. Another reason is that the baseline parser, or the modifier word generating sub-model P Mw , can not capture hierarchical semantic dependencies of internal structures of NCs due to the sparseness of bilexical dependencies. In our new parser, however, the selection preference model is able to build semantically preferable structures through word-class dependency statistics. For NCs like (n 1 , n 2 , n 3 ), where n i is a noun, dependency structures 7 \"Every way ambiguous\" constructions are those for which the number of analyses is the number of binary trees over the terminal elements. Prepositional phrase attachment, coordination, and nominal compounds are all \"every way ambiguous\" constructions. 8 Just as Klein et al. (see [8] ) said, one million words of training data just isn't enough. 9 Henceforth, [s1, s2] denotes a dependency structure, where s1 is a modifier word or its semantic class (C), and s2 is the head word. Coordination Disambiguation: Coordination is another kind of \"every way ambiguous\" construction. For coordination structures, the head word is meaningless. But that doesn't matter, since semantic dependency between the spurious head and modifier will be used to measure the meaning similarity of coordinated structures. Therefore, our selection preference model still works in coordination constructions. We have also found VP coordination ambiguity, which is similar to that observed by L&M. The latter VP in coordinated VPs is often parsed as an IP due to pro-drop by the baseline parser. That is, the coordinated structure VP is parsed as: V P 0 \u2192 V P 1 IP 2 . This parse will be penalized by the selection preference model because the hypothesis that the head word of IP 2 has a similar meaning to the head word of V P 1 under the grammatical relation V P 0 is infrequent.", | |
| "cite_spans": [ | |
| { | |
| "start": 109, | |
| "end": 110, | |
| "text": "7", | |
| "ref_id": "BIBREF6" | |
| }, | |
| { | |
| "start": 427, | |
| "end": 428, | |
| "text": "8", | |
| "ref_id": "BIBREF7" | |
| }, | |
| { | |
| "start": 1978, | |
| "end": 1979, | |
| "text": "8", | |
| "ref_id": "BIBREF7" | |
| }, | |
| { | |
| "start": 2006, | |
| "end": 2009, | |
| "text": "[8]", | |
| "ref_id": "BIBREF7" | |
| }, | |
| { | |
| "start": 2072, | |
| "end": 2073, | |
| "text": "9", | |
| "ref_id": "BIBREF8" | |
| } | |
| ], | |
| "ref_spans": [ | |
| { | |
| "start": 1015, | |
| "end": 1023, | |
| "text": "Figure 1", | |
| "ref_id": "FIGREF0" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Performance Improvement Analysis", | |
| "sec_num": "3.4" | |
| }, | |
| { | |
| "text": "The lack of overt morphological marking for transforming verbal words to nominal words in Chinese results in ambiguity between these two categories. L&M argued that the way to resolve this ambiguity is to look at more external context, like some function words, e.g. adverbial or prenominal modifiers, co-occurring with N V-ambiguous words. However, in some cases, N V-ambiguous words can be tagged correctly without external context. Chen et al. (see [16] ) studied the pattern of N V+noun, which will be analyzed as a predicate-object structure if N V is a verb and a modifiernoun structure if N V is a noun. They found that in most cases, this pattern can be parsed correctly without any external context. Furthermore, they argued that semantic preferences are helpful for the resolution of ambiguity between these two different structures. In our selection preference model, semantic preferences interweave with grammatical relations. These semantic dependencies impose constraints on the structure of the pattern N V+noun and therefore on the POS tag of N V. Figure 2 shows our new parser can correct N V mistagging errors occurring in the pattern of N V+noun.", | |
| "cite_spans": [ | |
| { | |
| "start": 452, | |
| "end": 456, | |
| "text": "[16]", | |
| "ref_id": "BIBREF15" | |
| } | |
| ], | |
| "ref_spans": [ | |
| { | |
| "start": 1064, | |
| "end": 1072, | |
| "text": "Figure 2", | |
| "ref_id": "FIGREF1" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "N V-ambiguous Tagging Disambiguation:", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "Smoothing: Besides the three ambiguity resolution noted above, semantic knowledge indeed helps alleviate the fundamental sparseness of the lexical dependency information available in the CTB. For many word pairs [mod,head] , whose count information is not available in the training data, the dependency statistics of head and modifier can still work through the semantic category of mod. During our manual analysis of performance improvement, many other structural ambiguities are addressed due to the smoothing function of semantic knowledge.", | |
| "cite_spans": [ | |
| { | |
| "start": 212, | |
| "end": 222, | |
| "text": "[mod,head]", | |
| "ref_id": null | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "N V-ambiguous Tagging Disambiguation:", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "Previous work on CTB parsing and their results are shown in table 5. Bikel and Chiang (see [14] ) used two different models on CTB, one based on the modified BBN model which is very similar to our baseline model, the other on Tree Insertion Grammar (TIG). While our baseline model used the same unknown word threshold with Bikel and Chiang but smaller beam width, our result outperforms theirs due to other features like distance, basic NP re-annotation used by our baseline model. Levy and Manning (see [12] ) used a factored model with rich re-annotations guided by error analysis. In the baseline model, we also used several re-annotations but find most re-annotations they suggested do not fit our model. The three parsing error types expounded above are also found by L&M. However, we used more efficient measures to keep our improved model from these errors.", | |
| "cite_spans": [ | |
| { | |
| "start": 91, | |
| "end": 95, | |
| "text": "[14]", | |
| "ref_id": "BIBREF13" | |
| }, | |
| { | |
| "start": 504, | |
| "end": 508, | |
| "text": "[12]", | |
| "ref_id": "BIBREF11" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Related Work on CTB Parsing", | |
| "sec_num": "4" | |
| }, | |
| { | |
| "text": "The work of Bikel thesis (see [10] ) emulated Collins' model and created a language package to Chinese parsing. He used subcat frames and an additional POS tagger for unseen words. Chiang and Bikel (see [17] ) used the EM algorithm on the same TIG-parser to improve the head percolation table for Chinese parsing. Both these two parsers used fine-tuned features recovered from the treebank that our model does not use. This leads to better results and indicates that there is still room of improvement for our model.", | |
| "cite_spans": [ | |
| { | |
| "start": 30, | |
| "end": 34, | |
| "text": "[10]", | |
| "ref_id": "BIBREF9" | |
| }, | |
| { | |
| "start": 203, | |
| "end": 207, | |
| "text": "[17]", | |
| "ref_id": "BIBREF16" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Related Work on CTB Parsing", | |
| "sec_num": "4" | |
| }, | |
| { | |
| "text": "We have shown that how semantic knowledge may be incorporated into a generative model for full parsing, which reaches 79.4% in CTB. Experimental results are quite consistent with our intuition. After the manual analysis of performance improvement, the working mechanism of semantic knowledge in the selection preference model is quite clear:", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Conclusions", | |
| "sec_num": "5" | |
| }, | |
| { | |
| "text": "1. Using semantic categories extracted from external dictionaries, the classbased selection preference model first generalizes standard bilexical dependencies, some of which are not available in training data, to word-class dependencies. These dependencies are neither too fine-grained nor too coarsegrained compared with bilexical and biclass dependencies, and really help to alleviate fundamental information sparseness in treebank. 2. Based on the generalized word-class pairs, semantic dependencies are captured and used to address different kinds of ambiguities, like nominal compounds, coordination construction, even N V-ambiguous words tagging.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Conclusions", | |
| "sec_num": "5" | |
| }, | |
| { | |
| "text": "Our experiments show that generative models have room for improvement by employing semantic knowledge. And that may be also true for discriminative models, since these models can easily incorporate richer features in a wellfounded fashion. This is the subject of our future work.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Conclusions", | |
| "sec_num": "5" | |
| }, | |
| { | |
| "text": "Here we use the modified head percolation table for Chinese from Xia (see[6]).2 Our distance definitions are different for termination symbol and non-termination symbol, which are similar to Klein and Manning (see[7]).", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "http://www.keenage.com/.4 The dictionary is recorded and expanded by Information Retrieval Laboratory, Harbin Institute of Technology.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "According to Thematic Role theory, this modifier has a PATIENT role.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "See http://www.cis.upenn.edu/ dbikel/software.html", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| } | |
| ], | |
| "back_matter": [ | |
| { | |
| "text": "This work was supported in part by National High Technology Research and Development Program under grant #2001AA114010 and #2003AA111010. We would like to acknowledge anonymous reviewers who provided helpful suggestions.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Acknowledgements", | |
| "sec_num": null | |
| } | |
| ], | |
| "bib_entries": { | |
| "BIBREF0": { | |
| "ref_id": "b0", | |
| "title": "Head-Driven Statistical Models for Natural Language Parsing", | |
| "authors": [ | |
| { | |
| "first": "Michael", | |
| "middle": [], | |
| "last": "Collins", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1999, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Michael Collins. 1999. Head-Driven Statistical Models for Natural Language Pars- ing. PhD thesis, University of Pennsylvania.", | |
| "links": null | |
| }, | |
| "BIBREF1": { | |
| "ref_id": "b1", | |
| "title": "Selection and Information: A Class-Based Approach to Lexical Relationships", | |
| "authors": [ | |
| { | |
| "first": "Philip", | |
| "middle": [], | |
| "last": "Stuart Resnik", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1993, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Philip Stuart Resnik. 1993. Selection and Information: A Class-Based Approach to Lexical Relationships. PhD thesis, University of Pennsylvania, Philadelphia, PA, USA.", | |
| "links": null | |
| }, | |
| "BIBREF2": { | |
| "ref_id": "b2", | |
| "title": "An Application of WordNet to Prepositional Attachement", | |
| "authors": [ | |
| { | |
| "first": "Sanda", | |
| "middle": [], | |
| "last": "Harabagiu", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1996, | |
| "venue": "Proceedings of ACL-96", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "360--363", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Sanda Harabagiu. 1996. An Application of WordNet to Prepositional Attachement. In Proceedings of ACL-96, June 1996, Santa Cruz CA, pages 360-363.", | |
| "links": null | |
| }, | |
| "BIBREF3": { | |
| "ref_id": "b3", | |
| "title": "Incorporating Knowledge in Natural Language Learning: A Case Study", | |
| "authors": [ | |
| { | |
| "first": "Yuval", | |
| "middle": [], | |
| "last": "Krymolowski", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Dan", | |
| "middle": [], | |
| "last": "Roth", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1998, | |
| "venue": "COLING-ACL'98 Workshop on Usage of Word-Net in Natural Language Processing Systems", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Yuval Krymolowski and Dan Roth. 1998. Incorporating Knowledge in Natural Lan- guage Learning: A Case Study. In COLING-ACL'98 Workshop on Usage of Word- Net in Natural Language Processing Systems,Montreal, Canada.", | |
| "links": null | |
| }, | |
| "BIBREF4": { | |
| "ref_id": "b4", | |
| "title": "Thesauruses for Prepositional Phrase Attachment", | |
| "authors": [ | |
| { | |
| "first": "Mark", | |
| "middle": [], | |
| "last": "Mclauchlan", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2004, | |
| "venue": "Proceedings of CoNLL-2004", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "73--80", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Mark McLauchlan. 2004. Thesauruses for Prepositional Phrase Attachment. In Proceedings of CoNLL-2004,Boston, MA, USA, 2004, pp. 73-80.", | |
| "links": null | |
| }, | |
| "BIBREF5": { | |
| "ref_id": "b5", | |
| "title": "Automatic Grammar Generation from Two Different Perspectives", | |
| "authors": [ | |
| { | |
| "first": "Fei", | |
| "middle": [], | |
| "last": "Xia", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1999, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Fei Xia. 1999. Automatic Grammar Generation from Two Different Perspectives. PhD thesis, University of Pennsylvania.", | |
| "links": null | |
| }, | |
| "BIBREF6": { | |
| "ref_id": "b6", | |
| "title": "Fast Exact Natural Language Parsing with a Factored Model", | |
| "authors": [ | |
| { | |
| "first": "Dan", | |
| "middle": [], | |
| "last": "Klein", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Christopher", | |
| "middle": [ | |
| "D" | |
| ], | |
| "last": "Manning", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2002, | |
| "venue": "Advances in Neural Information Processing Systems", | |
| "volume": "15", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Dan Klein and Christopher D. Manning. 2002. Fast Exact Natural Language Pars- ing with a Factored Model. In Advances in Neural Information Processing Systems 15 (NIPS-2002).", | |
| "links": null | |
| }, | |
| "BIBREF7": { | |
| "ref_id": "b7", | |
| "title": "Accurate Unlexicalized Parsing", | |
| "authors": [ | |
| { | |
| "first": "Dan", | |
| "middle": [], | |
| "last": "Klein", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Christopher", | |
| "middle": [ | |
| "D" | |
| ], | |
| "last": "Manning", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2003, | |
| "venue": "Proceedings of ACL-03", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Dan Klein and Christopher D. Manning. 2003. Accurate Unlexicalized Parsing. In Proceedings of ACL-03.", | |
| "links": null | |
| }, | |
| "BIBREF8": { | |
| "ref_id": "b8", | |
| "title": "Corpus variation and parser performance", | |
| "authors": [ | |
| { | |
| "first": "Daniel", | |
| "middle": [], | |
| "last": "Gildea", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2001, | |
| "venue": "Proceedings of EMNLP-01", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Daniel Gildea. 2001. Corpus variation and parser performance. In Proceedings of EMNLP-01, Pittsburgh, Pennsylvania.", | |
| "links": null | |
| }, | |
| "BIBREF9": { | |
| "ref_id": "b9", | |
| "title": "On the Parameter Space of Generative Lexicalized Statistical Parsing Models", | |
| "authors": [ | |
| { | |
| "first": "M", | |
| "middle": [], | |
| "last": "Daniel", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Bikel", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2004, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Daniel M. Bikel. 2004a. On the Parameter Space of Generative Lexicalized Statis- tical Parsing Models. PhD thesis, University of Pennsylvania.", | |
| "links": null | |
| }, | |
| "BIBREF10": { | |
| "ref_id": "b10", | |
| "title": "The Bracketing Guidelines for Chinese Treebank Project", | |
| "authors": [ | |
| { | |
| "first": "Nianwen", | |
| "middle": [], | |
| "last": "Xue", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Fei", | |
| "middle": [], | |
| "last": "Xia", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2000, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Nianwen Xue and Fei Xia. 2000. The Bracketing Guidelines for Chinese Treebank Project. Technical Report IRCS 00-08, University of Pennsylvania.", | |
| "links": null | |
| }, | |
| "BIBREF11": { | |
| "ref_id": "b11", | |
| "title": "Is it harder to parse Chinese, or the Chinese Treebank?", | |
| "authors": [ | |
| { | |
| "first": "Roger", | |
| "middle": [], | |
| "last": "Levy", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Christopher", | |
| "middle": [], | |
| "last": "Manning", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2003, | |
| "venue": "Proceedings of ACL-03", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Roger Levy and Christopher Manning. 2003. Is it harder to parse Chinese, or the Chinese Treebank? In Proceedings of ACL-03.", | |
| "links": null | |
| }, | |
| "BIBREF12": { | |
| "ref_id": "b12", | |
| "title": "Lexicalized Beam Thresholding Parsing with Prior and Boundary Estimates", | |
| "authors": [ | |
| { | |
| "first": "Deyi", | |
| "middle": [], | |
| "last": "Xiong", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Qun", | |
| "middle": [], | |
| "last": "Liu", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Shouxun", | |
| "middle": [], | |
| "last": "Lin", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2005, | |
| "venue": "Proceedings of the 6th Conference on Intelligent Text Processing and Computational Linguistics (CICLing)", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Deyi Xiong, Qun Liu and Shouxun Lin. 2005. Lexicalized Beam Thresholding Pars- ing with Prior and Boundary Estimates. In Proceedings of the 6th Conference on Intelligent Text Processing and Computational Linguistics (CICLing), Mexico City, Mexico, 2005.", | |
| "links": null | |
| }, | |
| "BIBREF13": { | |
| "ref_id": "b13", | |
| "title": "Two statistical parsing models applied to the chinese treebank", | |
| "authors": [ | |
| { | |
| "first": "M", | |
| "middle": [], | |
| "last": "Daniel", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "David", | |
| "middle": [], | |
| "last": "Bikel", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Chiang", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2000, | |
| "venue": "Proceedings of the Second Chinese Language Processing Workshop", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "1--6", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Daniel M. Bikel and David Chiang. 2000. Two statistical parsing models applied to the chinese treebank. In Proceedings of the Second Chinese Language Processing Workshop, pages 1-6.", | |
| "links": null | |
| }, | |
| "BIBREF14": { | |
| "ref_id": "b14", | |
| "title": "Intricacies of Collins' Parsing Model", | |
| "authors": [ | |
| { | |
| "first": "M", | |
| "middle": [], | |
| "last": "Daniel", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Bikel", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2004, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Daniel M. Bikel. 2004b. Intricacies of Collins' Parsing Model. to appear in Com- putational Linguistics.", | |
| "links": null | |
| }, | |
| "BIBREF15": { | |
| "ref_id": "b15", | |
| "title": "Resolving Ambiguities of Predicate-object and Modifier-noun Structures for Chinese V-N Patterns. in Chinese", | |
| "authors": [ | |
| { | |
| "first": "Kejian", | |
| "middle": [], | |
| "last": "Chen", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Weimei", | |
| "middle": [], | |
| "last": "Hong", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1996, | |
| "venue": "Communication of COLIPS", | |
| "volume": "6", | |
| "issue": "", | |
| "pages": "73--79", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Kejian Chen and Weimei Hong. 1996. Resolving Ambiguities of Predicate-object and Modifier-noun Structures for Chinese V-N Patterns. in Chinese. In Communi- cation of COLIPS, Vol.6, #2, pages 73-79.", | |
| "links": null | |
| }, | |
| "BIBREF16": { | |
| "ref_id": "b16", | |
| "title": "Recovering Latent Information in Treebanks", | |
| "authors": [ | |
| { | |
| "first": "David", | |
| "middle": [], | |
| "last": "Chiang", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Daniel", | |
| "middle": [ | |
| "M" | |
| ], | |
| "last": "Bikel", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2002, | |
| "venue": "proceedings of COLING", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "David Chiang and Daniel M. Bikel. 2002. Recovering Latent Information in Tree- banks. In proceedings of COLING,2002.", | |
| "links": null | |
| } | |
| }, | |
| "ref_entries": { | |
| "FIGREF0": { | |
| "num": null, | |
| "uris": null, | |
| "type_str": "figure", | |
| "text": "Nominal Compounds: The North Korean government's special envoy. a. is the incorrect flat parse, b. is the right one in corpus" | |
| }, | |
| "FIGREF1": { | |
| "num": null, | |
| "uris": null, | |
| "type_str": "figure", | |
| "text": "N V-ambiguity: a. implement plans (incorrect parse) versus b. implementation plans (corpus)" | |
| }, | |
| "TABREF0": { | |
| "content": "<table><tr><td/><td colspan=\"3\">Data HN1 HN2 HN3 CL1 CL2 CL3</td></tr><tr><td colspan=\"2\">words in train 9522</td><td>6040</td><td>6469</td></tr><tr><td colspan=\"2\">words in test 1824</td><td>1538</td><td>1581</td></tr><tr><td colspan=\"2\">words in both 1412</td><td>1293</td><td>1310</td></tr><tr><td colspan=\"4\">classes in train -1054 381 118 12 92 1033</td></tr><tr><td>classes in test</td><td>-</td><td colspan=\"2\">520 251 93 12 79 569</td></tr><tr><td colspan=\"2\">classes in both -</td><td colspan=\"2\">504 248 93 12 79 552</td></tr></table>", | |
| "html": null, | |
| "num": null, | |
| "text": "Sizes and coverage of words and semantic categories from different semantic knowledge sources", | |
| "type_str": "table" | |
| }, | |
| "TABREF1": { | |
| "content": "<table><tr><td colspan=\"4\">Baseline HN2+CL1+HR HN2+CL2+HR HN2+CL3+HR</td></tr><tr><td>F1(%) 78.5</td><td>79.2</td><td>79.4</td><td>79.3</td></tr></table>", | |
| "html": null, | |
| "num": null, | |
| "text": "Results for incorporating different semantic knowledge sources. The baseline parser is described in Sect. 2. in detail.Baseline HN1 HN2 HN3 CL1 CL2 CL3 F1(%) 78.5 78.6 79.1 78.9 77.5 78.7 78.8 Results for combinations of different semantic knowledge sources", | |
| "type_str": "table" | |
| }, | |
| "TABREF2": { | |
| "content": "<table><tr><td>Type</td><td colspan=\"2\">Count Percent(%)</td></tr><tr><td colspan=\"2\">Nominal Compound AR 19</td><td>38</td></tr><tr><td>Coordination AR</td><td>9</td><td>18</td></tr><tr><td>N V AR in N V+noun</td><td>6</td><td>12</td></tr><tr><td>Other AR</td><td>16</td><td>32</td></tr></table>", | |
| "html": null, | |
| "num": null, | |
| "text": "Frequency of parsing improvement types. AR represents ambiguity resolution.", | |
| "type_str": "table" | |
| }, | |
| "TABREF4": { | |
| "content": "<table><tr><td/><td>LP LR F1</td></tr><tr><td colspan=\"2\">Bikel and Chiang 2000 77.2 76.2 76.7</td></tr><tr><td colspan=\"2\">Levy and Manning 2003 78.4 79.2 78.8</td></tr><tr><td>Present work</td><td>80.1 78.7 79.4</td></tr><tr><td>Bikel Thesis 2004</td><td>81.2 78.0 79.6</td></tr><tr><td colspan=\"2\">Chiang and Bikel 2002 81.1 78.8 79.9</td></tr></table>", | |
| "html": null, | |
| "num": null, | |
| "text": "Previous Results on CTB parsing for sentences of length at most 40 words", | |
| "type_str": "table" | |
| } | |
| } | |
| } | |
| } |