Base_JSON/prefixI/json/inlg/2020.inlg-1.10.json

{
    "paper_id": "2020",
    "header": {
        "generated_with": "S2ORC 1.0.0",
        "date_generated": "2023-01-19T07:28:13.955002Z"
    },
    "title": "The CACAPO Dataset: A Multilingual, Multi-Domain Dataset for Neural Pipeline and End-to-End Data-to-Text Generation",
    "authors": [
        {
            "first": "Chris",
            "middle": [],
            "last": "Van Der Lee",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "Tilburg University",
                "location": {}
            },
            "email": ""
        },
        {
            "first": "Chris",
            "middle": [],
            "last": "Emmery",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "Tilburg University",
                "location": {}
            },
            "email": "c.d.emmery@uvt.nl"
        },
        {
            "first": "Sander",
            "middle": [],
            "last": "Wubben",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "Tilburg University",
                "location": {}
            },
            "email": "s.wubben@uvt.nl"
        },
        {
            "first": "Emiel",
            "middle": [],
            "last": "Krahmer",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "Tilburg University",
                "location": {}
            },
            "email": "e.j.krahmer@uvt.nl"
        }
    ],
    "year": "",
    "venue": null,
    "identifiers": {},
    "abstract": "This paper describes the CACAPO dataset, built for training both neural pipeline and endto-end data-to-text language generation systems. The dataset is multilingual (Dutch and English), and contains almost 10,000 sentences from human-written news texts in the sports, weather, stocks, and incidents domain, together with aligned attribute-value paired data. The dataset is unique in that the linguistic variation and indirect ways of expressing data in these texts reflect the challenges of real world NLG tasks.",
    "pdf_parse": {
        "paper_id": "2020",
        "_pdf_hash": "",
        "abstract": [
            {
                "text": "This paper describes the CACAPO dataset, built for training both neural pipeline and endto-end data-to-text language generation systems. The dataset is multilingual (Dutch and English), and contains almost 10,000 sentences from human-written news texts in the sports, weather, stocks, and incidents domain, together with aligned attribute-value paired data. The dataset is unique in that the linguistic variation and indirect ways of expressing data in these texts reflect the challenges of real world NLG tasks.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Abstract",
                "sec_num": null
            }
        ],
        "body_text": [
            {
                "text": "The current paper presents the Combinations of Aligned Data-Sentences from Naturally Produced Texts (hereafter: CACAPO) dataset; a dataset for data-to-text generation (the task of producing adequate, fluent natural language text from nonlinguistic structured data, such as database records, spreadsheets, knowledge graphs, tables, etc., Gatt and Krahmer, 2018) . The dataset contains sentences from automatically scraped news texts for the sports, weather, stock, and incidents domain in English and Dutch, aligned with relevant attributevalue paired data (see Figure 1 and Appendix A for examples). To our knowledge, this is the first dataset based on 'naturally occurring' humanwritten texts (i.e., texts that were not collected in a task-based setting), that covers various domains, as well as multiple languages.",
                "cite_spans": [
                    {
                        "start": 337,
                        "end": 360,
                        "text": "Gatt and Krahmer, 2018)",
                        "ref_id": "BIBREF14"
                    }
                ],
                "ref_spans": [
                    {
                        "start": 561,
                        "end": 569,
                        "text": "Figure 1",
                        "ref_id": "FIGREF0"
                    }
                ],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "Neural Natural Language Generation (NLG) is a promising technique, as neural NLG systems are not bound by any special-purpose mechanisms, and hence are argued to be easily adaptable to other domains and languages (Oraby et al., 2019; Puduppully et al., 2019; van der Lee et al., 2018 ). Yet despite this advantage, it can still be challenging to create a neural NLG system that achieves the same rich and detailed output as a well-designed traditional rule-based pipeline system (Novikova et al., 2017; van der Lee et al., 2018; Moryossef et al., 2019b) . This is because a large-scale parallel dataset (i.e., a dataset with aligned texts and relevant data) is required for training neural NLG systems, and such datasets are not a common natural occurrence. This limitation is especially persistent in neural pipeline architectures: neural architectures modeled after the 'traditional' pipeline architecture (Reiter and Dale, 2000) that sequentially performs tasks related to document planning, sentence planning and linguistic realization (Castro Ferreira et al., 2019) . These architectures require an explicit representation for every intermediate step. The (enriched) WebNLG dataset (Gardent et al., 2017a,b; Castro Ferreira et al., 2018) is presently the only other dataset viable for both end-to-end, as well as neural pipeline architectures.",
                "cite_spans": [
                    {
                        "start": 213,
                        "end": 233,
                        "text": "(Oraby et al., 2019;",
                        "ref_id": "BIBREF33"
                    },
                    {
                        "start": 234,
                        "end": 258,
                        "text": "Puduppully et al., 2019;",
                        "ref_id": "BIBREF40"
                    },
                    {
                        "start": 259,
                        "end": 283,
                        "text": "van der Lee et al., 2018",
                        "ref_id": "BIBREF22"
                    },
                    {
                        "start": 479,
                        "end": 502,
                        "text": "(Novikova et al., 2017;",
                        "ref_id": "BIBREF31"
                    },
                    {
                        "start": 503,
                        "end": 528,
                        "text": "van der Lee et al., 2018;",
                        "ref_id": "BIBREF22"
                    },
                    {
                        "start": 529,
                        "end": 553,
                        "text": "Moryossef et al., 2019b)",
                        "ref_id": "BIBREF30"
                    },
                    {
                        "start": 908,
                        "end": 931,
                        "text": "(Reiter and Dale, 2000)",
                        "ref_id": "BIBREF44"
                    },
                    {
                        "start": 1040,
                        "end": 1070,
                        "text": "(Castro Ferreira et al., 2019)",
                        "ref_id": "BIBREF5"
                    },
                    {
                        "start": 1187,
                        "end": 1212,
                        "text": "(Gardent et al., 2017a,b;",
                        "ref_id": null
                    },
                    {
                        "start": 1213,
                        "end": 1242,
                        "text": "Castro Ferreira et al., 2018)",
                        "ref_id": "BIBREF6"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "The present paper thus presents a new automatically scraped dataset that can be used for end-toend, as well as neural pipeline architectures. Furthermore, it describes a collection process inspired by Oraby et al. (2019) , where collection starts with the news reports and attribute-value datapoints are constructed from them, which also enables relatively low-effort extension and adaptation of the current dataset (Section 3). Characteristics of the dataset are described based on the methodology by Perez-Beltrachini and Gardent (2017) (Section 4) . Finally, a baseline is developed for the dataset using TGen (Du\u0161ek and Jur\u010d\u00ed\u010dek, 2015) (Section 5) .",
                "cite_spans": [
                    {
                        "start": 201,
                        "end": 220,
                        "text": "Oraby et al. (2019)",
                        "ref_id": "BIBREF33"
                    },
                    {
                        "start": 502,
                        "end": 538,
                        "text": "Perez-Beltrachini and Gardent (2017)",
                        "ref_id": "BIBREF37"
                    },
                    {
                        "start": 539,
                        "end": 550,
                        "text": "(Section 4)",
                        "ref_id": null
                    },
                    {
                        "start": 613,
                        "end": 639,
                        "text": "(Du\u0161ek and Jur\u010d\u00ed\u010dek, 2015)",
                        "ref_id": "BIBREF11"
                    },
                    {
                        "start": 640,
                        "end": 651,
                        "text": "(Section 5)",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "The full dataset is freely available for research purposes upon request, licensed under AusGoal Restrictive Licence. A 'thin' version of the dataset that contains the annotated data in combination with the URLs of the scraped texts and the scraping tools is publicly available via https://github. com/TallChris91/CACAPO-Dataset, licensed under CC BY-NC-SA.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "Value pitcherName CC Sabathia teamName Blue Jays teamName Yankees hitNumber hitless inningsPitched five \u2193 All CC Sabathia did was hold the Blue Jays hitless over the final five innings to give the Yankees a chance to rally. ",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Atribute",
                "sec_num": null
            },
            {
                "text": "Neural data-to-text NLG models have the ability to produce texts without requiring handwritten rules and templates, generating texts in a completely data-driven way. However, neural data-to-text NLG is struggling to overcome two critical bottlenecks, identified by Oraby et al. (2019) , that hamper the performance of the models: (1) a data bottleneck, a lack of (high quality, large scale) parallel data-text datasets; and (2) a control bottleneck, which they describe as an inability to control stylistic variation, but can be more broadly described as the inability to systematically control the generation process and the generated output of a neural system.",
                "cite_spans": [
                    {
                        "start": 265,
                        "end": 284,
                        "text": "Oraby et al. (2019)",
                        "ref_id": "BIBREF33"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Related work",
                "sec_num": "2"
            },
            {
                "text": "The field has started to address the data bottleneck issue recently, exhibited by an increase of parallel data-text corpora. E2E (Novikova et al., 2017) , and WebNLG (Gardent et al., 2017a,b) are two prime examples of this. Crowdsourcing techniques were employed for the creation of these datasets, meaning that humans were asked to write a text given an input meaning representation (MR). This makes it feasible to collect ample texts of good quality, but such techniques can quickly become a financial burden, and require significant effort from the researchers to design and assemble (Oraby et al., 2019) . This amount of time can be reduced as is shown by the construction of the ToTTo dataset , where humans edited existing Wikipedia phrases to reflect a given input MR, rather than writing text phrases from scratch. This still requires significant resources, however. Compiling a dataset via crowdsourcing usually ensures that texts are a direct verbalization of the aligned data, which limits the amount of noise and inaccuracies present in the datasets. However, peo- Tag  Entity  PATIENT-1 CC Sabathia  PATIENT-2 Blue Jays  PATIENT-3 hitless  PATIENT-4 five  PATIENT-5 Yankees   \u2193 All PATIENT-1 did was hold the PATIENT-2 PATIENT-3 over the final PATIENT-4 innings to give the PATIENT-5 a chance to rally. ple have increasingly started to criticize the realism of these datasets as they are usually not representative of real world scenarios and language use. 1 Verbalizations by crowdsource workers are different from how data is usually verbalized by professional journalists, for instance, whose focusbesides high fidelity-is also on producing fluent and enjoyable texts. Such a focus can result in more indirect descriptions of data or superfluous information. Generating texts from these indirect descriptions may be more challenging as the NLG systems need to learn how to abstract from the 'noise' in these datasets. A different but related problem is that the presence of superfluous descriptions in input make these neural systems more prone to 'hallucinations', i.e., producing output information that is not present in the input data (Reiter, 2018a) . However, having a system that performs well on such unedited texts might make it more attainable to develop systems that can be deployed in nonacademic settings, as these texts are representative of such settings. Companies for which data-totext systems may be especially relevant (i.e., press agencies, publishers, weather institutes, etc.), oftentimes have an extensive archive of historical data and human-written texts, that would contain similar types of 'noise' in their data representation. Therefore, it seems imperative to also pursue other dataset collection techniques-such as text and data collection-by scraping publicly available sources. Datasets that were compiled via this method have also seen a surge recently, with YelpNLG (Oraby et al., 2019) , RotoWire (Wiseman et al., 2017) , and RotoWire-inspired datasets like RotoWire-FG (Wang, 2019) and MLB (Puduppully et al., 2019) . Using this method 1 See, for instance, the discussion at https://twitter.com/yoavgo/status/ 1281971375029325824. enables data and text collection without having to spend as much time and resources as would be necessary with crowdsourcing techniques. However, most of the current automatically scraped datasets are for document-level texts, which generally requires different architectural approaches than the shorter sentence-level or phrase-level texts that are most commonly found in the datasets compiled via crowdsourcing. Furthermore, they are limited to one domain (restaurants, basketball, and baseball, for YelpNLG, RotoWire, and MLB respectively), and one language (English). This makes it difficult to train domain-invariant systems.",
                "cite_spans": [
                    {
                        "start": 129,
                        "end": 152,
                        "text": "(Novikova et al., 2017)",
                        "ref_id": "BIBREF31"
                    },
                    {
                        "start": 166,
                        "end": 191,
                        "text": "(Gardent et al., 2017a,b)",
                        "ref_id": null
                    },
                    {
                        "start": 587,
                        "end": 607,
                        "text": "(Oraby et al., 2019)",
                        "ref_id": "BIBREF33"
                    },
                    {
                        "start": 2155,
                        "end": 2170,
                        "text": "(Reiter, 2018a)",
                        "ref_id": "BIBREF42"
                    },
                    {
                        "start": 2916,
                        "end": 2936,
                        "text": "(Oraby et al., 2019)",
                        "ref_id": "BIBREF33"
                    },
                    {
                        "start": 2948,
                        "end": 2970,
                        "text": "(Wiseman et al., 2017)",
                        "ref_id": "BIBREF53"
                    },
                    {
                        "start": 3021,
                        "end": 3033,
                        "text": "(Wang, 2019)",
                        "ref_id": "BIBREF52"
                    },
                    {
                        "start": 3042,
                        "end": 3067,
                        "text": "(Puduppully et al., 2019)",
                        "ref_id": "BIBREF40"
                    }
                ],
                "ref_spans": [
                    {
                        "start": 1077,
                        "end": 1190,
                        "text": "Tag  Entity  PATIENT-1 CC Sabathia  PATIENT-2 Blue Jays  PATIENT-3 hitless  PATIENT-4 five  PATIENT-5 Yankees   \u2193",
                        "ref_id": "TABREF2"
                    }
                ],
                "eq_spans": [],
                "section": "Data bottleneck",
                "sec_num": "2.1"
            },
            {
                "text": "Furthermore, most existing datasets are constructed for end-to-end architectures, where the non-lingustic input is converted into natural language without explicit intermediate representations in between (Castro Ferreira et al., 2019) . By contrast, researchers have started to experiment with neural pipeline methods, in which the data conversion process happens via one or more explicit intermediate transformations (see, for instance, Castro Ferreira et al., 2019; Jiang et al., 2020; Moryossef et al., 2019a,b) . These methods enable the control over parts of the data-to-text conversion process, making it possible to develop hybrid (e.g. rule-based and neural) systems. Additionally, a direct comparison between end-to-end and pipeline approaches suggests that pipeline approaches lead to improved output quality, and decreases data hallucination and data omission; two challenges for datasets compiled using unedited texts from publicly available sources (Castro Ferreira et al., 2019) . However, pipeline architectures require a training dataset containing the intermediate representations in order to be trained. And, with the exception of the Enriched WebNLG dataset (Castro Ferreira et al., 2018) , there are currently no datasets facilitating such an approach. 2",
                "cite_spans": [
                    {
                        "start": 204,
                        "end": 234,
                        "text": "(Castro Ferreira et al., 2019)",
                        "ref_id": "BIBREF5"
                    },
                    {
                        "start": 418,
                        "end": 467,
                        "text": "(see, for instance, Castro Ferreira et al., 2019;",
                        "ref_id": null
                    },
                    {
                        "start": 468,
                        "end": 487,
                        "text": "Jiang et al., 2020;",
                        "ref_id": "BIBREF17"
                    },
                    {
                        "start": 488,
                        "end": 514,
                        "text": "Moryossef et al., 2019a,b)",
                        "ref_id": null
                    },
                    {
                        "start": 962,
                        "end": 992,
                        "text": "(Castro Ferreira et al., 2019)",
                        "ref_id": "BIBREF5"
                    },
                    {
                        "start": 1177,
                        "end": 1207,
                        "text": "(Castro Ferreira et al., 2018)",
                        "ref_id": "BIBREF6"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Control bottleneck",
                "sec_num": "2.2"
            },
            {
                "text": "The current work introduces the CACAPO dataset which addresses the aforementioned limitations of the existing datasets: it contains intermediate representations for discourse ordering, text structuring, lexicalization, referring expression generation, and textual realization for pipeline approaches such as the one by Castro Ferreira et al. (2019). Furthermore, it is a sentence-level dataset containing unedited sentences from news articles written by professional journalists and meteorologists (see Section 3 for details).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Current work",
                "sec_num": "2.3"
            },
            {
                "text": "Finally, many of the datasets that are commonly used currently lack domain diversity (Radev et al., 2020) and are solely constructed for the English language (with the exception of WebNLG, see Castro Ferreira et al., 2018; Shimorina et al., 2019) . The CACAPO dataset contains texts from the sports, weather, stocks, and incidents domain for both Dutch and English.",
                "cite_spans": [
                    {
                        "start": 85,
                        "end": 105,
                        "text": "(Radev et al., 2020)",
                        "ref_id": "BIBREF41"
                    },
                    {
                        "start": 193,
                        "end": 222,
                        "text": "Castro Ferreira et al., 2018;",
                        "ref_id": "BIBREF6"
                    },
                    {
                        "start": 223,
                        "end": 246,
                        "text": "Shimorina et al., 2019)",
                        "ref_id": "BIBREF50"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Current work",
                "sec_num": "2.3"
            },
            {
                "text": "3 The CACAPO dataset",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Current work",
                "sec_num": "2.3"
            },
            {
                "text": "Both the Dutch and English version of the CACAPO dataset contain the same four domains (sports, weather, stocks, and incidents) albeit with different events and hence also some topical variety between both languages. For each domain a scraping tool was used or custom built that either fully automatically collected relevant texts, or collected these texts with as little human effort as possible (e.g humans needed to copy the URLs, website source code, or needed to copy some aspects to a custombuilt tool). 3 The following texts were collected:",
                "cite_spans": [
                    {
                        "start": 510,
                        "end": 511,
                        "text": "3",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Collection methods",
                "sec_num": "3.1"
            },
            {
                "text": "\u2022 Dutch sports domain texts cover soccer match reports from the 15/16 and 16/17 season of the Dutch Eredivisie, the highest professional soccer league in The Netherlands. Texts were scraped from 10 professional news websites using Google search queries for all matches played during the 15/16 and 16/17 seasons (teams and play date). In total, 6,600 texts were scraped (2,101,338 tokens; 27,619 types).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Collection methods",
                "sec_num": "3.1"
            },
            {
                "text": "\u2022 Dutch stocks domain texts cover daily reports on stock exchanges, company stock listings, (crypto)currency exchange rates, and oil prices. These reports were collected from 49 different newspapers using Nexis Uni, 4 covering all reports from January 2019-January 2020. A total of 4,280 texts were collected (1,211,842 tokens; 22,685 types).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Collection methods",
                "sec_num": "3.1"
            },
            {
                "text": "\u2022 Dutch weather domain texts cover severaldaily short-term weather forecasts for The Netherlands from the Royal Netherlands Meteorological Institute (KNMI); the Dutch national weather service. These texts originate from the \"complete weather report\" prognosis, found on the KNMI website 5 . The weather reports were obtained for all of 2019, totalling 5,897 texts (1,099,556 tokens; 1,076 types).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Collection methods",
                "sec_num": "3.1"
            },
            {
                "text": "\u2022 Dutch incidents domain texts originate from Hendriks (2019) who collected data from https://www.hetongeluk.nl/; an online database for news articles about traffic incidents, which in total contains traffic incident reports from 139 websites from 2013 to 2019. This collection contains 1,600 texts (154,596 tokens; 8,919 types).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Collection methods",
                "sec_num": "3.1"
            },
            {
                "text": "\u2022 English sports domain texts cover baseball reports from the American MLB League, the top league in American professional baseball. The baseball reports were obtained using the scraper made available by Puduppully et al. 2019 \u2022 English stocks domain texts cover the same topics as the Dutch stocks domain texts. The texts were obtained using Google News by searching news items containing \"stock index\" and \"stock market\" in the period of January 2019-January 2020. 1,109 texts from 182 websites were collected (621,997 tokens; 23,216 types).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Collection methods",
                "sec_num": "3.1"
            },
            {
                "text": "\u2022 English weather domain texts cover weather forecasts for several countries (e.g., Canada, United States, India, Ireland). The weather forecasts were collected using Google News by searching news items containing \"weather forecast\" in the period of January 2019-2020. This resulted in a collection of 926 texts from 215 websites (341,622 tokens; 11,426 types).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Collection methods",
                "sec_num": "3.1"
            },
            {
                "text": "\u2022 English incidents domain texts cover gun violence incidents from the Gun Violence Archive, 7 a database on gun violence incidents, which in total contains 3,180 incident reports from 596 websites ranging from 2012 to 2019 (1,105,567 tokens; 26,968 types).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Collection methods",
                "sec_num": "3.1"
            },
            {
                "text": "Thus, in total 51,575 texts were collected via these different methods. For the CACAPO dataset, all texts above 325 words were discarded as most basic news reports typically do not exceed that amount of words (Asbreuk et al., 2017) , leaving 20,630 texts. 8 From this sample, 200 texts were randomly selected for each language and domain (a total of 1,600 texts; 12.89% of the text selection) to obtain a representative number of sentences while keeping the annotation load reasonable (see Section 3.2). These texts were automatically split into sentences using a sentence tokenizer. SpaCy (Honnibal and Montani, 2017) was used as a tokenizer for the Dutch part, and SoMaJo for the English part (Proisl and Uhrig, 2016) . Finally, the sentences were assigned to training, validation, and testing sets in a 76.5, 8.5, 15 ratio-the same ratio that (Novikova et al., 2017) used. All sentences occurring in the selected texts are part of the CACAPO dataset and the order of occurrence of the sentences in a text was preserved in the dataset.",
                "cite_spans": [
                    {
                        "start": 209,
                        "end": 231,
                        "text": "(Asbreuk et al., 2017)",
                        "ref_id": "BIBREF0"
                    },
                    {
                        "start": 695,
                        "end": 719,
                        "text": "(Proisl and Uhrig, 2016)",
                        "ref_id": "BIBREF39"
                    },
                    {
                        "start": 846,
                        "end": 869,
                        "text": "(Novikova et al., 2017)",
                        "ref_id": "BIBREF31"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Collection methods",
                "sec_num": "3.1"
            },
            {
                "text": "The data annotation process followed after the sample sentences were tokenized. Sentences were manually aligned with data using Prodigy 9 , a data annotation tool (Montani and Honnibal, 2018) , in a attribute-value pair format, done by two expert annotators. The annotators annotated a part of the dataset jointly (1,755 sentences), resulting in Cohen's \u03ba = 0.67 (substantial agreement; Landis and Koch, 1977 ) and a 70.92% agreement. This agreement was deemed high enough for a single coder per item approach for the rest of the dataset. One of the annotators developed the guidelines with a definition of each category and examples of passages belonging to that category resulting in relatively quick acquisition of the categories. Annotation took between 5 and 15 minutes per text on average.",
                "cite_spans": [
                    {
                        "start": 163,
                        "end": 191,
                        "text": "(Montani and Honnibal, 2018)",
                        "ref_id": "BIBREF28"
                    },
                    {
                        "start": 387,
                        "end": 408,
                        "text": "Landis and Koch, 1977",
                        "ref_id": "BIBREF20"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Data annotation",
                "sec_num": "3.2"
            },
            {
                "text": "All annotated attributes can be found in Appendix C. The amount of types that were annotated varied between 10 (Dutch/English stocks domain) and 76 (English sports domain). Which labels to annotate was decided upon by doing a practice set of 10 texts. All data labels are based on the 5 Ws and 1 H questions (Who, What, When, Where, Why, and How). As most journalism schools teach students to write news articles that focus on answers to the 5 Ws and 1 H questions (Canavilhas, 2007; Kussendrager et al., 2018) .",
                "cite_spans": [
                    {
                        "start": 465,
                        "end": 483,
                        "text": "(Canavilhas, 2007;",
                        "ref_id": "BIBREF4"
                    },
                    {
                        "start": 484,
                        "end": 510,
                        "text": "Kussendrager et al., 2018)",
                        "ref_id": "BIBREF19"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Data annotation",
                "sec_num": "3.2"
            },
            {
                "text": "'Who' data is for instance player and referee information for the sports domain (assistName, goalName, goalkeeperName, pitcherName, pitcher-Record, umpireName) and suspect and victim information for the incidents domain (suspec-tAge, suspectGender, victimBased, victimName, victimOccupation) . Examples of 'What' data are stock price increases and decreases for the stocks domain (stockChange, stockChangePercentage, stockPoints), and information about cloudiness, wind, and weather type for the weather domain (cloudAmount, gustChange, temperatureCelsius, weatherType). 'When' data types are the (next) match date for the sports domain (matchDate, matchTime, nextMatchDate), and the date/time that an incident occurred for the incidents domain (date-Time, accidentDate). 'Where' data is the stadium where a match is played for the sports domain (sta-diumPlayed, locationPlayed), or where weather events will happen for the weather domain (loca-tionArea). 'Why' data is for example the cause of a traffic incident for the incidents domain (inci-dentCause). And 'How' data can be information about the way a goal was scored or a ball was hit for the sports domain (goalType, strikingType), and how a traffic/shooting incident took place for the incidents domain (incidentType, shootingType).",
                "cite_spans": [
                    {
                        "start": 220,
                        "end": 291,
                        "text": "(suspec-tAge, suspectGender, victimBased, victimName, victimOccupation)",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Data annotation",
                "sec_num": "3.2"
            },
            {
                "text": "After the data annotation process was completed, the annotated data and collected texts were then used to create explicit intermediate representations suitable for neural pipeline architectures. The CACAPO dataset is saved in a similar XML format as the Enriched WebNLG dataset (Castro Ferreira et al., 2018) (see Figure 3 for an example) to enable effortless testing of systems designed for this dataset. This also means that the CACAPO dataset is suitable for pipeline systems that convert data into text using the same 5 sequential steps as Castro Ferreira et al. (2019), which follows the original pipeline architecture of (Reiter and Dale, 2000):",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 314,
                        "end": 322,
                        "text": "Figure 3",
                        "ref_id": "FIGREF3"
                    }
                ],
                "eq_spans": [],
                "section": "Intermediate representations",
                "sec_num": "3.3"
            },
            {
                "text": "1. Discourse Ordering is the task of determining in which order to present the data that should be verbalized in the tar-<e n t r y c a t e g o r y =\" E n g l i s h I n c i d e n t s \" e i d =\" I d 2 \" s i z e =\" 3 \"> <o r i g i n a l d a t a s e t> <o d a t a>v i c t i m A g e | 22\u2212y e a r\u2212o l d</ o d a t a> <o d a t a>v i c t i m S t a t u s | g r a z e d i n t h e t h i g h</ o d a t a> </ o r i g i n a l d a t a s e t> <l e x comment=\" good \" l i d =\" I d 1 \"> <s o r t e d d a t a s e t> <s e n t e n c e ID=\" 1 \"> <s d a t a>v i c t i m A g e | 22\u2212y e a r\u2212o l d</ s d a t a> <s d a t a>v i c t i m S t a t u s | g r a z e d i n t h e \u2192 t h i g h</ s d a t a> </ s e n t e n c e> </ s o r t e d d a t a s e t> <r e f e r e n c e s> <r e f e r e n c e e n t i t y =\"22\u2212y e a r\u2212o l d \" number=\" 1 \" \u2192 t a g =\"ENTITY\u22121\" \u2192 t y p e =\" d e s c r i p t i o n \">22\u2212y e a r\u2212o l d \u2192 </ r e f e r e n c e> <r e f e r e n c e e n t i t y =\" g r a z e d i n t h e t h i g h \" \u2192 number=\" 2 \" t a g =\"ENTITY\u22122\" \u2192 t y get text.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Intermediate representations",
                "sec_num": "3.3"
            },
            {
                "text": "This can be trained using the MRs found in the (alphabetically ordered) <originaldataset>, and the <sorteddataset> that is ordered based on the appearance of the MR in the sentence. This ordering is determined based on the string position information provided by Prodigy.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Intermediate representations",
                "sec_num": "3.3"
            },
            {
                "text": "2. Text Structuring is the task of organizing the ordered triples into paragraphs and sentences. The <sorteddataset> tag also contains sentence information relevant for the Text Structuring step. As the CACAPO dataset is a sentence-level dataset, Text Structuring is not a directly relevant step. Although the sentence information in the <sorteddataset> tag allows for extensions to phrase-level or paragraph-level instances.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Intermediate representations",
                "sec_num": "3.3"
            },
            {
                "text": "3. Lexicalization is the task of finding the words and phrases that describe the input data correctly (Reiter and Dale, 2000) . This means using the information found in <sorteddataset> to (ideally) generate the string in <lexicalization>, for this dataset. The string found in this tag is a delexicalized version of the original sentence (found in <text>). This <lexicalization> tag not only contains information to se-lect accurate words and phrases to describe an MR, but also contains information for the two steps further ahead in the pipeline. The ENTITY-[0-9] placeholders indicate where MRs should be realized. The entity number indicates which MR to realize based on the order in <sorteddataset>. Furthermore, the delexicalized string contains syntactical information. For (lemmatized) verbs, it stores aspect, mood, tense, voice and number in a VP tag. And it stores the form of determiners in a DT tag.",
                "cite_spans": [
                    {
                        "start": 102,
                        "end": 125,
                        "text": "(Reiter and Dale, 2000)",
                        "ref_id": "BIBREF44"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Intermediate representations",
                "sec_num": "3.3"
            },
            {
                "text": "Delexicalization was done by a script that matches the annotated data with the original string, using the string location information provided by Prodigy. The annotation of syntactical information and lemmatization was done using CoreNLP (Manning et al., 2014) for English, and DeepFrog 10 for Dutch.",
                "cite_spans": [
                    {
                        "start": 238,
                        "end": 260,
                        "text": "(Manning et al., 2014)",
                        "ref_id": "BIBREF26"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Intermediate representations",
                "sec_num": "3.3"
            },
            {
                "text": "task of generating the correct entities in a text (Krahmer and van Deemter, 2012) . In this step, a system can be trained to fill the ENTITY-[0-9] placeholders found in the <lexicalization> string with the data found in the <references> tag.",
                "cite_spans": [
                    {
                        "start": 50,
                        "end": 81,
                        "text": "(Krahmer and van Deemter, 2012)",
                        "ref_id": "BIBREF18"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Referring Expression Generation is the",
                "sec_num": "4."
            },
            {
                "text": "5. Textual Realization is the task of performing the final steps to convert the non-linguistic data into natural language text. For this dataset, this means converting the lemmatized verbs and determiners to a form that is congruent with the MR, using the VP and DT tags found in <lexicalization>.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Referring Expression Generation is the",
                "sec_num": "4."
            },
            {
                "text": "Of course, the dataset also lends itself for datato-text generation in an end-to-end fashion. For this, a system can be trained on the information in <originaldataset> and <text>.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Referring Expression Generation is the",
                "sec_num": "4."
            },
            {
                "text": "We compare the CACAPO dataset to the Enriched WebNLG dataset (Castro Ferreira et al., 2018; Gardent et al., 2017a,b) , as these datasets are comparable in the sense that both are multilingual, multidomain, and contain explicit intermediate steps that allow for neural pipeline architectures to be employed. However, they are different in the fact that WebNLG is constructed using crowdsourcing, while CACAPO is constructed using unedited texts 10 https://github.com/proycon/deepfrog scraped from publicly available sources. Similar to Novikova et al. (2017) we compare the two datasets on size, lexical richness, and sentence complexity.",
                "cite_spans": [
                    {
                        "start": 61,
                        "end": 91,
                        "text": "(Castro Ferreira et al., 2018;",
                        "ref_id": "BIBREF6"
                    },
                    {
                        "start": 92,
                        "end": 116,
                        "text": "Gardent et al., 2017a,b)",
                        "ref_id": null
                    },
                    {
                        "start": 535,
                        "end": 557,
                        "text": "Novikova et al. (2017)",
                        "ref_id": "BIBREF31"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Statistics",
                "sec_num": "4"
            },
            {
                "text": "Based on, Novikova et al. (2017) and Perez-Beltrachini and Gardent (2017) , we employ the following size metrics to compare the Enriched WebNLG dataset (Castro Ferreira et al., 2018) to our dataset (see Table 1 ):",
                "cite_spans": [
                    {
                        "start": 10,
                        "end": 32,
                        "text": "Novikova et al. (2017)",
                        "ref_id": "BIBREF31"
                    },
                    {
                        "start": 37,
                        "end": 73,
                        "text": "Perez-Beltrachini and Gardent (2017)",
                        "ref_id": "BIBREF37"
                    },
                    {
                        "start": 152,
                        "end": 182,
                        "text": "(Castro Ferreira et al., 2018)",
                        "ref_id": "BIBREF6"
                    }
                ],
                "ref_spans": [
                    {
                        "start": 203,
                        "end": 210,
                        "text": "Table 1",
                        "ref_id": "TABREF2"
                    }
                ],
                "eq_spans": [],
                "section": "Size",
                "sec_num": "4.1"
            },
            {
                "text": "\u2022 Number of instances: Absolute number of texts in the dataset (single sentences for CACAPO, single sentences and multi-sentence phrases for WebNLG). This gives a direct indication of the dataset size.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Size",
                "sec_num": "4.1"
            },
            {
                "text": "\u2022 Number of unique MRs: Number of different MRs appearing in the dataset (set of attribute-value paired data for CACAPO, set of RDF-triple data for WebNLG aligned to a text). Besides dataset size, this also gives an indication of training difficulty: more unique MRs means a greater challenge to train models on the data.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Size",
                "sec_num": "4.1"
            },
            {
                "text": "\u2022 Instances per MR: Average number of verbalizations for one MR. The more references for an MR appear in the training set, the better models can be trained to learn how to verbalize this MR.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Size",
                "sec_num": "4.1"
            },
            {
                "text": "\u2022 Slots per MR: Average number of data points (single attribute-value paired data for CACAPO, single RDF-triples for WebNLG) that compose an MR.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Size",
                "sec_num": "4.1"
            },
            {
                "text": "\u2022 Words per instance: Average number of words appearing in an instance (single sentences for CACAPO, single sentences and multi-sentence phrases for WebNLG).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Size",
                "sec_num": "4.1"
            },
            {
                "text": "\u2022 Words per sentence: Average number of words appearing in a sentence.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Size",
                "sec_num": "4.1"
            },
            {
                "text": "\u2022 Sentences per instance: Average number of sentences appearing in an instance.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Size",
                "sec_num": "4.1"
            },
            {
                "text": "The metrics in Table 1 show that the CACAPO dataset and Enriched WebNLG dataset are very similar in size, as displayed by the number of instances and number of unique MRs, with the CACAPO dataset being slightly bigger. Also, in terms of slots per MR, and words per reference, the CACAPO dataset and the Enriched WebNLG dataset seem comparable. However, on average, there are fewer references for MRs in the CACAPO dataset compared to the WebNLG dataset. This indicates that it would be more challenging for data-to-text generation systems to learn alignments between MRs and text for the CACAPO dataset compared to the WebNLG dataset.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 15,
                        "end": 22,
                        "text": "Table 1",
                        "ref_id": "TABREF2"
                    }
                ],
                "eq_spans": [],
                "section": "Size",
                "sec_num": "4.1"
            },
            {
                "text": "Following Novikova et al. 2017, we investigate various aspects of lexical richness by looking at traditional measures, such as the number of tokens, and types, and type-token ratio (TTR; see Table 2 ). And we include the more robust mean segmental TTR (MSTTR), which divides the dataset into equal segments of a given token length (here: 25 tokens) and calculates the average TTR of all these segments. Finally, we also include Lexical Sophistication (LS). Also known as Guiraud Advanced (Daller et al., 2003) which gauges the number of unique words in a dataset; another way to measure lexical richness. We calculate the Guiraud Advanced metric by taking the proportion of word types that are not in the top 2,000 most frequent words in large and diverse corpora for each language: the British National Corpus (British National Corpus, 2007), the SoNaR 500 corpus (Oostdijk et al., 2013) , and the German Internet corpus (Sharoff, 2006) , for the English and Dutch CACAPO dataset, with additional statistics for English and German WebNLG added for comparison, respectively. Each of these corpora contains a large amount of texts and covers a wide array of topics and domains. Therefore, we believe that their top 2,000 most frequent words are representative of the language. The number of tokens in Table 2 show that the texts of the CACAPO dataset are somewhat smaller than those found in the WebNLG dataset. However, supporting our expectations, the CACAPO dataset is the more lexically varied dataset of the two, as illustrated by the higher TTR and MSTTR scores, and the higher absolute number of types. The higher amount of lexical diversity found in the CACAPO dataset is a further indication that training a datato-text generation system to produce high quality output may be more challenging for this dataset. The lexical sophistication metric shows a similar proportion of infrequent words in the CACAPO and WebNLG dataset, which suggests that both datasets are similarly diverse in terms of the amount of nonstandard language found in the dataset. Also similar to Novikova et al. 2017, we have analyzed the appearance of bigrams and trigrams in the dataset. Focusing on (1) the proportion of bigrams and trigrams appearing only once in the CACAPO dataset and WebNLG dataset; and (2) on the average frequency of bigrams and trigrams of those that appear more than once. These metrics give further indication of lexical richness: a high amount of unique bigrams and trigrams, and a low average frequency for non-unique bigrams and trigrams makes it more challenging to train a neural data-to-text system.",
                "cite_spans": [
                    {
                        "start": 488,
                        "end": 509,
                        "text": "(Daller et al., 2003)",
                        "ref_id": "BIBREF8"
                    },
                    {
                        "start": 865,
                        "end": 888,
                        "text": "(Oostdijk et al., 2013)",
                        "ref_id": "BIBREF32"
                    },
                    {
                        "start": 922,
                        "end": 937,
                        "text": "(Sharoff, 2006)",
                        "ref_id": "BIBREF48"
                    }
                ],
                "ref_spans": [
                    {
                        "start": 191,
                        "end": 198,
                        "text": "Table 2",
                        "ref_id": "TABREF4"
                    },
                    {
                        "start": 1300,
                        "end": 1307,
                        "text": "Table 2",
                        "ref_id": "TABREF4"
                    }
                ],
                "eq_spans": [],
                "section": "Lexical Richness",
                "sec_num": "4.2"
            },
            {
                "text": "The results in Table 3 show further evidence that the English and Dutch versions of the CACAPO dataset are more lexically rich compared to the English and German versions of the WebNLG dataset. The CACAPO dataset has a much larger proportion of bigrams and trigrams that appear only once. Furthermore, of the bigrams and trigrams appearing more than once, the average frequency of bigrams and trigrams in the CACAPO dataset is much lower than for the WebNLG dataset.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 15,
                        "end": 22,
                        "text": "Table 3",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Lexical Richness",
                "sec_num": "4.2"
            },
            {
                "text": "To assess the complexity of sentences in the WebNLG and CACAPO datasets, we look at the Table 3 : Proportion of bigrams and trigrams occuring once, and average frequency of bigrams and trigrams that occur more than once.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 88,
                        "end": 95,
                        "text": "Table 3",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Sentence complexity",
                "sec_num": "4.3"
            },
            {
                "text": "revised Developmental Level scale (Rosenberg and Abbeduto, 1987; Covington et al., 2006) , also known as D-Level (similar to Novikova et al., 2017) . We used D-Level Analyser (Lu, 2009) to obtain the D-Level proportions for the English datasets, and T-Scan (Pander Maat et al., 2014) to find the D-Level proportions for the Dutch dataset. There are currently no tools to obtain D-Level for German, but it can be assumed that the composition of this dataset is similar to its English WebNLG counterpart, as the German WebNLG dataset is a close translation of that version (Castro Ferreira et al., 2018) . The D-Level scale contains 8 levels: level 0 being the simplest, and level 7 the most complex. Complexity is determined by, for instance, complex syntactic structures, subordinate clauses, and referring expressions. Table 4 shows sizable differences between the datasets in terms of complexity. The Dutch CACAPO dataset predominantly consists of simpler sentences (below level 4), while the English version of the dataset has a large portion of higher level sentences. The WebNLG resides somewhere in between those two in terms of complexity. This would mean that the Dutch version of the CACAPO dataset would be the least challenging for systems to learn the sentence structure of, and the English version of the dataset the most challenging.",
                "cite_spans": [
                    {
                        "start": 34,
                        "end": 64,
                        "text": "(Rosenberg and Abbeduto, 1987;",
                        "ref_id": "BIBREF45"
                    },
                    {
                        "start": 65,
                        "end": 88,
                        "text": "Covington et al., 2006)",
                        "ref_id": "BIBREF7"
                    },
                    {
                        "start": 125,
                        "end": 147,
                        "text": "Novikova et al., 2017)",
                        "ref_id": "BIBREF31"
                    },
                    {
                        "start": 175,
                        "end": 185,
                        "text": "(Lu, 2009)",
                        "ref_id": "BIBREF25"
                    },
                    {
                        "start": 257,
                        "end": 283,
                        "text": "(Pander Maat et al., 2014)",
                        "ref_id": "BIBREF34"
                    },
                    {
                        "start": 571,
                        "end": 601,
                        "text": "(Castro Ferreira et al., 2018)",
                        "ref_id": "BIBREF6"
                    }
                ],
                "ref_spans": [
                    {
                        "start": 820,
                        "end": 827,
                        "text": "Table 4",
                        "ref_id": "TABREF7"
                    }
                ],
                "eq_spans": [],
                "section": "Sentence complexity",
                "sec_num": "4.3"
            },
            {
                "text": "TGen, a sequence-to-sequence model using Attention (Du\u0161ek and Jur\u010d\u00ed\u010dek, 2015) , was used to establish a baseline on the CACAPO dataset. 11 The performance of TGen was evaluated on the test data of the CACAPO dataset using BLEU (Papineni et al., 2002) , NIST (Doddington, 2002) , ME-TEOR (Banerjee and Lavie, 2005) , ROUGE-L (Lin, 2004) , CIDEr (Vedantam et al., 2015) and 11 Parameters are provided in Appendix B. It should be noted that the system is only trained in an end-to-end fashion. BertScore (Table 5) . 12 The results show that the TGen baseline scores vary considerably across domains, as was to be expected. The Dutch Stocks subcorpus offers a positive outlier, which might have to do with the relatively few labels and consistent language of the domain. It should be noted that the same parametersoriginally used for the E2E challenge (Novikova et al., 2017) -were applied to all domains, which might mean that the model is too large and complex for some domains (such as Dutch Weather and English Incidents, where the texts are highly consistent translations of the data, and the domain only contains a small number of types), resulting in overfitting. In other cases, the dataset is arguably too small, which-combined with its lexical richnessmight make it difficult for a neural NLG model to be trained on. However, in all cases, parameter tuning, application of different models, and tokenization/delexicalization of the training texts (as done by Novikova et al., 2017) is likely to increase the text quality and automatic metrics scores. Additionally, it seems worthwhile to explore ways of semi-automatically extending the training corpora, as we hope to do in future work.",
                "cite_spans": [
                    {
                        "start": 51,
                        "end": 77,
                        "text": "(Du\u0161ek and Jur\u010d\u00ed\u010dek, 2015)",
                        "ref_id": "BIBREF11"
                    },
                    {
                        "start": 227,
                        "end": 250,
                        "text": "(Papineni et al., 2002)",
                        "ref_id": "BIBREF35"
                    },
                    {
                        "start": 258,
                        "end": 276,
                        "text": "(Doddington, 2002)",
                        "ref_id": "BIBREF10"
                    },
                    {
                        "start": 287,
                        "end": 313,
                        "text": "(Banerjee and Lavie, 2005)",
                        "ref_id": "BIBREF1"
                    },
                    {
                        "start": 324,
                        "end": 335,
                        "text": "(Lin, 2004)",
                        "ref_id": "BIBREF24"
                    },
                    {
                        "start": 344,
                        "end": 367,
                        "text": "(Vedantam et al., 2015)",
                        "ref_id": "BIBREF51"
                    },
                    {
                        "start": 372,
                        "end": 374,
                        "text": "11",
                        "ref_id": null
                    },
                    {
                        "start": 513,
                        "end": 515,
                        "text": "12",
                        "ref_id": null
                    },
                    {
                        "start": 848,
                        "end": 871,
                        "text": "(Novikova et al., 2017)",
                        "ref_id": "BIBREF31"
                    },
                    {
                        "start": 1465,
                        "end": 1487,
                        "text": "Novikova et al., 2017)",
                        "ref_id": "BIBREF31"
                    }
                ],
                "ref_spans": [
                    {
                        "start": 501,
                        "end": 510,
                        "text": "(Table 5)",
                        "ref_id": "TABREF9"
                    }
                ],
                "eq_spans": [],
                "section": "Baseline system performance",
                "sec_num": "5"
            },
            {
                "text": "This paper described the CACAPO dataset. A multilingual, multi-domain dataset that enables the use of neural pipeline architectures, as well as endto-end architectures. The dataset is comparable in size to the WebNLG dataset, and its lexical richness-due to the fact that the texts directly originate from journalistic articles-provides interesting challenges. Furthermore, the fact that these texts were derived from 'naturally occurring' texts means that there may be superfluous information, as well as indirect descriptions of the data in the text. This is challenging for NLG systems, as shown by the system performance scores when performing an end-to-end data-to-text task on the dataset using TGen (Du\u0161ek and Jur\u010d\u00ed\u010dek, 2015) . However, the dataset closely mirrors real-world scenarios in which companies oftentimes have large amounts of human-written texts that are not purposefully written for NLG applications, accompanied by corresponding data.",
                "cite_spans": [
                    {
                        "start": 706,
                        "end": 732,
                        "text": "(Du\u0161ek and Jur\u010d\u00ed\u010dek, 2015)",
                        "ref_id": "BIBREF11"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Conclusion",
                "sec_num": "6"
            },
            {
                "text": "Bias The fact that the CACAPO dataset is based on 'naturally occurring' data addresses the issue of datasets being not representative of real world NLP issues. However, it should also be noticed that having unedited texts in the dataset means that the biases from the original data are still present in the dataset and may lead to further generation of biased texts (Lepp\u00e4nen et al., 2020) . Therefore, texts generated with this dataset, as well as the texts in the dataset itself, could warrant more traditional linguistics-oriented text analysis research to investigate biases that might exist.",
                "cite_spans": [
                    {
                        "start": 366,
                        "end": 389,
                        "text": "(Lepp\u00e4nen et al., 2020)",
                        "ref_id": "BIBREF23"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Conclusion",
                "sec_num": "6"
            },
            {
                "text": "Evaluation NLG has recently increased its focus on evaluation and multiple researchers have argued that automatic metrics lack interpretability and do not correlate well with human judgments (see, for instance, Reiter, 2018b; . This might especially be an issue for this type of dataset, originating from texts that-besides informing-try to provide engaging texts to read, as evidenced by the high lexical richness and sentence complexity. Since journalists try to convey data in diverse ways, reference-based metrics such as BLEU (Papineni et al., 2002) , METEOR (Banerjee and Lavie, 2005), and ROUGE (Lin, 2004) might be especially ineffective to measure text quality. Van der Lee et al. 2018, for instance, found that BLEU scores were near zero for a similar dataset, while human evaluation showed the texts to be of reasonable quality. Recent learningbased metrics, such as RUSE (Shimanaka et al., 2018) , BertScore , MoverScore (Zhao et al., 2019) , and BLEURT (Sellam et al., 2020) might be more viable options, since they claim to capture semantic similarity. However, we discourage using this dataset as a leaderboard chasing game and recommend using various types of evaluation methods to evaluate systems trained on the CACAPO dataset (e.g., evaluating the results on the dataset using human and automatic metrics, and qualitative and quantitative research methods). Variety in evaluation methods ensures that the results obtained on this dataset are put into a broad perspective. This will give valuable insights into the systems trained on the dataset, as well as the characteristics of the dataset itself.",
                "cite_spans": [
                    {
                        "start": 211,
                        "end": 225,
                        "text": "Reiter, 2018b;",
                        "ref_id": "BIBREF43"
                    },
                    {
                        "start": 531,
                        "end": 554,
                        "text": "(Papineni et al., 2002)",
                        "ref_id": "BIBREF35"
                    },
                    {
                        "start": 602,
                        "end": 613,
                        "text": "(Lin, 2004)",
                        "ref_id": "BIBREF24"
                    },
                    {
                        "start": 883,
                        "end": 907,
                        "text": "(Shimanaka et al., 2018)",
                        "ref_id": "BIBREF49"
                    },
                    {
                        "start": 933,
                        "end": 952,
                        "text": "(Zhao et al., 2019)",
                        "ref_id": "BIBREF55"
                    },
                    {
                        "start": 966,
                        "end": 987,
                        "text": "(Sellam et al., 2020)",
                        "ref_id": "BIBREF46"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Conclusion",
                "sec_num": "6"
            },
            {
                "text": "The dataset creation method of this paper, where texts are collected first, and data is subsequently manually annotated for each text (Oraby et al., 2019) , also facilitates extensions to the dataset with relative ease. We make the tools to do so publicly available, so that anyone interested can extend the current dataset by annotating a selection of scraped texts that were not used for the definitive dataset. In future work, we would also like to extend the dataset to other languages and other domains (e.g. product reviews, movie descriptions, etc.). Furthermore, we would like to explore the possibility of BERT-based (Devlin et al., 2019) Information Extraction to automatically extend the size of the dataset in a semi-supervised fashion.",
                "cite_spans": [
                    {
                        "start": 134,
                        "end": 154,
                        "text": "(Oraby et al., 2019)",
                        "ref_id": "BIBREF33"
                    },
                    {
                        "start": 626,
                        "end": 647,
                        "text": "(Devlin et al., 2019)",
                        "ref_id": "BIBREF9"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Future work",
                "sec_num": null
            },
            {
                "text": "At least, datasets that start from data. Surface realization datasets such as the one employed in(Mille et al., 2019) can be seen as facilitating the pipeline approach.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "https://www.knmi.nl/nederland-nu/ weer/verwachtingen 6 http://www.espn.com/ 7 http://www.gunviolencearchive.org/",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "The full collection of unlabeled texts and the selection of unlabeled texts is freely available upon request-licensed under AusGoal Restrictive Licence-to facilitate extension of the dataset as well as other tasks, such as information extraction.9 https://prodi.gy/",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "METEOR and BertScore were calculated using the authors' provided scripts, while BLEU was calculated using SacreBLEU(Post, 2018), NIST using NLTK(Bird et al., 2009), and ROUGE-L and CIDEr using nlg-eval(Sharma et al., 2017).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            }
        ],
        "back_matter": [
            {
                "text": "We received support from RAAK-PRO SIA (2014-01-51PRO) and The Netherlands Organization for Scientific Research (NWO 360-89-050). We also want to thank the anonymous reviewers, Saar Hommes, Annemarie Nanne, Jeroen van de Nieuwenhof, Noa Reijnen, and Tess van der Zanden for their contributions.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Acknowledgements",
                "sec_num": null
            },
            {
                "text": "De wind is zwak tot matig en komt uit oost tot zuidoost. There will be a weak to moderate breeze coming from east to southeast. TGen training parameters as reported in (Novikova et al., 2017) : main-sequence-to-sequence model with attention. TGen training parameters as reported in (Novikova et al., 2017) : reranker.",
                "cite_spans": [
                    {
                        "start": 168,
                        "end": 191,
                        "text": "(Novikova et al., 2017)",
                        "ref_id": "BIBREF31"
                    },
                    {
                        "start": 282,
                        "end": 305,
                        "text": "(Novikova et al., 2017)",
                        "ref_id": "BIBREF31"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "\u2193",
                "sec_num": null
            },
            {
                "text": "Beam size 10 Reranker misfit penalty 100TGen decoder parameters as reported in (Novikova et al., 2017) .The parameters are the same as the TGen parameters for the E2E dataset (Novikova et al., 2017) . Raw strings are used for training and generation. Validation is performed on the reserved instances after each epoch using BLEU. Early stopping is applied if the top 3 BLEU results do not change for 5 epochs. , atBatNumber, baseNumber, baseReachedNumber, baseStolen, basesRan, batterHitsTries, batterName, batterScoreNumber, battersFacedNumber, battingAverage, battingLineupNumber, catchType, catcherName, competitionName, earnedRunsNumber, errorNumber, fielderName, fielderPosition, finalScore, gameNumber, gameTally, hasLostTeam, hasScored, hasWonTeam, hitNumber, homeAway, homeRunNumber, injuryType, inningNumber, inningScore, inningsPitched, isOut, leftOnBase, locationPlayed, managerName, matchDate, matchStreakNumber, matchStreakType, numberOfStarts, onBaseNumber, outNumber, pitchCount, pitchNumber, pitchResult, pitchResultNumber, pitchType, pitcherName, pitcherRecord, pitcherSaveRecord, pitchesTotalThrown, presidentName, retireNumber, runAverage, runNumber, scoreNumber, scoreTally, standingsGames, startsNumber, stealNumber, strikeNumber, strikeOutNumber, strikeTrajectory, strikingType, teamName, teamRecord, teamStandings, throwDirection, umpireName, umpireType, unearnedRunsNumber, walkNumber, winLossRecord, incidentCause, incidentLocation, incidentType, suspectAddress, suspectAge, suspectAmount, suspectDescription, suspectGender, suspectStatus, suspectVehicle, victimAddress, victimAge, victimAmount, victimDescription, victimGender, victimName, victimStatus, victimVehicle English incidents accidentAddress, accidentDate, hospitalName, numberOfRoundsFired, personnelArrivedTime, prisonName, shootingNumber, shootingType, suspectAge, suspectAgeGroup, suspectBased, suspectDescription, suspectGender, suspectHeight, suspectName, suspectNumber, suspectOccupation, suspectRace, suspectStatus, suspectVehicle, suspectWeapon, suspectWeight, takenToHospital, victimAge, victimAgeGroup, victimBased, victimGender, victimName, victimNumber, victimOccupation, victimRace, victimStatus, victimVehicle Labels of data types used in the CACAPO dataset per subdomain.",
                "cite_spans": [
                    {
                        "start": 79,
                        "end": 102,
                        "text": "(Novikova et al., 2017)",
                        "ref_id": "BIBREF31"
                    },
                    {
                        "start": 175,
                        "end": 198,
                        "text": "(Novikova et al., 2017)",
                        "ref_id": "BIBREF31"
                    },
                    {
                        "start": 410,
                        "end": 1424,
                        "text": ", atBatNumber, baseNumber, baseReachedNumber, baseStolen, basesRan, batterHitsTries, batterName, batterScoreNumber, battersFacedNumber, battingAverage, battingLineupNumber, catchType, catcherName, competitionName, earnedRunsNumber, errorNumber, fielderName, fielderPosition, finalScore, gameNumber, gameTally, hasLostTeam, hasScored, hasWonTeam, hitNumber, homeAway, homeRunNumber, injuryType, inningNumber, inningScore, inningsPitched, isOut, leftOnBase, locationPlayed, managerName, matchDate, matchStreakNumber, matchStreakType, numberOfStarts, onBaseNumber, outNumber, pitchCount, pitchNumber, pitchResult, pitchResultNumber, pitchType, pitcherName, pitcherRecord, pitcherSaveRecord, pitchesTotalThrown, presidentName, retireNumber, runAverage, runNumber, scoreNumber, scoreTally, standingsGames, startsNumber, stealNumber, strikeNumber, strikeOutNumber, strikeTrajectory, strikingType, teamName, teamRecord, teamStandings, throwDirection, umpireName, umpireType, unearnedRunsNumber, walkNumber, winLossRecord,",
                        "ref_id": null
                    },
                    {
                        "start": 1425,
                        "end": 2210,
                        "text": "incidentCause, incidentLocation, incidentType, suspectAddress, suspectAge, suspectAmount, suspectDescription, suspectGender, suspectStatus, suspectVehicle, victimAddress, victimAge, victimAmount, victimDescription, victimGender, victimName, victimStatus, victimVehicle English incidents accidentAddress, accidentDate, hospitalName, numberOfRoundsFired, personnelArrivedTime, prisonName, shootingNumber, shootingType, suspectAge, suspectAgeGroup, suspectBased, suspectDescription, suspectGender, suspectHeight, suspectName, suspectNumber, suspectOccupation, suspectRace, suspectStatus, suspectVehicle, suspectWeapon, suspectWeight, takenToHospital, victimAge, victimAgeGroup, victimBased, victimGender, victimName, victimNumber, victimOccupation, victimRace, victimStatus, victimVehicle",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Setting Value",
                "sec_num": null
            }
        ],
        "bib_entries": {
            "BIBREF0": {
                "ref_id": "b0",
                "title": "Basisboek journalistiek schrijven. Noordhoff Uitgevers",
                "authors": [
                    {
                        "first": "Henk",
                        "middle": [],
                        "last": "Asbreuk",
                        "suffix": ""
                    },
                    {
                        "first": "Addie",
                        "middle": [],
                        "last": "De Moor",
                        "suffix": ""
                    },
                    {
                        "first": "Esther",
                        "middle": [],
                        "last": "Van Der Meer",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Henk Asbreuk, Addie de Moor, and Esther van der Meer. 2017. Basisboek journalistiek schrijven. No- ordhoff Uitgevers.",
                "links": null
            },
            "BIBREF1": {
                "ref_id": "b1",
                "title": "METEOR: An automatic metric for MT evaluation with improved correlation with human judgments",
                "authors": [
                    {
                        "first": "Satanjeev",
                        "middle": [],
                        "last": "Banerjee",
                        "suffix": ""
                    },
                    {
                        "first": "Alon",
                        "middle": [],
                        "last": "Lavie",
                        "suffix": ""
                    }
                ],
                "year": 2005,
                "venue": "Proceedings of the ACL Workshop on Intrinsic and Extrinsic Evaluation Measures for Machine Translation and/or Summarization",
                "volume": "",
                "issue": "",
                "pages": "65--72",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Satanjeev Banerjee and Alon Lavie. 2005. METEOR: An automatic metric for MT evaluation with im- proved correlation with human judgments. In Pro- ceedings of the ACL Workshop on Intrinsic and Ex- trinsic Evaluation Measures for Machine Transla- tion and/or Summarization, pages 65-72, Ann Ar- bor, Michigan, USA. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF2": {
                "ref_id": "b2",
                "title": "Natural language processing with Python: Analyzing text with the natural language toolkit",
                "authors": [
                    {
                        "first": "Steven",
                        "middle": [],
                        "last": "Bird",
                        "suffix": ""
                    },
                    {
                        "first": "Ewan",
                        "middle": [],
                        "last": "Klein",
                        "suffix": ""
                    },
                    {
                        "first": "Edward",
                        "middle": [],
                        "last": "Loper",
                        "suffix": ""
                    }
                ],
                "year": 2009,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Steven Bird, Ewan Klein, and Edward Loper. 2009. Natural language processing with Python: Analyz- ing text with the natural language toolkit. \"O'Reilly Media, Inc.\".",
                "links": null
            },
            "BIBREF3": {
                "ref_id": "b3",
                "title": "British National Corpus",
                "authors": [],
                "year": 2007,
                "venue": "",
                "volume": "3",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "British National Corpus. 2007. British National Cor- pus, BNC XML Edition edition, volume 3. Dis- tributed by Oxford.",
                "links": null
            },
            "BIBREF4": {
                "ref_id": "b4",
                "title": "Web journalism: from the inverted pyramid to the tumbled pyramid. Biblioteca on-line de ci\u00eancias da comunica\u00e7\u00e3o",
                "authors": [
                    {
                        "first": "Jo\u00e3o",
                        "middle": [],
                        "last": "Canavilhas",
                        "suffix": ""
                    }
                ],
                "year": 2007,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Jo\u00e3o Canavilhas. 2007. Web journalism: from the in- verted pyramid to the tumbled pyramid. Biblioteca on-line de ci\u00eancias da comunica\u00e7\u00e3o.",
                "links": null
            },
            "BIBREF5": {
                "ref_id": "b5",
                "title": "Neural datato-text generation: A comparison between pipeline and end-to-end architectures",
                "authors": [
                    {
                        "first": "Chris",
                        "middle": [],
                        "last": "Thiago Castro Ferreira",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Van Der Lee",
                        "suffix": ""
                    },
                    {
                        "first": "Emiel",
                        "middle": [],
                        "last": "Emiel Van Miltenburg",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Krahmer",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Thiago Castro Ferreira, Chris van der Lee, Emiel van Miltenburg, and Emiel Krahmer. 2019. Neural data- to-text generation: A comparison between pipeline and end-to-end architectures. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing, Hong Kong, SAR. Associa- tion for Computational Linguistics.",
                "links": null
            },
            "BIBREF6": {
                "ref_id": "b6",
                "title": "Enriching the WebNLG corpus",
                "authors": [
                    {
                        "first": "Diego",
                        "middle": [],
                        "last": "Thiago Castro Ferreira",
                        "suffix": ""
                    },
                    {
                        "first": "Emiel",
                        "middle": [],
                        "last": "Moussallem",
                        "suffix": ""
                    },
                    {
                        "first": "Sander",
                        "middle": [],
                        "last": "Krahmer",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Wubben",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "Proceedings of the 11th International Conference on Natural Language Generation",
                "volume": "",
                "issue": "",
                "pages": "171--176",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/W18-6521"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Thiago Castro Ferreira, Diego Moussallem, Emiel Krahmer, and Sander Wubben. 2018. Enriching the WebNLG corpus. In Proceedings of the 11th Inter- national Conference on Natural Language Genera- tion, pages 171-176, Tilburg University, The Nether- lands. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF7": {
                "ref_id": "b7",
                "title": "How complex is that sentence? A proposed revision of the Rosenberg and Abbeduto D-Level Scale",
                "authors": [
                    {
                        "first": "Congzhou",
                        "middle": [],
                        "last": "Michael A Covington",
                        "suffix": ""
                    },
                    {
                        "first": "Cati",
                        "middle": [],
                        "last": "He",
                        "suffix": ""
                    },
                    {
                        "first": "Lorina",
                        "middle": [],
                        "last": "Brown",
                        "suffix": ""
                    },
                    {
                        "first": "John",
                        "middle": [],
                        "last": "Naci",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Brown",
                        "suffix": ""
                    }
                ],
                "year": 2006,
                "venue": "CASPR Research Report",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Michael A Covington, Congzhou He, Cati Brown, Lo- rina Naci, and John Brown. 2006. How complex is that sentence? A proposed revision of the Rosenberg and Abbeduto D-Level Scale. In CASPR Research Report 2006-01. University of Georgia Artificial In- telligence Center, Athens, GA.",
                "links": null
            },
            "BIBREF8": {
                "ref_id": "b8",
                "title": "Lexical richness in the spontaneous speech of bilinguals",
                "authors": [
                    {
                        "first": "Helmut",
                        "middle": [],
                        "last": "Daller",
                        "suffix": ""
                    },
                    {
                        "first": "Roeland",
                        "middle": [],
                        "last": "Van Hout",
                        "suffix": ""
                    },
                    {
                        "first": "Jeanine",
                        "middle": [],
                        "last": "Treffers-Daller",
                        "suffix": ""
                    }
                ],
                "year": 2003,
                "venue": "Applied linguistics",
                "volume": "24",
                "issue": "2",
                "pages": "197--222",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Helmut Daller, Roeland Van Hout, and Jeanine Treffers-Daller. 2003. Lexical richness in the spon- taneous speech of bilinguals. Applied linguistics, 24(2):197-222.",
                "links": null
            },
            "BIBREF9": {
                "ref_id": "b9",
                "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": "Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies",
                "volume": "1",
                "issue": "",
                "pages": "4171--4186",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/N19-1423"
                    ]
                },
                "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 Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Associ- ation for Computational Linguistics.",
                "links": null
            },
            "BIBREF10": {
                "ref_id": "b10",
                "title": "Automatic evaluation of machine translation quality using n-gram cooccurrence statistics",
                "authors": [
                    {
                        "first": "George",
                        "middle": [],
                        "last": "Doddington",
                        "suffix": ""
                    }
                ],
                "year": 2002,
                "venue": "Proceedings of the Second International Conference on Human Language Technology Research",
                "volume": "",
                "issue": "",
                "pages": "138--145",
                "other_ids": {
                    "DOI": [
                        "10.3115/1289189.1289273"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "George Doddington. 2002. Automatic evaluation of machine translation quality using n-gram co- occurrence statistics. In Proceedings of the Sec- ond International Conference on Human Language Technology Research, pages 138-145. Morgan Kauf- mann Publishers Inc.",
                "links": null
            },
            "BIBREF11": {
                "ref_id": "b11",
                "title": "Training a natural language generator from unaligned data",
                "authors": [
                    {
                        "first": "Ond\u0159ej",
                        "middle": [],
                        "last": "Du\u0161ek",
                        "suffix": ""
                    },
                    {
                        "first": "Filip",
                        "middle": [],
                        "last": "Jur\u010d\u00ed\u010dek",
                        "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": "451--461",
                "other_ids": {
                    "DOI": [
                        "10.3115/v1/P15-1044"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Ond\u0159ej Du\u0161ek and Filip Jur\u010d\u00ed\u010dek. 2015. Training a nat- ural language generator from unaligned data. In Pro- ceedings of the 53rd Annual Meeting of the Associa- tion for Computational Linguistics and the 7th Inter- national Joint Conference on Natural Language Pro- cessing (Volume 1: Long Papers), pages 451-461, Beijing, China. Association for Computational Lin- guistics.",
                "links": null
            },
            "BIBREF12": {
                "ref_id": "b12",
                "title": "Creating training corpora for NLG micro-planners",
                "authors": [
                    {
                        "first": "Claire",
                        "middle": [],
                        "last": "Gardent",
                        "suffix": ""
                    },
                    {
                        "first": "Anastasia",
                        "middle": [],
                        "last": "Shimorina",
                        "suffix": ""
                    },
                    {
                        "first": "Shashi",
                        "middle": [],
                        "last": "Narayan",
                        "suffix": ""
                    },
                    {
                        "first": "Laura",
                        "middle": [],
                        "last": "Perez-Beltrachini",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics",
                "volume": "1",
                "issue": "",
                "pages": "179--188",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/P17-1017"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Claire Gardent, Anastasia Shimorina, Shashi Narayan, and Laura Perez-Beltrachini. 2017a. Creating train- ing corpora for NLG micro-planners. In Proceed- ings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Pa- pers), pages 179-188, Vancouver, Canada. Associa- tion for Computational Linguistics.",
                "links": null
            },
            "BIBREF13": {
                "ref_id": "b13",
                "title": "The WebNLG challenge: Generating text from RDF data",
                "authors": [
                    {
                        "first": "Claire",
                        "middle": [],
                        "last": "Gardent",
                        "suffix": ""
                    },
                    {
                        "first": "Anastasia",
                        "middle": [],
                        "last": "Shimorina",
                        "suffix": ""
                    },
                    {
                        "first": "Shashi",
                        "middle": [],
                        "last": "Narayan",
                        "suffix": ""
                    },
                    {
                        "first": "Laura",
                        "middle": [],
                        "last": "Perez-Beltrachini",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "Proceedings of the 10th International Conference on Natural Language Generation",
                "volume": "",
                "issue": "",
                "pages": "124--133",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/W17-3518"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Claire Gardent, Anastasia Shimorina, Shashi Narayan, and Laura Perez-Beltrachini. 2017b. The WebNLG challenge: Generating text from RDF data. In Pro- ceedings of the 10th International Conference on Natural Language Generation, pages 124-133, San- tiago de Compostela, Spain. Association for Compu- tational Linguistics.",
                "links": null
            },
            "BIBREF14": {
                "ref_id": "b14",
                "title": "Survey of the state of the art in natural language generation: Core tasks, applications and evaluation",
                "authors": [
                    {
                        "first": "Albert",
                        "middle": [],
                        "last": "Gatt",
                        "suffix": ""
                    },
                    {
                        "first": "Emiel",
                        "middle": [],
                        "last": "Krahmer",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "Journal of Artificial Intelligence Research",
                "volume": "61",
                "issue": "",
                "pages": "65--170",
                "other_ids": {
                    "DOI": [
                        "10.1613/jair.5477"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Albert Gatt and Emiel Krahmer. 2018. Survey of the state of the art in natural language generation: Core tasks, applications and evaluation. Journal of Artifi- cial Intelligence Research, 61:65-170.",
                "links": null
            },
            "BIBREF15": {
                "ref_id": "b15",
                "title": "Retrieving, cleaning and analysing Dutch news articles about traffic accidents",
                "authors": [
                    {
                        "first": "Barry",
                        "middle": [],
                        "last": "Hendriks",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Barry Hendriks. 2019. Retrieving, cleaning and analysing Dutch news articles about traffic acci- dents. Master's thesis, University of Amsterdam, The Netherlands.",
                "links": null
            },
            "BIBREF16": {
                "ref_id": "b16",
                "title": "2017. spaCy 2: Natural language understanding with Bloom embeddings, convolutional neural networks and incremental parsing",
                "authors": [
                    {
                        "first": "Matthew",
                        "middle": [],
                        "last": "Honnibal",
                        "suffix": ""
                    },
                    {
                        "first": "Ines",
                        "middle": [],
                        "last": "Montani",
                        "suffix": ""
                    }
                ],
                "year": null,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Matthew Honnibal and Ines Montani. 2017. spaCy 2: Natural language understanding with Bloom embed- dings, convolutional neural networks and incremen- tal parsing. To appear.",
                "links": null
            },
            "BIBREF17": {
                "ref_id": "b17",
                "title": "PAN: Pipeline assisted neural networks model for data-to-text generation in social internet of things",
                "authors": [
                    {
                        "first": "Nan",
                        "middle": [],
                        "last": "Jiang",
                        "suffix": ""
                    },
                    {
                        "first": "Jing",
                        "middle": [],
                        "last": "Chen",
                        "suffix": ""
                    },
                    {
                        "first": "Ri-Gui",
                        "middle": [],
                        "last": "Zhou",
                        "suffix": ""
                    },
                    {
                        "first": "Changxing",
                        "middle": [],
                        "last": "Wu",
                        "suffix": ""
                    },
                    {
                        "first": "Honglong",
                        "middle": [],
                        "last": "Chen",
                        "suffix": ""
                    },
                    {
                        "first": "Jiaqi",
                        "middle": [],
                        "last": "Zheng",
                        "suffix": ""
                    },
                    {
                        "first": "Tao",
                        "middle": [],
                        "last": "Wan",
                        "suffix": ""
                    }
                ],
                "year": 2020,
                "venue": "Information Sciences",
                "volume": "530",
                "issue": "",
                "pages": "167--179",
                "other_ids": {
                    "DOI": [
                        "10.1016/j.ins.2020.03.080"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Nan Jiang, Jing Chen, Ri-Gui Zhou, Changxing Wu, Honglong Chen, Jiaqi Zheng, and Tao Wan. 2020. PAN: Pipeline assisted neural networks model for data-to-text generation in social internet of things. Information Sciences, 530:167-179.",
                "links": null
            },
            "BIBREF18": {
                "ref_id": "b18",
                "title": "Computational generation of referring expressions: A survey",
                "authors": [
                    {
                        "first": "Emiel",
                        "middle": [],
                        "last": "Krahmer",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Kees Van Deemter",
                        "suffix": ""
                    }
                ],
                "year": 2012,
                "venue": "Computational Linguistics",
                "volume": "38",
                "issue": "1",
                "pages": "173--218",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Emiel Krahmer and Kees van Deemter. 2012. Compu- tational generation of referring expressions: A sur- vey. Computational Linguistics, 38(1):173-218.",
                "links": null
            },
            "BIBREF19": {
                "ref_id": "b19",
                "title": "Dick Van der Lugt, and Ben Rogmans",
                "authors": [
                    {
                        "first": "Nico",
                        "middle": [],
                        "last": "Kussendrager",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "Basisboek journalistiek. Noordhoff",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Nico Kussendrager, Dick Van der Lugt, and Ben Rog- mans. 2018. Basisboek journalistiek. Noordhoff.",
                "links": null
            },
            "BIBREF20": {
                "ref_id": "b20",
                "title": "The measurement of observer agreement for categorical data",
                "authors": [
                    {
                        "first": "Richard",
                        "middle": [],
                        "last": "Landis",
                        "suffix": ""
                    },
                    {
                        "first": "Gary G",
                        "middle": [],
                        "last": "Koch",
                        "suffix": ""
                    }
                ],
                "year": 1977,
                "venue": "Biometrics",
                "volume": "",
                "issue": "",
                "pages": "159--174",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "J Richard Landis and Gary G Koch. 1977. The mea- surement of observer agreement for categorical data. Biometrics, pages 159-174.",
                "links": null
            },
            "BIBREF21": {
                "ref_id": "b21",
                "title": "Best practices for the human evaluation of automatically generated text",
                "authors": [
                    {
                        "first": "Chris",
                        "middle": [],
                        "last": "Van Der Lee",
                        "suffix": ""
                    },
                    {
                        "first": "Albert",
                        "middle": [],
                        "last": "Gatt",
                        "suffix": ""
                    },
                    {
                        "first": "Sander",
                        "middle": [],
                        "last": "Emiel Van Miltenburg",
                        "suffix": ""
                    },
                    {
                        "first": "Emiel",
                        "middle": [],
                        "last": "Wubben",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Krahmer",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "Proceedings of the 12th International Conference on Natural Language Generation",
                "volume": "",
                "issue": "",
                "pages": "355--368",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/W19-8643"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Chris van der Lee, Albert Gatt, Emiel van Miltenburg, Sander Wubben, and Emiel Krahmer. 2019. Best practices for the human evaluation of automatically generated text. In Proceedings of the 12th Interna- tional Conference on Natural Language Generation, pages 355-368, Tokyo, Japan. Association for Com- putational Linguistics.",
                "links": null
            },
            "BIBREF22": {
                "ref_id": "b22",
                "title": "Evaluating the text quality, human likeness and tailoring component of PASS: A Dutch data-to-text system for soccer",
                "authors": [
                    {
                        "first": "Chris",
                        "middle": [],
                        "last": "Van Der Lee",
                        "suffix": ""
                    },
                    {
                        "first": "Bart",
                        "middle": [],
                        "last": "Verduijn",
                        "suffix": ""
                    },
                    {
                        "first": "Emiel",
                        "middle": [],
                        "last": "Krahmer",
                        "suffix": ""
                    },
                    {
                        "first": "Sander",
                        "middle": [],
                        "last": "Wubben",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "Proceedings of the 27th International Conference on Computational Linguistics",
                "volume": "",
                "issue": "",
                "pages": "962--972",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Chris van der Lee, Bart Verduijn, Emiel Krahmer, and Sander Wubben. 2018. Evaluating the text quality, human likeness and tailoring component of PASS: A Dutch data-to-text system for soccer. In Proceed- ings of the 27th International Conference on Compu- tational Linguistics, pages 962-972, Santa Fe, New Mexico, USA. Association for Computational Lin- guistics.",
                "links": null
            },
            "BIBREF23": {
                "ref_id": "b23",
                "title": "Automated journalism as a source of and a diagnostic device for bias in reporting",
                "authors": [
                    {
                        "first": "Leo",
                        "middle": [],
                        "last": "Lepp\u00e4nen",
                        "suffix": ""
                    },
                    {
                        "first": "Hanna",
                        "middle": [],
                        "last": "Tuulonen",
                        "suffix": ""
                    },
                    {
                        "first": "Stefanie",
                        "middle": [],
                        "last": "Sir\u00e9n-Heikel",
                        "suffix": ""
                    }
                ],
                "year": 2020,
                "venue": "Media and Communication",
                "volume": "8",
                "issue": "3",
                "pages": "1--11",
                "other_ids": {
                    "DOI": [
                        "10.17645/mac.v8i3.3022"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Leo Lepp\u00e4nen, Hanna Tuulonen, Stefanie Sir\u00e9n-Heikel, et al. 2020. Automated journalism as a source of and a diagnostic device for bias in reporting. Media and Communication, 8(3):1-11.",
                "links": null
            },
            "BIBREF24": {
                "ref_id": "b24",
                "title": "ROUGE: A package for automatic evaluation of summaries",
                "authors": [
                    {
                        "first": "Chin-Yew",
                        "middle": [],
                        "last": "Lin",
                        "suffix": ""
                    }
                ],
                "year": 2004,
                "venue": "Text Summarization Branches Out",
                "volume": "",
                "issue": "",
                "pages": "74--81",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Chin-Yew Lin. 2004. ROUGE: A package for auto- matic evaluation of summaries. In Text Summariza- tion Branches Out, pages 74-81, Barcelona, Spain. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF25": {
                "ref_id": "b25",
                "title": "Automatic measurement of syntactic complexity in child language acquisition",
                "authors": [
                    {
                        "first": "Xiaofei",
                        "middle": [],
                        "last": "Lu",
                        "suffix": ""
                    }
                ],
                "year": 2009,
                "venue": "International Journal of Corpus Linguistics",
                "volume": "14",
                "issue": "1",
                "pages": "3--28",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Xiaofei Lu. 2009. Automatic measurement of syntac- tic complexity in child language acquisition. Inter- national Journal of Corpus Linguistics, 14(1):3-28.",
                "links": null
            },
            "BIBREF26": {
                "ref_id": "b26",
                "title": "The Stanford CoreNLP natural language processing toolkit",
                "authors": [
                    {
                        "first": "Christopher",
                        "middle": [
                            "D"
                        ],
                        "last": "Manning",
                        "suffix": ""
                    },
                    {
                        "first": "Mihai",
                        "middle": [],
                        "last": "Surdeanu",
                        "suffix": ""
                    },
                    {
                        "first": "John",
                        "middle": [],
                        "last": "Bauer",
                        "suffix": ""
                    },
                    {
                        "first": "Jenny",
                        "middle": [],
                        "last": "Finkel",
                        "suffix": ""
                    },
                    {
                        "first": "Steven",
                        "middle": [
                            "J"
                        ],
                        "last": "Bethard",
                        "suffix": ""
                    },
                    {
                        "first": "David",
                        "middle": [],
                        "last": "Mc-Closky",
                        "suffix": ""
                    }
                ],
                "year": 2014,
                "venue": "Association for Computational Linguistics (ACL) System Demonstrations",
                "volume": "",
                "issue": "",
                "pages": "55--60",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Christopher D. Manning, Mihai Surdeanu, John Bauer, Jenny Finkel, Steven J. Bethard, and David Mc- Closky. 2014. The Stanford CoreNLP natural lan- guage processing toolkit. In Association for Compu- tational Linguistics (ACL) System Demonstrations, pages 55-60.",
                "links": null
            },
            "BIBREF27": {
                "ref_id": "b27",
                "title": "The second multilingual surface realisation shared task (SR'19): Overview and evaluation results",
                "authors": [
                    {
                        "first": "Simon",
                        "middle": [],
                        "last": "Mille",
                        "suffix": ""
                    },
                    {
                        "first": "Anja",
                        "middle": [],
                        "last": "Belz",
                        "suffix": ""
                    },
                    {
                        "first": "Bernd",
                        "middle": [],
                        "last": "Bohnet",
                        "suffix": ""
                    },
                    {
                        "first": "Yvette",
                        "middle": [],
                        "last": "Graham",
                        "suffix": ""
                    },
                    {
                        "first": "Leo",
                        "middle": [],
                        "last": "Wanner",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "Proceedings of the 2nd Workshop on Multilingual Surface Realisation (MSR 2019)",
                "volume": "",
                "issue": "",
                "pages": "1--17",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/D19-6301"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Simon Mille, Anja Belz, Bernd Bohnet, Yvette Gra- ham, and Leo Wanner. 2019. The second mul- tilingual surface realisation shared task (SR'19): Overview and evaluation results. In Proceedings of the 2nd Workshop on Multilingual Surface Realisa- tion (MSR 2019), pages 1-17, Hong Kong, China. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF28": {
                "ref_id": "b28",
                "title": "Prodigy: A new annotation tool for radically efficient machine teaching",
                "authors": [
                    {
                        "first": "Ines",
                        "middle": [],
                        "last": "Montani",
                        "suffix": ""
                    },
                    {
                        "first": "Matthew",
                        "middle": [],
                        "last": "Honnibal",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "Artificial Intelligence",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Ines Montani and Matthew Honnibal. 2018. Prodigy: A new annotation tool for radically efficient machine teaching. Artificial Intelligence.",
                "links": null
            },
            "BIBREF29": {
                "ref_id": "b29",
                "title": "Improving quality and efficiency in planbased neural data-to-text generation",
                "authors": [
                    {
                        "first": "Amit",
                        "middle": [],
                        "last": "Moryossef",
                        "suffix": ""
                    },
                    {
                        "first": "Yoav",
                        "middle": [],
                        "last": "Goldberg",
                        "suffix": ""
                    },
                    {
                        "first": "Ido",
                        "middle": [],
                        "last": "Dagan",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "Proceedings of the 12th International Conference on Natural Language Generation",
                "volume": "",
                "issue": "",
                "pages": "377--382",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/W19-8645"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Amit Moryossef, Yoav Goldberg, and Ido Dagan. 2019a. Improving quality and efficiency in plan- based neural data-to-text generation. In Proceed- ings of the 12th International Conference on Nat- ural Language Generation, pages 377-382, Tokyo, Japan. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF30": {
                "ref_id": "b30",
                "title": "Step-by-step: Separating planning from realization in neural data-to-text generation",
                "authors": [
                    {
                        "first": "Amit",
                        "middle": [],
                        "last": "Moryossef",
                        "suffix": ""
                    },
                    {
                        "first": "Yoav",
                        "middle": [],
                        "last": "Goldberg",
                        "suffix": ""
                    },
                    {
                        "first": "Ido",
                        "middle": [],
                        "last": "Dagan",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies",
                "volume": "1",
                "issue": "",
                "pages": "2267--2277",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/N19-1236"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Amit Moryossef, Yoav Goldberg, and Ido Dagan. 2019b. Step-by-step: Separating planning from real- ization in neural data-to-text generation. In Proceed- ings of the 2019 Conference of the North American Chapter of the Association for Computational Lin- guistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 2267-2277, Min- neapolis, Minnesota. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF31": {
                "ref_id": "b31",
                "title": "The E2E dataset: New challenges for endto-end generation",
                "authors": [
                    {
                        "first": "Jekaterina",
                        "middle": [],
                        "last": "Novikova",
                        "suffix": ""
                    },
                    {
                        "first": "Ond\u0159ej",
                        "middle": [],
                        "last": "Du\u0161ek",
                        "suffix": ""
                    },
                    {
                        "first": "Verena",
                        "middle": [],
                        "last": "Rieser",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "Proceedings of the 18th Annual SIGdial Meeting on Discourse and Dialogue",
                "volume": "",
                "issue": "",
                "pages": "201--206",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/W17-5525"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Jekaterina Novikova, Ond\u0159ej Du\u0161ek, and Verena Rieser. 2017. The E2E dataset: New challenges for end- to-end generation. In Proceedings of the 18th An- nual SIGdial Meeting on Discourse and Dialogue, pages 201-206, Saarbr\u00fccken, Germany. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF32": {
                "ref_id": "b32",
                "title": "The construction of a 500-million-word reference corpus of contemporary written Dutch",
                "authors": [
                    {
                        "first": "Nelleke",
                        "middle": [],
                        "last": "Oostdijk",
                        "suffix": ""
                    },
                    {
                        "first": "Martin",
                        "middle": [],
                        "last": "Reynaert",
                        "suffix": ""
                    },
                    {
                        "first": "V\u00e9ronique",
                        "middle": [],
                        "last": "Hoste",
                        "suffix": ""
                    },
                    {
                        "first": "Ineke",
                        "middle": [],
                        "last": "Schuurman",
                        "suffix": ""
                    }
                ],
                "year": 2013,
                "venue": "Essential speech and language technology for Dutch",
                "volume": "",
                "issue": "",
                "pages": "219--247",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Nelleke Oostdijk, Martin Reynaert, V\u00e9ronique Hoste, and Ineke Schuurman. 2013. The construction of a 500-million-word reference corpus of contemporary written Dutch. In Essential speech and language technology for Dutch, pages 219-247. Springer, Berlin, Heidelberg.",
                "links": null
            },
            "BIBREF33": {
                "ref_id": "b33",
                "title": "Curate and generate: A corpus and method for joint control of semantics and style in neural NLG",
                "authors": [
                    {
                        "first": "Shereen",
                        "middle": [],
                        "last": "Oraby",
                        "suffix": ""
                    },
                    {
                        "first": "Vrindavan",
                        "middle": [],
                        "last": "Harrison",
                        "suffix": ""
                    },
                    {
                        "first": "Abteen",
                        "middle": [],
                        "last": "Ebrahimi",
                        "suffix": ""
                    },
                    {
                        "first": "Marilyn",
                        "middle": [],
                        "last": "Walker",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics",
                "volume": "",
                "issue": "",
                "pages": "5938--5951",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/P19-1596"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Shereen Oraby, Vrindavan Harrison, Abteen Ebrahimi, and Marilyn Walker. 2019. Curate and generate: A corpus and method for joint control of semantics and style in neural NLG. In Proceedings of the 57th An- nual Meeting of the Association for Computational Linguistics, pages 5938-5951, Florence, Italy. Asso- ciation for Computational Linguistics.",
                "links": null
            },
            "BIBREF34": {
                "ref_id": "b34",
                "title": "T-scan: A new tool for analyzing Dutch text",
                "authors": [
                    {
                        "first": "Rogier",
                        "middle": [],
                        "last": "Henk Pander Maat",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Kraf",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Van Den",
                        "suffix": ""
                    },
                    {
                        "first": "Nick",
                        "middle": [],
                        "last": "Bosch",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Dekker",
                        "suffix": ""
                    },
                    {
                        "first": "S",
                        "middle": [],
                        "last": "Van Gompel",
                        "suffix": ""
                    },
                    {
                        "first": "Ted",
                        "middle": [],
                        "last": "Kleijn",
                        "suffix": ""
                    },
                    {
                        "first": "K",
                        "middle": [],
                        "last": "Sanders",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Van Der",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Sloot",
                        "suffix": ""
                    }
                ],
                "year": 2014,
                "venue": "Computational Linguistics in The Netherlands journal",
                "volume": "4",
                "issue": "",
                "pages": "53--74",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Henk Pander Maat, Rogier Kraf, Antal van den Bosch, Nick Dekker, M van Gompel, S Kleijn, Ted Sanders, and K van der Sloot. 2014. T-scan: A new tool for analyzing Dutch text. Computational Linguistics in The Netherlands journal, 4:53-74.",
                "links": null
            },
            "BIBREF35": {
                "ref_id": "b35",
                "title": "BLEU: A method for automatic evaluation of machine translation",
                "authors": [
                    {
                        "first": "Kishore",
                        "middle": [],
                        "last": "Papineni",
                        "suffix": ""
                    },
                    {
                        "first": "Salim",
                        "middle": [],
                        "last": "Roukos",
                        "suffix": ""
                    },
                    {
                        "first": "Todd",
                        "middle": [],
                        "last": "Ward",
                        "suffix": ""
                    },
                    {
                        "first": "Wei-Jing",
                        "middle": [],
                        "last": "Zhu",
                        "suffix": ""
                    }
                ],
                "year": 2002,
                "venue": "Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics",
                "volume": "",
                "issue": "",
                "pages": "311--318",
                "other_ids": {
                    "DOI": [
                        "10.3115/1073083.1073135"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Kishore Papineni, Salim Roukos, Todd Ward, and Wei- Jing Zhu. 2002. BLEU: A method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Com- putational Linguistics, pages 311-318, Philadelphia, Pennsylvania, USA. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF36": {
                "ref_id": "b36",
                "title": "ToTTo: A controlled table-to-text generation dataset",
                "authors": [
                    {
                        "first": "P",
                        "middle": [],
                        "last": "Ankur",
                        "suffix": ""
                    },
                    {
                        "first": "Xuezhi",
                        "middle": [],
                        "last": "Parikh",
                        "suffix": ""
                    },
                    {
                        "first": "Sebastian",
                        "middle": [],
                        "last": "Wang",
                        "suffix": ""
                    },
                    {
                        "first": "Manaal",
                        "middle": [],
                        "last": "Gehrmann",
                        "suffix": ""
                    },
                    {
                        "first": "Bhuwan",
                        "middle": [],
                        "last": "Faruqui",
                        "suffix": ""
                    },
                    {
                        "first": "Diyi",
                        "middle": [],
                        "last": "Dhingra",
                        "suffix": ""
                    },
                    {
                        "first": "Dipanjan",
                        "middle": [],
                        "last": "Yang",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Das",
                        "suffix": ""
                    }
                ],
                "year": 2020,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:2004.14373"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Ankur P. Parikh, Xuezhi Wang, Sebastian Gehrmann, Manaal Faruqui, Bhuwan Dhingra, Diyi Yang, and Dipanjan Das. 2020. ToTTo: A controlled table-to-text generation dataset. arXiv preprint arXiv:2004.14373.",
                "links": null
            },
            "BIBREF37": {
                "ref_id": "b37",
                "title": "Analysing data-to-text generation benchmarks",
                "authors": [
                    {
                        "first": "Laura",
                        "middle": [],
                        "last": "Perez",
                        "suffix": ""
                    },
                    {
                        "first": "-",
                        "middle": [],
                        "last": "Beltrachini",
                        "suffix": ""
                    },
                    {
                        "first": "Claire",
                        "middle": [],
                        "last": "Gardent",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "Proceedings of the 10th International Conference on Natural Language Generation",
                "volume": "",
                "issue": "",
                "pages": "238--242",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/W17-3537"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Laura Perez-Beltrachini and Claire Gardent. 2017. Analysing data-to-text generation benchmarks. In Proceedings of the 10th International Conference on Natural Language Generation, pages 238-242, San- tiago de Compostela, Spain. Association for Compu- tational Linguistics.",
                "links": null
            },
            "BIBREF38": {
                "ref_id": "b38",
                "title": "A call for clarity in reporting BLEU scores",
                "authors": [
                    {
                        "first": "Matt",
                        "middle": [],
                        "last": "Post",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "Proceedings of the Third Conference on Machine Translation: Research Papers",
                "volume": "",
                "issue": "",
                "pages": "186--191",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Matt Post. 2018. A call for clarity in reporting BLEU scores. In Proceedings of the Third Conference on Machine Translation: Research Papers, pages 186- 191, Belgium, Brussels. Association for Computa- tional Linguistics.",
                "links": null
            },
            "BIBREF39": {
                "ref_id": "b39",
                "title": "SoMaJo: Stateof-the-art tokenization for German web and social media texts",
                "authors": [
                    {
                        "first": "Thomas",
                        "middle": [],
                        "last": "Proisl",
                        "suffix": ""
                    },
                    {
                        "first": "Peter",
                        "middle": [],
                        "last": "Uhrig",
                        "suffix": ""
                    }
                ],
                "year": 2016,
                "venue": "Proceedings of the 10th Web as Corpus Workshop",
                "volume": "",
                "issue": "",
                "pages": "57--62",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/W16-2607"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Thomas Proisl and Peter Uhrig. 2016. SoMaJo: State- of-the-art tokenization for German web and social media texts. In Proceedings of the 10th Web as Cor- pus Workshop, pages 57-62, Berlin. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF40": {
                "ref_id": "b40",
                "title": "Data-to-text generation with entity modeling",
                "authors": [
                    {
                        "first": "Ratish",
                        "middle": [],
                        "last": "Puduppully",
                        "suffix": ""
                    },
                    {
                        "first": "Li",
                        "middle": [],
                        "last": "Dong",
                        "suffix": ""
                    },
                    {
                        "first": "Mirella",
                        "middle": [],
                        "last": "Lapata",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics",
                "volume": "",
                "issue": "",
                "pages": "2023--2035",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/P19-1195"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Ratish Puduppully, Li Dong, and Mirella Lapata. 2019. Data-to-text generation with entity modeling. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 2023-2035, Florence, Italy. Association for Compu- tational Linguistics.",
                "links": null
            },
            "BIBREF41": {
                "ref_id": "b41",
                "title": "DART: Open-domain structured data record to text generation",
                "authors": [
                    {
                        "first": "Dragomir",
                        "middle": [],
                        "last": "Radev",
                        "suffix": ""
                    },
                    {
                        "first": "Rui",
                        "middle": [],
                        "last": "Zhang",
                        "suffix": ""
                    },
                    {
                        "first": "Amrit",
                        "middle": [],
                        "last": "Rau",
                        "suffix": ""
                    },
                    {
                        "first": "Abhinand",
                        "middle": [],
                        "last": "Sivaprasad",
                        "suffix": ""
                    },
                    {
                        "first": "Chiachun",
                        "middle": [],
                        "last": "Hsieh",
                        "suffix": ""
                    },
                    {
                        "first": "Nazneen",
                        "middle": [],
                        "last": "Fatema Rajani",
                        "suffix": ""
                    },
                    {
                        "first": "Xiangru",
                        "middle": [],
                        "last": "Tang",
                        "suffix": ""
                    },
                    {
                        "first": "Aadit",
                        "middle": [],
                        "last": "Vyas",
                        "suffix": ""
                    },
                    {
                        "first": "Neha",
                        "middle": [],
                        "last": "Verma",
                        "suffix": ""
                    },
                    {
                        "first": "Pranav",
                        "middle": [],
                        "last": "Krishna",
                        "suffix": ""
                    },
                    {
                        "first": "Yangxiaokang",
                        "middle": [],
                        "last": "Liu",
                        "suffix": ""
                    },
                    {
                        "first": "Nadia",
                        "middle": [],
                        "last": "Irwanto",
                        "suffix": ""
                    },
                    {
                        "first": "Jessica",
                        "middle": [],
                        "last": "Pan",
                        "suffix": ""
                    },
                    {
                        "first": "Faiaz",
                        "middle": [],
                        "last": "Rahman",
                        "suffix": ""
                    },
                    {
                        "first": "Ahmad",
                        "middle": [],
                        "last": "Zaidi",
                        "suffix": ""
                    },
                    {
                        "first": "Murori",
                        "middle": [],
                        "last": "Mutuma",
                        "suffix": ""
                    }
                ],
                "year": 2020,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:2007.02871"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Dragomir Radev, Rui Zhang, Amrit Rau, Abhinand Sivaprasad, Chiachun Hsieh, Nazneen Fatema Ra- jani, Xiangru Tang, Aadit Vyas, Neha Verma, Pranav Krishna, Yangxiaokang Liu, Nadia Irwanto, Jessica Pan, Faiaz Rahman, Ahmad Zaidi, Murori Mutuma, Yasin Tarabar, Ankit Gupta, Tao Yu, Yi Chern Tan, Xi Victoria Lin, Caiming Xiong, and Richard Socher. 2020. DART: Open-domain struc- tured data record to text generation. arXiv preprint arXiv:2007.02871.",
                "links": null
            },
            "BIBREF42": {
                "ref_id": "b42",
                "title": "Hallucination in neural nlg. Re",
                "authors": [
                    {
                        "first": "Ehud",
                        "middle": [],
                        "last": "Reiter",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Ehud Reiter. 2018a. Hallucination in neural nlg. Re- trieved from https://ehudreiter.com/2018/ 11/12/hallucination-in-neural-nlg/ on August 27, 2020.",
                "links": null
            },
            "BIBREF43": {
                "ref_id": "b43",
                "title": "A structured review of the validity of BLEU",
                "authors": [
                    {
                        "first": "Ehud",
                        "middle": [],
                        "last": "Reiter",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "Computational Linguistics",
                "volume": "",
                "issue": "",
                "pages": "1--12",
                "other_ids": {
                    "DOI": [
                        "10.1162/coli_a_00322"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Ehud Reiter. 2018b. A structured review of the validity of BLEU. Computational Linguistics, pages 1-12.",
                "links": null
            },
            "BIBREF44": {
                "ref_id": "b44",
                "title": "Building Natural Language Generation Systems",
                "authors": [
                    {
                        "first": "Ehud",
                        "middle": [],
                        "last": "Reiter",
                        "suffix": ""
                    },
                    {
                        "first": "Robert",
                        "middle": [],
                        "last": "Dale",
                        "suffix": ""
                    }
                ],
                "year": 2000,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "DOI": [
                        "10.1017/CBO9780511519857"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Ehud Reiter and Robert Dale. 2000. Building Natural Language Generation Systems. Cambridge Univer- sity Press, New York, NY, USA.",
                "links": null
            },
            "BIBREF45": {
                "ref_id": "b45",
                "title": "Indicators of linguistic competence in the peer group conversational behavior of mildly retarded adults",
                "authors": [
                    {
                        "first": "Sheldon",
                        "middle": [],
                        "last": "Rosenberg",
                        "suffix": ""
                    },
                    {
                        "first": "Leonard",
                        "middle": [],
                        "last": "Abbeduto",
                        "suffix": ""
                    }
                ],
                "year": 1987,
                "venue": "Applied Psycholinguistics",
                "volume": "8",
                "issue": "1",
                "pages": "19--32",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Sheldon Rosenberg and Leonard Abbeduto. 1987. In- dicators of linguistic competence in the peer group conversational behavior of mildly retarded adults. Applied Psycholinguistics, 8(1):19-32.",
                "links": null
            },
            "BIBREF46": {
                "ref_id": "b46",
                "title": "BLEURT: Learning robust metrics for text generation",
                "authors": [
                    {
                        "first": "Thibault",
                        "middle": [],
                        "last": "Sellam",
                        "suffix": ""
                    },
                    {
                        "first": "Dipanjan",
                        "middle": [],
                        "last": "Das",
                        "suffix": ""
                    },
                    {
                        "first": "Ankur",
                        "middle": [],
                        "last": "Parikh",
                        "suffix": ""
                    }
                ],
                "year": 2020,
                "venue": "Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics",
                "volume": "",
                "issue": "",
                "pages": "7881--7892",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Thibault Sellam, Dipanjan Das, and Ankur Parikh. 2020. BLEURT: Learning robust metrics for text generation. In Proceedings of the 58th Annual Meet- ing of the Association for Computational Linguistics, pages 7881-7892, Online. Association for Computa- tional Linguistics.",
                "links": null
            },
            "BIBREF47": {
                "ref_id": "b47",
                "title": "Relevance of unsupervised metrics in task-oriented dialogue for evaluating natural language generation",
                "authors": [
                    {
                        "first": "Shikhar",
                        "middle": [],
                        "last": "Sharma",
                        "suffix": ""
                    },
                    {
                        "first": "Layla",
                        "middle": [
                            "El"
                        ],
                        "last": "Asri",
                        "suffix": ""
                    },
                    {
                        "first": "Hannes",
                        "middle": [],
                        "last": "Schulz",
                        "suffix": ""
                    },
                    {
                        "first": "Jeremie",
                        "middle": [],
                        "last": "Zumer",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Shikhar Sharma, Layla El Asri, Hannes Schulz, and Jeremie Zumer. 2017. Relevance of unsupervised metrics in task-oriented dialogue for evaluating nat- ural language generation. CoRR, abs/1706.09799.",
                "links": null
            },
            "BIBREF48": {
                "ref_id": "b48",
                "title": "Creating general-purpose corpora using automated search engine queries",
                "authors": [
                    {
                        "first": "Serge",
                        "middle": [],
                        "last": "Sharoff",
                        "suffix": ""
                    }
                ],
                "year": 2006,
                "venue": "WaCky! Working papers on the Web as Corpus",
                "volume": "",
                "issue": "",
                "pages": "63--98",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Serge Sharoff. 2006. Creating general-purpose corpora using automated search engine queries. In Marco Baroni and Silvia Bernardini, editors, WaCky! Work- ing papers on the Web as Corpus, pages 63-98. GEDIT.",
                "links": null
            },
            "BIBREF49": {
                "ref_id": "b49",
                "title": "RUSE: Regressor Using Sentence Embeddings for Automatic Machine Translation Evaluation",
                "authors": [
                    {
                        "first": "Hiroki",
                        "middle": [],
                        "last": "Shimanaka",
                        "suffix": ""
                    },
                    {
                        "first": "Tomoyuki",
                        "middle": [],
                        "last": "Kajiwara",
                        "suffix": ""
                    },
                    {
                        "first": "Mamoru",
                        "middle": [],
                        "last": "Komachi",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "Proceedings of the Third Conference on Machine Translation: Shared Task Papers",
                "volume": "",
                "issue": "",
                "pages": "751--758",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/W18-6456"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Hiroki Shimanaka, Tomoyuki Kajiwara, and Mamoru Komachi. 2018. RUSE: Regressor Using Sen- tence Embeddings for Automatic Machine Transla- tion Evaluation. In Proceedings of the Third Confer- ence on Machine Translation: Shared Task Papers, pages 751-758, Belgium, Brussels. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF50": {
                "ref_id": "b50",
                "title": "Creating a corpus for Russian datato-text generation using neural machine translation and post-editing",
                "authors": [
                    {
                        "first": "Anastasia",
                        "middle": [],
                        "last": "Shimorina",
                        "suffix": ""
                    },
                    {
                        "first": "Elena",
                        "middle": [],
                        "last": "Khasanova",
                        "suffix": ""
                    },
                    {
                        "first": "Claire",
                        "middle": [],
                        "last": "Gardent",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "Proceedings of the 7th Workshop on Balto-Slavic Natural Language Processing",
                "volume": "",
                "issue": "",
                "pages": "44--49",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/W19-3706"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Anastasia Shimorina, Elena Khasanova, and Claire Gardent. 2019. Creating a corpus for Russian data- to-text generation using neural machine translation and post-editing. In Proceedings of the 7th Work- shop on Balto-Slavic Natural Language Processing, pages 44-49, Florence, Italy. Association for Com- putational Linguistics.",
                "links": null
            },
            "BIBREF51": {
                "ref_id": "b51",
                "title": "CIDEr: Consensus-based image description evaluation",
                "authors": [
                    {
                        "first": "R",
                        "middle": [],
                        "last": "Vedantam",
                        "suffix": ""
                    },
                    {
                        "first": "C",
                        "middle": [
                            "L"
                        ],
                        "last": "Zitnick",
                        "suffix": ""
                    },
                    {
                        "first": "D",
                        "middle": [],
                        "last": "Parikh",
                        "suffix": ""
                    }
                ],
                "year": 2015,
                "venue": "2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR)",
                "volume": "",
                "issue": "",
                "pages": "4566--4575",
                "other_ids": {
                    "DOI": [
                        "10.1109/CVPR.2015.7299087"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "R. Vedantam, C. L. Zitnick, and D. Parikh. 2015. CIDEr: Consensus-based image description evalua- tion. In 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4566-4575.",
                "links": null
            },
            "BIBREF52": {
                "ref_id": "b52",
                "title": "Revisiting challenges in datato-text generation with fact grounding",
                "authors": [
                    {
                        "first": "Hongmin",
                        "middle": [],
                        "last": "Wang",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "Proceedings of the 12th International Conference on Natural Language Generation",
                "volume": "",
                "issue": "",
                "pages": "311--322",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/W19-8639"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Hongmin Wang. 2019. Revisiting challenges in data- to-text generation with fact grounding. In Proceed- ings of the 12th International Conference on Nat- ural Language Generation, pages 311-322, Tokyo, Japan. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF53": {
                "ref_id": "b53",
                "title": "Challenges in data-to-document generation",
                "authors": [
                    {
                        "first": "Sam",
                        "middle": [],
                        "last": "Wiseman",
                        "suffix": ""
                    },
                    {
                        "first": "Stuart",
                        "middle": [],
                        "last": "Shieber",
                        "suffix": ""
                    },
                    {
                        "first": "Alexander",
                        "middle": [],
                        "last": "Rush",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing",
                "volume": "",
                "issue": "",
                "pages": "2253--2263",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/D17-1239"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Sam Wiseman, Stuart Shieber, and Alexander Rush. 2017. Challenges in data-to-document generation. In Proceedings of the 2017 Conference on Empiri- cal Methods in Natural Language Processing, pages 2253-2263, Copenhagen, Denmark. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF54": {
                "ref_id": "b54",
                "title": "BERTScore: Evaluating text generation with BERT",
                "authors": [
                    {
                        "first": "Tianyi",
                        "middle": [],
                        "last": "Zhang",
                        "suffix": ""
                    },
                    {
                        "first": "Varsha",
                        "middle": [],
                        "last": "Kishore",
                        "suffix": ""
                    },
                    {
                        "first": "Felix",
                        "middle": [],
                        "last": "Wu",
                        "suffix": ""
                    },
                    {
                        "first": "Q",
                        "middle": [],
                        "last": "Kilian",
                        "suffix": ""
                    },
                    {
                        "first": "Yoav",
                        "middle": [],
                        "last": "Weinberger",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Artzi",
                        "suffix": ""
                    }
                ],
                "year": 2020,
                "venue": "Proceedings of the Eighth International Conference on Learning Representations",
                "volume": "",
                "issue": "",
                "pages": "1--43",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q Weinberger, and Yoav Artzi. 2020. BERTScore: Evaluating text generation with BERT. In Pro- ceedings of the Eighth International Conference on Learning Representations, pages 1-43, Ethiopia, Addis Ababa. OpenReview.net.",
                "links": null
            },
            "BIBREF55": {
                "ref_id": "b55",
                "title": "MoverScore: Text generation evaluating with contextualized embeddings and earth mover distance",
                "authors": [
                    {
                        "first": "Wei",
                        "middle": [],
                        "last": "Zhao",
                        "suffix": ""
                    },
                    {
                        "first": "Maxime",
                        "middle": [],
                        "last": "Peyrard",
                        "suffix": ""
                    },
                    {
                        "first": "Fei",
                        "middle": [],
                        "last": "Liu",
                        "suffix": ""
                    },
                    {
                        "first": "Yang",
                        "middle": [],
                        "last": "Gao",
                        "suffix": ""
                    },
                    {
                        "first": "Christian",
                        "middle": [
                            "M"
                        ],
                        "last": "Meyer",
                        "suffix": ""
                    },
                    {
                        "first": "Steffen",
                        "middle": [],
                        "last": "Eger",
                        "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": "563--578",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/D19-1053"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Wei Zhao, Maxime Peyrard, Fei Liu, Yang Gao, Chris- tian M. Meyer, and Steffen Eger. 2019. MoverScore: Text generation evaluating with contextualized em- beddings and earth mover distance. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th Interna- tional Joint Conference on Natural Language Pro- cessing (EMNLP-IJCNLP), pages 563-578, Hong Kong, China. Association for Computational Lin- guistics.",
                "links": null
            }
        },
        "ref_entries": {
            "FIGREF0": {
                "text": "Example of a set of attribute-value pairs (top) and corresponding text (bottom).",
                "num": null,
                "type_str": "figure",
                "uris": null
            },
            "FIGREF1": {
                "text": "Mapping between tags and entities for the corresponding delexicalized template.",
                "num": null,
                "type_str": "figure",
                "uris": null
            },
            "FIGREF2": {
                "text": "p e =\" d e s c r i p t i o n \">g r a z e d i n t h e \u2192 t h i g h</ r e f e r e n c e> </ r e f e r e n c e s> <t e x t>A 22\u2212y e a r\u2212o l d was g r a z e d i n t h e t h i g h .</ t e x t> <t e m p l a t e>A ENTITY\u22121 was ENTITY\u22122 .</ t e m p l a t e> <l e x i c a l i z a t i o n>DT[ form = u n d e f i n e d ] A ENTITY\u22121 \u2192 VP [ a s p e c t = s i m p l e , t e n s e = p a s t , \u2192 v o i c e = a c t i v e , p e r s o n = n u l l , number= s i n g u l a r ] \u2192 be ENTITY\u22122 .</ l e x i c a l i z a t i o n> </ l e x> </ e n t r y>",
                "num": null,
                "type_str": "figure",
                "uris": null
            },
            "FIGREF3": {
                "text": "Example of an XML formatted data instance in the CACAPO dataset.",
                "num": null,
                "type_str": "figure",
                "uris": null
            },
            "TABREF2": {
                "html": null,
                "text": "Descriptive statistics for various size-related dimensions.",
                "content": "<table/>",
                "type_str": "table",
                "num": null
            },
            "TABREF4": {
                "html": null,
                "text": "Size and lexical diversity metrics.",
                "content": "<table/>",
                "type_str": "table",
                "num": null
            },
            "TABREF7": {
                "html": null,
                "text": "",
                "content": "<table/>",
                "type_str": "table",
                "num": null
            },
            "TABREF9": {
                "html": null,
                "text": "TGen results on the CACAPO dataset.",
                "content": "<table/>",
                "type_str": "table",
                "num": null
            }
        }
    }
}