Base_JSON/prefixE/json/evalnlgeval/2020.evalnlgeval-1.3.json

{
    "paper_id": "2020",
    "header": {
        "generated_with": "S2ORC 1.0.0",
        "date_generated": "2023-01-19T11:30:45.207337Z"
    },
    "title": "Evaluation rules! On the use of grammars and rule-based systems for NLG evaluation",
    "authors": [
        {
            "first": "Emiel",
            "middle": [],
            "last": "Van Miltenburg",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "Tilburg University",
                "location": {}
            },
            "email": "c.w.j.vanmiltenburg@tilburguniversity.edu"
        },
        {
            "first": "Chris",
            "middle": [],
            "last": "Van Der Lee",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "Tilburg University",
                "location": {}
            },
            "email": ""
        },
        {
            "first": "Thiago",
            "middle": [],
            "last": "Castro-Ferreira",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "Tilburg University",
                "location": {}
            },
            "email": ""
        },
        {
            "first": "Emiel",
            "middle": [],
            "last": "Krahmer",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "Tilburg University",
                "location": {}
            },
            "email": ""
        }
    ],
    "year": "",
    "venue": null,
    "identifiers": {},
    "abstract": "NLG researchers often use uncontrolled corpora to train and evaluate their systems, using textual similarity metrics, such as BLEU. This position paper argues in favour of two alternative evaluation strategies, using grammars or rule-based systems. These strategies are particularly useful to identify the strengths and weaknesses of different systems. We contrast our proposals with the (extended) WebNLG dataset, which is revealed to have a skewed distribution of predicates. We predict that this distribution affects the quality of the predictions for systems trained on this data. However, this hypothesis can only be thoroughly tested (without any confounds) once we are able to systematically manipulate the skewness of the data, using a rule-based approach.",
    "pdf_parse": {
        "paper_id": "2020",
        "_pdf_hash": "",
        "abstract": [
            {
                "text": "NLG researchers often use uncontrolled corpora to train and evaluate their systems, using textual similarity metrics, such as BLEU. This position paper argues in favour of two alternative evaluation strategies, using grammars or rule-based systems. These strategies are particularly useful to identify the strengths and weaknesses of different systems. We contrast our proposals with the (extended) WebNLG dataset, which is revealed to have a skewed distribution of predicates. We predict that this distribution affects the quality of the predictions for systems trained on this data. However, this hypothesis can only be thoroughly tested (without any confounds) once we are able to systematically manipulate the skewness of the data, using a rule-based approach.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Abstract",
                "sec_num": null
            }
        ],
        "body_text": [
            {
                "text": "Recent years have seen many Natural Language Generation (NLG) researchers move away from rule-based systems, and towards neural end-to-end systems. These systems are typically evaluated using textual similarity metrics like BLEU (Papineni et al., 2002) , METEOR (Banerjee and Lavie, 2005; Denkowski and Lavie, 2014) , or ROUGE (Lin, 2004) , on large corpora of crowd-sourced texts (e.g., the E2E dataset, Novikova et al. 2016 ; the WebNLG dataset, Gardent et al. 2017; or MS COCO, Lin et al. 2014) . This evaluation strategy tells us to what extent the generated texts are similar to the reference data, but it is often difficult to determine exactly what that resemblance buys us. By now it is well-known that BLEU correlates poorly with human ratings (Elliott and Keller, 2014; Kilickaya et al., 2017; Reiter, 2018; Sulem et al., 2018; Mathur et al., 2020) , but BLEU by itself also does not tell us anything about the strengths and weaknesses of a particular model, or model architecture. This paper argues that we need alternative (or at least additional) metrics to provide this kind of insight. We believe that rule-based approaches are well-suited for this task.",
                "cite_spans": [
                    {
                        "start": 229,
                        "end": 252,
                        "text": "(Papineni et al., 2002)",
                        "ref_id": "BIBREF41"
                    },
                    {
                        "start": 262,
                        "end": 288,
                        "text": "(Banerjee and Lavie, 2005;",
                        "ref_id": "BIBREF1"
                    },
                    {
                        "start": 289,
                        "end": 315,
                        "text": "Denkowski and Lavie, 2014)",
                        "ref_id": "BIBREF8"
                    },
                    {
                        "start": 327,
                        "end": 338,
                        "text": "(Lin, 2004)",
                        "ref_id": "BIBREF25"
                    },
                    {
                        "start": 405,
                        "end": 425,
                        "text": "Novikova et al. 2016",
                        "ref_id": "BIBREF38"
                    },
                    {
                        "start": 448,
                        "end": 468,
                        "text": "Gardent et al. 2017;",
                        "ref_id": "BIBREF18"
                    },
                    {
                        "start": 469,
                        "end": 497,
                        "text": "or MS COCO, Lin et al. 2014)",
                        "ref_id": null
                    },
                    {
                        "start": 753,
                        "end": 779,
                        "text": "(Elliott and Keller, 2014;",
                        "ref_id": "BIBREF11"
                    },
                    {
                        "start": 780,
                        "end": 803,
                        "text": "Kilickaya et al., 2017;",
                        "ref_id": "BIBREF23"
                    },
                    {
                        "start": 804,
                        "end": 817,
                        "text": "Reiter, 2018;",
                        "ref_id": "BIBREF48"
                    },
                    {
                        "start": 818,
                        "end": 837,
                        "text": "Sulem et al., 2018;",
                        "ref_id": "BIBREF54"
                    },
                    {
                        "start": 838,
                        "end": 858,
                        "text": "Mathur et al., 2020)",
                        "ref_id": "BIBREF28"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "1.1 Not just BLEU; also uncontrolled data BLEU is an easy target; it's a quick-and-dirty solution that ignores paraphrases and different-butvalid perspectives on the input data. But if we only look at the metrics, we miss the elephant in the room: the corpora we use to train NLG systems are the messy result of underspecified elicitation tasks, where annotators receive very little training as to what the outputs should look like (e.g., van Miltenburg 2016; van Miltenburg et al. 2017) . Ideally, we should want training data that conforms to a clear set of guidelines. Having clean data is a means to control the quality of the output of an NLG system. By using crowdsourcing, we have ceded that control to the crowd. 1 The problem with crowdsourcing, and particularly with elicitation tasks to create NLG corpora, is that quality control is difficult. And even if we can control the quality of the data, it is very hard to control the diversity of the generated texts. 2 This makes it harder to study the ability of NLG systems to generalise from the training data to unseen instances. We will argue (in Section 4) that we need a more systematic approach to produce NLG test benches. We believe 1 Although there are also benefits to having a more uncontrolled elicitation task. For example, having fewer constraints means that the resulting data will be more diverse.",
                "cite_spans": [
                    {
                        "start": 432,
                        "end": 459,
                        "text": "(e.g., van Miltenburg 2016;",
                        "ref_id": null
                    },
                    {
                        "start": 460,
                        "end": 487,
                        "text": "van Miltenburg et al. 2017)",
                        "ref_id": "BIBREF33"
                    },
                    {
                        "start": 721,
                        "end": 722,
                        "text": "1",
                        "ref_id": null
                    },
                    {
                        "start": 1199,
                        "end": 1200,
                        "text": "1",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "2 This is not just a problem in NLG. Freitag et al. (2020, and references therein) describe how human translators tend to produce translationese: translations that overly rely on the source text, resulting in less natural-sounding texts. This reduces the diversity of the evaluation data for Machine Translation (MT), which has strong effects on the evaluation metrics used in MT (Freitag et al., 2019) . The authors go on to show that we can improve the correlation between modern evaluation metrics and human ratings, by improving the reference data (in this case: asking linguists to generate more fluent and diverse translations). But of course, this kind of exercise is expensive and time-consuming. that a rule-based approach (combined with new or existing NLG data) would again be ideal.",
                "cite_spans": [
                    {
                        "start": 380,
                        "end": 402,
                        "text": "(Freitag et al., 2019)",
                        "ref_id": "BIBREF16"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "opportunities for rule-based approaches",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "The downside of end-to-end systems;",
                "sec_num": "1.2"
            },
            {
                "text": "There are many good reasons to develop end-to-end systems. For example, Du\u0161ek et al. (2020) found that, in the E2E-challenge (Novikova et al., 2017) , sequence-to-sequence models \"scored higher than other architectures on word-overlap-based metrics and human-rated naturalness.\" 3 However, given the above, we can also see the move away from rule-based systems as a means to evade responsibility for whatever output our NLG systems produce. And what better way to control and evaluate the output. . . than to use more rules? This paper presents some ways in which rules and rule-based systems can be used to improve current-day NLG research. 4",
                "cite_spans": [
                    {
                        "start": 72,
                        "end": 91,
                        "text": "Du\u0161ek et al. (2020)",
                        "ref_id": "BIBREF10"
                    },
                    {
                        "start": 125,
                        "end": 148,
                        "text": "(Novikova et al., 2017)",
                        "ref_id": "BIBREF37"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "The downside of end-to-end systems;",
                "sec_num": "1.2"
            },
            {
                "text": "Ideally, evaluation of NLG systems should be tied to the cognitive capacities those systems are claimed to possess (Schlangen, 2019 (Schlangen, , 2020 . 5 For example, one could evaluate whether a system is able to produce a grammatically correct sentence. Abilities like these can be formalised as a set of rules (cf. Chomsky's generative program; Chomsky 1965), and we could simply check whether the output of an NLG system conforms to a predefined grammar. Xie et al. (2019) do this using Flickinger's (2000; English Resource Grammar, which offers broad coverage of the English language. The DELPH-IN catalogue offers 3 Another advantage of neural end-to-end systems that is sometimes mentioned is development speed: if you have a training corpus, you can train an end-to-end system fairly quickly. But Reiter (2020) shows that this advantage is probably overstated (if not false). Elsewhere he remarks that 'effectively impossible to fix undesirable behaviour in a \"deep learning\" system' (Reiter, 2016) , meaning such a system would have to be re-trained if any changes need to be made to its output. This makes maintenance very time-consuming. 4 Code for this paper is available at: https://github. com/evanmiltenburg/EvaluationRules.",
                "cite_spans": [
                    {
                        "start": 115,
                        "end": 131,
                        "text": "(Schlangen, 2019",
                        "ref_id": "BIBREF51"
                    },
                    {
                        "start": 132,
                        "end": 150,
                        "text": "(Schlangen, , 2020",
                        "ref_id": "BIBREF52"
                    },
                    {
                        "start": 153,
                        "end": 154,
                        "text": "5",
                        "ref_id": null
                    },
                    {
                        "start": 492,
                        "end": 511,
                        "text": "Flickinger's (2000;",
                        "ref_id": "BIBREF13"
                    },
                    {
                        "start": 621,
                        "end": 622,
                        "text": "3",
                        "ref_id": null
                    },
                    {
                        "start": 806,
                        "end": 819,
                        "text": "Reiter (2020)",
                        "ref_id": "BIBREF49"
                    },
                    {
                        "start": 993,
                        "end": 1007,
                        "text": "(Reiter, 2016)",
                        "ref_id": "BIBREF46"
                    },
                    {
                        "start": 1150,
                        "end": 1151,
                        "text": "4",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Evaluation and cognitive capacities",
                "sec_num": "2"
            },
            {
                "text": "5 This may not always be the case. Or at least: not directly. For example, consider the question whether a system is userfriendly or pleasant to use. Some high-level properties are fairly subjective, and may best be evaluated using human ratings. Still one could argue that these properties may be decomposed into a set of different abilities. For example: using the correct register, being able to translate jargon into layman's terms, generating unambiguous descriptions. an overview of HPSG grammars (Pollard and Sag, 1994) that are available for other languages. 6 In related work, Bangalore et al. (2000) use automated parsers for evaluation, but they compare the parse trees for the system outputs with those of the reference data, and compute an accuracy metric.",
                "cite_spans": [
                    {
                        "start": 503,
                        "end": 526,
                        "text": "(Pollard and Sag, 1994)",
                        "ref_id": "BIBREF44"
                    },
                    {
                        "start": 567,
                        "end": 568,
                        "text": "6",
                        "ref_id": null
                    },
                    {
                        "start": 586,
                        "end": 609,
                        "text": "Bangalore et al. (2000)",
                        "ref_id": "BIBREF2"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Evaluation and cognitive capacities",
                "sec_num": "2"
            },
            {
                "text": "At this point, it is fair to say that not all languages are as well-resourced as English. Of course we can evaluate grammaticality if there is a relatively complete description of a language. But not everyone has that luxury. Moreover, as NLG researchers, we aren't just interested in grammaticality. Why should we care about grammars, then?",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Evaluation and cognitive capacities",
                "sec_num": "2"
            },
            {
                "text": "There are two ways to respond to this criticism. First, if you accept that NLG evaluation is a good use case for a broad-coverage parser, then that provides additional motivation to build a new or better parser (or to start talking to linguists in your area). Second, a grammar does not necessarily have to cover the entire language for it to be useful. It just needs to cover your domain of interest. One example grammar is provided by Van Miltenburg et al. (2018), who developed a context-free grammar to cover person-descriptions in the MS COCO dataset (a collection of images paired with image descriptions). The grammar has a set of production rules that describe as well as categorise the components of person-descriptions in the corpus. So the phrase 'the tall happy toddler' can be parsed as in Figure 1 . Van Miltenburg (2020) improved this grammar, and used it to evaluate the extent to which image captioning systems are able to generate different kinds of person-descriptions. This is another example of a cognitive ability that can be examined using a pre-determined set of rules. As an additional bonus, a complete characterisation of a domain such as PERSON-DESCRIPTIONS allows us to reflect on what kinds of outputs are desirable or not, for the system to generate.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 803,
                        "end": 811,
                        "text": "Figure 1",
                        "ref_id": "FIGREF0"
                    }
                ],
                "eq_spans": [],
                "section": "Evaluation and cognitive capacities",
                "sec_num": "2"
            },
            {
                "text": "3 Systems, models, and architectures Before we continue, it is important to recognise the difference between systems, models and architectures. We consider architectures to be abstract descriptions of all the components that make up a system. A system is a specific instance that implements an architecture. When a system is trained on a particular dataset, we can say that it has constructed a model for how the task should be carried out. These distinctions are important, because different NLG researchers may be interested in either systems, models, or architectures. Theoretically oriented researchers may be more interested in the properties of different architectures, whereas more applied researchers may be interested in the properties of different systems or models. In our experience, shared tasks are often misconstrued as a competition to see who can deliver the best model. This misses the point, because ideally the results of a shared task teach us about the strengths and weaknesses of different architectures. 7",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Evaluation and cognitive capacities",
                "sec_num": "2"
            },
            {
                "text": "Machine learning datasets are used to determine whether systems are able to generalise from experience to unseen situations (Mitchell, 1997) . To test this, researchers typically use separate training, development, and test sets. Different models are trained using the training set, the best model is selected using the development set, and then we evaluate its performance on the test set.",
                "cite_spans": [
                    {
                        "start": 124,
                        "end": 140,
                        "text": "(Mitchell, 1997)",
                        "ref_id": "BIBREF35"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Evaluating the ability to generalise",
                "sec_num": "4"
            },
            {
                "text": "Using different splits is necessary, but not sufficient for NLG tasks. We can see this when we look at the generation of weather forecasts, a popular topic in the NLG community (e.g. Gkatzia et al. 2016 and references therein). It is not good enough to only have a corpus where all inputs have the same weather but different place names. NLG models trained on such a corpus would only learn to produce a fixed weather template, where they should copy in the name from the input. An evaluation is only meaningful if there are clear differences in all (combinations of) variables, between training and test set. At the same time, the training data should also not contain so much variation that it's impossible to detect any pattern. It is an open question how much systematicity (and redundancy) there should be in the training data for NLG systems to learn how to perform any language generation task. Finally, it is important to have specific information about the output. For example: how many different ways are there to verbalise the same predicates, entities, numbers, dates and times? Without this information, it is impossible to say anything about the complexity of the task.",
                "cite_spans": [
                    {
                        "start": 183,
                        "end": 202,
                        "text": "Gkatzia et al. 2016",
                        "ref_id": "BIBREF20"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Requirements to measure generalisation",
                "sec_num": "4.1"
            },
            {
                "text": "We don't believe current datasets are sufficient to measure the extent to which systems are able to generalise from the training data, although some datasets do come close. WebNLG, for example, is a state-of-the-art dataset. It offers an excellent overview table (Table 1 in Gardent et al. 2017) describing properties of the input (e.g. number of different predicates, number of combinations of RDF triples, relations between the different triples) and output (e.g. number of sentences verbalising different amounts of triples).",
                "cite_spans": [
                    {
                        "start": 275,
                        "end": 295,
                        "text": "Gardent et al. 2017)",
                        "ref_id": "BIBREF18"
                    }
                ],
                "ref_spans": [
                    {
                        "start": 263,
                        "end": 271,
                        "text": "(Table 1",
                        "ref_id": "TABREF3"
                    }
                ],
                "eq_spans": [],
                "section": "Are current datasets sufficient?",
                "sec_num": "4.2"
            },
            {
                "text": "Still missing from the description of the WebNLG corpus is the distribution of different predicates. Figure 2 shows the frequency distribution of different labels in the training set (computed using the XML files from the enriched WebNLG dataset; Castro Ferreira et al. 2018). The plot reveals that the data is heavily skewed, with 76 predicates (out of 246) occurring fewer than 10 times, while the most frequent predicate ('country') occurs 2150 times. End-to-end systems will probably perform worse on the tail of the distribution (where example outputs are scarce) than on the head (where examples are plentiful).",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 101,
                        "end": 109,
                        "text": "Figure 2",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Are current datasets sufficient?",
                "sec_num": "4.2"
            },
            {
                "text": "On the output side, it is not clear from the original WebNLG corpus how many different possible lexicalisations there are for each predicate. 8 This is difficult to study with unstructured text output, but luckily the enriched WebNLG dataset converted the outputs into templates (see Table 2 below), which we can count. Table 1 shows a selection of predicates with different ratios of unique-to-total number of templates. One can imagine that it's much easier for a model to predict the template for a predicate with a ratio of 0.12, than for a predicate with a ratio of 1.00. After all: a lower ratio means that there are more examples for each unique template. The easiest situation would be one where there is a Figure 2 : Frequency of all predicates in the training split of the WebNLG corpus. The x-axis shows the frequency of the labels (range: 1-2150, where 'country' is the most frequent label). To improve readability, the y-axis only shows a selection of different labels at fixed intervals. ",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 284,
                        "end": 291,
                        "text": "Table 2",
                        "ref_id": "TABREF5"
                    },
                    {
                        "start": 320,
                        "end": 327,
                        "text": "Table 1",
                        "ref_id": "TABREF3"
                    },
                    {
                        "start": 715,
                        "end": 723,
                        "text": "Figure 2",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Are current datasets sufficient?",
                "sec_num": "4.2"
            },
            {
                "text": "Our discussion so far points us in the direction of more carefully planned corpora, with clear input distributions. Ideally the output language should also be controlled, so that it conforms to a set of guidelines for what appropriate output should look like. To really test the degree to which system performance depends on these variables (i.e. the distributions of input and output), having just one big training corpus isn't good enough. Rather, there should be different versions of the same corpus, so that we can manipulate different aspects of the training data, to see how each of those variables affects the outcome (performance on the test set).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "What do we need?",
                "sec_num": "4.3"
            },
            {
                "text": "5 How can we get there? Rules!",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "What do we need?",
                "sec_num": "4.3"
            },
            {
                "text": "Corpora constructed solely through human labor are not good enough, because they do not give us enough control over the data to carry out systematic experiments. For example: we only know the amount of different lexicalisations for a predicate after data collection has finished. We present two alternative approaches to generate evaluation data.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "What do we need?",
                "sec_num": "4.3"
            },
            {
                "text": "One way to construct a controlled corpus, is to use existing NLG systems to produce a large collection of texts in a particular domain. We can then train end-to-end systems on this data to produce similar texts. This has three major advantages, compared to the use of crowd-sourced data: 1. It is efficient and cost-effective to produce large corpora, since no human annotators are needed.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "From systems to synthetic datasets",
                "sec_num": "5.1"
            },
            {
                "text": "2. We can easily create different sub-corpora with very specific distributions of the input data, which would allow us to estimate the extent to which systems are able to generalise from low-frequent training examples.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "From systems to synthetic datasets",
                "sec_num": "5.1"
            },
            {
                "text": "3. It allows us to automatically evaluate the quality of the output in ways that are not possible (or very labor-intensive) with humangenerated data. The generate-and-train approach has recently been applied by Oraby et al. (2018, using the PER-SONAGE system; Mairesse and Walker 2010) to create a synthetic corpus of utterances in the RESTAU-RANT domain, where the authors controlled the personality of the utterances. Oraby et al. showed that it is possible for neural NLG models to distinguish style and content, and that models trained on their data were able to generate meaningful output with the desired personality traits.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "From systems to synthetic datasets",
                "sec_num": "5.1"
            },
            {
                "text": "The idea of training NLG-systems based on the output of other NLG systems is controversial. Ehud Reiter argues on his blog that this is just reverseengineering existing systems (Reiter, 2017) . This is a valid concern if the goal is to build an NLG system to be used in some application context. However, we are not concerned with any applications. Rather, we are interested in the core properties of different end-to-end architectures, and particularly the way those properties relate to learnability: to what extent a particular architecture can learn to generate natural language, based on a corpus with particular, controlled properties? Feasibility A natural question at this point is how to find existing systems to generate synthetic NLG corpora. One answer is simply to look for systems using SimpleNLG, since this is probably the most used realisation engine for NLG in academia. It may be possible to build a corpus generation tool that incorporates all different systems. To find these systems and assess the feasibility of our proposal, we used the Publish or Perish software 9 to retrieve all pub-9 Search carried out on the 17th of August, 2020, using the macOS GUI edition, version 7. lications on Google Scholar that cite the original SimpleNLG paper (Gatt and Reiter, 2009) . 10 We found 361 publications referring to SimpleNLG on Google Scholar, coming from a wide array of different venues. We are still in the process of analysing the results, but our impression is that only a small proportion of the reported systems is useful. Many are either unavailable, form part of a larger pipeline, or use proprietary/personal data (e.g. BT-Nurse; Hunter et al. 2012).",
                "cite_spans": [
                    {
                        "start": 177,
                        "end": 191,
                        "text": "(Reiter, 2017)",
                        "ref_id": "BIBREF47"
                    },
                    {
                        "start": 1267,
                        "end": 1290,
                        "text": "(Gatt and Reiter, 2009)",
                        "ref_id": "BIBREF19"
                    },
                    {
                        "start": 1293,
                        "end": 1295,
                        "text": "10",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "From systems to synthetic datasets",
                "sec_num": "5.1"
            },
            {
                "text": "Another approach is to construct our own templatedriven corpus generator, based on existing datasets. Additionally, the dataset also shows how different entities can be realised. This gives us all the ingredients to develop a rule-based system that can generate a corpus matching specific criteria (or indeed a collection of corpora that allow us to determine the ability to which end-to-end systems are able to generalise). With the templates and entity realisation options in hand, we can choose to make full use of all possible templates and realisations for all predicates and entities, or we can select only specific templates/realisations to have a particular distribution of the data. Here are the aspects that we imagine may be interesting to manipulate:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "From datasets to rules, and back again",
                "sec_num": "5.2"
            },
            {
                "text": "\u2022 The number of different templates/entities in the train, validation, and test sets. (Note that templates and entities may be manipulated separately from each other.)",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "From datasets to rules, and back again",
                "sec_num": "5.2"
            },
            {
                "text": "available from: https://harzing.com/resources/ publish-or-perish 10 This approach excludes many systems using SimpleNLG in a different language, e.g. Brazilian Portuguese (de Oliveira and Sripada, 2014), Dutch (de Jong, 2018), German (Bollmann, 2011; Braun et al., 2019) , French (Vaudry and Lapalme, 2013) , Galician (Cascallar-Fuentes et al., 2018) , Italian (Mazzei et al., 2016 ), or Spanish (Ramos-Soto et al., 2017 .",
                "cite_spans": [
                    {
                        "start": 234,
                        "end": 250,
                        "text": "(Bollmann, 2011;",
                        "ref_id": "BIBREF3"
                    },
                    {
                        "start": 251,
                        "end": 270,
                        "text": "Braun et al., 2019)",
                        "ref_id": "BIBREF4"
                    },
                    {
                        "start": 280,
                        "end": 306,
                        "text": "(Vaudry and Lapalme, 2013)",
                        "ref_id": "BIBREF55"
                    },
                    {
                        "start": 318,
                        "end": 350,
                        "text": "(Cascallar-Fuentes et al., 2018)",
                        "ref_id": "BIBREF5"
                    },
                    {
                        "start": 361,
                        "end": 381,
                        "text": "(Mazzei et al., 2016",
                        "ref_id": "BIBREF29"
                    },
                    {
                        "start": 382,
                        "end": 420,
                        "text": "), or Spanish (Ramos-Soto et al., 2017",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "From datasets to rules, and back again",
                "sec_num": "5.2"
            },
            {
                "text": "\u2022 The frequency with which those different templates/entities each occur. Is there a uniform distribution, or do some templates/entities occur more than others?",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "From datasets to rules, and back again",
                "sec_num": "5.2"
            },
            {
                "text": "\u2022 The overlap in terms of templates/entities between the train, validation, and test sets. Here we may also choose to generate multiple different test sets to accompany the same training set, to make evaluation more efficient.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "From datasets to rules, and back again",
                "sec_num": "5.2"
            },
            {
                "text": "\u2022 The ordering principles, and the number of different orders in which triples are realised in the output texts. (E.g. maintaining the input order, ordering triples alphabetically or based on their content.)",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "From datasets to rules, and back again",
                "sec_num": "5.2"
            },
            {
                "text": "\u2022 The segmentation principles, and the number of sentences that are used to realise a set of triples. (E.g. three triples per sentence; only one triple per sentence; segmentation depending on the predicate, the entities, or both.)",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "From datasets to rules, and back again",
                "sec_num": "5.2"
            },
            {
                "text": "\u2022 The amount of noise in the dataset. By default, there is no noise in the synthetic dataset, but we could add synthetic noise (i.e. knowingly introduce errors), to see how systems deal with the presence of noise in the data. This is similar to Du\u0161ek et al. (2019) , who systematically removed noise from the E2E dataset, to gauge the impact of erroneous meaning representations.",
                "cite_spans": [
                    {
                        "start": 245,
                        "end": 264,
                        "text": "Du\u0161ek et al. (2019)",
                        "ref_id": "BIBREF9"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "From datasets to rules, and back again",
                "sec_num": "5.2"
            },
            {
                "text": "Using data generated in this manner, we could answer questions like the following:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "From datasets to rules, and back again",
                "sec_num": "5.2"
            },
            {
                "text": "\u2022 How do skewness and diversity (of templates, referring expressions) influence the quality and diversity of the generated outputs?",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "From datasets to rules, and back again",
                "sec_num": "5.2"
            },
            {
                "text": "\u2022 How many minority examples are necessary before end-to-end models consider these a valid alternative to majority examples?",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "From datasets to rules, and back again",
                "sec_num": "5.2"
            },
            {
                "text": "\u2022 What kinds of generation rules are learnable by end-to-end systems? Which architectures are more apt to pick up on different kinds of systematic patterns in the data?",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "From datasets to rules, and back again",
                "sec_num": "5.2"
            },
            {
                "text": "There are two main challenges for this approach. The first challenge concerns multiple predicates.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Feasibility",
                "sec_num": null
            },
            {
                "text": "It is easy to see how textual output for singlepredicate inputs can automatically be generated (just fill in the empty slots in the template), but for inputs with multiple predicates the problem is more complex. The realisation of multiple predicates is not necessarily equal to the realisation of two single predicates, plus some text to link the two (e.g. the conjunction and). Indeed, Perez-Beltrachini et al. (2016) purposefully selected combinations of predicates that might lead to more concise solutions. E.g. combining Alan Bean, occupation, test pilot with Alan Bean, nationality, USA leads to the insertion of an adjective: Alan Bean was an American test pilot.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Feasibility",
                "sec_num": null
            },
            {
                "text": "Normally it would mean a large amount of human labor to find any systematicity in the corpus. To build a good NLG system, we need to know how to order the predicates; how to relate the predicates to each other; how to split up the information in different sentences; and how to realise sentences combining multiple predicates. However, for evaluation purposes, the exact answers to these questions aren't necessarily important. 11 What matters is that there is some output that conforms to a particular set of rules. The evaluation is just there to see if end-to-end systems are able to learn those rules. The exception here is when the ability to learn a specific kind of rule is in question. For example: can neural NLG systems learn to insert adjectives like American in the example above?",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Feasibility",
                "sec_num": null
            },
            {
                "text": "The second challenge concerns the distribution of the original corpus. As Table 1 shows, some predicates occur only three times. This limits the different kinds of corpora that we are able to produce. For example, it is not possible with just the WebNLG data alone to generate a corpus where there are more than three different lexicalisations for the FULLNAME predicate. Moreover, it is not even possible to generate more than nine different predicate-entity combinations (3 entities times 3 predicate-realisations). One way to address the distribution issue is to (semi-)automatically generate more examples by extracting triples from DBpedia (Auer et al., 2007; Lehmann et al., 2015) , and verbalising them using the predicate's lexicalization templates available in the enriched WebNLG dataset and a grammar-based NLG system (e.g., Mille et al. 2019) . As mentioned earlier, the data does not need to be perfect; it just needs to be consistent, so that learners are (in theory) able to infer rules from the data.",
                "cite_spans": [
                    {
                        "start": 645,
                        "end": 664,
                        "text": "(Auer et al., 2007;",
                        "ref_id": "BIBREF0"
                    },
                    {
                        "start": 665,
                        "end": 686,
                        "text": "Lehmann et al., 2015)",
                        "ref_id": "BIBREF24"
                    },
                    {
                        "start": 836,
                        "end": 854,
                        "text": "Mille et al. 2019)",
                        "ref_id": "BIBREF30"
                    }
                ],
                "ref_spans": [
                    {
                        "start": 74,
                        "end": 81,
                        "text": "Table 1",
                        "ref_id": "TABREF3"
                    }
                ],
                "eq_spans": [],
                "section": "Feasibility",
                "sec_num": null
            },
            {
                "text": "Ideally we would be able to control the data in such a way, that changes to individual variables happen ceteris paribus; i.e. with all other variables staying the same. But there are practical considerations we need to take into account:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "General feasibility",
                "sec_num": "5.3"
            },
            {
                "text": "\u2022 The number of times you can train a model, is limited by the size and complexity of that model. If it takes a long time to train the model, then it is not feasible to do this for tens or hundreds of different versions of the same training data. 12",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "General feasibility",
                "sec_num": "5.3"
            },
            {
                "text": "\u2022 This issue is further compounded by the fact that many models are randomly initialised. For a good estimation of system performance, the system needs to be trained multiple times on the exact same data. There is no universal solution to this problem, but it does help to have specific hypotheses about which factors might affect system performance, and to focus on those.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "General feasibility",
                "sec_num": "5.3"
            },
            {
                "text": "Next to increased control over the training data, the approaches proposed in this section have an additional benefit: because all training data has been generated using a rule-based approach, we can use those same rules to evaluate which rules were learned by the system, and which ones weren't. This is also the approach we described in Section 2. We could even split up the evaluation, to measure which templates, entity realisations, ordering rules, and segmentation rules the system acquired.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Assessing model performance",
                "sec_num": "5.4"
            },
            {
                "text": "One aspect we did not address yet is how to parse imperfect outputs. There are no guarantees that the output of end-to-end systems will conform to any of the rules through which the corpus was generated. Using a strict approach, we could say that faulty output just doesn't count; if it is flawed, the system simply did not fully learn the relevant rules. But perhaps we would also like to give partial credit to systems that almost learned how to perform the generation task. We leave this as a question for future research. 13",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Assessing model performance",
                "sec_num": "5.4"
            },
            {
                "text": "Since we intend to explore this approach in the future, and to encourage others to explore this space as well, we make a number of predictions:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Predictions",
                "sec_num": "5.5"
            },
            {
                "text": "1. Templates with a lower unique-to-total ratio (see Table 1 ) are easier to learn.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 53,
                        "end": 60,
                        "text": "Table 1",
                        "ref_id": "TABREF3"
                    }
                ],
                "eq_spans": [],
                "section": "Predictions",
                "sec_num": "5.5"
            },
            {
                "text": "2. The number of examples needed to successfully learn a template, depends on the amount of alternative templates that could also verbalise the same predicate, the amount of predicates in the corpus, and the size of the corpus.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Predictions",
                "sec_num": "5.5"
            },
            {
                "text": "3. It is easier to learn how to realise a predicate, if the arguments for that predicate are diverse. (If a predicate always occurs with the same arguments, they may be considered part of the template by the model.) 4. When combining multiple triples, conjunction (Susan is an astronaut and she is American) is easier to learn than insertion (Susan is an American astronaut). This list is not exhaustive; certainly many more predictions could be made about different combinations of the parameters we described above. But these hypotheses should serve as a starting point for future research. Initially we may want to see empirically whether the predictions hold up for popular architectures. A different avenue of research could be based on this evidence, to develop formal proofs about the properties of families of NLG architectures. We believe both are needed to inform NLG research and practice.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Predictions",
                "sec_num": "5.5"
            },
            {
                "text": "The output of rule-based systems is often said to be less fluent or natural than the output of end-to-end systems, and this claim is corroborated by the results of the E2E-challenge (Du\u0161ek et al., 2020) . It may thus be expected that any synthetically generated corpus will be less natural than humanproduced data, and the texts will probably have other shortcomings, too. However, the proposal in this paper is focused on determining what systems can or cannot learn from corpora with different properties. This means that, to some extent, the naturalness or fluency of the synthetic data does not really matter. What matters is that we learn how those different properties of the data affect the output of data-driven systems. We can then use those systems in other areas, knowing what they are capable of and what their limitations are. At that point, we need a different kind of evaluation (although rules are still valuable to check whether system output conforms to particular guidelines).",
                "cite_spans": [
                    {
                        "start": 182,
                        "end": 202,
                        "text": "(Du\u0161ek et al., 2020)",
                        "ref_id": "BIBREF10"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Limitations",
                "sec_num": "6"
            },
            {
                "text": "One might reasonably object here that the quality of the corpus does matter. How can you be sure that your synthetic data has the desired properties? Wouldn't that require some form of evaluation as well? We believe this concern could be addressed through unit-tests in the corpus generation code base. Because our proposal involves rule-based generation, the output should always be predictable.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Limitations",
                "sec_num": "6"
            },
            {
                "text": "We discussed the merits of (grammar) rules and rule-based systems in the context of NLG evaluation. Our conclusion is that there are clear benefits for practitioners who want to learn more about the architectures that they use for real-life applications. A concern that some may have, is that the real world is messy. Why should we solve toy problems like reverse-engineering rule-based systems? Our answer is two-fold. First, since our proposals involve synthetic data, we can make the data as clean or messy as we want. But because we have full control over the data, the evaluation will be much more informative about what systems can or cannot do. Second, a rule-based perspective is useful because it forces us to engage with the data. Looking at the WebNLG data, and all of the different templates that exist for each of the different predicates, one cannot help but ask: is this diversity really useful? Or should we try to reduce the diversity (e.g. formulating guidelines), to ensure the best possible outputs for our NLG systems? Messiness can be good or bad, and it is up to us to explore the impact of variation in NLG data.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Conclusion",
                "sec_num": "7"
            },
            {
                "text": "Huiyuan Xie, Tom Sherborne, Alexander Kuhnle, and Ann Copestake. 2019. Going beneath the surface: Evaluating image captioning for grammaticality, truthfulness and diversity. In Proceedings of Rigorous Evaluation of AI Systems 2019, collocated with The seventh AAAI Conference on Human Computation and Crowdsourcing.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Conclusion",
                "sec_num": "7"
            },
            {
                "text": "http://moin.delph-in.net/ GrammarCatalogue",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "For further discussion of shared tasks and leaderboards in NLP, see: Parra Escart\u00edn et al. 2017; Nissim et al. 2017; Rogers 2019; Ethayarajh and Jurafsky 2020.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "We limit ourselves to predicates here, but note that predicates are not the only part of the input that needs to be lexicalised.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "One might have multiple rules corresponding to different answers to each question. It would then be possible to experiment with different amounts of examples generated using the different rules.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "So next to environmental issues caused by computationally heavy approaches to NLP(Strubell et al., 2019), we can also say that such approaches are an obstacle to properly evaluate new systems.13 But note that the texts (and probably system outputs as well) are very predictable. This makes it interesting to explore whether metrics based on edit distance could work here, even though they have been shown to be inadequate 'in the wild.'",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            }
        ],
        "back_matter": [
            {
                "text": "We thank the anonymous reviewers for their feedback, which helped us refine the arguments laid out in this paper.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Acknowledgments",
                "sec_num": null
            }
        ],
        "bib_entries": {
            "BIBREF0": {
                "ref_id": "b0",
                "title": "Dbpedia: A nucleus for a web of open data",
                "authors": [
                    {
                        "first": "S\u00f6ren",
                        "middle": [],
                        "last": "Auer",
                        "suffix": ""
                    },
                    {
                        "first": "Christian",
                        "middle": [],
                        "last": "Bizer",
                        "suffix": ""
                    },
                    {
                        "first": "Georgi",
                        "middle": [],
                        "last": "Kobilarov",
                        "suffix": ""
                    },
                    {
                        "first": "Jens",
                        "middle": [],
                        "last": "Lehmann",
                        "suffix": ""
                    },
                    {
                        "first": "Richard",
                        "middle": [],
                        "last": "Cyganiak",
                        "suffix": ""
                    },
                    {
                        "first": "Zachary",
                        "middle": [],
                        "last": "Ives",
                        "suffix": ""
                    }
                ],
                "year": 2007,
                "venue": "The semantic web",
                "volume": "",
                "issue": "",
                "pages": "722--735",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "S\u00f6ren Auer, Christian Bizer, Georgi Kobilarov, Jens Lehmann, Richard Cyganiak, and Zachary Ives. 2007. Dbpedia: A nucleus for a web of open data. In The semantic web, pages 722-735. Springer.",
                "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. Association for Computational Lin- guistics.",
                "links": null
            },
            "BIBREF2": {
                "ref_id": "b2",
                "title": "Evaluation metrics for generation",
                "authors": [
                    {
                        "first": "Srinivas",
                        "middle": [],
                        "last": "Bangalore",
                        "suffix": ""
                    },
                    {
                        "first": "Owen",
                        "middle": [],
                        "last": "Rambow",
                        "suffix": ""
                    },
                    {
                        "first": "Steve",
                        "middle": [],
                        "last": "Whittaker",
                        "suffix": ""
                    }
                ],
                "year": 2000,
                "venue": "INLG'2000 Proceedings of the First International Conference on Natural Language Generation",
                "volume": "",
                "issue": "",
                "pages": "1--8",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Srinivas Bangalore, Owen Rambow, and Steve Whit- taker. 2000. Evaluation metrics for generation. In INLG'2000 Proceedings of the First International Conference on Natural Language Generation, pages 1-8, Mitzpe Ramon, Israel. Association for Compu- tational Linguistics.",
                "links": null
            },
            "BIBREF3": {
                "ref_id": "b3",
                "title": "Adapting SimpleNLG to German",
                "authors": [
                    {
                        "first": "Marcel",
                        "middle": [],
                        "last": "Bollmann",
                        "suffix": ""
                    }
                ],
                "year": 2011,
                "venue": "Proceedings of the 13th European Workshop on Natural Language Generation",
                "volume": "",
                "issue": "",
                "pages": "133--138",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Marcel Bollmann. 2011. Adapting SimpleNLG to Ger- man. In Proceedings of the 13th European Work- shop on Natural Language Generation, pages 133- 138, Nancy, France. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF4": {
                "ref_id": "b4",
                "title": "SimpleNLG-DE: Adapting Sim-pleNLG 4 to German",
                "authors": [
                    {
                        "first": "Daniel",
                        "middle": [],
                        "last": "Braun",
                        "suffix": ""
                    },
                    {
                        "first": "Kira",
                        "middle": [],
                        "last": "Klimt",
                        "suffix": ""
                    },
                    {
                        "first": "Daniela",
                        "middle": [],
                        "last": "Schneider",
                        "suffix": ""
                    },
                    {
                        "first": "Florian",
                        "middle": [],
                        "last": "Matthes",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "Proceedings of the 12th International Conference on Natural Language Generation",
                "volume": "",
                "issue": "",
                "pages": "415--420",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Daniel Braun, Kira Klimt, Daniela Schneider, and Flo- rian Matthes. 2019. SimpleNLG-DE: Adapting Sim- pleNLG 4 to German. In Proceedings of the 12th International Conference on Natural Language Gen- eration, pages 415-420, Tokyo, Japan. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF5": {
                "ref_id": "b5",
                "title": "Adapting SimpleNLG to Galician language",
                "authors": [
                    {
                        "first": "Andrea",
                        "middle": [],
                        "last": "Cascallar-Fuentes",
                        "suffix": ""
                    },
                    {
                        "first": "Alejandro",
                        "middle": [],
                        "last": "Ramos-Soto",
                        "suffix": ""
                    },
                    {
                        "first": "Alberto",
                        "middle": [
                            "Bugar\u00edn"
                        ],
                        "last": "Diz",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "Proceedings of the 11th International Conference on Natural Language Generation",
                "volume": "",
                "issue": "",
                "pages": "67--72",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Andrea Cascallar-Fuentes, Alejandro Ramos-Soto, and Alberto Bugar\u00edn Diz. 2018. Adapting SimpleNLG to Galician language. In Proceedings of the 11th In- ternational Conference on Natural Language Gener- ation, pages 67-72, Tilburg University, The Nether- lands. Association 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": {},
                "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": "Aspects of the Theory of Syntax",
                "authors": [
                    {
                        "first": "Noam",
                        "middle": [],
                        "last": "Chomsky",
                        "suffix": ""
                    }
                ],
                "year": 1965,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Noam Chomsky. 1965. Aspects of the Theory of Syntax. MIT Press.",
                "links": null
            },
            "BIBREF8": {
                "ref_id": "b8",
                "title": "Meteor universal: Language specific translation evaluation for any target language",
                "authors": [
                    {
                        "first": "Michael",
                        "middle": [],
                        "last": "Denkowski",
                        "suffix": ""
                    },
                    {
                        "first": "Alon",
                        "middle": [],
                        "last": "Lavie",
                        "suffix": ""
                    }
                ],
                "year": 2014,
                "venue": "Proceedings of the Ninth Workshop on Statistical Machine Translation",
                "volume": "",
                "issue": "",
                "pages": "376--380",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Michael Denkowski and Alon Lavie. 2014. Meteor uni- versal: Language specific translation evaluation for any target language. In Proceedings of the Ninth Workshop on Statistical Machine Translation, pages 376-380, Baltimore, Maryland, USA. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF9": {
                "ref_id": "b9",
                "title": "Semantic noise matters for neural natural language generation",
                "authors": [
                    {
                        "first": "Ond\u0159ej",
                        "middle": [],
                        "last": "Du\u0161ek",
                        "suffix": ""
                    },
                    {
                        "first": "David",
                        "middle": [
                            "M"
                        ],
                        "last": "Howcroft",
                        "suffix": ""
                    },
                    {
                        "first": "Verena",
                        "middle": [],
                        "last": "Rieser",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "Proceedings of the 12th International Conference on Natural Language Generation",
                "volume": "",
                "issue": "",
                "pages": "421--426",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Ond\u0159ej Du\u0161ek, David M. Howcroft, and Verena Rieser. 2019. Semantic noise matters for neural natural lan- guage generation. In Proceedings of the 12th Inter- national Conference on Natural Language Genera- tion, pages 421-426, Tokyo, Japan. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF10": {
                "ref_id": "b10",
                "title": "Evaluating the state-of-the-art of end-to-end natural language generation: The e2e nlg challenge",
                "authors": [
                    {
                        "first": "Ond\u0159ej",
                        "middle": [],
                        "last": "Du\u0161ek",
                        "suffix": ""
                    },
                    {
                        "first": "Jekaterina",
                        "middle": [],
                        "last": "Novikova",
                        "suffix": ""
                    },
                    {
                        "first": "Verena",
                        "middle": [],
                        "last": "Rieser",
                        "suffix": ""
                    }
                ],
                "year": 2020,
                "venue": "Computer Speech & Language",
                "volume": "59",
                "issue": "",
                "pages": "123--156",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Ond\u0159ej Du\u0161ek, Jekaterina Novikova, and Verena Rieser. 2020. Evaluating the state-of-the-art of end-to-end natural language generation: The e2e nlg challenge. Computer Speech & Language, 59:123 -156.",
                "links": null
            },
            "BIBREF11": {
                "ref_id": "b11",
                "title": "Comparing automatic evaluation measures for image description",
                "authors": [
                    {
                        "first": "Desmond",
                        "middle": [],
                        "last": "Elliott",
                        "suffix": ""
                    },
                    {
                        "first": "Frank",
                        "middle": [],
                        "last": "Keller",
                        "suffix": ""
                    }
                ],
                "year": 2014,
                "venue": "Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics",
                "volume": "",
                "issue": "",
                "pages": "452--457",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Desmond Elliott and Frank Keller. 2014. Comparing automatic evaluation measures for image descrip- tion. In Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics (Vol- ume 2: Short Papers), pages 452-457, Baltimore, Maryland. Association for Computational Linguis- tics.",
                "links": null
            },
            "BIBREF12": {
                "ref_id": "b12",
                "title": "Utility is in the eye of the user: A critique of nlp leaderboards",
                "authors": [
                    {
                        "first": "Kawin",
                        "middle": [],
                        "last": "Ethayarajh",
                        "suffix": ""
                    },
                    {
                        "first": "Dan",
                        "middle": [],
                        "last": "Jurafsky",
                        "suffix": ""
                    }
                ],
                "year": 2020,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Kawin Ethayarajh and Dan Jurafsky. 2020. Utility is in the eye of the user: A critique of nlp leaderboards.",
                "links": null
            },
            "BIBREF13": {
                "ref_id": "b13",
                "title": "On building a more efficient grammar by exploiting types",
                "authors": [
                    {
                        "first": "Dan",
                        "middle": [],
                        "last": "Flickinger",
                        "suffix": ""
                    }
                ],
                "year": 2000,
                "venue": "Nat. Lang. Eng",
                "volume": "6",
                "issue": "1",
                "pages": "15--28",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Dan Flickinger. 2000. On building a more efficient grammar by exploiting types. Nat. Lang. Eng., 6(1):15-28.",
                "links": null
            },
            "BIBREF14": {
                "ref_id": "b14",
                "title": "Accuracy vs. robustness in grammar engineering",
                "authors": [
                    {
                        "first": "Dan",
                        "middle": [],
                        "last": "Flickinger",
                        "suffix": ""
                    }
                ],
                "year": 2011,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Dan Flickinger. 2011. Accuracy vs. robustness in grammar engineering. In E. M. Bender and J. E.",
                "links": null
            },
            "BIBREF15": {
                "ref_id": "b15",
                "title": "Language from a cognitive perspective: Grammar, usage, and processing",
                "authors": [
                    {
                        "first": "Arnold",
                        "middle": [],
                        "last": "",
                        "suffix": ""
                    }
                ],
                "year": null,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "31--50",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Arnold, editors, Language from a cognitive perspec- tive: Grammar, usage, and processing, pages 31-50. Stanford: CSLI Publications.",
                "links": null
            },
            "BIBREF16": {
                "ref_id": "b16",
                "title": "APE at scale and its implications on MT evaluation biases",
                "authors": [
                    {
                        "first": "Markus",
                        "middle": [],
                        "last": "Freitag",
                        "suffix": ""
                    },
                    {
                        "first": "Isaac",
                        "middle": [],
                        "last": "Caswell",
                        "suffix": ""
                    },
                    {
                        "first": "Scott",
                        "middle": [],
                        "last": "Roy",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "Proceedings of the Fourth Conference on Machine Translation",
                "volume": "1",
                "issue": "",
                "pages": "34--44",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Markus Freitag, Isaac Caswell, and Scott Roy. 2019. APE at scale and its implications on MT evaluation biases. In Proceedings of the Fourth Conference on Machine Translation (Volume 1: Research Papers), pages 34-44, Florence, Italy. Association for Com- putational Linguistics.",
                "links": null
            },
            "BIBREF17": {
                "ref_id": "b17",
                "title": "Bleu might be guilty but references are not innocent",
                "authors": [
                    {
                        "first": "Markus",
                        "middle": [],
                        "last": "Freitag",
                        "suffix": ""
                    },
                    {
                        "first": "David",
                        "middle": [],
                        "last": "Grangier",
                        "suffix": ""
                    },
                    {
                        "first": "Isaac",
                        "middle": [],
                        "last": "Caswell",
                        "suffix": ""
                    }
                ],
                "year": 2020,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Markus Freitag, David Grangier, and Isaac Caswell. 2020. Bleu might be guilty but references are not innocent.",
                "links": null
            },
            "BIBREF18": {
                "ref_id": "b18",
                "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": {},
                "num": null,
                "urls": [],
                "raw_text": "Claire Gardent, Anastasia Shimorina, Shashi Narayan, and Laura Perez-Beltrachini. 2017. 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
            },
            "BIBREF19": {
                "ref_id": "b19",
                "title": "SimpleNLG: A realisation engine for practical applications",
                "authors": [
                    {
                        "first": "Albert",
                        "middle": [],
                        "last": "Gatt",
                        "suffix": ""
                    },
                    {
                        "first": "Ehud",
                        "middle": [],
                        "last": "Reiter",
                        "suffix": ""
                    }
                ],
                "year": 2009,
                "venue": "Proceedings of the 12th European Workshop on Natural Language Generation (ENLG 2009)",
                "volume": "",
                "issue": "",
                "pages": "90--93",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Albert Gatt and Ehud Reiter. 2009. SimpleNLG: A re- alisation engine for practical applications. In Pro- ceedings of the 12th European Workshop on Natural Language Generation (ENLG 2009), pages 90-93, Athens, Greece. Association for Computational Lin- guistics.",
                "links": null
            },
            "BIBREF20": {
                "ref_id": "b20",
                "title": "Natural language generation enhances human decision-making with uncertain information",
                "authors": [
                    {
                        "first": "Dimitra",
                        "middle": [],
                        "last": "Gkatzia",
                        "suffix": ""
                    },
                    {
                        "first": "Oliver",
                        "middle": [],
                        "last": "Lemon",
                        "suffix": ""
                    },
                    {
                        "first": "Verena",
                        "middle": [],
                        "last": "Rieser",
                        "suffix": ""
                    }
                ],
                "year": 2016,
                "venue": "Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics",
                "volume": "2",
                "issue": "",
                "pages": "264--268",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Dimitra Gkatzia, Oliver Lemon, and Verena Rieser. 2016. Natural language generation enhances human decision-making with uncertain information. In Pro- ceedings of the 54th Annual Meeting of the Associa- tion for Computational Linguistics (Volume 2: Short Papers), pages 264-268, Berlin, Germany. Associa- tion for Computational Linguistics.",
                "links": null
            },
            "BIBREF21": {
                "ref_id": "b21",
                "title": "Automatic generation of natural language nursing shift summaries in neonatal intensive care: Bt-nurse",
                "authors": [
                    {
                        "first": "James",
                        "middle": [],
                        "last": "Hunter",
                        "suffix": ""
                    },
                    {
                        "first": "Yvonne",
                        "middle": [],
                        "last": "Freer",
                        "suffix": ""
                    },
                    {
                        "first": "Albert",
                        "middle": [],
                        "last": "Gatt",
                        "suffix": ""
                    },
                    {
                        "first": "Ehud",
                        "middle": [],
                        "last": "Reiter",
                        "suffix": ""
                    },
                    {
                        "first": "Somayajulu",
                        "middle": [],
                        "last": "Sripada",
                        "suffix": ""
                    },
                    {
                        "first": "Cindy",
                        "middle": [],
                        "last": "Sykes",
                        "suffix": ""
                    }
                ],
                "year": 2012,
                "venue": "Artificial Intelligence in Medicine",
                "volume": "56",
                "issue": "3",
                "pages": "157--172",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "James Hunter, Yvonne Freer, Albert Gatt, Ehud Reiter, Somayajulu Sripada, and Cindy Sykes. 2012. Au- tomatic generation of natural language nursing shift summaries in neonatal intensive care: Bt-nurse. Ar- tificial Intelligence in Medicine, 56(3):157 -172.",
                "links": null
            },
            "BIBREF22": {
                "ref_id": "b22",
                "title": "Simplenlg-nl : Natural language generation for dutch",
                "authors": [
                    {
                        "first": "R",
                        "middle": [
                            "F"
                        ],
                        "last": "De",
                        "suffix": ""
                    },
                    {
                        "first": "Jong",
                        "middle": [],
                        "last": "",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "R.F. de Jong. 2018. Simplenlg-nl : Natural language generation for dutch.",
                "links": null
            },
            "BIBREF23": {
                "ref_id": "b23",
                "title": "Re-evaluating automatic metrics for image captioning",
                "authors": [
                    {
                        "first": "Mert",
                        "middle": [],
                        "last": "Kilickaya",
                        "suffix": ""
                    },
                    {
                        "first": "Aykut",
                        "middle": [],
                        "last": "Erdem",
                        "suffix": ""
                    },
                    {
                        "first": "Nazli",
                        "middle": [],
                        "last": "Ikizler-Cinbis",
                        "suffix": ""
                    },
                    {
                        "first": "Erkut",
                        "middle": [],
                        "last": "Erdem",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics",
                "volume": "1",
                "issue": "",
                "pages": "199--209",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Mert Kilickaya, Aykut Erdem, Nazli Ikizler-Cinbis, and Erkut Erdem. 2017. Re-evaluating automatic metrics for image captioning. In Proceedings of the 15th Conference of the European Chapter of the As- sociation for Computational Linguistics: Volume 1, Long Papers, pages 199-209, Valencia, Spain. Asso- ciation for Computational Linguistics.",
                "links": null
            },
            "BIBREF24": {
                "ref_id": "b24",
                "title": "Dbpedia-a large-scale, multilingual knowledge base extracted from wikipedia",
                "authors": [
                    {
                        "first": "Jens",
                        "middle": [],
                        "last": "Lehmann",
                        "suffix": ""
                    },
                    {
                        "first": "Robert",
                        "middle": [],
                        "last": "Isele",
                        "suffix": ""
                    },
                    {
                        "first": "Max",
                        "middle": [],
                        "last": "Jakob",
                        "suffix": ""
                    },
                    {
                        "first": "Anja",
                        "middle": [],
                        "last": "Jentzsch",
                        "suffix": ""
                    },
                    {
                        "first": "Dimitris",
                        "middle": [],
                        "last": "Kontokostas",
                        "suffix": ""
                    },
                    {
                        "first": "Pablo",
                        "middle": [
                            "N"
                        ],
                        "last": "Mendes",
                        "suffix": ""
                    },
                    {
                        "first": "Sebastian",
                        "middle": [],
                        "last": "Hellmann",
                        "suffix": ""
                    },
                    {
                        "first": "Mohamed",
                        "middle": [],
                        "last": "Morsey",
                        "suffix": ""
                    },
                    {
                        "first": "Patrick",
                        "middle": [],
                        "last": "Van Kleef",
                        "suffix": ""
                    },
                    {
                        "first": "S\u00f6ren",
                        "middle": [],
                        "last": "Auer",
                        "suffix": ""
                    }
                ],
                "year": 2015,
                "venue": "",
                "volume": "6",
                "issue": "",
                "pages": "167--195",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Jens Lehmann, Robert Isele, Max Jakob, Anja Jentzsch, Dimitris Kontokostas, Pablo N Mendes, Sebastian Hellmann, Mohamed Morsey, Patrick Van Kleef, S\u00f6ren Auer, et al. 2015. Dbpedia-a large-scale, mul- tilingual knowledge base extracted from wikipedia. Semantic web, 6(2):167-195.",
                "links": null
            },
            "BIBREF25": {
                "ref_id": "b25",
                "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
            },
            "BIBREF26": {
                "ref_id": "b26",
                "title": "Microsoft coco: Common objects in context",
                "authors": [
                    {
                        "first": "Tsung-Yi",
                        "middle": [],
                        "last": "Lin",
                        "suffix": ""
                    },
                    {
                        "first": "Michael",
                        "middle": [],
                        "last": "Maire",
                        "suffix": ""
                    },
                    {
                        "first": "Serge",
                        "middle": [],
                        "last": "Belongie",
                        "suffix": ""
                    },
                    {
                        "first": "James",
                        "middle": [],
                        "last": "Hays",
                        "suffix": ""
                    },
                    {
                        "first": "Pietro",
                        "middle": [],
                        "last": "Perona",
                        "suffix": ""
                    },
                    {
                        "first": "Deva",
                        "middle": [],
                        "last": "Ramanan",
                        "suffix": ""
                    },
                    {
                        "first": "Piotr",
                        "middle": [],
                        "last": "Doll\u00e1r",
                        "suffix": ""
                    },
                    {
                        "first": "C",
                        "middle": [
                            "Lawrence"
                        ],
                        "last": "Zitnick",
                        "suffix": ""
                    }
                ],
                "year": 2014,
                "venue": "Computer Vision -ECCV 2014",
                "volume": "",
                "issue": "",
                "pages": "740--755",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Doll\u00e1r, and C. Lawrence Zitnick. 2014. Microsoft coco: Common objects in context. In Computer Vision - ECCV 2014, pages 740-755, Cham. Springer Inter- national Publishing.",
                "links": null
            },
            "BIBREF27": {
                "ref_id": "b27",
                "title": "Towards personality-based user adaptation: psychologically informed stylistic language generation",
                "authors": [
                    {
                        "first": "Fran\u00e7ois",
                        "middle": [],
                        "last": "Mairesse",
                        "suffix": ""
                    },
                    {
                        "first": "Marilyn",
                        "middle": [
                            "A"
                        ],
                        "last": "Walker",
                        "suffix": ""
                    }
                ],
                "year": 2010,
                "venue": "User Modeling and User-Adapted Interaction",
                "volume": "20",
                "issue": "3",
                "pages": "227--278",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Fran\u00e7ois Mairesse and Marilyn A. Walker. 2010. To- wards personality-based user adaptation: psycholog- ically informed stylistic language generation. User Modeling and User-Adapted Interaction, 20(3):227- 278.",
                "links": null
            },
            "BIBREF28": {
                "ref_id": "b28",
                "title": "Tangled up in BLEU: Reevaluating the evaluation of automatic machine translation evaluation metrics",
                "authors": [
                    {
                        "first": "Nitika",
                        "middle": [],
                        "last": "Mathur",
                        "suffix": ""
                    },
                    {
                        "first": "Timothy",
                        "middle": [],
                        "last": "Baldwin",
                        "suffix": ""
                    },
                    {
                        "first": "Trevor",
                        "middle": [],
                        "last": "Cohn",
                        "suffix": ""
                    }
                ],
                "year": 2020,
                "venue": "Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics",
                "volume": "",
                "issue": "",
                "pages": "4984--4997",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Nitika Mathur, Timothy Baldwin, and Trevor Cohn. 2020. Tangled up in BLEU: Reevaluating the eval- uation of automatic machine translation evaluation metrics. In Proceedings of the 58th Annual Meet- ing of the Association for Computational Linguistics, pages 4984-4997, Online. Association for Computa- tional Linguistics.",
                "links": null
            },
            "BIBREF29": {
                "ref_id": "b29",
                "title": "SimpleNLG-IT: adapting SimpleNLG to Italian",
                "authors": [
                    {
                        "first": "Alessandro",
                        "middle": [],
                        "last": "Mazzei",
                        "suffix": ""
                    },
                    {
                        "first": "Cristina",
                        "middle": [],
                        "last": "Battaglino",
                        "suffix": ""
                    },
                    {
                        "first": "Cristina",
                        "middle": [],
                        "last": "Bosco",
                        "suffix": ""
                    }
                ],
                "year": 2016,
                "venue": "Proceedings of the 9th International Natural Language Generation conference",
                "volume": "",
                "issue": "",
                "pages": "184--192",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Alessandro Mazzei, Cristina Battaglino, and Cristina Bosco. 2016. SimpleNLG-IT: adapting SimpleNLG to Italian. In Proceedings of the 9th International Natural Language Generation conference, pages 184-192, Edinburgh, UK. Association for Compu- tational Linguistics.",
                "links": null
            },
            "BIBREF30": {
                "ref_id": "b30",
                "title": "A portable grammar-based nlg system for verbalization of structured data",
                "authors": [
                    {
                        "first": "Simon",
                        "middle": [],
                        "last": "Mille",
                        "suffix": ""
                    },
                    {
                        "first": "Stamatia",
                        "middle": [],
                        "last": "Dasiopoulou",
                        "suffix": ""
                    },
                    {
                        "first": "Leo",
                        "middle": [],
                        "last": "Wanner",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "Proceedings of the 34th ACM/SIGAPP Symposium on Applied Computing, SAC '19",
                "volume": "",
                "issue": "",
                "pages": "1054--1056",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Simon Mille, Stamatia Dasiopoulou, and Leo Wanner. 2019. A portable grammar-based nlg system for verbalization of structured data. In Proceedings of the 34th ACM/SIGAPP Symposium on Applied Com- puting, SAC '19, page 1054-1056, New York, NY, USA. Association for Computing Machinery.",
                "links": null
            },
            "BIBREF31": {
                "ref_id": "b31",
                "title": "Stereotyping and bias in the flickr30k dataset",
                "authors": [
                    {
                        "first": "",
                        "middle": [],
                        "last": "Emiel Van Miltenburg",
                        "suffix": ""
                    }
                ],
                "year": 2016,
                "venue": "Proceedings of Multimodal Corpora: Computer vision and language processing (MMC 2016)",
                "volume": "",
                "issue": "",
                "pages": "1--4",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Emiel van Miltenburg. 2016. Stereotyping and bias in the flickr30k dataset. In Proceedings of Multimodal Corpora: Computer vision and language processing (MMC 2016), pages 1-4.",
                "links": null
            },
            "BIBREF32": {
                "ref_id": "b32",
                "title": "How do image description systems describe people? a targeted assessment of system competence in the people domain",
                "authors": [
                    {
                        "first": "",
                        "middle": [],
                        "last": "Emiel Van Miltenburg",
                        "suffix": ""
                    }
                ],
                "year": 2020,
                "venue": "Proceedings of LANTERN. Association for Computational Linguistics",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Emiel van Miltenburg. 2020. How do image descrip- tion systems describe people? a targeted assessment of system competence in the people domain. In Proceedings of LANTERN. Association for Compu- tational Linguistics.",
                "links": null
            },
            "BIBREF33": {
                "ref_id": "b33",
                "title": "Cross-linguistic differences and similarities in image descriptions",
                "authors": [
                    {
                        "first": "Desmond",
                        "middle": [],
                        "last": "Emiel Van Miltenburg",
                        "suffix": ""
                    },
                    {
                        "first": "Piek",
                        "middle": [],
                        "last": "Elliott",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Vossen",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "Proceedings of the 10th International Conference on Natural Language Generation",
                "volume": "",
                "issue": "",
                "pages": "21--30",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Emiel van Miltenburg, Desmond Elliott, and Piek Vossen. 2017. Cross-linguistic differences and simi- larities in image descriptions. In Proceedings of the 10th International Conference on Natural Language Generation, pages 21-30, Santiago de Compostela, Spain. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF34": {
                "ref_id": "b34",
                "title": "Talking about other people: an endless range of possibilities",
                "authors": [
                    {
                        "first": "Desmond",
                        "middle": [],
                        "last": "Emiel Van Miltenburg",
                        "suffix": ""
                    },
                    {
                        "first": "Piek",
                        "middle": [],
                        "last": "Elliott",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Vossen",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "Proceedings of the 11th International Conference on Natural Language Generation",
                "volume": "",
                "issue": "",
                "pages": "415--420",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Emiel van Miltenburg, Desmond Elliott, and Piek Vossen. 2018. Talking about other people: an endless range of possibilities. In Proceedings of the 11th International Conference on Natural Lan- guage Generation, pages 415-420, Tilburg Univer- sity, The Netherlands. Association for Computa- tional Linguistics.",
                "links": null
            },
            "BIBREF35": {
                "ref_id": "b35",
                "title": "Machine learning",
                "authors": [
                    {
                        "first": "Tom",
                        "middle": [
                            "M"
                        ],
                        "last": "Mitchell",
                        "suffix": ""
                    }
                ],
                "year": 1997,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Tom M. Mitchell. 1997. Machine learning. McGraw- Hill.",
                "links": null
            },
            "BIBREF36": {
                "ref_id": "b36",
                "title": "Rob van der Goot, Hessel Haagsma, Barbara Plank, and Martijn Wieling",
                "authors": [
                    {
                        "first": "Malvina",
                        "middle": [],
                        "last": "Nissim",
                        "suffix": ""
                    },
                    {
                        "first": "Lasha",
                        "middle": [],
                        "last": "Abzianidze",
                        "suffix": ""
                    },
                    {
                        "first": "Kilian",
                        "middle": [],
                        "last": "Evang",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "Computational Linguistics",
                "volume": "43",
                "issue": "4",
                "pages": "897--904",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Malvina Nissim, Lasha Abzianidze, Kilian Evang, Rob van der Goot, Hessel Haagsma, Barbara Plank, and Martijn Wieling. 2017. Last words: Sharing is car- ing: The future of shared tasks. Computational Lin- guistics, 43(4):897-904.",
                "links": null
            },
            "BIBREF37": {
                "ref_id": "b37",
                "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": {},
                "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
            },
            "BIBREF38": {
                "ref_id": "b38",
                "title": "Crowd-sourcing NLG data: Pictures elicit better data",
                "authors": [
                    {
                        "first": "Jekaterina",
                        "middle": [],
                        "last": "Novikova",
                        "suffix": ""
                    },
                    {
                        "first": "Oliver",
                        "middle": [],
                        "last": "Lemon",
                        "suffix": ""
                    },
                    {
                        "first": "Verena",
                        "middle": [],
                        "last": "Rieser",
                        "suffix": ""
                    }
                ],
                "year": 2016,
                "venue": "Proceedings of the 9th International Natural Language Generation conference",
                "volume": "",
                "issue": "",
                "pages": "265--273",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Jekaterina Novikova, Oliver Lemon, and Verena Rieser. 2016. Crowd-sourcing NLG data: Pictures elicit bet- ter data. In Proceedings of the 9th International Nat- ural Language Generation conference, pages 265- 273, Edinburgh, UK. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF39": {
                "ref_id": "b39",
                "title": "Adapting SimpleNLG for Brazilian Portuguese realisation",
                "authors": [
                    {
                        "first": "Rodrigo",
                        "middle": [],
                        "last": "De Oliveira",
                        "suffix": ""
                    },
                    {
                        "first": "Somayajulu",
                        "middle": [],
                        "last": "Sripada",
                        "suffix": ""
                    }
                ],
                "year": 2014,
                "venue": "Proceedings of the 8th International Natural Language Generation Conference (INLG)",
                "volume": "",
                "issue": "",
                "pages": "93--94",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Rodrigo de Oliveira and Somayajulu Sripada. 2014. Adapting SimpleNLG for Brazilian Portuguese reali- sation. In Proceedings of the 8th International Natu- ral Language Generation Conference (INLG), pages 93-94, Philadelphia, Pennsylvania, U.S.A. Associa- tion for Computational Linguistics.",
                "links": null
            },
            "BIBREF40": {
                "ref_id": "b40",
                "title": "Controlling personality-based stylistic variation with neural natural language generators",
                "authors": [
                    {
                        "first": "Shereen",
                        "middle": [],
                        "last": "Oraby",
                        "suffix": ""
                    },
                    {
                        "first": "Lena",
                        "middle": [],
                        "last": "Reed",
                        "suffix": ""
                    },
                    {
                        "first": "Shubhangi",
                        "middle": [],
                        "last": "Tandon",
                        "suffix": ""
                    },
                    {
                        "first": "T",
                        "middle": [
                            "S"
                        ],
                        "last": "Sharath",
                        "suffix": ""
                    },
                    {
                        "first": "Stephanie",
                        "middle": [],
                        "last": "Lukin",
                        "suffix": ""
                    },
                    {
                        "first": "Marilyn",
                        "middle": [],
                        "last": "Walker",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "Proceedings of the 19th Annual SIGdial Meeting on Discourse and Dialogue",
                "volume": "",
                "issue": "",
                "pages": "180--190",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Shereen Oraby, Lena Reed, Shubhangi Tandon, Sharath T.S., Stephanie Lukin, and Marilyn Walker. 2018. Controlling personality-based stylistic varia- tion with neural natural language generators. In Pro- ceedings of the 19th Annual SIGdial Meeting on Dis- course and Dialogue, pages 180-190, Melbourne, Australia. Association for Computational Linguis- tics.",
                "links": null
            },
            "BIBREF41": {
                "ref_id": "b41",
                "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": {},
                "num": null,
                "urls": [],
                "raw_text": "Kishore Papineni, Salim Roukos, Todd Ward, and Wei- Jing Zhu. 2002. Bleu: a method for automatic eval- uation 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
            },
            "BIBREF42": {
                "ref_id": "b42",
                "title": "Ethical considerations in NLP shared tasks",
                "authors": [
                    {
                        "first": "Carla",
                        "middle": [],
                        "last": "Parra Escart\u00edn",
                        "suffix": ""
                    },
                    {
                        "first": "Wessel",
                        "middle": [],
                        "last": "Reijers",
                        "suffix": ""
                    },
                    {
                        "first": "Teresa",
                        "middle": [],
                        "last": "Lynn",
                        "suffix": ""
                    },
                    {
                        "first": "Joss",
                        "middle": [],
                        "last": "Moorkens",
                        "suffix": ""
                    },
                    {
                        "first": "Andy",
                        "middle": [],
                        "last": "Way",
                        "suffix": ""
                    },
                    {
                        "first": "Chao-Hong",
                        "middle": [],
                        "last": "Liu",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "Proceedings of the First ACL Workshop on Ethics in Natural Language Processing",
                "volume": "",
                "issue": "",
                "pages": "66--73",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Carla Parra Escart\u00edn, Wessel Reijers, Teresa Lynn, Joss Moorkens, Andy Way, and Chao-Hong Liu. 2017. Ethical considerations in NLP shared tasks. In Pro- ceedings of the First ACL Workshop on Ethics in Nat- ural Language Processing, pages 66-73, Valencia, Spain. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF43": {
                "ref_id": "b43",
                "title": "Building RDF content for data-to-text generation",
                "authors": [],
                "year": 2016,
                "venue": "Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers",
                "volume": "",
                "issue": "",
                "pages": "1493--1502",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Laura Perez-Beltrachini, Rania Sayed, and Claire Gar- dent. 2016. Building RDF content for data-to-text generation. In Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 1493-1502, Osaka, Japan. The COLING 2016 Organizing Com- mittee.",
                "links": null
            },
            "BIBREF44": {
                "ref_id": "b44",
                "title": "Head-Driven Phrase Structure Grammar. Head-driven Phrase Structure Grammar",
                "authors": [
                    {
                        "first": "Carl",
                        "middle": [],
                        "last": "Pollard",
                        "suffix": ""
                    },
                    {
                        "first": "Ivan",
                        "middle": [
                            "A"
                        ],
                        "last": "Sag",
                        "suffix": ""
                    }
                ],
                "year": 1994,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Carl. Pollard and Ivan A. Sag. 1994. Head-Driven Phrase Structure Grammar. Head-driven Phrase Structure Grammar. University of Chicago Press.",
                "links": null
            },
            "BIBREF45": {
                "ref_id": "b45",
                "title": "Adapting SimpleNLG to Spanish",
                "authors": [
                    {
                        "first": "Alejandro",
                        "middle": [],
                        "last": "Ramos-Soto",
                        "suffix": ""
                    },
                    {
                        "first": "Julio",
                        "middle": [],
                        "last": "Janeiro-Gallardo",
                        "suffix": ""
                    },
                    {
                        "first": "Alberto",
                        "middle": [
                            "Bugar\u00edn"
                        ],
                        "last": "Diz",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "Proceedings of the 10th International Conference on Natural Language Generation",
                "volume": "",
                "issue": "",
                "pages": "144--148",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Alejandro Ramos-Soto, Julio Janeiro-Gallardo, and Al- berto Bugar\u00edn Diz. 2017. Adapting SimpleNLG to Spanish. In Proceedings of the 10th International Conference on Natural Language Generation, pages 144-148, Santiago de Compostela, Spain. Associa- tion for Computational Linguistics.",
                "links": null
            },
            "BIBREF46": {
                "ref_id": "b46",
                "title": "Natural language generation and machine learning",
                "authors": [
                    {
                        "first": "Ehud",
                        "middle": [],
                        "last": "Reiter",
                        "suffix": ""
                    }
                ],
                "year": 2016,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Ehud Reiter. 2016. Natural language generation and machine learning. Ehud Reiter's blog, https:// ehudreiter.com/2016/12/12/nlg-and-ml/.",
                "links": null
            },
            "BIBREF47": {
                "ref_id": "b47",
                "title": "You need to understand your corpora! The Weathergov example",
                "authors": [
                    {
                        "first": "Ehud",
                        "middle": [],
                        "last": "Reiter",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Ehud Reiter. 2017. You need to understand your cor- pora! The Weathergov example. Ehud Reiter's blog, https://ehudreiter.com/2017/05/09/ weathergov/.",
                "links": null
            },
            "BIBREF48": {
                "ref_id": "b48",
                "title": "A structured review of the validity of BLEU",
                "authors": [
                    {
                        "first": "Ehud",
                        "middle": [],
                        "last": "Reiter",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "Computational Linguistics",
                "volume": "44",
                "issue": "3",
                "pages": "393--401",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Ehud Reiter. 2018. A structured review of the validity of BLEU. Computational Linguistics, 44(3):393- 401.",
                "links": null
            },
            "BIBREF49": {
                "ref_id": "b49",
                "title": "Is building neural nlg faster than rules nlg? no one knows, but i suspect not. Ehud Reiter's blog",
                "authors": [
                    {
                        "first": "Ehud",
                        "middle": [],
                        "last": "Reiter",
                        "suffix": ""
                    }
                ],
                "year": 2020,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Ehud Reiter. 2020. Is building neural nlg faster than rules nlg? no one knows, but i suspect not. Ehud Reiter's blog, https://ehudreiter.com/2020/ 05/11/is-building-neural-nlg-faster/.",
                "links": null
            },
            "BIBREF50": {
                "ref_id": "b50",
                "title": "How the transformers broke nlp leaderboards",
                "authors": [
                    {
                        "first": "Anna",
                        "middle": [
                            "Rogers"
                        ],
                        "last": "",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Anna Rogers. 2019. How the transformers broke nlp leaderboards. Posted on the Hacking Semantics blog: https://hackingsemantics.xyz/2019/ leaderboards/.",
                "links": null
            },
            "BIBREF51": {
                "ref_id": "b51",
                "title": "Language tasks and language games: On methodology in current natural language processing research",
                "authors": [
                    {
                        "first": "David",
                        "middle": [],
                        "last": "Schlangen",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "David Schlangen. 2019. Language tasks and language games: On methodology in current natural language processing research. CoRR, abs/1908.10747.",
                "links": null
            },
            "BIBREF52": {
                "ref_id": "b52",
                "title": "Targeting the benchmark: On methodology in current natural language processing research",
                "authors": [
                    {
                        "first": "David",
                        "middle": [],
                        "last": "Schlangen",
                        "suffix": ""
                    }
                ],
                "year": 2020,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "David Schlangen. 2020. Targeting the benchmark: On methodology in current natural language processing research. CoRR, abs/2007.04792.",
                "links": null
            },
            "BIBREF53": {
                "ref_id": "b53",
                "title": "Energy and policy considerations for deep learning in NLP",
                "authors": [
                    {
                        "first": "Emma",
                        "middle": [],
                        "last": "Strubell",
                        "suffix": ""
                    },
                    {
                        "first": "Ananya",
                        "middle": [],
                        "last": "Ganesh",
                        "suffix": ""
                    },
                    {
                        "first": "Andrew",
                        "middle": [],
                        "last": "Mccallum",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics",
                "volume": "",
                "issue": "",
                "pages": "3645--3650",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Emma Strubell, Ananya Ganesh, and Andrew McCal- lum. 2019. Energy and policy considerations for deep learning in NLP. In Proceedings of the 57th Annual Meeting of the Association for Computa- tional Linguistics, pages 3645-3650, Florence, Italy. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF54": {
                "ref_id": "b54",
                "title": "BLEU is not suitable for the evaluation of text simplification",
                "authors": [
                    {
                        "first": "Elior",
                        "middle": [],
                        "last": "Sulem",
                        "suffix": ""
                    },
                    {
                        "first": "Omri",
                        "middle": [],
                        "last": "Abend",
                        "suffix": ""
                    },
                    {
                        "first": "Ari",
                        "middle": [],
                        "last": "Rappoport",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing",
                "volume": "",
                "issue": "",
                "pages": "738--744",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Elior Sulem, Omri Abend, and Ari Rappoport. 2018. BLEU is not suitable for the evaluation of text sim- plification. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Process- ing, pages 738-744, Brussels, Belgium. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF55": {
                "ref_id": "b55",
                "title": "Adapting SimpleNLG for bilingual English-French realisation",
                "authors": [
                    {
                        "first": "Pierre-Luc",
                        "middle": [],
                        "last": "Vaudry",
                        "suffix": ""
                    },
                    {
                        "first": "Guy",
                        "middle": [],
                        "last": "Lapalme",
                        "suffix": ""
                    }
                ],
                "year": 2013,
                "venue": "Proceedings of the 14th European Workshop on Natural Language Generation",
                "volume": "",
                "issue": "",
                "pages": "183--187",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Pierre-Luc Vaudry and Guy Lapalme. 2013. Adapt- ing SimpleNLG for bilingual English-French reali- sation. In Proceedings of the 14th European Work- shop on Natural Language Generation, pages 183- 187, Sofia, Bulgaria. Association for Computational Linguistics.",
                "links": null
            }
        },
        "ref_entries": {
            "FIGREF0": {
                "num": null,
                "uris": null,
                "text": "Parse tree for the phrase: a tall happy toddler.",
                "type_str": "figure"
            },
            "TABREF3": {
                "text": "Number of templates for predicates in the '1triples' subset of the WebNLG corpus. Columns show the predicate, number of unique templates, total number of templates, and the ratio of unique-to-total number of predicates. This table shows a selection across the entire range of different ratios.",
                "content": "<table/>",
                "html": null,
                "num": null,
                "type_str": "table"
            },
            "TABREF4": {
                "text": "25.2877.7516. Software Triple: SAGE Publications, founder, Sara Miller McCune Text: Sara Miller McCune founded SAGE Publications. Template: ENT-1 founded ENT-2.",
                "content": "<table><tr><td>Mapping: ENT-1: Sara Miller McCune ENT-2: SAGE Publications</td></tr></table>",
                "html": null,
                "num": null,
                "type_str": "table"
            },
            "TABREF5": {
                "text": "Example from the extended WebNLG corpus.",
                "content": "<table/>",
                "html": null,
                "num": null,
                "type_str": "table"
            },
            "TABREF6": {
                "text": "",
                "content": "<table/>",
                "html": null,
                "num": null,
                "type_str": "table"
            }
        }
    }
}