Base_JSON/prefixN/json/nodalida/2021.nodalida-main.11.json

{
    "paper_id": "2021",
    "header": {
        "generated_with": "S2ORC 1.0.0",
        "date_generated": "2023-01-19T14:31:49.120926Z"
    },
    "title": "A Baseline Document Planning Method for Automated Journalism",
    "authors": [
        {
            "first": "Leo",
            "middle": [],
            "last": "Lepp\u00e4nen",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "University of Helsinki",
                "location": {}
            },
            "email": "leo.leppanen@helsinki.fi"
        },
        {
            "first": "Hannu",
            "middle": [],
            "last": "Toivonen",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "University of Helsinki",
                "location": {}
            },
            "email": "hannu.toivonen@helsinki.fi"
        }
    ],
    "year": "",
    "venue": null,
    "identifiers": {},
    "abstract": "In this work, we present a method for content selection and document planning for automated news and report generation from structured statistical data such as that offered by the European Union's statistical agency, Eurostat. The method is driven by the data and is highly topic-independent within the statistical dataset domain. As our approach is not based on machine learning, it is suitable for introducing news automation to the wide variety of domains where no training data is available. As such, it is suitable as a low-cost (in terms of implementation effort) baseline for document structuring prior to introduction of domainspecific knowledge.",
    "pdf_parse": {
        "paper_id": "2021",
        "_pdf_hash": "",
        "abstract": [
            {
                "text": "In this work, we present a method for content selection and document planning for automated news and report generation from structured statistical data such as that offered by the European Union's statistical agency, Eurostat. The method is driven by the data and is highly topic-independent within the statistical dataset domain. As our approach is not based on machine learning, it is suitable for introducing news automation to the wide variety of domains where no training data is available. As such, it is suitable as a low-cost (in terms of implementation effort) baseline for document structuring prior to introduction of domainspecific knowledge.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Abstract",
                "sec_num": null
            }
        ],
        "body_text": [
            {
                "text": "Automated generation of news texts from structured data -often referred to as 'automated journalism' (Graefe, 2016; D\u00f6rr, 2015; Caswell and D\u00f6rr, 2018) or 'news automation' (Linden, 2017; Sir\u00e9n-Heikel et al., 2019; Dierickx, 2019) -is of great interest to various news producers. It is seen as a way of 'providing efficiency, increasing output and aiding in reallocating resources to pursue quality journalism' (Sir\u00e9n-Heikel et al., 2019, p. 47) . While data-to-text NLG systems are still far from common especially among the smaller, regional news industry players, at least among the larger newsrooms the use of NLG approaches has clearly been established (Fanta, 2017) .",
                "cite_spans": [
                    {
                        "start": 101,
                        "end": 115,
                        "text": "(Graefe, 2016;",
                        "ref_id": "BIBREF13"
                    },
                    {
                        "start": 116,
                        "end": 127,
                        "text": "D\u00f6rr, 2015;",
                        "ref_id": "BIBREF5"
                    },
                    {
                        "start": 128,
                        "end": 151,
                        "text": "Caswell and D\u00f6rr, 2018)",
                        "ref_id": "BIBREF2"
                    },
                    {
                        "start": 173,
                        "end": 187,
                        "text": "(Linden, 2017;",
                        "ref_id": "BIBREF21"
                    },
                    {
                        "start": 188,
                        "end": 214,
                        "text": "Sir\u00e9n-Heikel et al., 2019;",
                        "ref_id": "BIBREF32"
                    },
                    {
                        "start": 215,
                        "end": 230,
                        "text": "Dierickx, 2019)",
                        "ref_id": "BIBREF4"
                    },
                    {
                        "start": 411,
                        "end": 445,
                        "text": "(Sir\u00e9n-Heikel et al., 2019, p. 47)",
                        "ref_id": null
                    },
                    {
                        "start": 658,
                        "end": 671,
                        "text": "(Fanta, 2017)",
                        "ref_id": "BIBREF8"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "While secrecy in the industry makes it difficult to establish the commercial reality as an outsider, the limited available evidence indicates that commercial automated journalism is mostly done using rule-based methods despite a surge of academic interest in increasingly complex neural methods for NLG (e.g. Puduppully et al., 2019; Ferreira et al., 2019) : Interviews of news automation users indicate that the employed methods are mostly based on templates (Sir\u00e9n-Heikel et al., 2019) , as are the few open source code repositories of real-world news automation systems (Yleisradio, 2018) . Indeed, some NLG industry experts believe that especially end-to-end neural models do not match customer needs at this time (Reiter, 2019) .",
                "cite_spans": [
                    {
                        "start": 309,
                        "end": 333,
                        "text": "Puduppully et al., 2019;",
                        "ref_id": "BIBREF26"
                    },
                    {
                        "start": 334,
                        "end": 356,
                        "text": "Ferreira et al., 2019)",
                        "ref_id": "BIBREF9"
                    },
                    {
                        "start": 460,
                        "end": 487,
                        "text": "(Sir\u00e9n-Heikel et al., 2019)",
                        "ref_id": "BIBREF32"
                    },
                    {
                        "start": 573,
                        "end": 591,
                        "text": "(Yleisradio, 2018)",
                        "ref_id": "BIBREF39"
                    },
                    {
                        "start": 718,
                        "end": 732,
                        "text": "(Reiter, 2019)",
                        "ref_id": "BIBREF30"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "Contributing factors include a lack of control (Reiter, 2019) ; issues with hallucination of nongrounded output (Nie et al., 2019; Du\u0161ek et al., 2019; Reiter, 2018) ; the difficulty in surgically correcting any issues identified in trained neural models beyond additional training; as well as the difficulty of establishing what the 'worst case' performance of a neural model is.",
                "cite_spans": [
                    {
                        "start": 47,
                        "end": 61,
                        "text": "(Reiter, 2019)",
                        "ref_id": "BIBREF30"
                    },
                    {
                        "start": 112,
                        "end": 130,
                        "text": "(Nie et al., 2019;",
                        "ref_id": "BIBREF25"
                    },
                    {
                        "start": 131,
                        "end": 150,
                        "text": "Du\u0161ek et al., 2019;",
                        "ref_id": "BIBREF7"
                    },
                    {
                        "start": 151,
                        "end": 164,
                        "text": "Reiter, 2018)",
                        "ref_id": "BIBREF29"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "In addition, we believe that that while neural NLG methods are theoretically highly transferable, the practical transferability of neural NLG solutions to many news domains is limited by a lack of training data. While newsrooms have extensive archives of news text, these are rarely associated with the matching data that is the 'input' for each piece of news text (E.g., MacKov\u00e1 and Sido, 2020, pp. 43-44, Kanerva et al., 2019, p. 247) . At the same time, the non-trainable methods for NLG, too, suffer from difficulties in transferability and reusability (Linden, 2017) .",
                "cite_spans": [
                    {
                        "start": 372,
                        "end": 383,
                        "text": "MacKov\u00e1 and",
                        "ref_id": "BIBREF22"
                    },
                    {
                        "start": 384,
                        "end": 436,
                        "text": "Sido, 2020, pp. 43-44, Kanerva et al., 2019, p. 247)",
                        "ref_id": null
                    },
                    {
                        "start": 557,
                        "end": 571,
                        "text": "(Linden, 2017)",
                        "ref_id": "BIBREF21"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "In this work, we investigate document planning (selecting what content and in what order should appear in the document) for structured, statistical data-to-text NLG in the context of automated journalism targeting human journalists. We are not in search of a perfect method, but rather something that is relatively easy to implement as a subdomainindependent baseline and which can then be enhanced with domain-specific processing later-on. Such a method would make it easier to introduce automated journalism solutions to completely new subdomains within the larger statistical data domain.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "When queried for insight into news structure, journalists and academics often recite the concept of the \"(inverted) news pyramid\", where the news article is structured so that the order in which information appears in the text reflects the journalist's belief about the importance of the piece of information (Thomson et al., 2008) . While the precise origin of the structure is not clear (P\u00f6ttker, 2003) , it has become so prototypical that it is held selfevident in the journalistic trade literature: \"Every journalist knows how to write a traditional news text: start with the most important thing and continue until you have either said everything relevant or the space reserved for the story runs out\" (Sulopuisto, 2018, translated from Finnish).",
                "cite_spans": [
                    {
                        "start": 309,
                        "end": 331,
                        "text": "(Thomson et al., 2008)",
                        "ref_id": "BIBREF36"
                    },
                    {
                        "start": 389,
                        "end": 404,
                        "text": "(P\u00f6ttker, 2003)",
                        "ref_id": "BIBREF27"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "A more rigorous analysis of the structures employed in 'hard' news is presented by White (1997) , who argues that hard news articles have an 'orbital' structure consisting of a nucleus which represents the main point of the article and satellites that give context and additional information about the nucleus. White (1997) assigns the role of the nucleus to the combination of the headline and the lead paragraph of the article, and describes the subsequent paragraphs as the satellites. White (1997) identifies five possible relations between a satellite and the nucleus: elaboration, cause-and-effect, justification, contextualization and apprisal. Thomson et al. (2008) , in turn, identify that the satellites can elaborate, reiterate, describe causes or consequences, contextualize or provide additional assessment. An important observation is that -as indicated by 'orbital' -these satellites are relatively freely reorderable without affecting readability or meaning. Together, these two observations indicate that a good document plan for hard news (1) priorizes more newsworthy items and (2) contains some overarching theme (exemplified by the nucleus) so that the text as a whole is coherent, i.e. the satellites are in some way related to the nucleus.",
                "cite_spans": [
                    {
                        "start": 83,
                        "end": 95,
                        "text": "White (1997)",
                        "ref_id": "BIBREF37"
                    },
                    {
                        "start": 311,
                        "end": 323,
                        "text": "White (1997)",
                        "ref_id": "BIBREF37"
                    },
                    {
                        "start": 489,
                        "end": 501,
                        "text": "White (1997)",
                        "ref_id": "BIBREF37"
                    },
                    {
                        "start": 652,
                        "end": 673,
                        "text": "Thomson et al. (2008)",
                        "ref_id": "BIBREF36"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "The relations identified by White (1997) and Thomson et al. (2008) are highly similar to those identified in the more general Rhetorical Structure Theory (RST) (Mann and Thompson, 1988) , which uses similar nucleus-satellite terminology. However, whereas White (1997) and Thomson et al. (2008) analyze news text on the level of paragraphs, RST can be applied on a more fine-grained level to much shorter text spans. As RST shows that similar relations can be applied on a sub-paragraph level, we hypothesize that a reasonably approximation of a news article might be constructed by applying White's (1997) orbital theory also within paragraphs, by considering the first sentence of the paragraph a nucleus, and the others as satellites.",
                "cite_spans": [
                    {
                        "start": 28,
                        "end": 40,
                        "text": "White (1997)",
                        "ref_id": "BIBREF37"
                    },
                    {
                        "start": 45,
                        "end": 66,
                        "text": "Thomson et al. (2008)",
                        "ref_id": "BIBREF36"
                    },
                    {
                        "start": 160,
                        "end": 185,
                        "text": "(Mann and Thompson, 1988)",
                        "ref_id": "BIBREF23"
                    },
                    {
                        "start": 255,
                        "end": 267,
                        "text": "White (1997)",
                        "ref_id": "BIBREF37"
                    },
                    {
                        "start": 272,
                        "end": 293,
                        "text": "Thomson et al. (2008)",
                        "ref_id": "BIBREF36"
                    },
                    {
                        "start": 591,
                        "end": 605,
                        "text": "White's (1997)",
                        "ref_id": "BIBREF37"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "Importantly, we interpret the orbital theory of news structuring to suggest that -as the satellites are freely orderable -the actual type of relation is not as important for document planning as knowing that some relation exists between the satellite and the nucleus. We hypothesize that while identifying whether a specific (RST) relation exists between two arbitrary pieces of information requires domain knowledge, an approximation of whether two arbitrary pieces of information are related in some way could be obtained by inspecting their similarity in a domain-independent fashion.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "That is, we expect that a piece of information regarding the US health care funding in 2020 is more likely to be related in some way to a piece of information discussing the US health care funding in 2020 than to another piece of information discussing the health care funding in Sweden in 1978. If a heuristic or similarity measure identifying such relations could be identified, it could be used together with some estimate of newsworthiness to construct paragraph and document plans that seek to maximize both the key aspects identified above: newsworthiness and the relatedness of the content.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "As noted in the introduction, there is a distinction between the theoretical and the practical transferability of neural processing methods. We believe that a good baseline document planning and content selection approach should avoid the need for training data present in the many of recently proposed document planning and content selection approaches. This rules out as unsuitable most recent work that are based on learning from an aligned corpus of data and human-written texts, such as Angeli et al. (2010) , Konstas and Lapata (2013) , Wiseman et al. (2017) , Zhang et al. (2017) , Li and Wan (2018) , Dou et al. (2018) and Puduppully et al. (2019) .",
                "cite_spans": [
                    {
                        "start": 492,
                        "end": 512,
                        "text": "Angeli et al. (2010)",
                        "ref_id": "BIBREF0"
                    },
                    {
                        "start": 515,
                        "end": 540,
                        "text": "Konstas and Lapata (2013)",
                        "ref_id": "BIBREF17"
                    },
                    {
                        "start": 543,
                        "end": 564,
                        "text": "Wiseman et al. (2017)",
                        "ref_id": "BIBREF38"
                    },
                    {
                        "start": 567,
                        "end": 586,
                        "text": "Zhang et al. (2017)",
                        "ref_id": "BIBREF40"
                    },
                    {
                        "start": 589,
                        "end": 606,
                        "text": "Li and Wan (2018)",
                        "ref_id": "BIBREF20"
                    },
                    {
                        "start": 609,
                        "end": 626,
                        "text": "Dou et al. (2018)",
                        "ref_id": "BIBREF6"
                    },
                    {
                        "start": 631,
                        "end": 655,
                        "text": "Puduppully et al. (2019)",
                        "ref_id": "BIBREF26"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "Outside of these trainable approaches, to our knowledge, most other document planning approaches are based on 'hand-engineered' (Konstas and Lapata, 2013), domain-specific methods. A highly relevant survey of various document planning methods is presented by Gkatzia (2016) . While these previous works are -to at least some degree -domain-specific, they establish concepts and ideas that are highly relevant for our goal. Both Hallett et al. (2006) and Gatt et al. (2009) describe a core set of information, called 'summary spine' or 'key events', that they hold as more important than the rest of the available information. They, as well as Banaee et al. (2013) , also employ a numeric estimate of importance. Demir et al. (2010) identify that content already selected for inclusion in the document plan affects how well suited so-far unselected content is for inclusion. Sripada et al. (2003) identify Gricean maxims (Grice, 1975) as providing requirements for document planning and content selection.",
                "cite_spans": [
                    {
                        "start": 259,
                        "end": 273,
                        "text": "Gkatzia (2016)",
                        "ref_id": "BIBREF12"
                    },
                    {
                        "start": 428,
                        "end": 449,
                        "text": "Hallett et al. (2006)",
                        "ref_id": "BIBREF15"
                    },
                    {
                        "start": 454,
                        "end": 472,
                        "text": "Gatt et al. (2009)",
                        "ref_id": "BIBREF11"
                    },
                    {
                        "start": 643,
                        "end": 663,
                        "text": "Banaee et al. (2013)",
                        "ref_id": "BIBREF1"
                    },
                    {
                        "start": 712,
                        "end": 731,
                        "text": "Demir et al. (2010)",
                        "ref_id": "BIBREF3"
                    },
                    {
                        "start": 874,
                        "end": 895,
                        "text": "Sripada et al. (2003)",
                        "ref_id": "BIBREF33"
                    },
                    {
                        "start": 920,
                        "end": 933,
                        "text": "(Grice, 1975)",
                        "ref_id": "BIBREF14"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "Our work on document planning is done in the context of a series of data-to-text NLG applications producing short highlights of structured statistical data. Importantly, the applications are intended to be deployed in contexts where they must be able to produce texts highlighting between 10 and 30 data points from datasets measured in 100.000s of data points. The resulting texts are intended to both alert journalists to potential news and to provide them with a starting place from which to write the final news text.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Context",
                "sec_num": "3"
            },
            {
                "text": "Our system, adapted from Lepp\u00e4nen et al. (2017a) , is based on a pipeline of components with dedicated responsibilities similar to those described by Reiter and Dale (2000) and Reiter (2007) . For this work, the relevant part of the architecture is the Document Planner component. This component receives as input two sets of message data structures, an example of which is shown in Table 1 . 1 The messages are extracted automatically from tables of statistical data obtained from Eurostat.",
                "cite_spans": [
                    {
                        "start": 25,
                        "end": 48,
                        "text": "Lepp\u00e4nen et al. (2017a)",
                        "ref_id": "BIBREF18"
                    },
                    {
                        "start": 150,
                        "end": 172,
                        "text": "Reiter and Dale (2000)",
                        "ref_id": "BIBREF31"
                    },
                    {
                        "start": 177,
                        "end": 190,
                        "text": "Reiter (2007)",
                        "ref_id": "BIBREF28"
                    },
                    {
                        "start": 393,
                        "end": 394,
                        "text": "1",
                        "ref_id": null
                    }
                ],
                "ref_spans": [
                    {
                        "start": 383,
                        "end": 390,
                        "text": "Table 1",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Context",
                "sec_num": "3"
            },
            {
                "text": "The core set contains messages that are known to be highly relevant to the generation task. Unlike the \"summary spine' of Hallett et al. (2006) , the set is unlinked and unordered, and not all members of the set are guaranteed to be included in the document plan. The expanded set, contains messages that can be, but are not guaranteed to be, relevant for the document. Expressed using the terminology from Section 2, we assume that only messages in the core set can be nuclei, while messages from either set can be satellites.",
                "cite_spans": [
                    {
                        "start": 122,
                        "end": 143,
                        "text": "Hallett et al. (2006)",
                        "ref_id": "BIBREF15"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Context",
                "sec_num": "3"
            },
            {
                "text": "These core and expanded sets are determined automatically from user input. When requesting a new text, the user of the system must define a dataset the text is to be generated from, for example the consumer price data available from Eurostat. This dataset is then divided into the core set and the expanded set by the user when they select what country the generated text should focus on. For example, if the user were to select that the text should discuss French consumer prices, the core set would contain all data from the consumer price dataset that pertains directly to France, while the rest of the consumer price dataset (including data pertaining to the UK, Finland, Croatia, etc.) would be set as the expanded set.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Context",
                "sec_num": "3"
            },
            {
                "text": "We estimate each message's 'newsworthiness' using the Interquartile Range based method described by Lepp\u00e4nen et al. (2017b) with the values scaled to have mean 0 and standard deviation 1 for the purposes of this computation. The resulting value is conceptually similar to 'importance' of Gatt et al. (2009) and 'risk' of Banaee et al. (2013). The IQR based method compares each data point in turn to a larger distribution, giving it higher scores the further it is from the area between the first and the third quartile of the larger distribution. Values between the quartiles are given a minimal, uniform, score that is dependent on the shape of the distribution. In other words, higher IQR values indicate that the value is more of an outlier compared to the rest of related data in the dataset. As such, it captures a degree of 'unexpectedness', which is an important aspect of newsworthiness (Galtung and Ruge, 1965 ).",
                "cite_spans": [
                    {
                        "start": 100,
                        "end": 123,
                        "text": "Lepp\u00e4nen et al. (2017b)",
                        "ref_id": "BIBREF19"
                    },
                    {
                        "start": 288,
                        "end": 306,
                        "text": "Gatt et al. (2009)",
                        "ref_id": "BIBREF11"
                    },
                    {
                        "start": 896,
                        "end": 919,
                        "text": "(Galtung and Ruge, 1965",
                        "ref_id": "BIBREF10"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Context",
                "sec_num": "3"
            },
            {
                "text": "We do not use the domain-specific parts of the method described by Lepp\u00e4nen et al. (2017b) . That is, we make no value judgement of whether messages pertaining to French consumer prices are more newsworthy than messages pertaining to Croatian consumer prices, nor do we make judgements of whether changes in the price of education are more or less newsworthy than changes in the price of alcohol and tobacco. However, we do weight the scores so that messages with the timestamp field being closer to present receive higher weights, as recency is an important aspect of newsworthiness. While we have described our method for computing the newsworthiness value in some detail, we emphasize that for the rest of this article we only assume that the newsworthiness values are non-negative and that higher values indicate higher newsworthiness.",
                "cite_spans": [
                    {
                        "start": 67,
                        "end": 90,
                        "text": "Lepp\u00e4nen et al. (2017b)",
                        "ref_id": "BIBREF19"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Context",
                "sec_num": "3"
            },
            {
                "text": "More crucially for the method described be-low, we specify that the value type fields (which describe how the messages' values are to be interpreted) contain members of a hierarchical taxonomy of data types represented as colon-separated hierarchies of labels. For example, the value type field value health:cost:hc2:mio eur would indicate that the number in the value field is the amount of money (cost), measured in millions of euros (mio eur), spent by some nation (as defined by the location and location type fields) on rehabilitative care (hc2) in some time period (as defined by the timestamp and timestamp type fields) and that this is part of the larger health care topic (health). In our case, these labels are automatically established from the headers of the input data tables.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Context",
                "sec_num": "3"
            },
            {
                "text": "The goal of document structuring is to produce a three-level tree-structure with ordered children. The root node corresponds to the document as a whole and the mid-level structures correspond to paragraphs. The leaves are the messages selected for inclusion in the document. While the messages have not yet, at this stage, been associated with any linguistic structures, they can be conceptualized as being phrases or very short sentences. We are thus concurrently determining both the content and the structure the document.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Context",
                "sec_num": "3"
            },
            {
                "text": "We emphasize that our applications are employed in domains where they must be able to select some 10-30 messages from a pool of potential messages numbering in 100,000s. Given infinite computational resources, it would be preferential to construct all possible document plans and then score them in some fashion. This, however, is infeasible given the size of the search space. Previously, other authors have employed, for example, stochastic searches with significantly smaller search spaces (Mellish et al., 1998) . Indeed, some kind of a beam search approach could be very useful in smartly searching a subset of the search space. However, we have thus far been unable to identify a document-level metric that adequately balances the 'total amount of newsworthiness' in a text with the length of the text, a requirement for beam search.",
                "cite_spans": [
                    {
                        "start": 493,
                        "end": 515,
                        "text": "(Mellish et al., 1998)",
                        "ref_id": "BIBREF24"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Context",
                "sec_num": "3"
            },
            {
                "text": "Based on the above considerations, our main goal is to identify a widely applicable method for content selection and document planning that matches the following requirements: Again, we emphasize that our goal is not to identify a method that is optimal for any specific scenario, but rather to determine a baseline method that is adequate for a broad spectrum of applications and sub-domains.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Research Objective",
                "sec_num": "4"
            },
            {
                "text": "Optimally, we would wish to produce some sort of a globally optimal document plan. However, as discussed above, this would entail significant computational costs and require a scoring function applicable to the document as a whole. As such, we propose a method for producing document plans in a greedy, linear, and iterative fashion. At every stage, decisions are made considering only a limited local context, thus avoiding the need for a method of determining the global quality of the document plan, thus fulfilling REQ1 ('The method needs to be highly performant'). The document's overall theme, in our use case, is selected by the user who initiates the generation task. In initiating the task, the users selects both a dataset and a focus location. The generation process then derives the core messages and expanded messages sets (the inputs to the Document Planner, see Section 3) so that both sets discuss the dataset",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "A Baseline Approach to Document Planning",
                "sec_num": "5"
            },
            {
                "text": "Example value where What location the fact relates to Finland where type",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Description",
                "sec_num": null
            },
            {
                "text": "What the type of the location is country timestamp",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Description",
                "sec_num": null
            },
            {
                "text": "The time (or time range) the fact relates to 2020M05 timestamp type The type of the timestamp month value A (usually) numeric value 0.01 value type",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Description",
                "sec_num": null
            },
            {
                "text": "Interpretation of value cphi:hicp2015:cp-hi02:rt01 newsworthiness An estimate of how newsworthy the message is 1 Table 1 : An example of a message. The hypothetical message states that in the fifth month of 2020, in Finland, the consumer price index, using the year 2015 as the start of the index, of alcoholic beverages and tobacco changed by 0.01 points with respect to the value of the index during the previous month. indicated by the user (i.e. messages from other datasets are not generated) and that the core set contains messages pertaining to the user's indicated focus location, while messages pertaining to all other locations are in the expanded set. This fulfills REQ3 ('The document should have a theme'). This step is also independent of the specific subdomain, thus fulfilling REQ2 ('The method should not be dependent on domain knowledge'). This step thus fulfills all the relevant requirements. Next, we'll describe how both the first and subsequent paragraphs can be planned in a way consistent with the requirements defined above.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 113,
                        "end": 120,
                        "text": "Table 1",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Description",
                "sec_num": null
            },
            {
                "text": "At the start of the document planning process, we select the most newsworthy message from the core messages set to act as the nucleus (n 1 ) of the first paragraph (p 1 ). This nucleus establishes the theme of the first paragraph as follows: We inspect the value type field of this first nucleus n 1 , and retrieve a prefix Prefix(n 1 ). The prefix is the least amount of colon-separated labels wherein the total amount of prefixes in the core set is greater than the minimal amount of paragraphs a document can have, in our case two. In our case, as a consequence of our label hierarchy, this is always the first three colon-separated units. For the message shown in Table 1 , the prefix would thus be cphi:hicp2015:cp-hi02, meaning that the first paragraph's theme would be the prices of alcoholic beverages and tobacco. This fulfills REQ5, 'the paragraphs should have distinct themes related to the document theme' for the first paragraph.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 668,
                        "end": 675,
                        "text": "Table 1",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Planning the First Paragraph",
                "sec_num": "5.1"
            },
            {
                "text": "Next, the first paragraph is completed with satellites from the union of the core messages and the expanded messages sets. These satellites are initially filtered so that only messages that have the same prefix as the nucleus n i are considered in paragraph p i to fulfill REQ8 ('All messages should relate to the paragraph theme'). The satellites are then selected in a linear, greedy, and iterative manner to fulfill REQ1.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Planning the First Paragraph",
                "sec_num": "5.1"
            },
            {
                "text": "For selecting the k'th satellite to a partially constructed paragraph already containing k \u2212 1 satellites and one nucleus, we consider both the newsworthiness of the available messages (REQ9), as well as how well they would fit the already constructed segment (REQ8). Observing only the newsworthiness would produce a highly incoherent narrative, whereas focusing only on the narrative risks leaving out highly important information.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Planning the First Paragraph",
                "sec_num": "5.1"
            },
            {
                "text": "Following the reasoning in Section 2, we assume that two subsequent messages are more likely to form a good narrative if they are similar. As such, we need a method for weighing the message's newsworthiness by the similarity of the message to the last message of the under-construction paragraph, thus balancing the requirements of REQ8 and REQ9. In terms of the message objects described in Table 1 , it seems to us that the intuitive aspects of similarity are related to the degree of similarity within the 'meta' fields such as timestamp, location and value type.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 392,
                        "end": 399,
                        "text": "Table 1",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Planning the First Paragraph",
                "sec_num": "5.1"
            },
            {
                "text": "For the timestamp and location fields, we can state that two messages that have identical values in the fields are more similar that two messages that are otherwise the same but have distinct values for said fields. We call this the contextual similarity of the messages, and the fields the contextual fields (F c ), as these fields provide us access to the larger context in which the value and value type fields can be interpreted. Contextual similarity captures the notion that it is likely better to follow a fact about French healthcare spending in 2020 with another piece of information about France in 2020, rather than about Austria in 1990.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Planning the First Paragraph",
                "sec_num": "5.1"
            },
            {
                "text": "In more precise terms, we propose the following weighing scheme for contextual similarity: The similarity sim c (A, B) of two messaged A and B is the product of weights w f > 1 for each field f among the contextual fields F c , where both A and B have the same value for the field:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Planning the First Paragraph",
                "sec_num": "5.1"
            },
            {
                "text": "sim c (A, B) = {f \u2208Fc|A.f =B.f } w f (1)",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Planning the First Paragraph",
                "sec_num": "5.1"
            },
            {
                "text": "This value strictly increases as more fields are shared between A and B. We explicitly define the similarity to be zero if there are no fields f where A and B share a value. If w f is a uniform value for all fields f , this scheme is completely domain-agnostic. Setting different weights w f for each field f \u2208 F c allows for encoding some domain knowledge about which fields are the most important for the text, thus providing a method for producing more tailored texts at the cost of slightly violating REQ2. In our case study, we set w timestamp = 1.1 and w location = 1.5. The above consideration of similarity still ignores valuable information available from the value type field, which describes how the value in the value field is to be interpreted. Denoting health:cost:hc2:mio eur (the cost of rehabilitative care in millions of euros) by T 1 , consider its similarity to T 2 = health:cost:hc2:eur hab, the cost of rehabilitative care as euros per inhabitant, and T 3 = health:cost:hc41:mio eur, the cost of health care related imaging services in millions of euros. Intuitively, T 1 and T 2 are thematically closer than T 1 and T 3 . We model this similarity between two facts A and B simply as",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Planning the First Paragraph",
                "sec_num": "5.1"
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "sim t (A, B) = 1 s(A, B)",
                        "eq_num": "(2)"
                    }
                ],
                "section": "Planning the First Paragraph",
                "sec_num": "5.1"
            },
            {
                "text": "where s (A, B) is the length -in colon-separated units -of the unshared suffix between A and B's value type fields. That is, s(T 1 , T 2 ) = 1 whereas s(T 1 , T 3 ) = 2. We specify that sim",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 8,
                        "end": 14,
                        "text": "(A, B)",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Planning the First Paragraph",
                "sec_num": "5.1"
            },
            {
                "text": "t (\u2022, \u2022)",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Planning the First Paragraph",
                "sec_num": "5.1"
            },
            {
                "text": "is zero for all pairs without any shared prefix. Our formulation of sim t (\u2022, \u2022) was influenced by the observation that in our context the messages' value type values have a constant number of colon-separated segments. In cases where the lengths of the value type values differ, an alternative formulation of",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Planning the First Paragraph",
                "sec_num": "5.1"
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "sim t (A, B) = 2p(A, B) (A) + (B)",
                        "eq_num": "(3)"
                    }
                ],
                "section": "Planning the First Paragraph",
                "sec_num": "5.1"
            },
            {
                "text": "where (\u2022) provides the length of the value type value, and p(\u2022, \u2022) is the length of shared prefix between A and B, both measured as colon-separated units, might be preferable if also more complex. When considering whether the k'th satellite s k i of paragraph p i should be a specific candidate c \u2208 C, where C is all so far unused messages, we can combine the similarity metrics with the newsworthiness of c into a general fitness value as follows:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Planning the First Paragraph",
                "sec_num": "5.1"
            },
            {
                "text": "fit(c, x) = c.newsworthiness \u00d7 sim c (c, x) \u00d7 sim t (c, x) \u00d7 set penalty(c)",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Planning the First Paragraph",
                "sec_num": "5.1"
            },
            {
                "text": "The set penalty(c) factor depends on whether the message originates from the core messages set, or the extended messages set. For messages originating from the core message set, the penalty is 1. For messages originating from the extended messages set, the penalty is 1 dist+1 , where dist is the distance from the previous core message.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Planning the First Paragraph",
                "sec_num": "5.1"
            },
            {
                "text": "The final score describing how good of an addition c would be as the kth satellite of the ith paragraph s k i is then obtained by taking the average of fitnesses of c in relation to both the nucleus n i and the previous satellite s k\u22121 i by computing:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Planning the First Paragraph",
                "sec_num": "5.1"
            },
            {
                "text": "score(c, n i , s k\u22121 i ) = fit(c, n i ) + fit(c, s k\u22121 i ) 2",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Planning the First Paragraph",
                "sec_num": "5.1"
            },
            {
                "text": "This maximizes the newsworthiness of the paragraph's contents (fulfilling REQ9, 'all messages should be newsworthy'), while also enforcing relatedness to the theme of the paragraph (fulfilling REQ8, 'all messages should relate to the paragraph theme') by measuring against the nucleus and with the inclusion of the set penalty. By continuously measuring against the previously selected satellite, the procedure also allows for interludes to e.g. discuss highly newsworthy information related to but not strictly about the paragraph's main topic, or 'thematic drift'. It thus fulfills REQ10 ('Within each paragraph, the messages should be presented in an order that produces a coherent narrative') while also paying attention to the pyramid model of news (See Section 2).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Planning the First Paragraph",
                "sec_num": "5.1"
            },
            {
                "text": "Using score, the highest scoring candidate c top = arg max c\u2208C score(c, n i , s k\u22121 i ) is then compared to both an absolute threshold t abs and the newsworthiness of the nucleus n i multiplied by relative threshold value t rel . Provided that the maximal paragraph length has not been reached, the top candidate message c top is appended to the paragraph p i as the k'th satellite s k i in the document plan provided that either score(c",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Planning the First Paragraph",
                "sec_num": "5.1"
            },
            {
                "text": "top , n i , s k\u22121 i ) \u2265 t abs or score(c top , n i , s k\u22121 i ) \u2265 t rel \u00d7 n i .newsworthiness.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Planning the First Paragraph",
                "sec_num": "5.1"
            },
            {
                "text": "These thresholds ensure that the paragraph does not stray into minutiae, whether considered in absolute terms or in relation to the nucleus of the paragraph. In cases where the minimum paragraph length has not been reached, the thresholds are ignored and the top candidate is always appended. This accounts for REQ7 ('The paragraphs should not be excessively long or short').",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Planning the First Paragraph",
                "sec_num": "5.1"
            },
            {
                "text": "The above considerations take into account several free parameters, namely the maximal and minimal paragraph lengths as well as the threshold values t rel and t abs . In our case study, we selected the minimal and maximal paragraph lengths as 2 and 5 messages empirically by trialing out various values and observing the resulting texts. These should, naturally, be based on the genre of text and the target audience. For the threshold values we selected 0.2 and 0.5, respectively, using the same method as with the paragraph lengths above. Both the thresholds and the minimal and maximal paragraph lengths should be viewed as (manually) tuneable hyperparameters.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Planning the First Paragraph",
                "sec_num": "5.1"
            },
            {
                "text": "We then proceed to generate further paragraphs in a manner highly similar to that used when planning the first paragraph. The only distinction is that, when selecting the nucleus n i for a subsequent paragraph p i , we obtain the message from the core messages set with a highest newsworthiness value that has a prefix (theme) not yet discussed among the previously planned paragraphs p 1 -p i\u22121 :",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Planning Subsequent Paragraphs",
                "sec_num": "5.2"
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "n i = arg max c\u2208C c.newsworthiness",
                        "eq_num": "(4)"
                    }
                ],
                "section": "Planning Subsequent Paragraphs",
                "sec_num": "5.2"
            },
            {
                "text": "where",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Planning Subsequent Paragraphs",
                "sec_num": "5.2"
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "C = c \u2208 CoreM essages|Prefix(c) \u2208 {Prefix(n k )|k \u2208 [1..i \u2212 1]}",
                        "eq_num": "(5)"
                    }
                ],
                "section": "Planning Subsequent Paragraphs",
                "sec_num": "5.2"
            },
            {
                "text": "This ensures that the different paragraphs are highly newsworthy, thus fulfilling REQ6, while also fulfilling REQ5 for having distinct themes for the different paragraphs.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Planning Subsequent Paragraphs",
                "sec_num": "5.2"
            },
            {
                "text": "As when constructing the subsequent paragraphs, the total length of the document also needs to be considered. To fulfill REQ4 ('The document should have multiple paragraphs but not be excessively long'), we employ a variation of the method described in the previous section for ending individual paragraphs. A maximal length (in our case, 3 paragraphs) ensures that the document is not allowed to grow beyond reason, whereas a minimal length (for us, 2 paragraphs) ensures that the document is not unreasonably short. After the minimal length has been reached (but not yet the maximal length), a new paragraph is only started if the nucleus of the potential paragraph has a newsworthiness value that is at least 30 % of the newsworthiness value of the first nucleus of the document. This, as with the satellites, ensures that the the document does not stray into minutiae, balancing REQs 4 and 6. the maximal and minimal lengths, as well as the 30 % threshold, were determined by manual fine-tuning and should be viewed as tuneable hyperparameters.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Planning Subsequent Paragraphs",
                "sec_num": "5.2"
            },
            {
                "text": "The method described above was implemented in a larger NLG application producing news alerts for journalists from datasets provided by Eurostat. A variation of the same application was also developed with a simplified document planner. In this simplified planner, the planner always selects the maximally newsworthy available message as the message without any early stopping threshold. Nuclei are selected from the core messages set, while satellites can be from either set. Contrasting our proposed method with this simplified method enables us to evaluate the importance of narrative coherence in the generated texts. The larger application is multilingual, but the evaluation was conducted using English language texts.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Evaluation",
                "sec_num": "6"
            },
            {
                "text": "Three experts were recruited from the Finnish News Agency STT, a national European news agency, to evaluate documents on the consumer price indices in five different European nations. For all nations, the judges were shown variants produced by both our proposed method and the simplified method. One of the selected countries is the country the news agency is based in, with the assumption that the judges would have high amounts of world knowledge they would be able to use in evaluating these texts. Another variant pair describes a country that is both relatively small and geographically remote (but still within EU), with the assumption that the journalists are unlikely to",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Evaluation",
                "sec_num": "6"
            },
            {
                "text": "In June 2020, in Estonia, the monthly growth rate of the harmonized consumer price index for the category 'education' was 30.8 points. It was 30.7 percentage points more than the EU average. In July 2020, it was 0.4 percentage points less than the EU average. It was -0.4 points. In May 2020, the yearly growth rate of the harmonized consumer price index for the category 'education' was -20.5 points. It was 21.9 percentage points less than the EU average.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Consumer Prices in Estonia",
                "sec_num": null
            },
            {
                "text": "In August 2020, the monthly growth rate of the harmonized consumer price index for the category 'housing, water, electricity, gas and other fuels' was 2.5 points. It was 2.3 percentage points more than the EU average. In North Macedonia, it was 3 percentage points more than the EU average. It was 3.2 points. Estonia had the 3rd highest monthly growth rate of the harmonized consumer price index for the category 'housing, water, electricity, gas and other fuels' across the observed countries. In Sweden, the monthly growth rate of the harmonized consumer price index for the category 'housing, water, electricity, gas and other fuels' was 3.1 points. have much world knowledge about this country's consumer prices. The three other countries were selected from among those bordering the first country, with the assumption that the journalists would have some, but not much, world knowledge relating to these countries. The final output texts were not inspected prior to selecting the countries.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Consumer Prices in Estonia",
                "sec_num": null
            },
            {
                "text": "All of the texts used in the evaluation were generated from a copy of the same underlying Eurostat dataset, entitled 'Harmonised index of consumer prices -monthly data [ei cphi m]' 2 downloaded in September 2020. It contains country-level data regarding the harmonized consumer prices indices, and their change over time, for various EU nations starting from January 1996. We preprocess the data by adding monthly rankings (i.e. determine what country had the greatest, the second greatest, etc. value for a specific index category during any specific month) and comparisons to the EU average values.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Consumer Prices in Estonia",
                "sec_num": null
            },
            {
                "text": "As the evaluation was focused on document planning and content selection, the larger system was simplified in some respects, e.g., to not conduct complex aggregation. This was done to minimize the effect of later stages of the generation process on the evaluation. As a result, the language in the evaluated documents was relatively stilted, as exemplified by Figure 1 . The only manual alteration was the addition of headings to indicate the texts' intended themes.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 360,
                        "end": 368,
                        "text": "Figure 1",
                        "ref_id": "FIGREF0"
                    }
                ],
                "eq_spans": [],
                "section": "Consumer Prices in Estonia",
                "sec_num": null
            },
            {
                "text": "The judges did not receive any direct compensation but their employer, the news agency, is a member of the EU-wide EMBEDDIA research project within which parts of this work was conducted. The evaluations were conducted online. The judges were first provided with some basic information on the type of documents they were to read (i.e. that the texts are intended to be news alerts for journalists, rather than publication ready news texts), the length of the task, etc. All instructions were in the judges' native language, in this case Finnish. The judges were not told which texts were produced by which variants nor how many variants were being tested. Following this, the judges were shown the documents one by one. For each document, the judges were asked to indicate their agreement with the following statements (translated from Finnish): Q1: The text matches the heading Q2: The text is coherent Q3: The text lacks some pertinent information Q4: The text contains unnecessary information Q5: The text has a suitable length For Q1-Q4, the judges indicated their agreement on a 7-point Likert scale ranging from 1 ('completely disagree') to 7 ('completely agree'). For Q5, the answers were provided on 5-point scale ranging from 1 ('clearly too short') to 3 ('length is suitable') to 5 ('clearly too long'). In addition, the judges were able to provide textual feedback for each individual text, as well as for the evaluation task as a whole. The judges' answers to Q1 -Q5, are aggregated in Table 2 .",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 1498,
                        "end": 1505,
                        "text": "Table 2",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Consumer Prices in Estonia",
                "sec_num": null
            },
            {
                "text": "The results indicate that the proposed method statistically significantly increases the document's coherence (Q2, mean 4.33 vs. 1.60, median 5 vs 2), the matching of the document's content to the document's theme (Q1, mean 4.40 vs. 1.80, median 5 vs 2), and produces documents of more suitable length (Q5, mean 2.93 vs. 4.07, median 3 vs 4, with 3 being best). The proposed method also seems Our method",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Consumer Prices in Estonia",
                "sec_num": null
            },
            {
                "text": "Median Mean SD. Median Mean SD. p MWU",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Statement",
                "sec_num": null
            },
            {
                "text": "(1-7, \u2191) 5 4.40 1.64 2 1.80 0.41 < 0.001* Q2 (1-7, \u2191) 5 4.33 1.76 2 1.60 0.51 < 0.001* Q3 (1-7, \u2193) 4 4.47 1.81 6 5.80 1.42 0.049 Q4 (1-7, \u2193) 5 5.13 1.55 6 6.33 0.62 0.024 Q5 (1-5, 3 best) 3 2.93 0.59 4 4.07 0.70 < 0.001* Table 2 : Results obtained during the evaluation. Parentheses indicate answer ranges and whether the higher (\u2191), lower (\u2193) or middle values are to be interpreted as the best. The p M W U column contains the (uncorrected) p-value of a two-sided Mann-Whitney U test. An asterisk indicates the p-value is statistically significant also after applying a Bonferroni correction to account for multiple tests.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 221,
                        "end": 228,
                        "text": "Table 2",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Q1",
                "sec_num": null
            },
            {
                "text": "to result in less unnecessary information being included in the document (Q4, mean 5.13 vs 6.33, median 5 vs 6), and in the text missing less necessary information (Q3, mean 4.47 vs 5.80, median 4 vs 6), but these effects are not statistically significant after correcting for multiple comparisons with the Bonferroni correction. We hypothesize this difference would become significant in a larger-scale evaluation.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Q1",
                "sec_num": null
            },
            {
                "text": "The free-form textual feedback provided by the judges, as expected, indicates that the texts could be further improved. For example, in the case of the text shown in Figure 1 , the judges called for a sentence explicitly noting that North Macedonia had the highest monthly growth rate. In addition, they noted it might be better to produce distinct, even shorter, texts as 'news alerts' while reserving the evaluated texts for use as a starting point when the journalist starts writing.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 166,
                        "end": 174,
                        "text": "Figure 1",
                        "ref_id": "FIGREF0"
                    }
                ],
                "eq_spans": [],
                "section": "Q1",
                "sec_num": null
            },
            {
                "text": "In this work, we have identified a need for, and proposed, a widely applicable baseline document planning method for generating journalistic texts from statistical datasets. Our method is based on observations on the similarities between the orbital theory of news structure (White, 1997) and Rhetorical Structure Theory (Mann and Thompson, 1988) . While our proposed method is likely to fall short of the performance of subdomain-specific planning methods, results indicate that it achieves adequate performance while fulfilling a set of requirements identified based on the larger application domain of news generation.",
                "cite_spans": [
                    {
                        "start": 275,
                        "end": 288,
                        "text": "(White, 1997)",
                        "ref_id": "BIBREF37"
                    },
                    {
                        "start": 293,
                        "end": 346,
                        "text": "Rhetorical Structure Theory (Mann and Thompson, 1988)",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Conclusions",
                "sec_num": "7"
            },
            {
                "text": "The concrete implementation details are somewhat more complex. We omit details irrelevant for this work.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "Available for download and browsing from http://appsso.eurostat.ec.europa.eu/ nui/show.do?dataset=ei_cphi_m",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            }
        ],
        "back_matter": [],
        "bib_entries": {
            "BIBREF0": {
                "ref_id": "b0",
                "title": "A simple domain-independent probabilistic approach to generation",
                "authors": [
                    {
                        "first": "Gabor",
                        "middle": [],
                        "last": "Angeli",
                        "suffix": ""
                    },
                    {
                        "first": "Percy",
                        "middle": [],
                        "last": "Liang",
                        "suffix": ""
                    },
                    {
                        "first": "Dan",
                        "middle": [],
                        "last": "Klein",
                        "suffix": ""
                    }
                ],
                "year": 2010,
                "venue": "Proceedings of the 2010 Conference on Empirical Methods in Natural Language Processing",
                "volume": "",
                "issue": "",
                "pages": "502--512",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Gabor Angeli, Percy Liang, and Dan Klein. 2010. A simple domain-independent probabilistic approach to generation. In Proceedings of the 2010 Confer- ence on Empirical Methods in Natural Language Processing, pages 502-512.",
                "links": null
            },
            "BIBREF1": {
                "ref_id": "b1",
                "title": "Towards NLG for physiological data monitoring with body area networks",
                "authors": [
                    {
                        "first": "",
                        "middle": [],
                        "last": "Hadi Banaee",
                        "suffix": ""
                    },
                    {
                        "first": "Amy",
                        "middle": [],
                        "last": "Mobyen Uddin Ahmed",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Loutfi",
                        "suffix": ""
                    }
                ],
                "year": 2013,
                "venue": "14th European Workshop on Natural Language Generation",
                "volume": "",
                "issue": "",
                "pages": "193--197",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Hadi Banaee, Mobyen Uddin Ahmed, and Amy Loutfi. 2013. Towards NLG for physiological data moni- toring with body area networks. In 14th European Workshop on Natural Language Generation, Sofia, Bulgaria, August 8-9, 2013, pages 193-197.",
                "links": null
            },
            "BIBREF2": {
                "ref_id": "b2",
                "title": "Automated journalism 2.0: Event-driven narratives: From simple descriptions to real stories",
                "authors": [
                    {
                        "first": "David",
                        "middle": [],
                        "last": "Caswell",
                        "suffix": ""
                    },
                    {
                        "first": "Konstantin",
                        "middle": [],
                        "last": "D\u00f6rr",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "Journalism practice",
                "volume": "12",
                "issue": "4",
                "pages": "477--496",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "David Caswell and Konstantin D\u00f6rr. 2018. Automated journalism 2.0: Event-driven narratives: From sim- ple descriptions to real stories. Journalism practice, 12(4):477-496.",
                "links": null
            },
            "BIBREF3": {
                "ref_id": "b3",
                "title": "A discourse-aware graph-based contentselection framework",
                "authors": [
                    {
                        "first": "Seniz",
                        "middle": [],
                        "last": "Demir",
                        "suffix": ""
                    },
                    {
                        "first": "Sandra",
                        "middle": [],
                        "last": "Carberry",
                        "suffix": ""
                    },
                    {
                        "first": "Kathleen",
                        "middle": [
                            "F"
                        ],
                        "last": "Mc-Coy",
                        "suffix": ""
                    }
                ],
                "year": 2010,
                "venue": "Proceedings of the 6th International Natural Language Generation Conference",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Seniz Demir, Sandra Carberry, and Kathleen F. Mc- Coy. 2010. A discourse-aware graph-based content- selection framework. In Proceedings of the 6th In- ternational Natural Language Generation Confer- ence.",
                "links": null
            },
            "BIBREF4": {
                "ref_id": "b4",
                "title": "Why news automation fails",
                "authors": [
                    {
                        "first": "Laurence",
                        "middle": [],
                        "last": "Dierickx",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "Computation+ Journalism Symposium",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Laurence Dierickx. 2019. Why news automation fails. In Computation+ Journalism Symposium, Miami, FL.",
                "links": null
            },
            "BIBREF5": {
                "ref_id": "b5",
                "title": "Mapping the field of algorithmic journalism",
                "authors": [
                    {
                        "first": "Konstantin",
                        "middle": [],
                        "last": "Nicholas D\u00f6rr",
                        "suffix": ""
                    }
                ],
                "year": 2015,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Konstantin Nicholas D\u00f6rr. 2015. Mapping the field of algorithmic journalism. Digital journalism.",
                "links": null
            },
            "BIBREF6": {
                "ref_id": "b6",
                "title": "Data2text studio: Automated text generation from structured data",
                "authors": [
                    {
                        "first": "Longxu",
                        "middle": [],
                        "last": "Dou",
                        "suffix": ""
                    },
                    {
                        "first": "Guanghui",
                        "middle": [],
                        "last": "Qin",
                        "suffix": ""
                    },
                    {
                        "first": "Jinpeng",
                        "middle": [],
                        "last": "Wang",
                        "suffix": ""
                    },
                    {
                        "first": "Jin-Ge",
                        "middle": [],
                        "last": "Yao",
                        "suffix": ""
                    },
                    {
                        "first": "Chin-Yew",
                        "middle": [],
                        "last": "Lin",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "Proc. 2018 Conference on Empirical Methods in Natural Language Processing",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Longxu Dou, Guanghui Qin, Jinpeng Wang, Jin-Ge Yao, and Chin-Yew Lin. 2018. Data2text studio: Automated text generation from structured data. In Proc. 2018 Conference on Empirical Methods in Natural Language Processing.",
                "links": null
            },
            "BIBREF7": {
                "ref_id": "b7",
                "title": "Semantic noise matters for neural natural language generation",
                "authors": [
                    {
                        "first": "Ond\u0159ej",
                        "middle": [],
                        "last": "Du\u0161ek",
                        "suffix": ""
                    },
                    {
                        "first": "M",
                        "middle": [],
                        "last": "David",
                        "suffix": ""
                    },
                    {
                        "first": "Verena",
                        "middle": [],
                        "last": "Howcroft",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "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.",
                "links": null
            },
            "BIBREF8": {
                "ref_id": "b8",
                "title": "Putting Europe's robots on the map: automated journalism in news agencies",
                "authors": [
                    {
                        "first": "Alexander",
                        "middle": [],
                        "last": "Fanta",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "Reuters Institute Fellowship Paper",
                "volume": "",
                "issue": "",
                "pages": "2017--2026",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Alexander Fanta. 2017. Putting Europe's robots on the map: automated journalism in news agencies. Reuters Institute Fellowship Paper, pages 2017-09.",
                "links": null
            },
            "BIBREF9": {
                "ref_id": "b9",
                "title": "Neural data-to-text generation: A comparison between pipeline and end-to-end architectures",
                "authors": [
                    {
                        "first": "Chris",
                        "middle": [],
                        "last": "Thiago Castro Ferreira",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Van Der Lee",
                        "suffix": ""
                    },
                    {
                        "first": "Emiel",
                        "middle": [],
                        "last": "Emiel Van Miltenburg",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Krahmer",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:1908.09022"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Thiago Castro Ferreira, Chris van der Lee, Emiel van Miltenburg, and Emiel Krahmer. 2019. Neu- ral data-to-text generation: A comparison be- tween pipeline and end-to-end architectures. arXiv preprint arXiv:1908.09022.",
                "links": null
            },
            "BIBREF10": {
                "ref_id": "b10",
                "title": "The structure of foreign news: The presentation of the Congo, Cuba and Cyprus crises in four Norwegian newspapers",
                "authors": [
                    {
                        "first": "Johan",
                        "middle": [],
                        "last": "Galtung",
                        "suffix": ""
                    },
                    {
                        "first": "Mari",
                        "middle": [],
                        "last": "Holmboe Ruge",
                        "suffix": ""
                    }
                ],
                "year": 1965,
                "venue": "Journal of peace research",
                "volume": "2",
                "issue": "1",
                "pages": "64--90",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Johan Galtung and Mari Holmboe Ruge. 1965. The structure of foreign news: The presentation of the Congo, Cuba and Cyprus crises in four Norwegian newspapers. Journal of peace research, 2(1):64-90.",
                "links": null
            },
            "BIBREF11": {
                "ref_id": "b11",
                "title": "From data to text in the neonatal intensive care unit: Using NLG technology for decision support and information management",
                "authors": [
                    {
                        "first": "Albert",
                        "middle": [],
                        "last": "Gatt",
                        "suffix": ""
                    },
                    {
                        "first": "Francois",
                        "middle": [],
                        "last": "Portet",
                        "suffix": ""
                    },
                    {
                        "first": "Ehud",
                        "middle": [],
                        "last": "Reiter",
                        "suffix": ""
                    },
                    {
                        "first": "Jim",
                        "middle": [],
                        "last": "Hunter",
                        "suffix": ""
                    },
                    {
                        "first": "Saad",
                        "middle": [],
                        "last": "Mahamood",
                        "suffix": ""
                    },
                    {
                        "first": "Wendy",
                        "middle": [],
                        "last": "Moncur",
                        "suffix": ""
                    },
                    {
                        "first": "Somayajulu",
                        "middle": [],
                        "last": "Sripada",
                        "suffix": ""
                    }
                ],
                "year": 2009,
                "venue": "Ai Communications",
                "volume": "22",
                "issue": "3",
                "pages": "153--186",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Albert Gatt, Francois Portet, Ehud Reiter, Jim Hunter, Saad Mahamood, Wendy Moncur, and Somayajulu Sripada. 2009. From data to text in the neonatal in- tensive care unit: Using NLG technology for deci- sion support and information management. Ai Com- munications, 22(3):153-186.",
                "links": null
            },
            "BIBREF12": {
                "ref_id": "b12",
                "title": "Content selection in data-totext systems: A survey",
                "authors": [
                    {
                        "first": "Dimitra",
                        "middle": [],
                        "last": "Gkatzia",
                        "suffix": ""
                    }
                ],
                "year": 2016,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Dimitra Gkatzia. 2016. Content selection in data-to- text systems: A survey. arXiv preprint. Available at https://arxiv.org/abs/1610.08375.",
                "links": null
            },
            "BIBREF13": {
                "ref_id": "b13",
                "title": "Guide to automated journalism",
                "authors": [
                    {
                        "first": "Andreas",
                        "middle": [],
                        "last": "Graefe",
                        "suffix": ""
                    }
                ],
                "year": 2016,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Andreas Graefe. 2016. Guide to automated journalism.",
                "links": null
            },
            "BIBREF14": {
                "ref_id": "b14",
                "title": "Logic and conversation",
                "authors": [
                    {
                        "first": "P",
                        "middle": [],
                        "last": "Herbert",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Grice",
                        "suffix": ""
                    }
                ],
                "year": 1975,
                "venue": "Speech acts",
                "volume": "",
                "issue": "",
                "pages": "41--58",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Herbert P Grice. 1975. Logic and conversation. In Speech acts, pages 41-58. Brill.",
                "links": null
            },
            "BIBREF15": {
                "ref_id": "b15",
                "title": "Summarisation and visualisation of e-health data repositories",
                "authors": [
                    {
                        "first": "Catalina",
                        "middle": [],
                        "last": "Hallett",
                        "suffix": ""
                    },
                    {
                        "first": "Richard",
                        "middle": [],
                        "last": "Power",
                        "suffix": ""
                    },
                    {
                        "first": "Donia",
                        "middle": [],
                        "last": "Scott",
                        "suffix": ""
                    }
                ],
                "year": 2006,
                "venue": "UK E-Science All-Hands Meeting",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Catalina Hallett, Richard Power, and Donia Scott. 2006. Summarisation and visualisation of e-health data repositories. In UK E-Science All-Hands Meeting.",
                "links": null
            },
            "BIBREF16": {
                "ref_id": "b16",
                "title": "Template-free data-to-text generation of Finnish sports news",
                "authors": [
                    {
                        "first": "Jenna",
                        "middle": [],
                        "last": "Kanerva",
                        "suffix": ""
                    },
                    {
                        "first": "Samuel",
                        "middle": [],
                        "last": "R\u00f6nnqvist",
                        "suffix": ""
                    },
                    {
                        "first": "Riina",
                        "middle": [],
                        "last": "Kekki",
                        "suffix": ""
                    },
                    {
                        "first": "Tapio",
                        "middle": [],
                        "last": "Salakoski",
                        "suffix": ""
                    },
                    {
                        "first": "Filip",
                        "middle": [],
                        "last": "Ginter",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "Proceedings of the 22nd Nordic Conference on Computational Linguistics",
                "volume": "",
                "issue": "",
                "pages": "242--252",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Jenna Kanerva, Samuel R\u00f6nnqvist, Riina Kekki, Tapio Salakoski, and Filip Ginter. 2019. Template-free data-to-text generation of Finnish sports news. In Proceedings of the 22nd Nordic Conference on Com- putational Linguistics, pages 242-252, Turku, Fin- land. Link\u00f6ping University Electronic Press.",
                "links": null
            },
            "BIBREF17": {
                "ref_id": "b17",
                "title": "Inducing document plans for concept-to-text generation",
                "authors": [
                    {
                        "first": "Ioannis",
                        "middle": [],
                        "last": "Konstas",
                        "suffix": ""
                    },
                    {
                        "first": "Mirella",
                        "middle": [],
                        "last": "Lapata",
                        "suffix": ""
                    }
                ],
                "year": 2013,
                "venue": "Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing",
                "volume": "",
                "issue": "",
                "pages": "1503--1514",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Ioannis Konstas and Mirella Lapata. 2013. Induc- ing document plans for concept-to-text generation. In Proceedings of the 2013 Conference on Empiri- cal Methods in Natural Language Processing, pages 1503-1514.",
                "links": null
            },
            "BIBREF18": {
                "ref_id": "b18",
                "title": "Data-driven news generation for automated journalism",
                "authors": [
                    {
                        "first": "Leo",
                        "middle": [],
                        "last": "Lepp\u00e4nen",
                        "suffix": ""
                    },
                    {
                        "first": "Myriam",
                        "middle": [],
                        "last": "Munezero",
                        "suffix": ""
                    },
                    {
                        "first": "Mark",
                        "middle": [],
                        "last": "Granroth-Wilding",
                        "suffix": ""
                    },
                    {
                        "first": "Hannu",
                        "middle": [],
                        "last": "Toivonen",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "Proceedings of the 10th International Conference on Natural Language Generation",
                "volume": "",
                "issue": "",
                "pages": "188--197",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Leo Lepp\u00e4nen, Myriam Munezero, Mark Granroth- Wilding, and Hannu Toivonen. 2017a. Data-driven news generation for automated journalism. In Pro- ceedings of the 10th International Conference on Natural Language Generation, pages 188-197.",
                "links": null
            },
            "BIBREF19": {
                "ref_id": "b19",
                "title": "Finding and expressing news from structured data",
                "authors": [
                    {
                        "first": "Leo",
                        "middle": [],
                        "last": "Lepp\u00e4nen",
                        "suffix": ""
                    },
                    {
                        "first": "Myriam",
                        "middle": [],
                        "last": "Munezero",
                        "suffix": ""
                    },
                    {
                        "first": "Stefanie",
                        "middle": [],
                        "last": "Sir\u00e9n-Heikel",
                        "suffix": ""
                    },
                    {
                        "first": "Mark",
                        "middle": [],
                        "last": "Granroth-Wilding",
                        "suffix": ""
                    },
                    {
                        "first": "Hannu",
                        "middle": [],
                        "last": "Toivonen",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "Proceedings of the 21st International Academic Mindtrek Conference",
                "volume": "",
                "issue": "",
                "pages": "174--183",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Leo Lepp\u00e4nen, Myriam Munezero, Stefanie Sir\u00e9n- Heikel, Mark Granroth-Wilding, and Hannu Toivo- nen. 2017b. Finding and expressing news from structured data. In Proceedings of the 21st Interna- tional Academic Mindtrek Conference, pages 174- 183. ACM.",
                "links": null
            },
            "BIBREF20": {
                "ref_id": "b20",
                "title": "Point precisely: Towards ensuring the precision of data in generated texts using delayed copy mechanism",
                "authors": [
                    {
                        "first": "Liunian",
                        "middle": [],
                        "last": "Li",
                        "suffix": ""
                    },
                    {
                        "first": "Xiaojun",
                        "middle": [],
                        "last": "Wan",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "Proceedings of the 27th International Conference on Computational Linguistics",
                "volume": "",
                "issue": "",
                "pages": "1044--1055",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Liunian Li and Xiaojun Wan. 2018. Point precisely: Towards ensuring the precision of data in generated texts using delayed copy mechanism. In Proceed- ings of the 27th International Conference on Com- putational Linguistics, pages 1044-1055, Santa Fe, New Mexico, USA. ACL.",
                "links": null
            },
            "BIBREF21": {
                "ref_id": "b21",
                "title": "Decades of Automation in the Newsroom: Why are there still so many jobs in journalism? Digital",
                "authors": [
                    {
                        "first": "Carl-Gustav",
                        "middle": [],
                        "last": "Linden",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "Journalism",
                "volume": "5",
                "issue": "2",
                "pages": "123--140",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Carl-Gustav Linden. 2017. Decades of Automation in the Newsroom: Why are there still so many jobs in journalism? Digital Journalism, 5(2):123-140.",
                "links": null
            },
            "BIBREF22": {
                "ref_id": "b22",
                "title": "The robotic reporter in the Czech News Agency: Automated journalism and augmentation in the newsroom. Communication Today",
                "authors": [
                    {
                        "first": "Veronika",
                        "middle": [],
                        "last": "Mackov\u00e1",
                        "suffix": ""
                    },
                    {
                        "first": "Jakub",
                        "middle": [],
                        "last": "Sido",
                        "suffix": ""
                    }
                ],
                "year": 2020,
                "venue": "",
                "volume": "11",
                "issue": "",
                "pages": "36--53",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Veronika MacKov\u00e1 and Jakub Sido. 2020. The robotic reporter in the Czech News Agency: Au- tomated journalism and augmentation in the news- room. Communication Today, 11(1):36-53.",
                "links": null
            },
            "BIBREF23": {
                "ref_id": "b23",
                "title": "Rhetorical structure theory: Toward a functional theory of text organization. Text-Interdisciplinary Journal for the Study of Discourse",
                "authors": [
                    {
                        "first": "C",
                        "middle": [],
                        "last": "William",
                        "suffix": ""
                    },
                    {
                        "first": "Sandra",
                        "middle": [
                            "A"
                        ],
                        "last": "Mann",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Thompson",
                        "suffix": ""
                    }
                ],
                "year": 1988,
                "venue": "",
                "volume": "8",
                "issue": "",
                "pages": "243--281",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "William C Mann and Sandra A Thompson. 1988. Rhetorical structure theory: Toward a functional the- ory of text organization. Text-Interdisciplinary Jour- nal for the Study of Discourse, 8(3):243-281.",
                "links": null
            },
            "BIBREF24": {
                "ref_id": "b24",
                "title": "Experiments using stochastic search for text planning",
                "authors": [
                    {
                        "first": "Chris",
                        "middle": [],
                        "last": "Mellish",
                        "suffix": ""
                    },
                    {
                        "first": "Alistair",
                        "middle": [],
                        "last": "Knott",
                        "suffix": ""
                    },
                    {
                        "first": "Jon",
                        "middle": [],
                        "last": "Oberlander",
                        "suffix": ""
                    },
                    {
                        "first": "Mick O'",
                        "middle": [],
                        "last": "Donnell",
                        "suffix": ""
                    }
                ],
                "year": 1998,
                "venue": "Natural Language Generation",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Chris Mellish, Alistair Knott, Jon Oberlander, and Mick O'Donnell. 1998. Experiments using stochas- tic search for text planning. In Natural Language Generation.",
                "links": null
            },
            "BIBREF25": {
                "ref_id": "b25",
                "title": "A simple recipe towards reducing hallucination in neural surface realisation",
                "authors": [
                    {
                        "first": "Feng",
                        "middle": [],
                        "last": "Nie",
                        "suffix": ""
                    },
                    {
                        "first": "Jin-Ge",
                        "middle": [],
                        "last": "Yao",
                        "suffix": ""
                    },
                    {
                        "first": "Jinpeng",
                        "middle": [],
                        "last": "Wang",
                        "suffix": ""
                    },
                    {
                        "first": "Rong",
                        "middle": [],
                        "last": "Pan",
                        "suffix": ""
                    },
                    {
                        "first": "Chin-Yew",
                        "middle": [],
                        "last": "Lin",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics",
                "volume": "",
                "issue": "",
                "pages": "2673--2679",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Feng Nie, Jin-Ge Yao, Jinpeng Wang, Rong Pan, and Chin-Yew Lin. 2019. A simple recipe towards re- ducing hallucination in neural surface realisation. In Proceedings of the 57th Annual Meeting of the Asso- ciation for Computational Linguistics, pages 2673- 2679.",
                "links": null
            },
            "BIBREF26": {
                "ref_id": "b26",
                "title": "Data-to-text generation with content selection and planning",
                "authors": [
                    {
                        "first": "Ratish",
                        "middle": [],
                        "last": "Puduppully",
                        "suffix": ""
                    },
                    {
                        "first": "Li",
                        "middle": [],
                        "last": "Dong",
                        "suffix": ""
                    },
                    {
                        "first": "Mirella",
                        "middle": [],
                        "last": "Lapata",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "Proc. 33rd AAAI Conference on Artificial Intelligence",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Ratish Puduppully, Li Dong, and Mirella Lapata. 2019. Data-to-text generation with content selection and planning. In Proc. 33rd AAAI Conference on Arti- ficial Intelligence.",
                "links": null
            },
            "BIBREF27": {
                "ref_id": "b27",
                "title": "News and its communicative quality: The inverted pyramid-when and why did it appear?",
                "authors": [
                    {
                        "first": "Horst",
                        "middle": [],
                        "last": "P\u00f6ttker",
                        "suffix": ""
                    }
                ],
                "year": 2003,
                "venue": "Journalism Studies",
                "volume": "4",
                "issue": "4",
                "pages": "501--511",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Horst P\u00f6ttker. 2003. News and its communicative qual- ity: The inverted pyramid-when and why did it ap- pear? Journalism Studies, 4(4):501-511.",
                "links": null
            },
            "BIBREF28": {
                "ref_id": "b28",
                "title": "An architecture for data-to-text systems",
                "authors": [
                    {
                        "first": "Ehud",
                        "middle": [],
                        "last": "Reiter",
                        "suffix": ""
                    }
                ],
                "year": 2007,
                "venue": "Proceedings of the Eleventh European Workshop on Natural Language Generation",
                "volume": "",
                "issue": "",
                "pages": "97--104",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Ehud Reiter. 2007. An architecture for data-to-text systems. In Proceedings of the Eleventh European Workshop on Natural Language Generation, pages 97-104. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF29": {
                "ref_id": "b29",
                "title": "Hallucination in neural NLG",
                "authors": [
                    {
                        "first": "Ehud",
                        "middle": [],
                        "last": "Reiter",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "2020--2023",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Ehud Reiter. 2018. Hallucination in neural NLG. https://ehudreiter.com/2018/11/12/ hallucination-in-neural-nlg/. Accessed: 2020-03-02.",
                "links": null
            },
            "BIBREF30": {
                "ref_id": "b30",
                "title": "ML is used more if it does not limit control",
                "authors": [
                    {
                        "first": "Ehud",
                        "middle": [],
                        "last": "Reiter",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "2020--2027",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Ehud Reiter. 2019. ML is used more if it does not limit control. https://ehudreiter.com/2019/ 08/15/ml-limits-control/. Accessed: 2020- 07-25.",
                "links": null
            },
            "BIBREF31": {
                "ref_id": "b31",
                "title": "Building natural language generation systems",
                "authors": [
                    {
                        "first": "Ehud",
                        "middle": [],
                        "last": "Reiter",
                        "suffix": ""
                    },
                    {
                        "first": "Robert",
                        "middle": [],
                        "last": "Dale",
                        "suffix": ""
                    }
                ],
                "year": 2000,
                "venue": "Studies in Natural Language Processing",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Ehud Reiter and Robert Dale. 2000. Building natu- ral language generation systems. Studies in Natural Language Processing. Cambridge University Press.",
                "links": null
            },
            "BIBREF32": {
                "ref_id": "b32",
                "title": "Unboxing news automation: Exploring imagined affordances of automation in news journalism",
                "authors": [
                    {
                        "first": "Stefanie",
                        "middle": [],
                        "last": "Sir\u00e9n-Heikel",
                        "suffix": ""
                    },
                    {
                        "first": "Leo",
                        "middle": [],
                        "last": "Lepp\u00e4nen",
                        "suffix": ""
                    },
                    {
                        "first": "Carl-Gustav",
                        "middle": [],
                        "last": "Lind\u00e9n",
                        "suffix": ""
                    },
                    {
                        "first": "Asta",
                        "middle": [],
                        "last": "B\u00e4ck",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "Nordic Journal of Media Studies",
                "volume": "1",
                "issue": "1",
                "pages": "47--66",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Stefanie Sir\u00e9n-Heikel, Leo Lepp\u00e4nen, Carl-Gustav Lind\u00e9n, and Asta B\u00e4ck. 2019. Unboxing news au- tomation: Exploring imagined affordances of au- tomation in news journalism. Nordic Journal of Me- dia Studies, 1(1):47-66.",
                "links": null
            },
            "BIBREF33": {
                "ref_id": "b33",
                "title": "Generating English summaries of time series data using the Gricean maxims",
                "authors": [
                    {
                        "first": "G",
                        "middle": [],
                        "last": "Somayajulu",
                        "suffix": ""
                    },
                    {
                        "first": "Ehud",
                        "middle": [],
                        "last": "Sripada",
                        "suffix": ""
                    },
                    {
                        "first": "Jim",
                        "middle": [],
                        "last": "Reiter",
                        "suffix": ""
                    },
                    {
                        "first": "Jin",
                        "middle": [],
                        "last": "Hunter",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Yu",
                        "suffix": ""
                    }
                ],
                "year": 2003,
                "venue": "Proceedings of the ninth ACM SIGKDD international conference on Knowledge discovery and data mining",
                "volume": "",
                "issue": "",
                "pages": "187--196",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Somayajulu G Sripada, Ehud Reiter, Jim Hunter, and Jin Yu. 2003. Generating English summaries of time series data using the Gricean maxims. In Proceed- ings of the ninth ACM SIGKDD international con- ference on Knowledge discovery and data mining, pages 187-196.",
                "links": null
            },
            "BIBREF34": {
                "ref_id": "b34",
                "title": "Uutisia kortti kerrallaan",
                "authors": [
                    {
                        "first": "Olli",
                        "middle": [],
                        "last": "Sulopuisto",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Olli Sulopuisto. 2018. Uutisia kortti kerrallaan.",
                "links": null
            },
            "BIBREF36": {
                "ref_id": "b36",
                "title": "Objectivity\" and \"hard news\" reporting across cultures: Comparing the news report in English, French, Japanese and Indonesian journalism",
                "authors": [
                    {
                        "first": "A",
                        "middle": [],
                        "last": "Elizabeth",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Thomson",
                        "suffix": ""
                    },
                    {
                        "first": "R",
                        "middle": [
                            "R"
                        ],
                        "last": "Peter",
                        "suffix": ""
                    },
                    {
                        "first": "Philip",
                        "middle": [],
                        "last": "White",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Kitley",
                        "suffix": ""
                    }
                ],
                "year": 2008,
                "venue": "Journalism studies",
                "volume": "9",
                "issue": "2",
                "pages": "212--228",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Elizabeth A Thomson, Peter RR White, and Philip Kit- ley. 2008. \"Objectivity\" and \"hard news\" reporting across cultures: Comparing the news report in En- glish, French, Japanese and Indonesian journalism. Journalism studies, 9(2):212-228.",
                "links": null
            },
            "BIBREF37": {
                "ref_id": "b37",
                "title": "Death, disruption and the moral order: the narrative impulse in mass-media 'hard news' reporting. Genres and institutions: Social processes in the workplace and school",
                "authors": [
                    {
                        "first": "Peter",
                        "middle": [],
                        "last": "White",
                        "suffix": ""
                    }
                ],
                "year": 1997,
                "venue": "",
                "volume": "101",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Peter White. 1997. Death, disruption and the moral or- der: the narrative impulse in mass-media 'hard news' reporting. Genres and institutions: Social processes in the workplace and school, 101:133.",
                "links": null
            },
            "BIBREF38": {
                "ref_id": "b38",
                "title": "Challenges in data-to-document generation",
                "authors": [
                    {
                        "first": "Sam",
                        "middle": [],
                        "last": "Wiseman",
                        "suffix": ""
                    },
                    {
                        "first": "Stuart",
                        "middle": [],
                        "last": "Shieber",
                        "suffix": ""
                    },
                    {
                        "first": "Alexander",
                        "middle": [],
                        "last": "Rush",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "Proc. 2017 Conference on Empirical Methods in Natural Language Processing",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Sam Wiseman, Stuart Shieber, and Alexander Rush. 2017. Challenges in data-to-document generation. In Proc. 2017 Conference on Empirical Methods in Natural Language Processing.",
                "links": null
            },
            "BIBREF39": {
                "ref_id": "b39",
                "title": "Avoin voitto",
                "authors": [
                    {
                        "first": "Yleisradio",
                        "middle": [],
                        "last": "",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Yleisradio. 2018. Avoin voitto. https://github. com/Yleisradio/avoin-voitto.",
                "links": null
            },
            "BIBREF40": {
                "ref_id": "b40",
                "title": "Positionaware attention and supervised data improve slot filling",
                "authors": [
                    {
                        "first": "Yuhao",
                        "middle": [],
                        "last": "Zhang",
                        "suffix": ""
                    },
                    {
                        "first": "Victor",
                        "middle": [],
                        "last": "Zhong",
                        "suffix": ""
                    },
                    {
                        "first": "Danqi",
                        "middle": [],
                        "last": "Chen",
                        "suffix": ""
                    },
                    {
                        "first": "Gabor",
                        "middle": [],
                        "last": "Angeli",
                        "suffix": ""
                    },
                    {
                        "first": "Christopher",
                        "middle": [
                            "D"
                        ],
                        "last": "Manning",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing",
                "volume": "",
                "issue": "",
                "pages": "35--45",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Yuhao Zhang, Victor Zhong, Danqi Chen, Gabor An- geli, and Christopher D. Manning. 2017. Position- aware attention and supervised data improve slot filling. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 35-45, Copenhagen, Denmark. ACL.",
                "links": null
            }
        },
        "ref_entries": {
            "FIGREF0": {
                "type_str": "figure",
                "text": "Example output regarding Eurostat statistics on consumer prices. The text contains 12 messages, selected from among 207,210 messages available during generation.",
                "num": null,
                "uris": null
            }
        }
    }
}