{ "paper_id": "2020", "header": { "generated_with": "S2ORC 1.0.0", "date_generated": "2023-01-19T07:28:43.597317Z" }, "title": "DaMata: A Robot-Journalist Covering the Brazilian Amazon Deforestation", "authors": [ { "first": "Andr\u00e9", "middle": [], "last": "Luiz", "suffix": "", "affiliation": {}, "email": "" }, { "first": "Rosa", "middle": [], "last": "Teixeira", "suffix": "", "affiliation": { "laboratory": "Laboratory for Experimentation in Translation", "institution": "Federal University of Minas Gerais", "location": {} }, "email": "" }, { "first": "Jo\u00e3o", "middle": [], "last": "Gabriel", "suffix": "", "affiliation": {}, "email": "" }, { "first": "Moura", "middle": [], "last": "Campos", "suffix": "", "affiliation": { "laboratory": "", "institution": "University of S\u00e3o", "location": { "settlement": "Paulo" } }, "email": "" }, { "first": "Rossana", "middle": [], "last": "Cunha", "suffix": "", "affiliation": { "laboratory": "Laboratory for Experimentation in Translation", "institution": "Federal University of Minas Gerais", "location": {} }, "email": "rossanacunha@ufmg.br" }, { "first": "Thiago", "middle": [ "Castro" ], "last": "Ferreira", "suffix": "", "affiliation": { "laboratory": "Laboratory for Experimentation in Translation", "institution": "Federal University of Minas Gerais", "location": {} }, "email": "thiagocf05@ufmg.br" }, { "first": "Adriana", "middle": [ "Silvina" ], "last": "Pagano", "suffix": "", "affiliation": { "laboratory": "Laboratory for Experimentation in Translation", "institution": "Federal University of Minas Gerais", "location": {} }, "email": "apagano@ufmg.br" }, { "first": "Fabio", "middle": [ "Gagliardi" ], "last": "Cozman", "suffix": "", "affiliation": { "laboratory": "", "institution": "University of S\u00e3o", "location": { "settlement": "Paulo" } }, "email": "fgcozman@usp.br" } ], "year": "", "venue": null, "identifiers": {}, "abstract": "This demo paper introduces DaMata, a robotjournalist covering deforestation in the Brazilian Amazon. The robot-journalist is based on a pipeline architecture of Natural Language Generation, which yields multilingual daily and monthly reports based on the public data provided by DETER, a real-time deforestation satellite monitor developed and maintained by the Brazilian National Institute for Space Research (INPE). DaMata automatically generates reports in Brazilian Portuguese and English and publishes them on the Twitter platform. Corpus and code are publicly available. 1", "pdf_parse": { "paper_id": "2020", "_pdf_hash": "", "abstract": [ { "text": "This demo paper introduces DaMata, a robotjournalist covering deforestation in the Brazilian Amazon. The robot-journalist is based on a pipeline architecture of Natural Language Generation, which yields multilingual daily and monthly reports based on the public data provided by DETER, a real-time deforestation satellite monitor developed and maintained by the Brazilian National Institute for Space Research (INPE). DaMata automatically generates reports in Brazilian Portuguese and English and publishes them on the Twitter platform. Corpus and code are publicly available. 1", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Abstract", "sec_num": null } ], "body_text": [ { "text": "Robot-Journalism is one of the most promising Natural Language Generation (NLG) applications thanks to the high volume of structured datastreams available nowadays, which enables automated systems to report recurrent information with high-fidelity and low latency. Such automation can unburden journalists from tedious data reporting tasks, thus enabling human agents to devote efforts to more investigative coverage (Graefe, 2016) .", "cite_spans": [ { "start": 417, "end": 431, "text": "(Graefe, 2016)", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "The health of the Amazon rainforest has drawn global news coverage attention lately, when soaring deforestation raised public awareness worldwide (Escobar, 2020) . Accurate information about deforested land in the Brazilian part of this territory is yielded by the country's National Institute for Space Research (INPE) 2 through several satellites and other monitoring systems. One of the most effective is DETER, a real-time deforestation warning surveillance system (Diniz et al., 2015) .", "cite_spans": [ { "start": 146, "end": 161, "text": "(Escobar, 2020)", "ref_id": "BIBREF1" }, { "start": 469, "end": 489, "text": "(Diniz et al., 2015)", "ref_id": "BIBREF0" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "Despite being publicly available 3 , DETER's deforestation data output is in graphical and numeri-1 https://github.com/BotsDoBem/DEMO_ INPE_COVID 2 http://www.inpe.br/ 3 http://terrabrasilis.dpi.inpe.br cal style, which renders it less accessible to general audiences and demands coverage by human journalists specialized in the domain. To address this issue, we introduce DaMata, a robot-journalist based on an NLG pipeline architecture (Reiter and Dale, 2000; Gatt and Krahmer, 2018) , which generates daily and monthly reports about deforestation in the Brazilian Amazon, both in Brazilian Portuguese 4 and English 5 , and publishes them on Twitter.", "cite_spans": [ { "start": 438, "end": 461, "text": "(Reiter and Dale, 2000;", "ref_id": "BIBREF7" }, { "start": 462, "end": 485, "text": "Gatt and Krahmer, 2018)", "ref_id": "BIBREF3" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "Unlike novel end-to-end systems, which may hallucinate content (Moryossef et al., 2019) and may be problematic for sensitive domains, DaMata follows the pipeline architecture proposed by Ferreira et al. (2019), which converts non-linguistic data into text in 6 steps: Content Selection, Discourse Ordering, Text Structuring, Lexicalization, Referring Expression Generation and Textual Realization. We have added a seventh step to DaMata -publication -responsible for sharing the generated news on Twitter. This kind of architecture, depicted in Figure 1 , allows for trustworthy output as well as easier access and maintenance of the sub-modules and easier multi-domain and multi-language applications. The grammar used by the model was built by first running the content selection step in previous data, generating 14 non-linguistic monthly reports and 25 daily ones. These non-linguistic reports were then manually verbalized and the input and output representations for each pipeline module were manually annotated. This process resulted in a list of possible discourse orders, text structures, lexicalizations, and referring expressions. When deployed, each module draws on the selected combination of templates using rule-based approaches.", "cite_spans": [ { "start": 63, "end": 87, "text": "(Moryossef et al., 2019)", "ref_id": "BIBREF6" } ], "ref_spans": [ { "start": 545, "end": 553, "text": "Figure 1", "ref_id": "FIGREF0" } ], "eq_spans": [], "section": "System Overview", "sec_num": "2" }, { "text": "The next sections describe each module. Content Selection This module is responsible for selecting relevant messages to be verbalized in the reports. Satellite data are gathered, processed, and interpreted by INPE, which provides this information through DETER API 6 . These data come in a structured format, a JSON file comprising pre-defined features, e.g., deforestation area, cause, date, city, state, etc., which are mapped to DaMata's database. The information is structured as intent-attribute-value messages:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "System Overview", "sec_num": "2" }, { "text": "DAILY ALERT(area=\"5.91\",city=\"None\",uc= \"CHICO MENDES EXTRACTIVIST RESERVE\",day= \"25\",month=\"3\",year=\"2019\",state=\"Acre\", location=\"amazon\",daily accumulation=\"1\")", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "System Overview", "sec_num": "2" }, { "text": "We adopt a rule-based approach to content selection. The system checks for monthly and daily updates on deforestation. For a monthly report, DaMata always verbalizes news on the amount of total deforestation in the Brazilian Amazon and the main cause for deforestation in the corresponding time period. Next, the system verbalizes month and annual variation in case of an increase in these values compared to the previous period. Finally, it reports news on the state, city, and nature protected area with the highest deforestation rates in the respective month, if the corresponding deforested area was greater than the established threshold (e.g., the sum of mean value and standard deviation of the deforestation areas time series).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "System Overview", "sec_num": "2" }, { "text": "For a daily report, our robot-journalist selects data on the amount of deforested area and the main cause of deforestation in a given city or nature protected area as well as the total area hitherto deforested in the respective month for the given region. The following is an example of the content selection module output: Discourse Ordering According to Jurafsky and Martin (2019) , ordering events in discourse to construe a logical timeline enhances clarity and reader comprehension. Based on the list of possible intent orderings collected from our corpus (for instance 8 ordering templates for daily reports and 10 for monthly reports), we arrange our discourse messages by a rule-based approach. By using discourse ordering templates based on human generated texts, we ensure conveying the relative importance of each piece of information, as it is usually done in journalistic texts. For example, for the messages selected for a daily report -daily deforested area, main cause of deforestation and total deforestation in the month, a likely outcome order would be:", "cite_spans": [ { "start": 356, "end": 382, "text": "Jurafsky and Martin (2019)", "ref_id": "BIBREF5" } ], "ref_spans": [], "eq_spans": [], "section": "System Overview", "sec_num": "2" }, { "text": "DAILY ALERT \u2192 CAUSE \u2192 TOTAL DEFORESTATION", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "System Overview", "sec_num": "2" }, { "text": "Text Structuring This step is responsible for structuring the information in sentences and paragraphs (Gatt and Krahmer, 2018; Ferreira et al., 2019) , bearing in mind the character limitation on Twitter (280 characters per tweet) and how intents should be ordered. As in the previous phase, our robot-journalist draws on the annotated corpus structures and also decides which structuring template to use by a rule-based approach. For our daily report example, this step would return:", "cite_spans": [ { "start": 102, "end": 126, "text": "(Gatt and Krahmer, 2018;", "ref_id": "BIBREF3" }, { "start": 127, "end": 149, "text": "Ferreira et al., 2019)", "ref_id": "BIBREF2" } ], "ref_spans": [], "eq_spans": [], "section": "System Overview", "sec_num": "2" }, { "text": "
DAILY ALERT CAUSE
TOTAL DEFORESTATION