Summarizing Legal Rulings: Comparative Experiments

Proceedings of Recent Advances in Natural Language Processing, pages 313–322, Varna, Bulgaria, Sep 2–4, 2019. https://doi.org/10.26615/978-954-452-056-4_036 313 Summarizing Legal Rulings: Comparative Experiments Diego de Vargas Feijo Institute of Informatics Federal University of Rio Grande do Sul http://www.inf.ufrgs.br/ dvfeijo@inf.ufrgs.br Viviane Pereira Moreira Institute of Informatics Federal University of Rio Grande do Sul http://www.inf.ufrgs.br/ viviane@inf.ufrgs.br Abstract In the context of text summarization, texts in the legal domain have peculiarities re- lated to their length and to their special- ized vocabulary. Recent neural network- based approaches can achieve high-quality scores for text summarization. However, these approaches have been used mostly for generating very short abstracts for news articles. Thus, their applicability to the legal domain remains an open is- sue. In this work, we experimented with ten extractive and four abstractive models in a real dataset of legal rulings. These models were compared with an extractive baseline based on heuristics to select the most relevant parts of the text. Our re- sults show that abstractive approaches sig- nificantly outperform extractive methods in terms of ROUGE scores. 1 Introduction Text summarization is the task of producing a con- densed representation of an input text, keeping the most relevant information. A summary should be concise, fluent, and contains paraphrased versions of the input text with a reduced length. There are two approaches to text summariza- tion. The first, known as extractive, works by se- lecting entire sentences directly from the source text. This has been the most widely used solution for a number of years (Luhn, 1958; Edmundson, 1969; Erkan and Radev, 2004a; Mihalcea and Ta- rau, 2004). Extractive methods typically work by simply (i) scoring phrases or sentences to deter- mine the most relevant; and (ii) selecting the top scoring sentences to compose the summary. Of- ten, the sentences are arranged in the same order of occurrence in the original text in an attempt to preserve the ideas and meanings of the sentences. The scoring function should capture how well the selected sentences represent the text and cover the topics present in the text. The lack of connectives may cause the impression that the generated sum- mary does not have a logical flow. The second approach, known as abstractive, aims to extract the main concepts or ideas from the text and generate a new condensed version differ- ent from the original. In this case, the model must learn how to write sentences in a logical flow. Us- ing this approach, it is possible to paraphrase the original and use words that did not occur in the source. This is much more similar to the way a human would create a summary. Most recent re- search has focused on this approach. With the increasing availability of data, the need for summarization is felt in many areas. This is especially true in the legal area as the texts are usually lengthy. Law operators are expected to keep updated with important information rang- ing from news, jurisprudence changes, and rulings from many courts. This important information amounts …

Content truncated to 500 words