The Impact of Rule-Based Text Generation on the Quality of Abstractive Summaries

Proceedings of Recent Advances in Natural Language Processing, pages 1275–1284, Varna, Bulgaria, Sep 2–4, 2019. https://doi.org/10.26615/978-954-452-056-4_146 1275 The Impact of Rule-Based Text Generation on the Quality of Abstractive Summaries Tatiana Vodolazova, Elena Lloret Dept. of Software and Computing Systems University of Alicante Apdo. de Correos 99 E-03080, Alicante, Spain {tvodolazova,elloret}@dlsi.ua.es Abstract In this paper we describe how an abstrac- tive text summarization method improved the informativeness of automatic sum- maries by integrating syntactic text simpli- fication, subject-verb-object concept fre- quency scoring and a set of rules that transform text into its semantic represen- tation. We analyzed the impact of each component of our approach on the qual- ity of generated summaries and tested it on DUC 2002 dataset. Our experi- ments showed that our approach outper- formed other state-of-the-art abstractive methods while maintaining acceptable lin- guistic quality and redundancy rate. 1 Introduction Rapid growth of digital information increases the need for automatic text summarization methods that can digest large amounts of textual data, such as scientific articles, blogs and news articles to extract concise and relevant information from them. Text summarization methods can be classi- fied into abstractive and extractive ones (Nenkova and McKeown, 2012). Extractive methods com- pose summaries from the most salient sentences of the original document. In contrast, abstractive methods generate novel or partially novel text us- ing such techniques as sentence compression, fu- sion, calculation of path scores in graphs or, natu- ral language generation tools such as SimpleNLG (Gupta and Gupta, 2018). They involve an inter- mediate step of deep linguistic analysis and an ab- stract semantic representation of the data. Extrac- tive techniques have been intensively researched for over half a century and, according to some studies, “have more or less achieved their peak performance” (Mehta, 2016). Over the past few years interest in the field of text summarization has shifted towards abstrac- tive methods and quickly produced a large variety of approaches. Gupta and Gupta (2018) classify them broadly into methods based on the structure, semantics and deep learning with neural networks. The main advantage of semantic-based ap- proaches over deep learning ones lies in their in- dependence from a large training corpus. Most of the available datasets for deep learning belong to the domain of news text that further restricts the application of these methods to other domains. However, semantic-based approaches rely on a parser to transform text to its semantic represen- tation and, therefore, a poor parser performance will reduce the quality of generated summaries. Another limitation of the deep learning methods comes from the fact that they rely on statistical co- occurrence of words and are prone to semantic and grammatical errors. This is something that a reli- able parser could help to avoid. Structure-based methods, such as template and ontology based ones reveal other weaknesses. Template-based methods lack diversity. At the same time, ontology based ones rely on a time- consuming task of creating an ontology by a hu- man expert. However, they provide highly co- herent summaries and …

Content truncated to 500 words