Combining Thai EDUs: Principle and Implementation

Combining Thai EDUs: Principle and Implementation Chanatip Saetia Supawat Taerungruang Tawunrat Chalothorn Kasikorn Labs Kasikorn Business Technology Group (KBTG) Nontaburi, Thailand chanatip.sae@kbtg.tech, supawat.t@kbtg.tech, tawunrat.c@kbtg.tech Abstract Due to the lack of explicit end-of-sentence marker in Thai, Elementary Discourse Units (EDUs) are usually preferred over sentences as the basic linguistic units for processing Thai language. However, some segmented EDUs lack of structural or se- mantic information. To obtain a well-form unit, which represents a complete idea, this paper proposes combining EDUs with rhe- torical relations selected depending on our proposed syntactic and semantic criteria. The combined EDUs can be then used without considering other parts of the text. Moreover, we also annotated data with the criteria. After that, we trained a deep learn- ing model inspired by coreference resolu- tion and dependency parsing models. As a result, our model achieves the F1 score of 82.72%. 1 Introduction Generally, sentences have served as the basic linguistic units required by many tasks in NLP (e.g., text summarization and question answer- ing) for processing long bodies of text (Mihalcea, 2004; Raiman & Miller, 2017; Van Lierde & Chow, 2019). Basic linguistic units must contain complete propositional content that represents a single piece of idea. However, in Thai language, the sentence cannot clearly speci- fy the boundary (Intasaw & Aroonmanakun, 2013, p. 491; Lertpiya et al., 2018) since there is no explicit end-of-sentence marker like a period in English. For this reason, prior works (Ketui et al., 2015; Singkul et al., 2019; Sinthupoun & Sornil, 2010; Sukvaree et al., 2004) have suggested to use the Elementary Discourse Unit (EDU), which is the basic unit in discourse based on Rhetorical Structure Theory (RST) (Mann & Thompson, 1988; Taboada & Mann, 2006), as a processing unit for Thai language. However, a single EDU, on its own, may not contain com- plete information to be understandable. General- ly, EDUs are segmented based solely on syntac- tic criteria (Carlson & Marcu, 2001; Intasaw & Aroonmanakun, 2013; Ketui et al., 2012) and used alongside RST relation (which contains semantic and structural information) to represent complete discourse structure. Figure 1. EDU that is separated due to embed- ded clause To highlight that the EDU segmentation based on syntactic criteria creates individual EDUs with incomplete in structure and meaning, an example is shown in Figure 1. The figure il- lustrates RST d-tree (Morey et al., 2018) of EDU, which is separated by an embedded clause. A matrix clause [ผู้ชายไปท างานเมื่อเช้านี้] ‘A man went to work this morning’, which should be one unit of EDU, is separated into two units, [ผู้ชาย] ‘A man’ and [ไปท างานเมื่อเช้านี้] ‘went to work this morn- ing’, because there is embedded clause [ที่เป็นทนาย] ‘who is a lawyer’ modifying a noun [ผู้ชาย] ‘A man’ in matrix clause. As such, combining incomplete EDUs is necessary for obtaining a complete idea that is to represent through a well-formed EDU. This paper states the syntactic and semantic criteria for considering rhetorical relations used to combine EDUs into a well-formed EDU, which represents a …

Content truncated to 500 words