Discourse Element Identification in Student Essays based on Global and Local Cohesion

Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 2255–2261, Lisbon, Portugal, 17-21 September 2015. c⃝2015 Association for Computational Linguistics. Discourse Element Identification in Student Essays based on Global and Local Cohesion Wei Song†, Ruiji Fu‡, Lizhen Liu†, Ting Liu§ †Information Engineering, Capital Normal University, Bejing 100048, China ‡Iflytek Research Beijing, Beijing 100083, China §Harbin Institute of Technology, Harbin 150001, China {wsong, lzliu}@cnu.edu.cn, rjfu@iflytek.com, tliu@ir.hit.edu.cn Abstract We present a method of using cohesion to improve discourse element identification for sentences in student essays. New fea- tures for each sentence are derived by con- sidering its relations to global and local cohesion, which are created by means of cohesive resources and subtopic coverage. In our experiments, we obtain significant improvements on identifying all discourse elements, especially of +5% F1 score on thesis and main idea. The analysis shows that global cohesion can better capture the- sis statements. 1 Introduction Automatic discourse analysis of student essays can benefit many downstream applications such as essay rating, text organization assessment and writing instruction. In this paper we focus on iden- tifying discourse elements for sentences in persua- sive essays written by Chinese high school stu- dents. Discourse elements represent the contribu- tions that sentences can make to text organization. Typical discourse elements and their functions in persuasive writing are summarized in Table 1. Previous work mainly exploits the properties of a sentence itself or adjacent sentences for this task. In this work, we explore cohesion to express re- lations among sentences through the whole text. Cohesion can be defined as a set of resources link- ing within a text that organize the text together (Halliday and Hasan, 1976). It can be achieved through the use of reference, ellipsis, substitution, conjunction and lexical cohesion. Among them, lexical cohesion has been widely used for mod- eling local coherence and applied to related ap- plications (Barzilay and Elhadad, 1999; Barzilay and Lapata, 2008; Galley et al., 2003; Hsueh et al., 2006; Filippova and Strube, 2006). Since cohesion Element Definition Introduction (I) introduces the background and/or grabs readers’ attention Prompt (P) restates or summarize the prompt Thesis (T) states the author’s main claim on the issue for which he/she is arguing Main idea (M) asserts foundational ideas or aspects that are related to the thesis Supporting idea (S) provides evidence to explain or support the thesis and main ideas Conclusion (C) concludes the whole essay or one of the main ideas Other (O) doesn’t fit into the above elements or makes no meaningful contribution Table 1: Definitions of discourse elements. is closely related to the structure of text (Morris and Hirst, 1991), it motivates us to explore similar techniques for discourse element identification. In addition, its ease of implementation is also attrac- tive. Other options for representing text structure such as full-text discourse parsers (Marcu, 2000) may be not available or don’t have satisfied per- formance, especially for non-English languages. However, modeling local coherence alone is not adequate to distinguish discourse elements in per- …

Content truncated to 500 words