Implicit Discourse Relation Recognition by Selecting Typical Training Examples

Proceedings of COLING 2012: Technical Papers, pages 2757–2772, COLING 2012, Mumbai, December 2012. Implicit Discourse Relation Recognition by Selecting Typical Training Examples Xun Wang1 Sujian Li1 Jiwei Li1 Wenjie Li2 (1) Key Laboratory of Computational Linguistics, Peking University, Ministry of Education, CHINA (2) Department of Computing, Hong Kong Polytechnic University {xunwang,lisujian,bdlijiwei}@pku.edu.cn, cswjli@comp.polyu.edu.hk ABSTRACT Implicit discourse relation recognition is a challenging task in the natural language processing field, but important to many applications such as question answering, summarization and so on. Previous research used either artificially created implicit discourse relations with connectives removed from explicit relations or annotated implicit relations as training data to detect the possible implicit relations, and do not further discern which examples are fit to be training data. This paper is the first time to apply a different typical/atypical perspective to select the most suitable discourse relation examples as training data. To differentiate typical and atypical examples for each discourse relation, a novel single centroid clustering algorithm is proposed. With this typical/atypical distinction, we aim to recognize those easily identified discourse relations more precisely so as to promote the performance of the implicit relation recognition. The experimental results verify that the proposed new method outperforms the state-of-the-art methods. KEYWORDS : Discourse relation recognition, single centroid clustering, implicit discourse relation.  Corresponding author. 2757 1 Introduction It is widely agreed that sentences/clauses are usually not understood in isolation, but in relation to their neighbouring sentences/clauses. The task of discourse relation recognition is to identify and label the relations between sentences/clauses, which is fundamental to many natural language processing applications such as question answering, automatic summarization and so on. Discourse relations, such as comparison and causal relations, can be divided into explicit and implicit relations by the presence or absence of discourse connectives (e.g., but, because et. al.). Previous study indicates that the presence of discourse markers can greatly help relation recognition and the most general senses (i.e., comparison, contingency, temporal and expansion ) can be disambiguated with 93% accuracy based solely on the discourse connectives (Pitler et al., 2008). On the other hand, the absence of explicit textual cues makes it very difficult to identify the implicit discourse relations. Thus, recently discourse relation recognition research puts more efforts to meet the challenges in implicit discourse relation recognition. Existing work mainly focused on exploiting various linguistic features to learn the implicit discourse relation classifiers based on the training data collected (Wellner, Pustejovsky and Havasi, 2006; Pitler, Louis and Nenkova, 2009; Lin, Kan and Ng 2009; Wang, Su and Tan, 2010). Most useful linguistic features (such as word pairs) are extracted from the local context, which is usually determined as Argument 1 (Arg1, the first sentence/clause) plus Argument 2 (Arg2, the second sentence/clause). Like the other related work in the literature, in this paper, we focus on the recognition of local implicit discourse relations, i.e. only the two arguments are examined. To collect training data, the state-of-the-art methods normally start from the artificial/real perspective and simply make use of the …

Content truncated to 500 words