Discourse Relation Recognition by Comparing Various Units of Sentence Expression with Recursive Neural Network

Discourse Relation Recognition by Comparing Various Units of Sentence Expression with Recursive Neural Network Atsushi Otsuka, Toru Hirano, Chiaki Miyazaki, Ryo Masumura, Ryuichiro Higashinaka, Toshiro Makino and Yoshihiro Matsuo NTT Media Intelligence Laboratories NTT Corporation {otsuka.atsushi, hirano.tohru, miyazaki.chiaki, masumura.ryo, higashinaka.ryuichiro, makino.toshiro, matsuo.yoshihiro}@lab.ntt.co.jp Abstract We propose a method for implicit discourse relation recognition using a recursive neural network (RNN). Many previous studies have used the word-pair feature to compare the meaning of two sentences for implicit dis- course relation recognition. Our proposed method differs in that we use various-sized sentence expression units and compare the meaning of the expressions between two sen- tences by converting the expressions into vec- tors using the RNN. Experiments showed that our method significantly improves the accu- racy of identifying implicit discourse relations compared with the word-pair method. 1 Introduction Discourse relation recognition is a technique to iden- tify the type of discourse relation between two sen- tences. Because discourse relation contributes to the coherence of sentences, it has potential applications in many natural language processing (NLP) tasks. For example, in text summarization, it makes sum- mary documents more consistent by using discourse relations and structures (Gerani et al., 2014). Simi- larly, in conversational systems (Higashinaka et al., 2014), discourse relations can help the system select contextually appropriate system utterances. Discourse relations are categorized into explicit and implicit relations. Explicit relations have a dis- course marker such as a connective, making them easy to identify with a high degree of accuracy (Pitler and Nenkova, 2009). Implicit discourse rela- tions, in contrast, have no discourse marker between sentences. Previous studies have proposed many methods for implicit discourse recognition, among them reasoning-based (Sugiura et al., 2013) and pattern-based (Saito et al., 2006) methods. Many of these earlier studies (Marcu and Echihabi, 2002; Lin et al., 2009; Pitler et al., 2009; Wang et al., 2012; Lan et al., 2013; Biran and McKeown, 2013; Ruther- ford and Xue, 2014) focused on using word pairs or their derivative features. For example, take the two following sentences: A1 : I like summer. B1 : I prefer winter. In this case, we can easily identify the relation as “comparison” by focusing on the word pair “sum- mer - winter”. However, there is emerging evidence that word pairs may no longer have a role to play in implicit discourse relation recognition (Park and Cardie, 2012). This is because identification is not always possible by using just word pairs. When we consider the following sentences, A2 : I got soaked by the sudden rain yesterday. B21: Did you forget your umbrella at the office? B22: The rain was so heavy that my umbrella was useless. discourse A2 − B21 and A2 − B22 have different relations. discourse A2 − B21 is causal relation: B21 explains the reason for A1, and A2 − B22 is expansion relation: B22 is a supplemental explana- tion about the “sudden rain” in A2. Nevertheless, the same word pair “soaked - umbrella” can be ex- PACLIC 29 63 29th Pacific Asia …

Content truncated to 500 words