Integrating Tree Structures and Graph Structures with Neural Networks to Classify Discussion Discourse Acts

Proceedings of the 27th International Conference on Computational Linguistics, pages 3806–3818 Santa Fe, New Mexico, USA, August 20-26, 2018. 3806 Integrating Tree Structures and Graph Structures with Neural Networks to Classify Discussion Discourse Acts Yasuhide Miura†,‡ yasuhide.miura@fujixerox.co.jp Ryuji Kano† kano.ryuji@fujixerox.co.jp Motoki Taniguchi† motoki.taniguchi@fujixerox.co.jp Tomoki Taniguchi† taniguchi.tomoki@fujixerox.co.jp Shotaro Misawa† misawa.shotaro@fujixerox.co.jp Tomoko Ohkuma† ohkuma.tomoko@fujixerox.co.jp †Fuji Xerox Co., Ltd. ‡Tokyo Institute of Technology Abstract We proposed a model that integrates discussion structures with neural networks to classify dis- course acts. Several attempts have been made in earlier works to analyze texts that are used in various discussions. The importance of discussion structures has been explored in those works but their methods required a sophisticated design to combine structural features with a classifier. Our model introduces tree learning approaches and a graph learning approach to directly capture discussion structures without structural features. In an evaluation to classify discussion discourse acts in Reddit, the model achieved improvements of 1.5% in accuracy and 2.2 in F1 score com- pared to the previous best model. We further analyzed the model using an attention mechanism to inspect interactions among different learning approaches. 1 Introduction With the recent growth of social news sites and debate portals, people today discuss various topics online. These discussions include valuable public opinions of crowds. However, automated analyses of them are often difficult, requiring understanding of textual contents and discussion structures. By extending approaches taken in the analysis of spoken dialogue acts (Stolcke et al., 2000; Bunt et al., 2010), there have been a number of attempts to analyze discussions with discourse acts. Discussions in emails (Cohen et al., 2004; Carvalho and Cohen, 2005; Carvalho and Cohen, 2006; Hu et al., 2009; Omuya et al., 2013), newsgroups (Wang et al., 2007), technical forums (Kim et al., 2010b; Wang et al., 2011; Bhatia et al., 2012; Liu et al., 2017) and social news (Zhang et al., 2017) are targeted in earlier studies. The automatic classification of these discourse acts can improve tasks like information access and summarization for discussions. The importance of discussion structures for analyzing discourse acts has already been recognized in earlier studies. Relational features (Carvalho and Cohen, 2005), link features (Hu et al., 2009), and struc- tural features (Wang et al., 2007; Kim et al., 2010a; Kim et al., 2010b; Wang et al., 2011; Bhatia et al., 2012; Zhang et al., 2017; Liu et al., 2017) are combined with a probabilistic graphical model, a struc- tured prediction model, or sequential classification models. These approaches have achieved promising results in classifying discourse acts, but they require a sophisticated design to integrate structural features into a classification model. In this paper, we propose a model that integrates discussion structures with neural networks to classify discourse acts. Recently, neural networks have shown their effectiveness at capturing tree structures (Socher et al., 2011; Socher et al., 2014; Tai et al., 2015) and graph structures (Defferrard et al., 2016; Kipf and Welling, 2017). Our model introduces tree learning approaches and a graph learning approach to directly capture …

Content truncated to 500 words