Modeling Dialogue Acts with Content Word Filtering and Speaker Preferences

Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 2179–2189 Copenhagen, Denmark, September 7–11, 2017. c⃝2017 Association for Computational Linguistics Modeling Dialogue Acts with Content Word Filtering and Speaker Preferences Yohan Jo and Michael Miller Yoder and Hyeju Jang and Carolyn P. Ros´e Language Technologies Institute Carnegie Mellon University {yohanj,yoder,hyejuj,cprose}@cs.cmu.edu Abstract We present an unsupervised model of dia- logue act sequences in conversation. By modeling topical themes as transitioning more slowly than dialogue acts in conver- sation, our model de-emphasizes content- related words in order to focus on con- versational function words that signal di- alogue acts. We also incorporate speaker tendencies to use some acts more than others as an additional predictor of dia- logue act prevalence beyond temporal de- pendencies. According to the evaluation presented on two dissimilar corpora, the CNET forum and NPS Chat corpus, the ef- fectiveness of each modeling assumption is found to vary depending on characteris- tics of the data. De-emphasizing content- related words yields improvement on the CNET corpus, while utilizing speaker ten- dencies is advantageous on the NPS cor- pus. The components of our model com- plement one another to achieve robust per- formance on both corpora and outperform state-of-the-art baseline models. 1 Introduction Dialogue acts (DAs), or speech acts, represent the intention behind an utterance in conversa- tion to achieve a conversational goal (Austin, 1975). Modeling conversations as structured DA sequences is a step toward the automated under- standing of dialogue, useful for dialogue agents (Traum, 1999; Louwerse et al., 2002) and the processing of informal online conversational data (Misra and Walker, 2013; Vosoughi and Roy, 2016). Distributions of DAs can also be used as predictors of conversational outcome measures such as student learning in tutoring systems (Lit- man and Forbes-Riley, 2006) and engagement in meetings (Wrede and Shriberg, 2003). Unsuper- vised models for DA recognition may substitute or aid costly human annotation. We present an unsu- pervised model of DA sequences in conversation that overcomes limitations of prior models. The first improvement our model offers is sep- arating out content-related words to emphasize words more relevant to DAs. DAs are associ- ated more closely with style and function words such as discourse markers and light verbs than with content words, which are more related to the propositional content (Erkens and Janssen, 2008; O’Shea et al., 2012). However, separating out con- tent words is not standard in our field. For ex- ample, in some rule-based semantic and pragmatic parsing, the content and function of dialogue acts are not formally distinguished in the formalization (Becker et al., 2011), especially in domain-specific applications in dialogue systems (Gavald`a, 2004). A separation between content and function is use- ful for making cross-domain or cross-task gener- alizations about conversational processes. Our model filters out content words by imple- menting the assumption that conversations pro- ceed against a backdrop of underlying topics that transition more slowly than DAs or that are con- stant throughout. Based on a difference in tran- sition speed, two …

Content truncated to 500 words