Adaptation of Hierarchical Structured Models for Speech Act Recognition in Asynchronous Conversation

Proceedings of NAACL-HLT 2019, pages 1326–1336 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 1326 Adaptation of Hierarchical Structured Models for Speech Act Recognition in Asynchronous Conversation Tasnim Mohiuddin1∗and Thanh-Tung Nguyen1⋆ and Shafiq Joty1,2⋆ ⋆Nanyang Technological University, Singapore †Salesforce Research Asia, Singapore {mohi0004@e.,NG0155NG@e.,srjoty@}ntu.edu.sg Abstract We address the problem of speech act recog- nition (SAR) in asynchronous conversations (forums, emails). Unlike synchronous conver- sations (e.g., meetings, phone), asynchronous domains lack large labeled datasets to train an effective SAR model. In this paper, we propose methods to effectively leverage abun- dant unlabeled conversational data and the available labeled data from synchronous do- mains. We carry out our research in three main steps. First, we introduce a neural architec- ture based on hierarchical LSTMs and condi- tional random fields (CRF) for SAR, and show that our method outperforms existing methods when trained on in-domain data only. Second, we improve our initial SAR models by semi- supervised learning in the form of pretrained word embeddings learned from a large unla- beled conversational corpus. Finally, we em- ploy adversarial training to improve the results further by leveraging the labeled data from synchronous domains and by explicitly mod- eling the distributional shift in two domains. 1 Introduction With the ever-increasing popularity of Internet and mobile technologies, communication media like emails and forums have become an inte- gral part of people’s daily life where they discuss events, issues and experiences. Participants inter- act with each other asynchronously in these me- dia by writing at different times, generating a type of conversational discourse that is different from synchronous conversations such as meeting and phone conversations (Louis and Cohen, 2015). In the course of the interactions, the participants per- form certain communicative acts like asking ques- tions, requesting information, or suggesting some- thing, which are known as speech acts (Austin, ∗All authors contibuted equally. C1:hoping to do the XinJiang Tibet Highway. [Statement] Has anyone done it? [Question] Am hoping to hire a 4-wheel drive. [Statement] I know the roads are bad and I would need an experienced driver and guide - any recommendations? [Question] C2:I never done this routine,however i been to Xinjiang twice,in my opinion the local people not friendly, not safe to do this. [Response] I still have relative stay in Xinjiang, however don’t know what they can offer for help... [Response] C3:I’m not sure if travelling overland from Xinjiang to Tibet is officially legal yet. [Response] You might want to post your question on the North- East Asia branch of Lonely Planet’s ThornTree forum for more (useful) answers. [Suggestion] C4:a frend and i are trying this route as well, we will likely be in urumuqi and northern part of xinjiang from 8th apr to end apr; looking at doing the xin jiang tibet highway from end apr. (truncated) [Statement] contact me at [email] if you want to hook up for possible transport sharing [Suggestion] cheers. [Polite] Figure 1: Example of speech acts in a forum thread. 1962). For example, consider the …

Content truncated to 500 words