Joint Learning of Dialog Act Segmentation and Recognition in Spoken Dialog Using Neural Networks

Proceedings of the The 8th International Joint Conference on Natural Language Processing, pages 704–712, Taipei, Taiwan, November 27 – December 1, 2017 c⃝2017 AFNLP Joint Learning of Dialog Act Segmentation and Recognition in Spoken Dialog Using Neural Networks Tianyu Zhao and Tatsuya Kawahara {zhao, kawahara}@sap.ist.i.kyoto-u.ac.jp Abstract Dialog act segmentation and recognition are basic natural language understanding tasks in spoken dialog systems. This pa- per investigates a unified architecture for these two tasks, which aims to improve the model’s performance on both of the tasks. Compared with past joint models, the pro- posed architecture can (1) incorporate con- textual information in dialog act recogni- tion, and (2) integrate models for tasks of different levels as a whole, i.e. dialog act segmentation on the word level and dialog act recognition on the segment level. Ex- perimental results show that the joint train- ing system outperforms the simple cascad- ing system and the joint coding system on both dialog act segmentation and recogni- tion tasks. 1 Introduction Recently the burst of interactive assistants and chatbots leads to an increasing interest of dialog systems. Natural language understanding (NLU), as an important component of dialog system, is usually responsible for dialog act (DA) or dialog intent tagging, where text classification techniques are necessary. Dialog act (also speech act) is a representation of the meaning of a sentence at the level of illocutionary force (Stolcke et al., 2000). For instance, a sentence “How is the weather?” belongs to the dialog act class Question. For an- other sentence, “The weather is quite good to- day.”, if it follows a previous Question sentence, it should be an Answer to the question. Otherwise it is likely to be a Statement. Therefore, DA recog- nition requires us to understand the sentence from semantic, pragmatic and syntactic aspects, and its context plays an important role as well. Words hi my name is Erica Segment E I I I E DA Greeting Statement Table 1: DA segmentation and recognition. The prerequisite for DA recognition is to split a sequence of words into segments, each of which corresponds to one DA unit. Especially for spo- ken dialog systems, NLU is based on Automatic Speech Recognition (ASR) hypotheses or tran- scripts, in which we cannot make any assump- tion of punctuation such as periods, commas and question marks for segmentation. Therefore DA segmentation becomes essential for spoken dialog systems. As the example given in Table 1 shows, a long utterance is firstly split into two segments “hi” and “my name is Erica”, to which DA tags “Greeting” and “Statement” are assigned after- wards. DA segmentation is a sequence labeling task and an “IE” tag coding scheme is adopted to describe segment boundaries, where “I” denotes “inside” of a segment and “E” denotes the “end”. While the DA segmentation is a pre-process of DA recognition, recognition of DA in the sequence helps segmentation. Thus, DA segmentation and recognition are two highly related tasks. We can expect that joint learning of these two tasks can improve …

Content truncated to 500 words