A Vietnamese Dialog Act Corpus Based on ISO 24617-2 standard

A Vietnamese Dialog Act Corpus Based on ISO 24617-2 Standard Thi-Lan Ngo1,2, Khac Linh Pham2, Hiadeki Takeda3 1University of Information and Communication Technology (ICTU), Thainguyen Vietnam 2University of Engineering and Technology (UET-VNU), Hanoi Vietnam 3National Institute of Informatics (NII), Tokyo Japan ntlan@ictu.edu.vn, 15021790@vnu.edu.vn, takeda@nii.ac.jp Abstract The voice-based human-machine interaction systems such as personal virtual assistants, chat-bots, and contact centers are becoming increasingly popular. In this trend, conversation mining research also is getting the attention of many researchers. Standardized data play an important role in conversation mining. In this paper, we present a new Vietnamese corpus annotated for dialog acts using the ISO 24617-2 standard (2012), for emotions using Ekman’s six primitives (1972), and for sentiment using the tags “positive”, “negative” and “neutral”. Emotion and sentiment are tagged at functional segment level. We show how the corpus is constructed and provide a brief statistical description of the data. This is the first Vietnamese dialog act corpus. Keywords: Vietnamese corpus, dialog act corpus, ISO 24617–2 standard, conversational mining. 1. Introduction In recent years, an extremely rapid progress in speech processing and recognition technology has led the emer- gence of voice-based human-machine interaction systems such as mobile virtual assistants, automatic contact cen- tres, and chat-bots. These applications accommodate dif- ferent purposes but they all need the ability to understand conversations while interacting with the user. Therefore, with the development trend of intelligent systems using the conversational interface, studies on conversation min- ing such as conversation structure analysis, conversation topic modelling, user intent understanding, and user emo- tion or satisfaction identification have attracted the atten- tion of many researchers. In these research, standardized dialog act corpora are the foundation. It is widely ac- cepted that dialog act annotation is very valuable in further understanding of interaction structure, and in the design of artificial spoken or text dialogue (Wrede and Shriberg, 2003; Stolcke et al., 2006). There were several dialog act corpora available to the research community like as TRAINS (Traum, 1996), VERBMOBIL (Alexandersson et al., 1998), SWBD-DAMSL (Jurafsky et al., 1997), MRDA (Shriberg et al., 2004), AMI (McCowan et al., 2005) and so on. However, different corpora often apply a differ- ent scheme or modifying the existing scheme for dialog act annotation to serve task-specific needs. This creates a hardship in comparison of results and conclusions obtained when using different approaches due to a wide scatter of data in terms of the used annotation. Currently, ISO 24617- 2 standard (ISO, 2012) is seen as “lingua franca” for dialog act annotation (Chowdhury et al., 2016; Bunt et al., 2012). Experimental studies in DBOX (Amanova et al., 2016) and DialogBank (Wijnhoven, 2016) corpora and the DIT theory (Bunt, 2009), of the DIT++ annotation scheme that under- lies the ISO 24617-2 standard have shown good effects of the ISO standard on dialog act annotation. Thus, in our work, we build a Vietnamese spoken corpus, called ViDA corpus, annotated dialog act using the ISO 24617–2 stan- dard (ISO, 2012). Addition, we emotion tagging for func- tional segments (FSs) …

Content truncated to 500 words