Rapid Prototyping of Robust Language Understanding Modules for Spoken Dialogue Systems

Rapid Prototyping of Robust Language Understanding Modules for Spoken Dialogue Systems †Yuichiro Fukubayashi, †Kazunori Komatani, ‡Mikio Nakano, ‡Kotaro Funakoshi, ‡Hiroshi Tsujino, †Tetsuya Ogata, †Hiroshi G. Okuno †Graduate School of Informatics, Kyoto University Yoshida-Hommachi, Sakyo, Kyoto 606-8501, Japan {fukubaya,komatani}@kuis.kyoto-u.ac.jp {ogata,okuno}@kuis.kyoto-u.ac.jp ‡Honda Research Institute Japan Co., Ltd. 8-1 Honcho, Wako, Saitama 351-0188, Japan nakano@jp.honda-ri.com {funakoshi,tsujino}@jp.honda-ri.com Abstract Language understanding (LU) modules for spoken dialogue systems in the early phases of their development need to be (i) easy to construct and (ii) robust against vari- ous expressions. Conventional methods of LU are not suitable for new domains, be- cause they take a great deal of effort to make rules or transcribe and annotate a suf- ficient corpus for training. In our method, the weightings of the Weighted Finite State Transducer (WFST) are designed on two levels and simpler than those for conven- tional WFST-based methods. Therefore, our method needs much fewer training data, which enables rapid prototyping of LU mod- ules. We evaluated our method in two dif- ferent domains. The results revealed that our method outperformed baseline methods with less than one hundred utterances as training data, which can be reasonably prepared for new domains. This shows that our method is appropriate for rapid prototyping of LU modules. 1 Introduction The language understanding (LU) of spoken dia- logue systems in the early phases of their devel- opment should be trained with a small amount of data in their construction. This is because large amounts of annotated data are not available in the early phases. It takes a great deal of effort and time to transcribe and provide correct LU results to a Figure 1: Relationship between our method and con- ventional methods large amount of data. The LU should also be robust, i.e., it should be accurate even if some automatic speech recognition (ASR) errors are contained in its input. A robust LU module is also helpful when col- lecting dialogue data for the system because it sup- presses incorrect LU and unwanted behaviors. We developed a method of rapidly prototyping LU mod- ules that is easy to construct and robust against var- ious expressions. It makes LU modules in the early phases easier to develop. Several methods of implementing an LU mod- ule in spoken dialogue systems have been proposed. Using grammar-based ASR is one of the simplest. Although its ASR output can easily be transformed into concepts based on grammar rules, complicated grammars are required to understand the user’s ut- terances in various expressions. It takes a great deal of effort to the system developer. Extracting con- 210 Figure 2: Example of WFST for LU cepts from user utterances by keyword spotting or heuristic rules has also been proposed (Seneff, 1992) where utterances can be transformed into concepts without major modifications to the rules. However, numerous complicated rules similarly need to be manually prepared. Unfortunately, neither method is robust against ASR errors. To cope with these problems, corpus-based (Su- doh and Tsukada, 2005; He and Young, 2005) and Weighted Finite State …

Content truncated to 500 words