Robust Dialog Management with N-Best Hypotheses Using Dialog Examples and Agenda

Proceedings of ACL-08: HLT, pages 630–637, Columbus, Ohio, USA, June 2008. c⃝2008 Association for Computational Linguistics Robust Dialog Management with N-best Hypotheses Using Dialog Examples and Agenda Cheongjae Lee, Sangkeun Jung and Gary Geunbae Lee Pohang University of Science and Technology Department of Computer Science and Engineering Pohang, Republic of Korea {lcj80,hugman,gblee}@postech.ac.kr Abstract This work presents an agenda-based approach to improve the robustness of the dialog man- ager by using dialog examples and n-best recognition hypotheses. This approach sup- ports n-best hypotheses in the dialog man- ager and keeps track of the dialog state us- ing a discourse interpretation algorithm with the agenda graph and focus stack. Given the agenda graph and n-best hypotheses, the system can predict the next system actions to maximize multi-level score functions. To evaluate the proposed method, a spoken dia- log system for a building guidance robot was developed. Preliminary evaluation shows this approach would be effective to improve the ro- bustness of example-based dialog modeling. 1 Introduction Development of spoken dialog systems involves hu- man language technologies which must cooperate to answer user queries. Since the performance in human language technologies such as Automatic Speech Recognition (ASR) and Natural Language Understanding (NLU)1 have been improved, this ad- vance has made it possible to develop spoken dialog systems for many different application domains. Nevertheless, there are major problems for practi- cal spoken dialog systems. One of them which must be considered by the Dialog Manager (DM) is the error propagation from ASR and NLU modules. In 1Through this paper, we will use the term natural language to include both spoken language and written language general, errors in spoken dialog systems are preva- lent due to errors in speech recognition or language understanding. These errors can cause the dialog system to misunderstand a user and in turn lead to an inappropriate response. To avoid these errors, a basic solution is to improve the accuracy and robust- ness of the recognition and understanding processes. However, it has been impossible to develop perfect ASR and NLU modules because of noisy environ- ments and unexpected input. Therefore, the devel- opment of robust dialog management has also been one of the most important goals in research on prac- tical spoken dialog systems. In the dialog manager, a popular method to deal with these errors is to adopt dialog mechanisms for detecting and repairing potential errors at the con- versational level (McTear et al., 2005; Torres et al., 2005; Lee et al., 2007). In human-computer com- munication, the goal of error recovery strategy is to maximize the user’s satisfaction of using the sys- tem by guiding for the repair of the wrong informa- tion by human-computer interaction. On the other hand, there are different approaches to improve the robustness of dialog management using n-best hy- potheses. Rather than Markov Decision Processes (MDPs), partially observable MDPs (POMDPs) po- tentially provide a much more powerful framework for robust dialog modeling since they consider n- best hypotheses to estimate the distribution of the belief …

Content truncated to 500 words