Profile Research Publications Tutorials
Predicate-argument structure analysis is one important component of dialogue system that can understand semantic information. One motivation of this research is building an analyzer which can be adapted (retrained) with partially annotated corpora. The key idea is pointwise, constructing the analyzer with the appropriate information. Pointwise classifier does not refer to the estimated information such as parsed dependency information, and directly uses the features used in the estimation of the preliminary information.
Previous dialogue systems understand the intention of the user by using the pre-defined intention set, however, in case of information navigation, focus of the user is also important. User focus is defined as focused object in dialogue, like an attentional state of the user. The proposed navigation system understand the user intention and the user focus to manage the dialogue module activation, I also proposed a framework of dialogue management for non-task-oriented dialogue system. The framework uses user focus information to track the user intention in non-task-oriented dialogue, and improves the appropriate action selection of dialogue system.
My approach is hybrid of conventional rule-based dialogue management and statistical dialogue management using partially observable Markov decision process (POMDP). The proposed method converts the rule-based dialogue manager to intention dependency graph (IDG), and enables to convert the rule-based one to the POMDP-based one by using the restriction of rules.
Language model is a component of automatic speech recognition. For spoken dialogue system, it is important to adapt the language model for its style and domain. The proposed method selects appropriate texts for the training of language model for the system by using semantic information. The resultant language model works better than the baseline language model based on lexical selection method on not only recognition accuracy but also dialogue-level accuracy.
I developed a spoken dialogue system that extracts exactly or partially matched information from Web texts. The information structure for the information extraction is defined by semantic (predicate-argument: P-A) structures. This framework defines domain dependent important semantic patterns automatically, and presents relevant information by using semantic similarity between the user query and information candidates.