Dynamic Bayesian Networks and Variable Length Genetic Algorithm for Dialogue Act Recognition
The recognition of dialogue act is a task of crucial importance for the processing of natural language in many applications such as dialogue system. However, it is one of the most challenging problems. The current dialogue act recognition models, namely cue-based models, are based on machine learnin...
No Tags, Be the first to tag this record!
|The recognition of dialogue act is a task of crucial importance for the processing of natural language in many applications such as dialogue system. However, it is one of the most challenging problems. The current dialogue act recognition models, namely cue-based models, are based on machine learning techniques, particularly statistical ones. Despite the success of the cue-based models, they still have serious drawbacks. Among them are, inadequate representation of dialogue context, intra-utterance and inter-utterances independencies assumptions, inaccurate estimation of the recognition accuracy and suboptimality of the lexical cues selection approaches.
Motivating by these drawbacks, this research proposes a new model of dialogue act recognition in which dynamic Bayesian machine learning is applied to induce dynamic Bayesian networks models from task-oriented dialogue corpus using sets of lexical cues selected automatically by means of new variable length genetic algorithm. In achieving this, the research is planned in three main stages. In the initial stage, the dynamic Bayesian networks models are constructed based on a set of lexical cues selected tentatively from the dialogue corpus. The results are compared with the results of static Bayesian networks and naïve bayes. The results confirm the merits of using dynamic Bayesian networks for dialogue act recognition.
In the second stage, the previous ranking approaches are investigated for the selection of lexical cues. The main drawbacks of these approaches are highlighted, and based on that an alternative approach is proposed. The proposed approach consists of preparation phase and selection phase. The preparation phase transforms the original dialogue corpus into phrases space. In the selection phase, a new variable length genetic algorithm is applied to select the lexical cues. The results of the proposed approach are compared with the results of the ranking approaches. The results provide experimental evidences on the ability of the proposed approach to avoid the drawbacks of the ranking approaches.
In the final stage; the dynamic Bayesian networks models are redesigned using the lexical cues generated from the proposed lexical cues selection approaches. The results confirm the effectiveness of proposed approaches for the design of dialogue act recognition model.