Advanced search


Carlos D. Martínez-Hinarejos, José-Miguel Benedí. A comparative study between generative and discriminative statistical models for unsegmented dialogue annotation. Proceedings of IberSpeech 2014, 2014. pp. 178-187.

Abstract. Dialogue systems employ the dialogue strategy to define is behaviour in their interaction with users. The dialogue strategy is usually based on models whose parameters are estimated from dialogues annotated in terms of Dialogue Acts. Therefore, dialogue annotation is necessary to obtain dialogue systems, but manual annotation is hard to achieve and automatic annotation is desirable to obtain at least a draft annotation. The annotation problem can be formulated as an statistical optimisation problem on a sequence of turns. Some previous annotation works assumed the segmentation of turns into relevant subsequences (segments), but this segmentation is not usually available. Probabilistic annotation can be based on different statistical models. In this work, we compare the performance of two different paradigms: generative and discriminative models. These models are applied and compared in unsegmented dialogue annotation of two dialogue corpora of different nature.