Analysis of the Dirichlet process mixture model with application to dialogue act classification
|Authors:||Bakhtiari Koohsorkhi, Alireza|
|Abstract:||Recognition of user intentions is one of the most challenging problems in the design of dialogue systems. These intentions are usually coded in terms of Dialogue Acts (Following Austin’s work on speech act theory), where a functional role is assigned to each utterance of a conversation. Manual annotation of dialogue acts is both time consuming and expensive, therefore there is a huge interest in systems which are able to automatically annotate dialogue corpora. In this thesis, we propose a nonparametric Bayesian approach for the automatic classification of dialogue acts. We make use of the Dirichlet Process Mixture Model (DPMM), within which each of the components is governed by a Dirichlet-Multinomial distribution. Two novel approaches for hyperparameter estimation in these distributions are also introduced. Results of the application of this model to the DIHANA corpus shows that the DPMM can successfully recover the true number of DA labels with high precision|
|Document Type:||Mémoire de maîtrise|
|Open Access Date:||17 April 2018|
|Collection:||Thèses et mémoires|
All documents in CorpusUL are protected by Copyright Act of Canada.