Probabilistic dialogue models with prior domain knowledge
SIGDIAL '12 Proceedings of the 13th Annual Meeting of the Special Interest Group on Discourse and Dialogue
| Hi-index | 0.00 |
A dialogue-based interface for information systems is considereda potentially very useful approach to information access. A keystep in computer processing of natural-language dialogues isdialogue-act (DA) recognition. In this paper, we apply afeature-based classification approach for DA recognition, by usingthe maximum entropy (ME) method to build a classifier for labelingutterances with DA tags. The ME method has the advantage that alarge number of heterogeneous features can be flexibly combined inone classifier, which can facilitate feature selection. A uniquecharacteristic of our approach is that it does not need to modelthe prior probability of DAs directly, and thus avoids the use of adiscourse grammar. This simplifies the implementation of theclassifier and improves the efficiency of DA recognition, withoutsacrificing the classification accuracy. We evaluate the classifierusing a large data set based on the Switchboard corpus. Encouragingperformance is observed; the highest classification accuracyachieved is 75.03%. We also propose a heuristic to address theproblem of sparseness of the data set. This problem has resulted inpoor classification accuracies of some DA types that have very lowoccurrence frequencies in the data set. Preliminary evaluationshows that the method is effective in improving the macroaverageclassification accuracy of the ME classifier. © 2008 WileyPeriodicals, Inc.