Chinese dialogue intention classification based on multi-model ensemble

Abstract

In dialogue systems, understanding the user utterances is crucial for providing appropriate responses. A traditional dialogue act classification (DA) task is to classify each user reply into 'ACCEPT, REJECT, PROPOSE, and others'. In contrast, in this paper, we define the DA task on multiple round conversations between humans. The re-defined task is to classify a full dialogue according to the intention of one participant. We term this task as intention classification (IC). We, then, propose a hybrid neural network-based ensemble model for solving this problem. Two novel ensemble schemes are introduced for combining the classification results or features from various classifiers. One is ensembling features from each individual classifier using stacking, and we term this scheme as SFE. The other is adding wrong examples' weight to loss functions of each individual classifier using the AdaBoost scheme, and we term this scheme as MN-Ada. We have empirically examined the performance of the proposed ensemble schemes by using three popular deep neural networks, as well as one newly modified networks for IC. Extensive experiments have been conducted on a Chinese dialogue corpus. Our model can achieve state-of-the-art accuracy on the experimental dialogue corpus.