Improving the classification of call center service dialogue with key utterences

In the field of customer service management, classifying service dialogues to different business labels is beneficial for managers to improve their service quality. However, the size of labeled service dialogue dataset in real scenarios is usually small due to the expensive labeling cost, which makes it difficult to fully train the supervised classification models. Moreover, the service dialogue usually contains chitchat which can be regarded as the noise affecting the classification performance. Existing text classification methods fail to address above two issues simultaneously. Hence, in this paper, we propose a dialogue classification algorithm that strengthens the influence of the business-related utterances in the dialogue and use them as the key utterances to improve the classification. Firstly, we propose key utterance labels that can indicate which utterances in the dialogue are key utterances. Then, we propose the dialogue classification model that is based on the key utterance labels and logistic regression, namely KU-LR. The KU-LR can learn the key utterance patterns and increase the importance of key utterances in the dialogue, and then the KU-LR makes more accurate decisions for dialogue classification. The experimental results on real-world dataset show that the KU-LR method outperforms other baselines when the training dataset is small.