Knowledge-Based Chat Detection with False Mention Discrimination

Chat detection is critical for recently emerged personal intelligent assistants (PIA), which can be seen as a hybrid of domain-specific task-oriented spoken dialogue systems and open-domain non-task-oriented ones. Recent advances have attempted to utilize external domain knowledge to enhance utterance semantics understanding and can contribute to chat detection. However, it also inevitably introduces false mention (i.e., token spans being misidentified as entity mentions) in Chat utterances, causing performance to degrade. To deal with this issue, this paper proposes a new model for knowledge-based chat detection with false mention discrimination (FMD-KChat). A two-stage pipeline is adopted, which contains an additional neural network-based classifier in the first stage for distinguishing the false mentions and a feature fusion gate in the chat detection stage for combining the contextual representation with the external knowledge feature based on the false mention discrimination probability. Experiments on the SMP-ECDT benchmark corpus show the well performance of the proposed model.