A Deep Learning Methodology for Semantic Utterance Classification in Domain-Specific Dialogue Systems

Recent advances in deep learning approaches have transformed fields such as natural language and image processing. In particular, these new advances have the potential to transform dialogue systems which are traditionally implemented using language model based or boosting approaches. In this work, we have proposed a deep learning framework to perform semantic utterance classification (SUC) for use in domain-specific dialogue systems. Deep learning has only recently been used for SUC but has not been used with domain-specific word embeddings or dialogue systems. Semantic classifiers need to account for a variety of instances where the utterance for the semantic domain class varies. In order to capture the candidate relationships between the semantic class and the word sequence in an utterance, we have proposed a shallow convolutional neural network (CNN) that uses domain-specific word embeddings, that has been initialized using word2vec for determining semantic similarity of words. These embeddings can remain static, be updated during training or can even be created from scratch for the particular intent determination task at hand. Finally, these methodologies have been integrated into a library for easy deployment into existing platforms with dialogue systems. Experimental results obtained on two different use cases demonstrate the effectiveness of shallow neural networks for SUC. The methods produce superior classification accuracy comparable to existing benchmarks. We also demonstrate our framework in a real-world medical training system.