End-to-End Question Answering Models for Goal-Oriented Dialog Learning

The task of Next Utterance Classification in dialog learning highly resembles that of Question Answering, but there has not been much attention to applying models across the two fields, especially not in more practical dialog modeling scenarios. Hence, this paper presents end-to-end Question Answering (QA) style models for the goal-oriented dialog system task in Dialog System Technology Challenges (DSTC) 7. We first provide a comprehensive quantitative and qualitative analysis of the newly introduced Advising dataset and show the heavy reliance of the data on an external Knowledge Base (KB) of course offerings. Based on such analysis, we model the dialogs with popular approaches from both dialog and QA literature, and show that QA methods perform comparably well to the former, despite they were designed for a fairly different task. We mainly compare Hierarchical RNNs (Serban et al. 2017) and a modified version of the BiDAF model (Seo et al. 2017a). We then employ large-scale KB query methods from DrQA (Chen et al. 2017) to incorporate external knowledge into the dialog. Furthermore, contrary to a recent previous work (Tao et al. 2018), we show that Embeddings from Language Models (ELMo) (Peters et al. 2018) do significantly improve the performance in dialog systems without fine-tuning.