Speaker-Informed Time-and-Content-Aware Model for Contextual Natural Language Understanding

Abstract

This dissertation proposes speaker-informed content-aware and time-aware models that capture relevant historical utterances based on awareness of both content and time of all historical utterances using speaker information for natural language understanding (NLU) of the current utterance. By improving the awareness of important historical utterances for NLU of the current utterance, the proposed models are improved in accuracy of NLU. Specifically, the proposed approach for capturing important historical utterances is decomposed into two main axes, ``Awareness" and ``Attention Level". Awareness considers the distance of a historical utterance from the current utterance, or content-similarity between the utterances to assess the importance of the historical utterance. Attention level considers the speaker identity of a historical utterance to assess the importance of the utterance. Additionally, this dissertation proposes speaker-modeling methods that are tailored for Transformer-based general-purpose pretrained models like XLNet. These methods are implemented on the top of XLNet to verify the effectiveness of them. To evaluate the proposed models, experiments were conducted on two benchmark datasets, the fourth dialog state tracking challenge (DSTC 4) and Loqui. On the experiments, the proposed models achieved state-of-the-art F1 scores on both datasets. Also the results of experiments demonstrate that the proposed methods are effective to improve accuracy of NLU. Finally, detailed analysis of the results is provided.