Date: 2023-03-03 / 4:00 ~ 5:00 PM
Location: MSC W301
We propose exploring and experimenting with different contextual information in the embedding space to address the challenges of integrating context into dialogue systems. The work develops models that overcome limitations of state-of-the-art models in terms of token encoding and fusion of arbitrary forms of context. Additionally, diarization methods are explored to resolve speaker ID errors in dialogue data training. The proposed models are evaluated on retrieval-based and generation-based dialogue systems, incorporating different forms of context, including previous conversation utterances, semantically similar response candidates, and domain information. The performance of the contextual response ranking model exceeded state-of-the-art models, showing potential for various forms of context incorporation. The generative model built on top of Blenderbot integrates previous conversation utterances and stacked questions, achieving an average satisfaction score of 3.5 out of 5 in real conversations. Topical clustering and diversity are also investigated to improve conversational dialogue models.