Toward Few-Shot Response Generation for Conversational Agents

One of the major impediments to the deployment of task-oriented response generation models in real-world scenarios is a lack of labeled in-domain training data. In this thesis, we explore a multi-task learning approach in which we pre-train a model with unlabeled dialogues, and then fine-tuned in a few-shot setting to generate factually accurate and natural agent responses.

We formulate our primary task as utterance paraphrasing, where a template-based agent utterance in the context of the preceding dialogue is converted into a more natural-sounding one. We adapt unlabeled dialogues as pre-training data for this task by using automatically generated paraphrases of the last agent utterance as templates to predict the utterance in the context of the preceding dialogue turns. Since the real templates lack the contextual information embedded in the paraphrases, we introduce an auxiliary task, next sentence prediction, where the last turn of a dialogue is masked in the input and used as the predicted gold output.

Evaluation across several conversational datasets shows that our pre-training strategy improves the quality of responses in an in-domain context, indicating that leveraging unlabeled datasets can be an effective method for improving the quality of responses.