Seeking information about products and services is an important activity of online consumers before making a purchase decision. Inspired by recent research on conversational reading comprehension (CRC) on formal documents, this paper studies the task of leveraging knowledge from a huge amount of reviews to answer multi-turn questions from consumers or users. Questions spanning multiple turns in a dialogue enables users to ask more specific questions that are hard to ask within a single question as in traditional machine reading comprehension (MRC). In this paper, we first build a dataset and then propose a novel task-adaptation approach to encoding the formulation of CRC task into a pre-trained language model. This task-adaptation approach is unsupervised and can greatly enhance the performance of the end CRC task that has only limited supervision. Experimental results show that the proposed approach is highly effective and has competitive performance as supervised approach. We plan to release the datasets and the code in May 2019.

本研究通过对话式阅读理解，研究了利用评论构建一个能够回答在线商业潜在消费者多轮问题的代理的新任务。该研究首先构建了一个评论CRC数据集，然后在语言模型（例如BERT）预训练和领域特定的微调之间提出了一种新的任务感知预调优步骤。该预调优不需要数据注释，但可以极大地提高我们的最终任务的性能。实验结果表明，所提出的方法效果显著，并具有与受监督方法竞争的性能。

综述会话式阅读理解