Question-answer pairing from IM conversations via message merging and reply-to prediction

摘要

Preparing question-answer pairs from conversation logs (chat logs) is often considered a prerequisite for downstream dialogue tasks such as response generation and response selection tasks. In this paper, we study a task called reply-to prediction, which can be used to prepare question-answer pairs. Unlike other works, our data comes from the instant messaging (IM) platform where participants could split long sentences into short utterances and send them in multiple messages. We consider a task called message merging task which aims to determine whether those messages need to be merged or not before generating message pairs for reply-to prediction task. The theory behind this task is similar to, yet different from reply-to prediction task in which thistask uses the messages from the same speaker to predict whether these two messages are related or not. We propose a CONTEXT-AOA model to include the context (previous dialogue) as additional input apart from pairwise messages. Our experiments show that our proposed model outperforms both single-turn (pairwise) conversation models and multi-turn (context-aware) conversation models on message merging task and achieves a close performance compares to other multi-turn models on reply-to prediction for manually labeled data and outperforms other models when using heuristic labeled data.