In multi-turn dialogue generation, responses are not only related to the topic and background of the context but also related to words and phrases in the sentences of the context. However, currently widely used hierarchical dialog models solely rely on context representations from the utterance-level encoder, ignoring the sentence representations output by the word-level encoder. This inevitably results in a loss of information while decoding and generating. In this paper, we propose a new dialog model X-ReCoSa to tackle this problem which aggregates multi-scale context information for hierarchical dialog models. Specifically, we divide the generation decoder into upper and lower parts, namely the intention part and the generation part. Firstly, the intention part takes context representations as input to generate the intention of the response. Then the generation part generates words depending on sentence representations. Therefore, the hierarchical information has been fused into response generation. we conduct experiments on the English dataset DailyDialog. Experimental results exhibit that our method outperforms baseline models on both automatic metric-based and human-based evaluations.

本文提出一种名为X-ReCoSa的新型对话模型，旨在解决当前设置仅仅依赖于话语级编码器的上下文表示以忽略句子级编码器输出的句子表示而导致的信息丢失问题，通过聚合多尺度上下文信息，将分层信息与响应生成融合，从而有效提高模型表现。

X-ReCoSa:多尺度上下文聚合用于多轮对话生成