A time-specific dialogue model with reranking methods using a measure of time-dependency

Abstract

Chat-oriented dialogue systems have been drastically improved by large-scale encoder-decoder models because large-scale datasets enable the systems to generate various responses. In recent years, it has been found that dialogue systems can improve user impressions by generating responses specific to various types of information. One type of information is time, but it is not known whether time-specific responses improve user impressions like other types of information because time information has no content. In this paper, we examine whether a time-specific dialogue model (time-specific dialogue model) can improve user impressions. Specifically, we propose a quantitative measure of the time-dependency of responses and construct the time-specific dialogue model by reranking the responses using the proposed measure. We verified the effectiveness of the time-specific dialogue model through subjective evaluation.