SCAKE: Simultaneous Contextualization and Keyword Extraction

Abstract

In utterance processing, it is crucial to refer to the context. However, the dominant dialogue systems pay attention to only one input sentence, ignoring a whole dialogue context. This paper proposes SCAKE, a method that dynamically estimates context from ongoing dialogue. The characteristic of SCAKE is that it simultaneously estimates keywords in the dialogue and reflects them on context estimation, which makes it possible to resolve mutual dependence between the context and interpretation. Experiment results showed that SCAKE achieved better context estimations and keyword extractions, improving over existing algorithms.