Meeting Support System with Topic Estimation by Embeddings via Predicate-Argument Structure Analysis

Abstract

We develop a system that supports meetings in business operations interactively. The goal of the system is to provide insights from real-time conversations through natural language understanding of utterances. Our system allows developers to design a specific pipeline for different forms of meetings such as discussions and daily scrums. The pipeline can include various natural language processing techniques including predicate argument structure analysis. In order to apply natural language processing techniques to utterances, we need a topic estimation method that is applicable to varying topics and can be constructed from small set of examples. To deal with the above problem, we developed a score-based topic estimation method, which utilizes predicate argument structures and word embeddings. This paper demonstrates the proposed method successfully improves estimation performance from only small sets of topic utterances.