Context-Dependent Synonym and Concept Extraction for Dialogue Systems Training

Abstract

Modern scenario-based approaches to the construction of dialogue systems require extraction of domain-dependent synonyms and concepts. For example, making rules for rule-based dialogue systems is greatly facilitated by the ability to automatically generate synonyms for keywords. Automatic detection of domain concepts allows to make a list of slots that need to be filled to create service scenarios in goal-oriented dialogues. This paper describes the method of unsupervised synonym and concept extraction from natural language texts. Our method includes context-dependent algorithms (based on applying word embeddings) and does not require using tagged data and external resources.