Antonym Disambiguation for a German-language Conversational Intelligent Tutoring System

Published: 01 Jan 2021, Last Modified: 16 Sept 2024PerCom Workshops 2021EveryoneRevisionsBibTeXCC BY-SA 4.0
Abstract: Conversational agent systems often rely on semantic similarity measures for capturing the meaning of the user's textual input. Among the popular similarity measures are those based on word embeddings. State-of-the-art solutions show good results at measuring the semantic relatedness, but can have problems with accurately estimating the semantic similarity. Antonyms are an example of such words that despite being semantically related have opposite meanings. In this work, related to the development of a conversational intelligent tutoring system for students in nursing subjects, we address the problem of antonyms disambiguation by performing an empirical evaluation of the performance of different scores for German antonyms and synonyms detection. The scores are calculated based on word embeddings-based similarity measure. The results show that the investigated scores are not significantly different, indicating that only the relatedness instead of the meaning of the words was captured. This poses a problem for the development of a conversational agent that needs to capture the meaning of the input text. To address this problem, we discuss strategies for improving the performance by integrating knowledge about the semantic structure of the words.
Loading