Distributional Models for Semantic Relations: A Study on Hyponymy and Antonymy
Giulia Benotto · Zenodo (CERN European Organization for Nuclear Research) · 2023
The overall purpose of this research project is to study and explore the potential and limitations of the distributional approach with respect to lexical semantics. Analyzing paradigmatic relations such as synonymy, antonymy, hypernymy, hyponymy and co-hyponymy while using distributional methods could be very interesting and really challenging. For instance, state of the art semantic distributional methods have difficulties in distinguishing these relations. This is because these relations tend to distribute, within the texts, in a very similar way. We first carried out a study about current state of the art distributional measures developed for the recognition of paradigmatic semantic relations, trying to assess the degree of success reported by such measures. We then used the data we gathered to develop new distributional measures apt to classify the hyponymy/hypernymy relation and the antonymy relation. Not only did we prove that distributional methods can indeed be used in order to classify, in a non-supervised way, these relations; we also developed measures, especially conceived for recognizing and classifying hypernymy and antonymy, that performs better than state of the art ones.