Investigation of Transitivity Relation in Natural Language Inference

Abstract

Motivation of this work is a data-centric approach of improving model accuracy by improving data quality instead of improving model architecture. The idea is to improve dataset with transitivity relations to help machine learning model learn such dependencies. Alongside with enriching dataset investigate how good is previously trained model in catching such relations. So, basically study can be divided into two main parts investigating dataset and investigating machine learning model trained on such datasets. It was found that the existing model catches transitive dependencies well. It was also investigated that “entailment” relation is more directional that “contradiction” and “neutral”. © 2023 Copyright for this paper by its authors.