Lexico-Semantic Relation Classification with Multilingual Finetuning

Following the research line of lexico-semantic relation induction based on Language Models (LM) [4, 1, 5], this work analyses how the multilingual fine-tuning of a LM might improve its performance on a relation classification task. The model is trained on a set of word pairs which share a known relation (e.g., Big–small, antonymy). Then, it is used to classify the relation between an unseen word pair. Our hypothesis is that enriching this training with multilingual information benefits the classification of lexico-semantic relations that are common in several languages.