Evaluating Contextualized Vectors from both Large Language Models and Compositional Strategies”
In this article, we compare contextualized vectors derived from large language models with those generated by means of dependency-based compositional techniques. For this purpose, we make use of a word-in-context similarity task. As all experiments are conducted for the Galician language, we created a new Galician evaluation dataset for this specific semantic task. The results show that compositional vectors derived from syntactic approaches based on selectional preferences are competitive with the contextual embeddings derived from neural-based large language models.
keywords: Large Language Models, Contextual Word Embeddings, Comositionality, Semantic Similarity, Selection Preferences, Syntactic Dependencies