Revisiting Word Embeddings in the LLM Era
This work addresses the NLP research community by clarifying the comparative strengths of LLM-induced versus classical embeddings, though it is incremental as it builds on existing embedding methods.
The paper investigates whether the performance gains of word embeddings from Large Language Models (LLMs) are due to scale or inherent differences from classical models like Word2Vec and SBERT, finding that LLMs excel in decontextualized tasks with tighter semantic clustering but classical models often outperform in contextualized sentence similarity.
Large Language Models (LLMs) have recently shown remarkable advancement in various NLP tasks. As such, a popular trend has emerged lately where NLP researchers extract word/sentence/document embeddings from these large decoder-only models and use them for various inference tasks with promising results. However, it is still unclear whether the performance improvement of LLM-induced embeddings is merely because of scale or whether underlying embeddings they produce significantly differ from classical encoding models like Word2Vec, GloVe, Sentence-BERT (SBERT) or Universal Sentence Encoder (USE). This is the central question we investigate in the paper by systematically comparing classical decontextualized and contextualized word embeddings with the same for LLM-induced embeddings. Our results show that LLMs cluster semantically related words more tightly and perform better on analogy tasks in decontextualized settings. However, in contextualized settings, classical models like SimCSE often outperform LLMs in sentence-level similarity assessment tasks, highlighting their continued relevance for fine-grained semantics.