Transformer-based large language models (LLMs) rely on contextual embeddings which generate different (continuous) representations for the same token depending on its surrounding context. Nonetheless, words and tokens typically have a limited number of senses (or meanings). We propose multi-sense embeddings as a drop-in replacement for each token in order to capture the range of their uses in a language. To construct a sense embedding dictionary, we apply a clustering algorithm to embeddings generated by an LLM and consider the cluster centers as representative sense embeddings. In addition, we propose a novel knowledge distillation method that leverages the sense dictionary to learn a smaller student model that mimics the senses from the much larger base LLM model, offering significant space and inference time savings, while maintaining competitive performance. Via thorough experiments on various benchmarks, we showcase the effectiveness of our sense embeddings and knowledge distillation approach.

Learning word embeddings is an essential topic in natural language processing. Most existing works use a vast corpus as a primary source while training, but this requires massive time and space for data pre-processing and model training. We propose a new model, HG2Vec, that learns word embeddings utilizing only dictionaries and thesauri. Our model reaches the state-of-art on multiple word similarity and relatedness benchmarks. We demonstrate that dictionaries and thesauri are effective resources to learn word embeddings. In addition, we exploit a new context-focused loss that models transitive relationships between word pairs and balances the performance between similarity and relatedness benchmarks, yielding superior results.