On the Semantics of LM Latent Space: A Vocabulary-defined Approach

In the realm of deep learning, understanding the latent space of language models (LMs) like transformers is crucial for refining their performance and interpretability. However, existing analyses often fall short in providing absolute and model-centric insights into LM semantics, and neglect essential aspects of LM adaption. In response, we introduce a pioneering method called vocabulary-defined semantics, which establishes a fixed reference frame within the LM latent space, ensuring absolute semantic analysis grounded in LM vocabulary. Our approach transcends prior relative analyses, leveraging LM vocabulary for model-centric insights. Furthermore, we propose a novel technique to compute logits, emphasizing differentiability and local isotropy, and introduce a neural clustering module for semantically calibrating data representations during LM adaptation. Through extensive experiments across diverse text understanding datasets, our approach surpasses state-of-the-art methods of retrieval-augmented generation and parameters-efficient finetuning, showcasing its efficacy and broad applicability. Our findings not only shed light on LM mechanics but also offer practical solutions for enhancing LM performance and interpretability.