Abstract
By testing with Bidirectional Encoder Representations from Transformers (BERT), this paper investigates whether Large Language Models (LLMs) exhibit compositionality. Compositionality, the principle that complex meanings are systematically built from components and their structure, is central to philosophy, linguistics, and cognitive science. Traditional theories like Jerry Fodor’s Language of Thought Hypothesis (LoTH) conceptualize cognition in terms of rule-governed symbolic operations. In contrast, LLMs use distributed representations. In addition to the success of LLMs in recent years, there are debates on whether LLMs’ success is due to the true capture of compositionality or merely mimicking it. To address this issue, this paper tests BERT’s embedding vectors by investigating whether the model detects silenced structures in elliptical sentences. If BERT encodes the silenced structure with measurable similarity to their explicitly stated counterparts, it would be a sign for compositional understanding. Based on the measurements of cosine similarity and Euclidean distance between embedding vectors, the findings suggest that BERT does exhibit signs of compositionality but in an inconsistent way. These results add to the broader debate on LLMs’ compositionality and indicate the need for further study into the analysis of LLMs’ natural language processing.
Details
Presentation Type
Paper Presentation in a Themed Session
Theme
Communications and Linguistic Studies
KEYWORDS
Philosophy of Mind, Functionalism, Language of Thought, Large Language Models
