Conceptual Combination in Large Language Models: Uncovering Implicit Relational Interpretations in Compound Words...

Area of Science:

• Cognitive Science
• Computational Linguistics
• Psycholinguistics

Background:

• Large language models (LLMs) are explored as potential models of human semantic processing.
• Conceptual combination, crucial for understanding compound words, is a key area for evaluating semantic models.

Purpose of the Study:

• To investigate the capacity of BERT-base and Llama-2-13b large language models (LLMs) to capture the implicit meaning of compound words.
• To assess how LLMs represent the decomposition and semantic relations within compound words, comparing existing and novel forms.

Main Methods:

• Utilized LLMs to generate contextualized embeddings for compound words and their plausible paraphrased interpretations.
• Analyzed the relationship between the plausibility of paraphrases (rated by humans) and the changes in word embeddings.

Main Results:

• A decrease in embedding change correlated with higher paraphrase plausibility for both existing and novel compounds.
• LLMs showed weaker performance on novel compounds compared to existing ones, with older distributional models sometimes outperforming them.

Conclusions:

• LLMs offer a promising computational tool for approximating the cognitive structure of compound words, aligning with psycholinguistic theories.
• Further research is needed to enhance LLM capabilities for understanding novel conceptual combinations and implicit semantic relations.