The relationship between language and perception is a foundational subject in cognitive science Most languages lack dedicated olfactory vocabularies. Odor descriptions are often ambiguous and that presents a challenge for understanding olfaction through language analyses.
Traditional approaches to investigating odor vocabularies have limitations in terms of dataset size, semantic relationships, and inclusion of comprehensive olfactory words. We address these limitations by leveraging recent advancements in Natural Language Processing (NLP) that range from early models like Word2Vec to contextual models like BERT, and finally to large language models (LLMs) such as ChatGPT and GPT-4. We focus on evaluating the capacity of different generations of language models to capture olfactory-semantic relationships. We meticulously assess four prominent language models—Word2Vec, FastText, BERT, and ChatGPT—using various configurations. We compare the resulting semantic spaces for olfactory vocabulary with three distinct ratings-based datasets, each representing different facets of olfactory-semantic representations. The evaluation includes the famous Dravnieks dataset as well as two novel datasets involving perceptual odor-pair ratings and imagined odor-pair ratings. We show that (1) screening text corpora for odor terms can enable learning word embeddings that resemble human ratings, but that (2) LLMs, such as Chat-GPT, resemble what humans imagine odors to smell like, rather than human odor perception.
The findings of this study shed light on the capabilities of an NLP approach to capture olfactory information and contribute to the understanding of perceptual and semantic representations of odors. Moreover, we show the potential use of AI models as substitutes for human participants in generating olfactory-related responses, providing resources for researchers and practitioners in the field of olfaction.