Leveraging Context for Perceptual Prediction Using Word Embeddings.

Area of Science:

• Cognitive Science
• Artificial Intelligence
• Computational Linguistics

Background:

• Word embeddings from large language corpora are used to represent linguistic meaning.
• Debate exists on whether these embeddings capture perceptual qualities of concepts, impacting theories of embodied semantics.
• Inferring perceptual properties from language alone could suggest language aids conceptual knowledge acquisition.

Purpose of the Study:

• To investigate if Transformer-based language models, using contextualized word embeddings, improve prediction of perceptual concept qualities.
• To assess the ability of word embeddings to predict human ratings of concept brightness and shape.
• To examine the impact of context on perceptual prediction for both noun and adjective-noun phrase embeddings.

Main Methods:

• Generated decontextualized and contextualized Word2Vec and BERT embeddings for a large set of concepts.
• Compared the predictive performance of these embeddings against human ratings for concept brightness and shape.
• Conducted two experiments focusing on noun representations and adjective-noun phrase representations.

Main Results:

• Shape properties were predicted with high accuracy from word embeddings.
• Brightness properties were predicted with more modest accuracy.
• Adding context to word embeddings showed a limited impact on predicting perceptual features for both nouns and adjective-noun phrases.

Conclusions:

• Word embeddings effectively encode shape information but are less effective for brightness.
• Contextualization in language models has a limited effect on predicting perceptual qualities from word embeddings.
• Findings contribute to understanding the scope of embodiment in human conceptual processing and language model interpretability.