Fine-tuning Semantic Vectors With Semantic Fluency Data

Semantic vectors derived from training on large text corpora (e.g., word2vec, BERT) are widely used as a methodological tool to model similarity of concepts. Recent work has demonstrated that a small amount of human training data can be used to fine-tune these vectors for modeling specific tasks. For example, human ratings of pairwise similarity can be used to estimate a set of dimensional weights, and these weights can improve estimates of human similarity ratings for held-out pairs. We applied this methodology to the semantic fluency task (listing items from a category) and find that category- specific weights can be used to identify the semantic category of a fluency list. The results have methodological implications for modeling retrieval in semantic fluency tasks, estimating semantic representations, and identifying semantic clusters and switches in fluency data.