Poster C26, Friday, August 17, 10:30 am – 12:15 pm, Room 2000AB
Evaluating experiential models of word semantics relative to distributional and taxonomic models
Leonardo Fernandino1, Lisa Conant1, Colin Humphries1, Jeffrey Binder1;1Department of Neurology, Medical College of Wisconsin
Automatic semantic priming in lexical decision is an objective measure of how similar the brain perceives the meanings of the prime and the target word to be. We have previously shown that a model of concept representation based on a set of experiential attributes (CREA) predicts automatic semantic priming for nouns (Fernandino et al., 2016). Here, we replicate that result with different prime-target combinations and evaluate the predictive power of two CREA models relative to models based on word co-occurrence (LSA and HAL) and on taxonomic relations (WordNet). Twenty-six subjects completed the study. Targets consisted of 210 English nouns and 210 matched pseudowords. Primes were 630 real nouns (some nouns also appeared as targets in different trials). All target nouns and their respective primes were previously rated on a set of 65 experiential attributes through Amazon Mechanical Turk (Binder et al., 2016). Words and pseudowords were matched for letter length, bigram and trigram frequency, and orthographic neighborhood size. Primes and targets were matched for word length. Forward associationpriming effects were precluded by excluding any pairs with nonzero values in the USF Free Association Norms. The study consisted of two sessions, approximately one week apart. Each target word was presented once in each session, each time preceded by a different prime. The priming effect for a given target word was calculated as the difference in response times between the two sessions. Each trial started with a central fixation cross (duration jittered 1-2 sec), followed by the prime (150 ms), a mask (hash marks, 50 ms), and the target (2 sec). The prime was presented in lowercase and the target in uppercase letters. Trials were presented in a different pseudorandomized order for each participant, and the order of presentation of the 2 primes for a given target was counterbalanced across participants. Participants were instructed to ignore the prime and make a speeded lexical decision on the target. For each model, semantic similarity between prime and target was estimated as the cosine between the vector representations of the two words. Prime-target pairs represented a continuous distribution of word similarity values ranging from 0 to 1. We evaluated two versions of the experiential attributes model, one including all 65 attributes (CREA-65) and one including only 12 (CREA-12) attributes. We compared these models with LSA, HAL, and three measures of word similarity based on WordNet (path length [WN-PL], Leacock-Chodorow [WN-LC], and Wu-Palmer [WN-WP). All models showed highly significant correlations with priming (all p < .0001, across trials and across participants). Pearson’s r were as follows: CREA-65, .46; CREA-28, .46; LSA, .34; HAL, .37; WN-PL, .33; WN-LC, .35; WN-WP, .39. A repeated-measures T-test across participants showed that the mean correlation with priming was significantly higher for the CREA models than for LSA, HAL, WN-PL, and WN-LC (all p < .005). These results indicate that experiential models of word semantics can predict semantic priming effects more accurately than several widely used models based on word co-occurrence patterns or in human-curated taxonomies, at least for nouns.
Topic Area: Meaning: Lexical Semantics