A Brain-Based Componential Model of Semantic Representation Correctly Classifies Words into Superordinate Categories Leonardo Fernandino1, Colin J. Humphries1, Lisa L. Conant1, Rutvik Desai2, Jeffrey R. Binder1 1Language
Introduction
Imaging Laboratory, Department of Neurology, Medical College of Wisconsin, Milwaukee, WI 2Department of Psychology, University of South Carolina, Columbia, SC
Examples: Mean Category Ratings
Classification Results: Model Performance
• Classical feature-based theories of concept categorization make little contact with neurobiology or learning mechanisms. • We investigated whether a model based on embodied features, rooted in known brain systems, succeeds in classifying words into superordinate categories. • The model consists of 65 attributes related to sensory, motor, spatial, temporal, affective, social, and cognitive processes (Figure 1; see poster E25 for details on model features and participant ratings).
Methods Materials • 302 English nouns from 10 categories 30 animals (e.g., ant, duck, fish, elephant, turtle) 17 food items (e.g., bread, cheese, chocolate, honey, coffee) 43 human occupations (e.g., actor, author, banker, businessman, soldier) 20 musical instruments (e.g., accordion, banjo, drum, flute, piano) 40 locations (e.g., airport, bar, beach, church, farm) 30 plants (e.g., asparagus, chestnut, dandelion, flower, ivy) 35 social situations (e.g., applause, battle, carnival, funeral, protest) 27 tools (e.g., camera, key, pencil, scissors, umbrella) 20 vehicles (e.g., bicycle, boat, bus, cab, subway) 40 abstract (e.g., advantage, hierarchy, luck, peace, truth)
Ratings Method • Data crowdsourced using Amazon Mechanical Turk • English speakers, US accounts • 1 session = 65 queries for a single target word • 16,743 sessions collected from 1743 participants • Average 28.6 ratings (65-attribute vectors) per word
Figure 1. Ratings for 4 of the semantic categories used in the analysis, averaged across items within each category.
Classification analysis
• Classification implemented through logistic regression • All 65 attribute ratings used as predictors • Leave-one-word-out cross-validation procedure:
• For each category, the model was trained to discriminate between words • •
from that category and all other words Model trained on 301 words and tested on remaining word For each word, the model generated a membership probability for each category, and assigned it to the category with highest probability
Simulating categorical deficits by “lesioning” the model
• Attributes were “lesioned” by setting their ratings to zero during testing • Models evaluated by computing the average Luce’s Choice index for each category
Conclusions • A componential model of word semantics based on embodied features was able to classify previously unseen words into superordinate categories
• “Lesioning” the model produced category-specific deficits in word categorization
• Category-specific semantic impairments observed in stroke patients may be explained by embodied componential models
• Future studies will seek to validate the model through behavioral and neuroimaging experiments
Contact: Leo Fernandino:
[email protected] Sponsored by IARPA grant FA8650-14-C-7357, and by R01 NS033576 and R01 DC10783.
