Poster D39, Wednesday, August 21, 2019, 5:15 – 7:00 pm, Restaurant Hall
Disentangling semantic association from semantic composition in the LATL
Jixing Li1, Julien Dirani1, Liina Pylkkänen1,2;1New York University Abu Dhabi, 2New York University
Introduction Although the process for composing two meanings (e.g., “coffee cup”) is conceptually different from associating two words together in memory (e.g., “coffee, cup”), the brain regions supporting semantic composition have also been implicated for associative encoding. For example, compositional phrases such as “red boat” elicit an increased left anterior temporal lobe (LATL) activity as compared to non-compositional phrases (e.g., Bemis & Pylkkänen, 2011; Pylkkänen et al., 2014), yet this region is also involved in associative memory tasks such as face-location associations (e.g., Nieuwenhuis et al., 2012). The current study applies distributional semantic models to disentangle semantic association from semantic composition, and tests whether the LATL is indeed sensitive to semantic composition or simply tracks the association between words. Methods 24 right-handed native English speakers (12 female, M=21.8 years) participated in the study. The experiment was a 2x2 design with associative strength (Low, High) and compositionality (List, Comp) as factors. Associative strength between the two words was determined by the cosine similarity score between the word embeddings based on the GloVe model (Pennington et al., 2014). High associative phrases (e.g., “French/France cheese”) have a cosine similarity score greater than 0.3 and low associative pairs (e.g., “Korean/Korea cheese”) have a cosine similarity score lower than 0.15. The composition condition consisted of country adjective and food noun pairs (e.g., “French/Korean cheese”) and the list condition consisted of country noun and food noun pairs (e.g., “France/Korea cheese”). The critical stimuli (i.e., “cheese”) were held constant cross all four conditions. In each trial, participants indicated whether the target picture matched the preceding words. MEG data were recorded at 1000 Hz (200 Hz low-pass filter), noise reduced and epoched from 100 ms before to 600 ms after the onset of the critical word. A main analysis was conducted in the LATL (BA38) which has previously been implicated for both associative encoding and composition. MEG activity was averaged over all sources within the ROI. A non-parametric cluster permutation test (Maris & Oostenveld, 2007) with 10,000 permutations was performed to identify temporal clusters during which the localized activity differed significantly between conditions (p<0.05). Results The 2x2 ANOVA yielded a highly significant cluster for association in the LATL from 227 to 264 ms (p=0.0012), within which the low-association phrases had greater activity than the high-association pairs. A main effect of composition was also found from 313 to 332 ms (p=0.013), where the adjective-noun phrases elicited higher activity than the list conditions. No interaction effect was found for association and composition. We further examined the association and combination activity within subsets of the data and found that within the low-association condition, combinatory effect was significant from 319 to 331 ms (p =0.048); within the list condition, association was significant from 214 to 263 ms (p=0.0014). Conclusion Our results suggest that the LATL is modulated by both semantic association and semantic composition. Specifically, low-association phrases induced increased LATL activity at ~220 ms after the word onset, whereas the combinatory effects came later at ~310 ms.
Themes: Meaning: Combinatorial Semantics, Meaning: Lexical Semantics
Method: Electrophysiology (MEG/EEG/ECOG)