Presentation

Search Abstracts | Symposia | Slide Sessions | Poster Sessions | Poster Slams

Verbs’ selectional preferences modulate N400 response in sentence processing

Poster B40 in Poster Session B and Reception, Thursday, October 6, 6:30 - 8:30 pm EDT, Millennium Hall

Chiebuka Ohams¹, Shohini Bhattasali¹, Philip Resnik¹; ¹University of Maryland

Introduction: We use naturalistic EEG data to investigate neural responses to selectional preference, a lexical property reflecting how restrictive a verb is about the semantic category of its objects. Prior work shows argument structure guides sentence processing e.g., Boland (2005, 1993); Shapiro et al. (1991), Trueswell et al. (1993), McRae et al. (1998). Here we investigate selectional preference strength (SPS) for direct objects as a top-down predictive cue during sentence processing. Methods: We used preprocessed EEG data from The Alice Datasets (Bhattasali et al., 2020). Participants(n=33) listened to the first chapter of Alice’s Adventure in Wonderland audiobook for approximately 13 mins(2,129 words;84 sentences). 475 verbs in total were identified using the SpaCy POS tagger. Excluding modals, auxiliaries, gerunds, there are 325 verbs attested in the story. We operationalized selectional preferences of a verb as SPS (Resnik, 1996), specifically the Kullback-Liebler divergence between the Wordnet (Miller, 1995) supersense distributions for objects conditioned and not conditioned on the verb. These distributions were approximated using verb-direct object pairs from COCA corpus (Davies, 2008). In contrast to word-based surprisal, SPS represents the degree of constraint placed by the verb on the semantic category of its direct object. We use Eelbrain (Brodbeck et al., 2022) and multivariate temporal response functions (mTRF) (Brodbeck & Simon, 2020) for data analysis. In the baseline model, we controlled for acoustics, word frequency, sequential processing (using a 5gram language model), and a categorical predictor for verbs. The second model had the same predictors plus SPS.Results: In a mass-univariate related-measures t-test, we observe a significant difference between the model with SPS predictor and the baseline mTRF model. Additionally, in a spatio-temporal cluster based t-test we observe a significant negative peak at 350 ms. Discussion: Are expectations about argument structure tracked differently versus other kinds of expectations? Bhattasali & Hale (2019) show when verbs are encountered, processing related to selectional constraints is localized differently from processing related to other expectations. Here we ask, is there also evidence of a distinct role for those verbs’ constraints based on the time course of processing? Our results suggest yes, since SPS of verbs is clearly providing additional predictive information. The N400 response we observe potentially involves information integration: SPS measures the quantity of information a verb provides about the semantic category of its object; higher SPS indicates the verb is more selective about its objects’ semantic category and provides additional information about what that upcoming object could be, in contrast to verbs with low selectional preference strength. This view can be interpreted as consistent with prior work connecting higher information integration cost with N400 effects (Lau et al. 2008, Frank et al., 2013). Alternatively, with highly selective verbs we could be observing preactivation for the object as reflected by the N400 (Chow et al., 2016; DeLong et al., 2005). Conclusion: Our results suggest that implicit knowledge about the semantic class of verbs’ objects plays a distinct role during sentence processing and selectivity of verbs gives rise to an N400 response.

Topic Areas: Computational Approaches, Syntax