My Account

Poster D37, Wednesday, August 21, 2019, 5:15 – 7:00 pm, Restaurant Hall

Bag of words precedes bag of arguments: The time course of computing argument identity in sentence comprehension

Chia-Hsuan Liao1, Wing-Yee Chow2, Ellen Lau1;1University of Maryland, College Park, 2University College London

Unpredictable words typically elicit larger N400 responses than predictable words, but ‘role-reversed’ sentences (‘The waitress that the customer served…”) are a notable exception. This observation has often been taken to provide insight into the speed or accuracy of argument structure computation. Chow et al. (2015) proposed initial verb prediction was driven by arguments in the same clause as the verb (the ‘bag-of-argument’ mechanism), even though it takes up to 1200 ms for argument role information to impact predictions of the verb (Momma et al., 2015). Here we focus on mapping the time course of one part of this computation that must either precede or co-occur with argument role assignment: identifying which noun phrases are arguments of the clause at all. We did this by extending the standard paradigm to include structures containing a clause boundary like (‘The waitress thought that the customer served…”), and evaluating sensitivity to this boundary on the N400 to the verb in two experiments that varied in stimulus-onset asynchrony. Notably, these experiments were conducted in Mandarin Chinese instead of English because Mandarin allows tight matching of word position across the relevant conditions, and is known to show English-like N400 insensitivity to role reversal (Chow & Phillips, 2013; Chow et al., 2018). Our materials were sentences modified from Chow et al. 2018. In the Baseline condition, the two arguments were still in the same clause (The millionaire ba the servant fired, meaning: the millionaire fired the servant). The predictability of the target verb was 40% based on the result of offline cloze norming. In the Complement condition, the second argument was the subject of a separate complement clause (the millionaire thought the servant fired …), which reduced the predictability of the target verb to 0%. Sentences were presented with RSVP, with 20% of the sentences followed by a comprehension question. In this design, facilitation due to simple associations between nouns and verb would be equal across conditions, and an N400 predictability difference should only be observed if comprehenders have enough time to identify the arguments based on the clause boundary. In Experiment 1 (n=33), with a presentation rate of 800ms/word, we did observe an N400 difference between conditions. This suggests that the grouping of arguments into clauses impacts verb predictions fairly rapidly, perhaps more rapidly than role information (Momma, et al., 2015; Chow et al., 2018). In Experiment 2 (n=38) we aimed to identify a lower time limit for this “bag-of-arguments” mechanism by speeding up the presentation rate, now using 600ms/word. With this slightly faster rate, we no longer observed N400 differences. These findings suggest that 600ms was not long enough for the parser to compute argument identity information; at this point in time, the parser relies on simple word associations, or “bag-of-words” mechanism to predict the verb. It takes up to 800ms for the “bag-of-argument” mechanism to exert its effects, such that an argument outside of the clause domain no longer constrains predictions of the verb in the embedded clause.

Themes: Meaning: Combinatorial Semantics, Syntax
Method: Electrophysiology (MEG/EEG/ECOG)

Back