Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Jungle tiger, tiger jungle: Tracking the feature constellation of phrases and their constituent parts using MEG during serial and parallel presentation
Poster Session C, Saturday, September 13, 11:00 am - 12:30 pm, Field House
Simone Krogh1, Liina Pylkkänen1; 1New York University
A hallmark of human language is the ability to construct complex meanings from simpler pieces using hierarchical structures. Noun-noun compounds exemplify this at a basic level: Different pairings of the same lexical material yield distinct concepts like VALLEY TIGER (animate) and TIGER VALLEY (inanimate) based on the head noun (Kamp & Partee, 1995). Unlike prior research focusing on the combinatory operations themselves (e.g., Maran et al., 2022; Pylkkänen, 2019), we ask how composition affects the neural representations of individual lexical items. Specifically, we use animacy decodability as a proxy for how neural representations evolve depending on functional roles (TIGER as head noun: CHINESE VALLEY TIGER; TIGER as nominal modifier: CHINESE TIGER VALLEY) and the type of initial modifier (MARBLE VALLEY TIGER; a representation of an animal). The test data for our logistic regression classifier are brain responses to our all-caps, low-probability three-word phrases and corresponding controls with ungrammatical modifier ordering (e.g., VALLEY CHINESE TIGER) during magnetoencephalography (MEG) recordings. All stimuli are shown word-by-word in Rapid Serial Visual Presentation (RSVP) as well as flashed all-at-once in Rapid Parallel Visual Presentation (RPVP; Snell & Grainger, 2017). RPVP has been widely used in recent years to study Sentence Superiority Effects, i.e., facilitated processing for structured vis-à-vis unstructured representations. Regardless of presentation mode, each trial stimulus is followed by a briefly flashed target stimulus that either matches or differs from the target stimulus by one word. The training data for our classifier are brain responses to individual animal and location exemplars (e.g., TIGER and VALLEY) recorded in a separate block. Once trained, classifier accuracy on RSVP trials over time is assessed word-by-word to determine if the animacy of individual lexical items and/or the entire phrase is represented neurally. The relative decodability of nominal modifier vs. head noun animacy allows us to assess if hierarchical structure impacts the processing depth of lexical items. MARBLE VALLEY TIGER-trials, with their animate-to-inanimate conceptual shift, further allows us to test animacy decodability when animacy differs between the head noun and full constituent. Exploratorily, we examine temporal decodability of animacy in RPVP trials and spatial decodability of animacy in both presentation modes. As expected, our online behavioral pilot (n = 25) reveals clear instances of a structural superiority effect across presentation modes, a sanity check that supports the validity of our experimental design. MEG data collection and analysis are underway (projected n = 35), with expected completion by September. If the parser is sensitive to hierarchical structure, we expect successful classification only for head nouns, not nominal modifiers. This would indicate that neural representations of lexical items are not accessed in their entirety unless necessary. As for MARBLE VALLEY TIGER-trials, we expect the neural data to first reflect the animacy feature of the head noun and then of the full constituent if processing proceeds in an incremental fashion (Schumacher, 2013). Should the temporal decoding accuracies pattern differently across RSVP and RPVP, this would suggest that input seriality constrains the feature constellation of phrases and their constituent parts.
Topic Areas: Syntax and Combinatorial Semantics,