You are viewing the SNL 2017 Archive Website. For the latest information, see the Current Website.

Poster C54, Thursday, November 9, 10:00 – 11:15 am, Harborview and Loch Raven Ballrooms

Direct Encoding of Semantic and Orthographic Neighborhood Reveals The Organization of Lexical Access

Jona Sassenhagen1, Benjamin Gagl1,2, Christian J. Fiebach1,2;1Goethe University Frankfurt, 2IDeA Center for Indidivudal Development and Adaptive Education, Frankfurt

Many visual word recognition theories propose a more or less parallel sequence of processes mapping from (orthographic, phonological) word form to lexical and semantic representations. For example, Interactive Activation models (e.g. McClelland & Rumelhart, 1981) propose that individual letters contribute bottom-up activation to corresponding word forms, each activating the corresponding semantic features; for this reason, the word "HOUSE" also activates semantic features like "small", "grey", "likes cheese" because "MOUSE" is a close word-form neighbor, at a Levenshtein Distance/LD (Yarkoni et al., 2008) of 1. Typically, such models are studied in context/target pair studies (e.g. priming), or by comparing words with rich vs. sparse neighborhoods. We instead directly model such spreading-activation effects to explain the full time course of brain activity elicited by reading individual words. Specifically, we predict brain activity by assuming that words with similar word forms and words with similar meanings elicit similar brain responses, and compare predictions to real EEG data from a large word reading study (Dufau et al., 2015; 960 words, n=75). First, we construct pairwise (word-to-word) distance matrices based on either LD or semantic similarity. Orthographic distance allows us to predict early brain activity (before 200ms; r > .2, p < .01), but also brain activity in the N400 time window (~350ms; r > .3, p < .01). Semantic distance allows us to predict only late activity (>300ms; r > .3, p < .01). Then, we model spreading-activation of semantic features under partial activation by word forms that are orthographically similar to the target word. For this we sum up high-dimensional semantic features (W2V word vectors; Mikolov, 2012) for each word weighted by their respective LD to the actually presented word; that is, we aggregate the semantic features of orthographically similar words. This model predicts brain activity in an intermediate window (150-400ms). Importantly, after ~400ms, this neighborhood-based model does not add to the predictive success of a semantics-only model (built on the meaning of just the target word), indicating only one word remains activated in this time. Thus, with word-to-word distance-based encoding, we directly delineate: (1), an early time window where lexical access is initiated. Here, word forms are (partially) activated by (partial) orthographic matches; these in turn activate semantic features (ca. 250-450ms). (2), a later time window (>400ms) where only one word meaning is represented and lateral-inhibition-like processes have suppressed irrelevant partially-activated concepts. (3) Finally, we found that words with similar word forms elicit similar brain activity well into the N400 time window, indicating that even at latter stages of word reading, spreading activation-like processing occurs in a part of semantic memory organized by both meaning and form. Our findings support parallel-processing models, constraining them in some parameters (e.g. establishing lower bounds on semantic memory processes sensitive to orthographic neighborhood). This opens up new venues for studying the time courses of orthographic processing, semantic/orthographic neighborhood spreading-activation, and exclusive lexical access, while avoiding crucial confounds.

Topic Area: Meaning: Lexical Semantics

Back to Poster Schedule