Poster B70, Thursday, August 16, 3:05 – 4:50 pm, Room 2000AB

How event probability impacts sentence processing: Modeling N400 amplitudes during reversal anomalies

Milena Rabovsky1, James L. McClelland2;1Freie Universitaet Berlin, 2Stanford University

Introduction. Traditional accounts hold that during sentence processing, syntax is used to build a linguistic structure, and retrieved word meanings are placed into syntactically specified slots within this structure to derive a compositional representation of sentence meaning. However, accumulating evidence calls this perspective into question. For instance, comprehenders sometimes interpret canonical role-reversed sentences (e.g., “The dog was bitten by the man.”) in line with event probabilities instead of syntactic conventions (Ferreira et al., 2003). Another piece of evidence comes from the N400 component which is typically increased in sentences with semantic anomalies, but has been found to be small in sentences with reversal anomalies (e.g., “The fox on the poacher hunted.”; literal translation from Dutch; Van Herten et al., 2005). Such findings have been taken to indicate a temporary “semantic illusion” (of the poacher hunting the fox) consistent with event probabilities. Methods. We use a neural network model implementing a fundamentally different perspective on sentence comprehension to simulate a broad range of N400 effects (Rabovsky, Hansen, & McClelland, 2016). Specifically, in the Sentence Gestalt (SG) model (St. John & McClelland, 1990), word meanings are not retrieved from memory prior to semantic integration; instead incoming words serve as ‘cues to meaning’, updating a probabilistic representation of sentence meaning which is jointly constrained by event probability and syntax, and where syntactic cues may be overridden when event probability constraints are strong. The model’s N400 correlate is the semantic update, i.e. the update in the representation of sentence meaning induced by the new incoming word. To simulate the kind of reversal anomalies described above, we trained the model with sentences presented in Dutch word order, and extended the training environment with scenarios set up to align with the characteristics of the materials used in the target experiment. Results. When presenting the model with sentences corresponding to those in a reversal anomaly experiment, it captures the empirical N400 data. Specifically, simulated N400 amplitudes are only slightly increased in reversal anomaly sentences (“the fox on the poacher hunted.”) as compared to the corresponding control sentences (“the poacher on the fox hunted.”) while they are considerably larger in incongruent sentences (“the poacher on the fox planted.”). Probing the model’s internal representations during the processing of the reversal anomalies reveals that the small N400 effect does not indicate a clear-cut “semantic illusion”. Instead, the conflict between the constraints imposed by word order and event probability induces a state of uncertainty, which begins at the presentation of the second noun and is not resolved by the presentation of the verb. Discussion. Our results reveal a new way of thinking about the small N400 effect in reversal anomalies, shedding light on the possible unresolved state of mind in the initial processing of such sentences. Subsequent controlled processes, possibly reflected in P600 amplitudes, may be required to resolve the conflict between competing cues.

Topic Area: Computational Approaches