Proceedings of the Annual Meeting of the Cognitive Science Society

Recurrent neural network (RNN) models of sentence process-ing have recently displayed a remarkable ability to learn as-pects of structure comprehension, as evidenced by their abilityto account for reading times on sentences with local syntac-tic ambiguities (i.e., garden-path effects). Here, we investi-gate if these models can also simulate the effect of semanticappropriateness of the ambiguity’s readings. RNN-based esti-mates of surprisal of the disambiguating verb of sentences withan NP/S-coordination ambiguity (as in ‘The wizard guards theking and the princess protects ...’) show identical patters to hu-man reading times on the same sentences: Surprisal is higheron ambiguous structures than on their disambiguated counter-parts and this effect is weaker, but not absent, in cases of poorthematic fit between the verb and its potential object (‘Theteacher baked the cake and the baker made ...’). These resultsshow that an RNN is able to simultaneously learn about struc-tural and semantic relations between words and suggest thatgarden-path phenomena may be more closely related to wordpredictability than traditionally assumed.