In sentence comprehension, it is widely assumed (Gibson 2000, Lewis & Vasishth, 2005) that the distance between linguistic co-dependents affects the latency of dependency resolution: the longer the distance, the longer the retrieval time (the distance-based account). An alternative theory of dependency resolution difficulty is the direct-access model (McElree et al., 2003); this model assumes that retrieval times are a mixture of two distributions: one distribution represents successful retrieval and the other represents an initial failure to retrieve the correct dependent, followed by a reanalysis that leads to successful retrieval. The time needed for a successful retrieval is independent of the dependency distance (cf. the distance-based account), but reanalyses cost extra time, and the proportion of failures increases with increasing dependency distance. We implemented a series of increasingly complex hierarchical Bayesian models to compare the distance-based account and the direct-access model; the latter was implemented as a hierarchical finite mixture model with heterogeneous variances for the two mixture distributions. We evaluated the models using two published data-sets on Chinese relative clauses which have been used to argue in favour of the distance account, but this account has found little support in subsequent work (e.g., J\"ager et al., 2015). The hierarchical finite mixture model, i.e., an implementation of direct-access, is shown to provide a superior account of the data than the distance account.
View on arXiv