Statistical sequence and parsing models for descriptive linguistics and psycholinguistics

Abstract

This study shows that using computational linguistic models is beneficial for descriptive linguistics and psycholinguistics. It applies two models to various English genres and learner language: 1) surprisal and 2) a syntactic parser, allowing us to investigate the role of ambiguity and the interplay between idiom and syntax principles. We find that surprisal and ambiguity are higher for learner language, while parser scores and model fit are lower. In addition, the random application of alternations leads to more ambiguous sentences. Failures to generate optimal orderings in the sense of relevance theory, such as nonnative-like utterances by language learners exhibit, increase processing load, both for human and automatic processors. As human and automatic parsing difficulties correlate, we suggest syntactic parsers as psycholinguistic processing models.