On the Role of Low-level Linguistic Levels for Reading Time Prediction

It has been shown that complexity metrics, computed by a syntactic parser, is a predictor of human reading time, which is an approximation of human sentence comprehension difficulty. Nevertheless, parsers usually take as input sentences that have already been processed or even manually annotated. We propose to study a more realistic scenario, where the various processing levels (tokenization, PoS and morphology tagging, lemmatization, syntactic parsing and sentence segmentation) are predicted incrementally from raw text. To this end, we propose a versatile modeling framework, we call the Reading Machine, that performs all such linguistic tasks and allows to incorporate cognitive constrains such as incrementality. We illustrate the behavior of this setting through a case study where we test the hypothesis that the complexity metrics computed at different processing levels predicts human reading difficulty, and that when cognitive constraints are applied to the machine (e.g., incrementality), it yields better predictions