Cognitively Plausible Computational Models of Lexical Processing Can Explain Variance in Human Word Predictions and Reading Times
Abstract
Lexical processing times can yield valuable insights about structure in language and the cognitive processes that enable the use of language. Being able to estimate lexical processing times enables us to estimate readability and reading times of any text.
It has been claimed that lexical processing times of words are influenced by word occurrence frequencies as well as the context it appears in (McDonald and Shillcock 2001; Baayen 2010). The context might be important because of predictive processes that enable quicker lexical processing (Christiansen and Chater 2016).
In the present paper, the effects of morphosyntactic predictions on lexical processing times are investigated using two computational models. These computational models are trained to predict upcoming part-of-speech tags based on preceding part-of-speech tags and their predictions are compared with human predictions and human reading times from the PROVO corpus (Luke and Christianson 2018).
A recurrent neural network is able to explain variance in human prediction errors whereas the Rescorla-Wagner model performs less well. The Rescorla-Wagner prediction associations do however explain more variance in human reading times.
Moreover, the Rescorla-Wagner model associations explain more variance in gaze durations than human prediction errors. The human prediction errors and the Recorla-Wagner model associations combined explain most variance (Adj.
R2=0.7192=0.719^2 = 0.719) in reading times, which indicates that the part-of-speech tag-based Rescorla-Wagner model associations contain complementary information to explicit human predictions about lexical processing times.