A Computational Approach to Recontextualization in Human Reading Behavior

Abstract

Many aspects of human reading behavior remain unexplained. A human's ability to recontextualize their understanding of a sentence is a particularly difficult phenomena to explain due to the large variety of linguistic and motor factors that affect how and when a human recontextualizes. We propose a computational approach to investigating properties of recontextualization. First, we perform analyses of experimental eye movement data to better understand how well linguistic, motor, and computationally-derived metrics predict a human reader's likelihood to regress througout a text. Finding that metrics associated with a reader's decreasing confidence in their understanding of a sentence are strongly correlated with a reader's likelihood to regress, our investigation suggests a close relationship between reader regressive eye movements and recontextualization that is explainable in part by modern language models. Pursuing this relationship further, we formulate, train, and test a novel bidirectional attention language model that directly incorporates an ability to recontextualize as it receives new input tokens. We find that this novel model performs better than GPT-2 on certain linguistic ambiguities, reflecting that the ability to maintain bidirectional attention allows the model to recontextualize and update its understanding of the sentence. We claim that this ability makes the model more representative of human reading behavior. Testing this claim, we connect our dataset analyses with our novel model by extracting metrics from our bidirectional model and finding that they are indeed predictive of human regressive eye movements. This finding suggests that bidirectional attention might provide the model with a human-like ability to recontextualize while dynamically engaging with a text. We hope that this thesis will provide a foundation for future research into the use of computational language models for understanding and modelling human reading behavior.