A Principled Approach to Feature Selection in Models of Sentence Processing

Area of Science:

• Psycholinguistics
• Computational Linguistics
• Cognitive Science

Background:

• Cue-based memory retrieval is a key framework for human language comprehension.
• General retrieval cues may not fully explain processing difficulties like illusions of plausibility.
• Handpicking lexically specific features for retrieval is challenging and lacks principled methods.

Purpose of the Study:

• To develop a principled method for creating lexically specific retrieval cues.
• To improve models of human language comprehension by incorporating distributed lexical features.
• To quantitatively compare different parsing models.

Main Methods:

• Utilized word embedding methods to generate distributed lexical feature representations.
• Developed distributed retrieval cue vectors to encode retrieval-relevant information.
• Calculated the similarity between feature and cue vectors as a measure of plausibility.
• Analyzed eye-tracking data to correlate plausibility with total reading times.

Main Results:

• The similarity between feature and cue vectors significantly predicted total reading times.
• The proposed method successfully captured retrieval interference effects.
• Distributed lexical features provide a more nuanced understanding of dependency resolution.

Conclusions:

• Distributed lexical features and cue vectors offer a principled and effective approach to modeling language comprehension.
• This method enhances the predictive power of parsing models for reading times.
• Facilitates quantitative comparisons between diverse computational models of language processing.