Causality and signalling of garden-path sentences

Area of Science:

• Mathematical foundations of language processing
• Computational linguistics and psycholinguistics
• Application of sheaf theory to natural language

Background:

• Sheaves and presheaves are mathematical tools for organizing data.
• Presheaves have previously modeled lexical ambiguity in natural language.
• Syntactic ambiguity and garden-path sentences pose challenges for current models like surprisal.

Purpose of the Study:

• To model syntactic ambiguities using presheaves.
• To investigate the phenomenon of garden-pathing in human language parsing.
• To develop a new predictive measure for parsing difficulty.

Main Methods:

• Utilizing presheaves to represent syntactic ambiguities.
• Computing the 'degree of signalling' using probabilities from BERT (Bidirectional Encoder Representations from Transformers).
• Evaluating the model's predictions on two psycholinguistic datasets.

Main Results:

• The degree of signalling successfully distinguishes between easy and hard garden-path sentences.
• The new measure shows a larger garden-path effect compared to surprisal in one dataset.
• Improved prediction accuracies for human reading times in syntactic ambiguity resolution.

Conclusions:

• Presheaves offer a robust framework for modeling syntactic ambiguity.
• The degree of signalling is a more effective predictor of parsing difficulty than surprisal.
• This work advances computational models of human language processing.