Selectional constraints: an information-theoretic model and its computational realization

Area of Science:

• Computational Linguistics
• Cognitive Science
• Psycholinguistics

Background:

• Traditional models of selectional constraints face empirical challenges and struggle with the graded nature of semantic anomaly.
• Existing theories often rely on definitional approaches to word meaning, which can be problematic.

Purpose of the Study:

• To propose a new, minimalist information-theoretic model for selectional constraints.
• To provide a computational account of how selectional constraints can be learned from data.
• To explore the role of selectional constraints in verb meaning acquisition.

Main Methods:

• Developed a model with a generic taxonomic concept representation.
• Formalized selectional constraints probabilistically using observable co-occurrence frequencies.
• Implemented a computational model that learns constraints from natural text.
• Evaluated model predictions against human judgments and syntactic realization patterns.

Main Results:

• The information-theoretic model successfully learns selectional constraints.
• The model's predictions align with human judgments on semantic anomaly.
• The approach accommodates the degree of semantic anomaly and avoids definitional issues.

Conclusions:

• Selectional constraints can be effectively modeled using information theory and co-occurrence data.
• This model offers a learnable and empirically viable alternative to traditional selection restrictions.
• The findings contribute to understanding verb meaning acquisition and computational semantics.