Using Neural Networks to Generate Inferential Roles for Natural Language

Area of Science:

• Computational Linguistics
• Cognitive Science
• Artificial Intelligence

Background:

• Neural networks are widely applied to linguistic studies, including phonology, morphology, syntax, and semantics.
• The precise requirements for neural network models to account for semantic understanding, particularly of complex expressions like sentences, remain unclear.
• A key aspect of semantic understanding may involve predicting subsequent sentences based on a given sentence, defining its inferential role.

Purpose of the Study:

• To investigate whether a neural network model can be trained to generate inferential roles for sentences.
• To empirically evaluate the model's performance in predicting sentence entailments.
• To compare human and model-generated entailments and explore theoretical implications for semantic cognition.

Main Methods:

• Training a tree-structured neural network model on the Stanford Natural Language Inference (SNLI) dataset.
• Evaluating the model's entailment prediction accuracy on a held-out test set.
• Conducting a study comparing human plausibility ratings of model-generated versus human-generated entailments.

Main Results:

• The neural network model demonstrated the ability to generate simple inferential roles.
• The model achieved measurable accuracy in predicting sentence entailments on unseen data.
• Human plausibility ratings provided insights into the quality of model-generated entailments.

Conclusions:

• Understanding linguistic expressions involves the ability to draw inferences, suggesting inference should be central to semantic models.
• Neural network models can be trained to capture aspects of inferential roles in language.
• This work proposes a revision to semantic theories, emphasizing inferential capabilities over simple representational mapping.