Language models follow polarity rules when handling negation in sentences
Polarity Composition Hypothesis: Causal Evidence from Negation in NLI
Computation and Language
Summary
Understanding negation in language is tricky for AI models that try to figure out if one sentence implies another. The authors propose a model called the Polarity Composition Hypothesis that explains how these models process negation. They show that language models mostly behave in ways that match this hypothesis. They also found specific parts inside the models that handle this kind of reasoning, and these parts work across different ways of expressing negation.
What this means in practice
- •For natural language processing engineers: Improve language model components that handle negation for better sentence inference tasks.
- •For ai system integrators: Use identified model subspaces to enhance negation understanding in AI applications across various text forms.
Authors
Nhi Hoai Doan, Tatsuya Hiraoka, Munachiso Samuel Nwadike, Kentaro Inui
Abstract
We investigate how language models process negation in natural language inference (NLI). We propose the \textit{Polarity Composition Hypothesis}, a high-level causal model of this computation. Behavioral evaluations show that language-model predictions largely conform to the proposed hypothesis. Using Distributed Alignment Search, we identify causally effective subspaces aligned with the variables in the polarity-composition hypothesis. These representations transfer across negation forms and partially to a subset of MNLI. We release the dataset and code.