Language models are not naysayers: An analysis of language models on negation benchmarks

TL;DR

This paper evaluates whether current large language models can effectively understand and handle negation, revealing significant limitations in their ability to process this fundamental linguistic feature.

Contribution

It provides a comprehensive analysis of LLMs' performance on negation benchmarks, highlighting their insensitivity and reasoning failures regarding negation.

Findings

LLMs show insensitivity to negation presence

They struggle with lexical semantics of negation

They fail to reason correctly under negation

Abstract

Negation has been shown to be a major bottleneck for masked language models, such as BERT. However, whether this finding still holds for larger-sized auto-regressive language models (``LLMs'') has not been studied comprehensively. With the ever-increasing volume of research and applications of LLMs, we take a step back to evaluate the ability of current-generation LLMs to handle negation, a fundamental linguistic phenomenon that is central to language understanding. We evaluate different LLMs -- including the open-source GPT-neo, GPT-3, and InstructGPT -- against a wide range of negation benchmarks. Through systematic experimentation with varying model sizes and prompts, we show that LLMs have several limitations including insensitivity to the presence of negation, an inability to capture the lexical semantics of negation, and a failure to reason under negation.