Conservative Bias in Large Language Models: Measuring Relation Predictions

TL;DR

This paper investigates the conservative bias in large language models during relation extraction, highlighting its frequency, trade-offs, and implications for information loss and model behavior.

Contribution

It introduces the concept of Hobson's choice to quantify conservative bias and systematically evaluates its prevalence across datasets and prompts.

Findings

Conservative bias occurs twice as often as hallucination.

Models often default to No_Relation, leading to information loss.

Semantic similarity measures quantify bias behavior.

Abstract

Large language models (LLMs) exhibit pronounced conservative bias in relation extraction tasks, frequently defaulting to No_Relation label when an appropriate option is unavailable. While this behavior helps prevent incorrect relation assignments, our analysis reveals that it also leads to significant information loss when reasoning is not explicitly included in the output. We systematically evaluate this trade-off across multiple prompts, datasets, and relation types, introducing the concept of Hobson's choice to capture scenarios where models opt for safe but uninformative labels over hallucinated ones. Our findings suggest that conservative bias occurs twice as often as hallucination. To quantify this effect, we use SBERT and LLM prompts to capture the semantic similarity between conservative bias behaviors in constrained prompts and labels generated from semi-constrained and open-ended prompts.