Causal Inference-Guided Bias Correction and Hallucination Suppression for Trustworthy Text Summarization

Abstract

We propose a bias correction and illusion suppression method that integrates causal inference. This method treats the summary generation process as a conditional generation decision influenced by potential confounding factors. By constructing a target distribution for causal correction, it weakens the systematic influence of corpus distribution and narrative style on information selection. Furthermore, it introduces a reweighting strategy based on propensity scoring to calibrate bias-related conditions during the learning phase, improving the consistency and robustness of the generation mechanism across different contexts. To reduce the hallucination risk caused by unfounded completion and over-certainty, the method explicitly establishes the alignment relationship between the summary statement and the input evidence fragments during generation and incorporates consistency constraints to make the output more inclined towards facts and entity information supported by the original text. Simultaneously, the method constructs semantically equivalent input variants through counterfactual perturbations and uses distribution consistency constraints to suppress the model's sensitivity to non-causal surface factors, thereby reducing generation bias caused by style differences and rewriting. This paper compares and analyzes the proposed method with related methods under a unified evaluation index system. The results show that the proposed method performs more balancedly in terms of semantic similarity, factual consistency, and entity alignment, and can effectively control the risks associated with hallucination. This verifies the role of causal correction and evidence consistency constraints in improving the credibility and interpretability of the abstract.