Semantic Adaptive Fusion with Trust Signals for Fake News Detection

Abstract

Due to the rapid proliferation of deceptive information disseminated via digital platforms, there is a need to construct scalable, robust and mathematical models to detect fake news automatically. Even though Large Language Models (LLMs) show remarkable ability to solve contextual semantic reasoning tasks, the current state-of-the-art hybrid models require static approaches to feature fusion, which cannot handle different claim structures adaptively. In this paper, we introduce AQFND (Adaptive and Trust-aware Fake News Detector), an end-to-end model that combines dense contextual semantic features provided by a frozen LLM encoder (Qwen2.5) with statistical lexical features (TF-IDF) using a complexity-aware dynamic gating mechanism. With the aim to increase the level of reliability of the decision-making process, we incorporate the idea of a trust-aware inference engine based on Shannon Entropy and confidence thresholding (τ = 0.65). The proposed framework is evaluated on three widely-used benchmark datasets covering complex political claims, extensive fact-checking articles, and entertainment news: LIAR, PolitiFact, and GossipCop. With the LIAR dataset, AQFND attains an F1-score of 76.20% (accuracy 71.86%, ROC-AUC 76.96%, and MCC 0.4215), while on the PolitiFact dataset, AQFND gets an F1-score of 71.31% (accuracy 73.01%, ROC-AUC 81.42%, and MCC 0.4668). Also, on the GossipCop dataset, AQFND has an F1-score of 97.19% (accuracy 97.17%, ROC-AUC 97.85%, and MCC 0.9435).