This paper aims to construct a multidisciplinary theoretical framework for online disinformation dissemination and summarizes key features for identification, while evaluating the effectiveness of evaluation-index-based methods in large language models (LLMs)-driven disinformation identification.
Grounded in communication, psychological and sociological theories, a theoretical framework is developed by deconstructing disinformation dissemination into four components: information source, content, release entity and receiver. The system integrates key features used for identifying disinformation in existing research and applications, with the GPT-4 model applied to study the identification of online disinformation. To ensure robust results, this paper compares the identification results of the GPT-4 model with no evaluation indicators; the ERNIE-3.5-128K and GLM-4-PLUS models with and without evaluation indicators; and with the results of the GPT-4 model with evaluation indicators.
The proposed framework highlights the critical role of multidimensional evaluation metrics in enhancing identification accuracy. The experimental results show that the evaluation index constructed in this paper significantly improves the performance of LLMs in identifying disinformation compared to evaluation index-free.
This paper integrates interdisciplinary perspectives to advance the theoretical understanding of disinformation dissemination and provides actionable insights for optimizing automated identification systems. The empirical validation of evaluation metrics offers a scalable methodology for improving LLM-based disinformation identification tools in practice.
