Running summarizing tools on a new article

Abstract: To address the hallucination problems such as factual errors, citation forgery, parameter bias, and logical inconsistencies that easily arise in generative writing models for generating power technology reports, dispatch operation briefings, equipment defect analyses, load forecast descriptions, and procedure summaries, this paper proposes a multi-source knowledge verification-based hallucination mitigation algorithm for generative writing models. This method decomposes the generated text into sentence-level factual statements, extracts equipment entities, operating times, numerical indicators, fault events, procedure clauses, and citation information, and retrieves candidate evidence from the power grid knowledge base, dispatch procedures, equipment ledgers, defect logs, academic databases, and local document repositories. Furthermore, a joint scoring mechanism based on semantic consistency, evidence coverage, source credibility, cross-source support, timeliness, and citation traceability is constructed to assess the hallucination risk of text fragments. For high-risk segments, the system completes closed-loop correction through partial rewriting, evidence supplementation, citation replacement, and deletion of low-confidence content. Experimental results show that the proposed method outperforms the original LLM, traditional RAG, self-consistency checks, and single-source verification methods in tasks such as dispatch report generation, equipment fault description, load forecast interpretation, and procedure summary. The average hallucination rate is reduced from 26.7% to 8.1%, the fact accuracy rate is increased from 72.6% to 91.4%, and the citation matching rate is increased from 59.8% to 88.6%. The research results demonstrate that multi-source knowledge verification can effectively enhance the factual consistency, traceability, and engineering usability of generative writing models in power industry text generation.