Lightweight Large Model Hallucination Detection: A Non-Neural Network Approach Based on TF-IDF and Wikipedia Evidence Retrieval

This article presents a lightweight large model hallucination detection framework that does not require neural networks. It corely uses TF-IDF and cosine similarity, combined with Wikipedia evidence retrieval to verify factual claims in LLM outputs. This approach compares the credibility performance of three open-source models: Llama-2, Mistral-7B, and Qwen-2. It features lightweight design and strong interpretability, providing a feasible hallucination detection path for resource-constrained scenarios.

The "hallucination" problem of large language models (LLMs) is a major obstacle to their application in fact-sensitive scenarios (such as healthcare and law), as models generate incorrect but seemingly plausible information. Most existing hallucination detection methods rely on neural network models, which have limitations like high computational cost, need for large amounts of labeled data, and difficulty in explaining decision-making bases. Therefore, developing lightweight and interpretable solutions has important practical value.

This approach adopts a three-stage pipeline design:

1. Claim Extraction: Identify factual claims that need verification from LLM outputs;
2. Evidence Retrieval: Query Wikipedia to obtain relevant evidence documents (based on its wide coverage and high structuring features);
3. Similarity Verification: Calculate the matching degree between claims and evidence via TF-IDF vectorization and cosine similarity. The advantages of TF-IDF include interpretability (weights correspond to the importance of vocabulary), efficient computation (no GPU required), and no need for training data.

The experiment compares three mainstream open-source models: Llama-2 (a classic Meta model known for its security), Mistral-7B (a European efficient architecture model), and Qwen-2 (the latest version of Alibaba's Tongyi Qianwen). A hybrid verification strategy is adopted: direct matching (semantic similarity between claims and evidence), context verification (overall consistency between paragraphs and evidence), and multi-source cross-verification (consistency across multiple evidence documents) to enhance verification robustness.

Significance of Lightweight Design: It can run on ordinary servers or edge devices, suitable for resource-constrained enterprises, data-sensitive local deployments, and real-time application scenarios; Interpretability: When marking suspected hallucinations, it can display keyword matching degrees, evidence documents, and similarity values, which is conducive to human-machine collaboration; Trade-offs of Wikipedia: Its advantages are wide coverage and timely updates, while limitations include insufficient coverage in professional fields and possible errors. However, the framework supports replacing with other knowledge sources.

Current Limitations: As a bag-of-words model, TF-IDF cannot capture subtle semantic differences (e.g., easy misjudgment between "Apple Inc." and "apple fruit"); Wikipedia has insufficient coverage of emerging/niche topics. Improvement Directions: Introduce lightweight semantic models to supplement semantic blind spots; integrate multiple knowledge sources to improve coverage; adopt hierarchical verification strategies for claims of different complexities.

This open-source project provides practical references for the community, proving that classic information retrieval technologies still have unique value in the era dominated by neural networks (interpretable, efficient, no training data required). The hallucination problem of large language models is difficult to eliminate in the short term, but lightweight detection tools can reduce risks. Such low-cost and highly interpretable tools will play an important role in the AI security ecosystem.