VerbaNexAI at SemEval-2025 Task 3: Fact Retrieval with Google Snippets for LLM Context Filtering to identify Hallucinations

Anderson Morillo, Edwin Puertas, Juan Carlos Martinez SantosACL 2025

TL;DR

This paper tackles the critical problem of "hallucinations" in Large Language Models (LLMs) by developing systems that can identify fabricated or non-factual information in LLM outputs. They propose two main approaches: one leveraging LLMs with Google snippets for context retrieval and another using traditional NLP techniques, both aiming to improve the reliability of LLM-generated text.

What Problem Does This Paper Address?

Large Language Models (LLMs) are powerful tools, but they have a significant drawback: they can generate convincing-sounding but entirely false information, a phenomenon known as "hallucination." This makes it difficult to trust LLM outputs for factual accuracy. SemEval-2025 Task 3 specifically aims to address this by identifying "hallucination spans" – the exact parts of a generated text that are factually incorrect. The VerbaNexAI team's work focuses on mitigating these hallucinations in English text by incorporating external factual evidence.

Methods

The VerbaNexAI team presents two distinct approaches:

  • LLM-based System: This system uses a multi-stage process. First, it retrieves relevant context by performing Google searches and then semantically ranks the resulting snippets to select the most pertinent information. Next, a specially formatted prompt, optimized using a one-shot chain-of-thought strategy, guides an LLM to extract hallucinations from the generated responses. Finally, data cleaning and token extraction methods refine the identified candidates.
  • NLP-based System: This approach utilizes similar context retrieval and snippet ranking mechanisms as the LLM system. However, instead of relying on an LLM for hallucination detection, it employs traditional NLP techniques. This involves data cleaning, token extraction, and a lexical comparison between the generated text and the top snippets. Part-of-Speech (PoS) analysis is used to filter for important tokens (nouns, proper nouns, verbs, numbers, and others) to avoid false positives.

Datasets

The paper mentions the use of two datasets for English language evaluation:

  • An English test dataset with 50 labeled examples for initial evaluation.
  • A separate test dataset with 150 examples for performance assessment.
  • A validation dataset comprising 154 unlabeled examples, which were evaluated using the platform proposed by Vázquez et al. (2025), employing Intersection over Union (IoU) and Spearman Correlation metrics.

The paper does not mention any newly introduced benchmarks.

Key Insights and Ideas

  • Google Snippets as a Factual Anchor: The core idea of using Google snippets as an external source of factual evidence is a practical approach to grounding LLM outputs. This directly addresses the lack of inherent factual knowledge in some LLM generations.
  • Chain-of-Thought Prompting for LLMs: The successful application of chain-of-thought prompting with one-shot learning for hallucination detection highlights its effectiveness in guiding LLMs to perform complex reasoning tasks.
  • Complementary Strengths of LLM and NLP: The paper demonstrates that while LLM-based systems can achieve competitive results, traditional NLP methods can still offer valuable qualitative insights, suggesting potential for hybrid approaches.
  • Challenges in Snippet Retrieval: The reliance on external web scraping (Google snippets) introduces practical challenges, such as the need for IP rotation and the possibility of unavailable snippets, which can impact system performance.

Key Results

The VerbaNexAI team's LLM-based system achieved competitive results in SemEval-2025 Task 3, ranking 25th out of 41 teams for Intersection over Union (IoU) and 28th out of 41 teams for Spearman correlation on the English dataset. The NLP approach, while not achieving the same quantitative performance, provided valuable qualitative insights. The paper notes that the best-performing model during evaluation was the LLM system using deepseek-r1-distill-llama-70b.

Limitations and Future Work

The authors acknowledge several limitations:

  • Data Availability and Snippet Retrieval: The system's performance is affected by the availability of Google snippets. Issues like IP bans due to excessive requests and unavailable snippets were encountered.
  • Prompt Optimization: While the chain-of-thought approach showed promise, further optimization of prompts could lead to improved efficiency and accuracy.

Future work suggested includes:

  • Improving the snippet extraction system to ensure greater data availability.
  • Optimizing prompts by testing different variations.
  • Exploring enhanced snippet extraction and fine-tuning strategies.
  • Investigating potential biases in Google snippets and LLM responses, and how the filtering process might introduce its own biases.

Summary updated

BibTeX
@inproceedings{morillo2025verbanexai,
  author = {Anderson Morillo and Edwin Puertas and Juan Carlos Martinez Santos},
  title = {{VerbaNexAI at SemEval-2025 Task 3: Fact Retrieval with Google Snippets for LLM Context Filtering to identify Hallucinations}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2025},
}
Figure