Triple-R: Automatic Reasoning for Fact Verification Using Language Models

Mohammadamin Kanaani, Sajjad Dadkhah, Ali A. GhorbaniCOLING 2024

TL;DR

This paper introduces Triple-R, a novel framework for automatic fact verification that addresses the limitations of existing methods by incorporating external web knowledge and generating human-understandable explanations. By combining retrieval, ranking, and reasoning components, Triple-R aims to improve transparency and accuracy in fact-checking, making AI-driven verification more trustworthy.

What Problem Does This Paper Address?

The proliferation of online social media has made it a primary news source for many, but it also serves as a fertile ground for misinformation and fake news. While manual fact-checking is crucial, it's a resource-intensive task. Existing automated fact-verification systems often struggle with a lack of transparency, failing to provide clear explanations for their decisions, which erodes user trust. Furthermore, these systems often lack the ability to leverage external, up-to-date information, which is essential for verifying claims about rapidly evolving events. This paper tackles these challenges by proposing a framework that not only verifies claims but also explains why a claim is deemed true or false, using the vast resources of the internet as an external knowledge base.

Methods

The Triple-R framework operates in three stages:

  • Retriever: This component leverages the internet as an external knowledge source. It uses search engines (specifically the Bing Search API) to find relevant documents for a given claim by searching for the claim verbatim. The retrieved web pages are then scraped and divided into paragraphs.
  • Ranker: To filter out noise and identify the most pertinent information, this stage scores the retrieved paragraphs based on their relevance to the claim. It employs a hybrid approach combining a term-based model (Okapi BM25 algorithm) and a neural network model (DistilBERT). The DistilBERT model is used in a zero-shot manner to encode claims and paragraphs, and their similarity is calculated using an inner product. The Okapi BM25 algorithm, a well-established information retrieval method, is also incorporated to enhance relevance scoring. A weighted combination of these scores is used to select the top-k most relevant paragraphs as evidence.
  • Reasoner: This is the core of the explainability aspect. It utilizes a causal language model (specifically, a fine-tuned Llama 2 7b-chat model) to generate both a natural language reason and a label (true/false) for the claim, based on the retrieved evidence. To overcome the challenge of datasets lacking explanations, the authors propose a method to create a synthetic dataset. This involves using a large language model (GPT-3.5-Turbo) to generate reasons for claims based on provided evidence. For claims where the model initially fails to predict the correct label, a "rationalization" technique is employed, where the model is given the correct label as a hint to generate a reason. This process allows for the creation of a dataset suitable for fine-tuning the language model for both reason and label generation. The fine-tuning process focuses on minimizing the loss for reason and label tokens.

Datasets

The primary dataset used for evaluation is the LIAR dataset (Wang, 2017). This dataset consists of claims categorized into six labels: true, mostly true, half true, barely true, false, and pants-on-fire. The claims are sourced from the PolitiFact website. For the purpose of the Retriever, the search engine was restricted to retrieve documents only from the PolitiFact domain.

A synthetic dataset for reason generation was created using GPT-3.5-Turbo, leveraging the evidence and claims from the LIAR dataset.

Key Insights and Ideas

  • Leveraging the Web as a Dynamic Knowledge Source: The paper highlights the critical need to access up-to-date information, especially for claims related to current events. Using the internet as an external knowledge base is a robust strategy to overcome the limitations of static training datasets.
  • Hybrid Ranking for Enhanced Evidence Selection: Combining a statistical method like Okapi BM25 with a neural network model like DistilBERT provides a more comprehensive approach to ranking evidence, balancing keyword matching with contextual understanding.
  • Synthetic Data Generation for Explainability: The approach to creating a synthetic dataset with generated reasons is a key innovation. This addresses the common problem of lacking explainable data, enabling the fine-tuning of language models for generating human-understandable justifications.
  • Rationalization for Improved Reason Generation: The "rationalization" technique, where the model is guided by the correct label to generate a reason, is a clever way to improve the model's ability to explain its decisions, particularly for challenging cases.
  • Smaller Models Can Outperform Larger Ones: The study demonstrates that a fine-tuned smaller model (Llama 2 7b-chat) can outperform a larger, more general-purpose model (GPT-3.5-Turbo) on this task, suggesting the power of targeted fine-tuning.

Key Results

  • The Triple-R framework, specifically the fine-tuned Llama 2 (7b) model, achieved an accuracy of 42.72 ± 0.19% on the LIAR dataset.
  • This performance surpassed other models in the ablation study, including BERT without evidence (29.3% accuracy) and BERT with evidence (40.29% accuracy).
  • The fine-tuned Llama 2 (7b) model also outperformed GPT-3.5-Turbo (37.1% accuracy) in terms of accuracy, precision, recall, and F1 score, despite having significantly fewer parameters.
  • The rationalization method was crucial for improving reason generation, enabling the model to generate reasons for nearly 97.8% of the dataset, compared to around 36% without it.
  • Augmenting context with evidence significantly enhanced fact verification system effectiveness, as shown by the BERT model results.

Limitations and Future Work

  • Limitations: The paper notes that the GPU VRAM was barely sufficient for fine-tuning the Llama 2 model, necessitating the use of parameter-efficient techniques like QLoRA. While the model achieved good accuracy, there were instances where the model mainly predicted labels as next-sentence predictions, and only once did it generate a label not covered by predefined labels. In the test set, 19 samples out of 1283 did not return any search engine results, but Llama 2 still demonstrated comparable accuracy on these samples.
  • Future Work: The authors suggest exploring more innovative methods for generating queries that are directly related to the claim's content, rather than just verbatim searches. They also emphasize the need for techniques to evaluate and select the most appropriate answer when language models generate multiple responses, ensuring they incorporate more evidence and provide better explanations. Finally, they plan to focus on developing methods to verify the reliability and credibility of the sources retrieved from the web.

Summary updated

BibTeX
@inproceedings{kanaani2024triple,
  author = {Mohammadamin Kanaani and Sajjad Dadkhah and Ali A. Ghorbani},
  title = {{Triple-R: Automatic Reasoning for Fact Verification Using Language Models}},
  booktitle = {International Conference on Computational Linguistics},
  year = {2024},
}
Figure