TL;DR
This paper introduces FinGrAct, a novel framework for evaluating the "actionability" of explanations in automated fact-checking systems. Actionability, defined as how well an explanation identifies errors, provides corrections, and supports them with evidence, is crucial for user trust and understanding but has been largely unaddressed. FinGrAct achieves superior correlation with human judgments compared to existing SOTA evaluators and exhibits significantly less ego-centric bias.
What Problem Does This Paper Address?
Automated Fact-Checking (AFC) systems aim to verify claims, but their explanations often lack clarity and trustworthiness. A key aspect of a good explanation, termed "actionability," is its ability to pinpoint factual errors, offer correct information, and back it up with verifiable sources. While other explanation qualities like coherence and fluency have been explored, actionability has remained an underexplored but critical desideratum in AFC. The lack of a robust evaluation method for actionability hinders progress in developing more effective and trustworthy AFC systems.
Methods
FinGrAct employs a fine-grained, divide-and-conquer approach to evaluate actionability. It breaks down the evaluation into three distinct stages:
- Error Segmentation and Correction: An LLM decomposes a false claim into atomic subclaims, identifies factual errors based on provided evidence, and proposes corrections. This stage outputs structured data detailing false subclaims, error reasons, and corrections.
- Explanation Evaluation: This stage verifies if the errors and corrections identified in the previous step are explicitly present in the given explanation. It outputs boolean judgments for error detection, correction, and the presence of supporting links.
- Source Evaluation: This stage assesses the relevance and supportiveness of any provided web links. It utilizes a novel URL Content Retriever (UCR) component that scrapes, cleans, and summarizes web page content before feeding it to the LLM for evaluation. This allows for real-time verification of external sources, rather than relying solely on the LLM's potentially outdated internal knowledge.
The overall actionability score is then computed based on categorized levels of error detection, correction, and supporting link quality, scaled to a Likert score.
Datasets
The authors constructed a novel dataset for evaluating actionability by merging and sampling from two existing sources:
- Dai et al. (2022): This dataset provided counterfactual (CF) data, which was categorized into six types based on error detection, correction, and source presence.
- Kotonya and Toni (2020b): This dataset provided explanations with varying degrees of actionability for public health claims.
The combined dataset was augmented with explanations generated by LLMs (LLAMA-7B, Mistral-7B, and GPT-4) to analyze ego-centric bias. The final dataset contains 203 random instances with diverse actionability levels, including both counterfactual and non-counterfactual explanations. Human annotators rated the actionability of explanations on a scale from 0 to 5.
Key Insights and Ideas
- Actionability is Crucial but Underexplored: The paper highlights that actionability is a critical but overlooked aspect of explainable AFC, essential for user trust and understanding.
- Fine-Grained Evaluation is Key: FinGrAct's structured, multi-stage approach (claim breakdown, error detection, correction, source verification) significantly improves evaluation accuracy and correlation with human judgments compared to holistic methods.
- URL Content Retrieval (UCR) Enhances Accuracy: Incorporating a UCR component to fetch and process external web content dramatically improves the performance of actionability evaluation, as it grounds the LLM's assessment in real-time information.
- LLMs Benefit from Modular Tasks: FinGrAct's success suggests that breaking down complex evaluation tasks into simpler, modular subtasks (e.g., separate steps for error detection, correction, and link verification) leads to better performance for LLM-based evaluators.
- Ego-Centric Bias is a Real Concern: The study confirms that LLM-based evaluators exhibit ego-centric bias, favoring their own generated outputs. FinGrAct demonstrates significantly less of this bias compared to other SOTA evaluators.
Key Results
- Superior Correlation with Human Judgments: FinGrAct achieved the highest Pearson and Kendall Tau correlations with human actionability ratings among all tested SOTA evaluators, both with and without the UCR component. For instance, with UCR, FinGrAct achieved a Pearson correlation of 0.46, outperforming Prometheus (0.328) and G-Eval (0.147).
- Reduced Ego-Centric Bias: FinGrAct exhibited the lowest ego-centric bias, with only 8.4% of biased samples, compared to Prometheus (26%) and G-Eval (48.7%).
- UCR Improves Performance: The inclusion of the UCR component consistently improved the correlation scores for all evaluated methods, demonstrating the importance of real-time source verification.
- Framework Matters More Than Model Size: Experiments with open-source LLMs (LLaMA-3.1 and Mistral-7B) using FinGrAct's prompts showed that the structured framework itself leads to better performance, even with smaller models, compared to using larger models with less structured prompts.
Limitations and Future Work
- UCR Limitations: The current UCR component is limited to processing textual content and struggles with images or JavaScript-heavy pages, leading to potential misinterpretations. Future work aims to develop a multimodal UCR.
- Zero-Shot Approach: All experiments were conducted using a zero-shot learning approach due to the high cost of fine-tuning large LLMs. Future work may explore cost-effective fine-tuning strategies.
- Focus on Ego-Centric Bias: The study primarily focused on ego-centric bias. Future research will explore other biases, such as cross-model biases, to gain a more comprehensive understanding of LLM evaluation systems.
Summary updated