TL;DR
This paper tackles the challenge of ensuring factual consistency in automatically generated summaries, a critical issue as LLMs become more prevalent. The authors introduce SIFiD, a novel method that filters irrelevant document sentences before LLM evaluation, significantly improving efficiency and accuracy. Their findings highlight GPT-4's superior performance over GPT-3.5 and the effectiveness of SIFiD in boosting factual consistency detection.
What Problem Does This Paper Address?
Document summarization, while increasingly sophisticated thanks to Large Language Models (LLMs), faces a persistent challenge: ensuring the generated summaries are factually consistent with the original source document. Inconsistencies, often subtle and difficult for traditional methods to detect, can lead to misinformation. While LLMs show promise, early attempts at using them for inconsistency detection have been hampered by limitations in prompt adherence and the lack of tailored methodologies. This paper aims to reassess and improve LLM-based factual inconsistency detection by evaluating the latest GPT models and proposing a more efficient and effective approach.
Methods
The core contribution of this paper is the SIFiD (Summary Inconsistency Detection with Filtered Document) method. SIFiD addresses the computational burden and potential noise introduced by feeding entire documents to LLMs for inconsistency checks. The process involves:
- Relevance Matrix Calculation: For each document-summary sentence pair, a relevance score is computed. This can be done using either:
- Entailment Scorer: Leveraging Natural Language Inference (NLI) models to assess how well a summary sentence is supported by a document sentence, considering both entailment and contradiction.
- Semantic Similarity Scorer: Using sentence embeddings to measure the cosine similarity between document and summary sentences.
- Sentence Filtering: A relevance matrix is constructed, and then max pooling is applied to identify the highest relevance score for each document sentence. A threshold ($\beta$) is used to filter out sentences deemed less relevant, along with their immediate neighbors to maintain contextual continuity. This results in a
Dfiltereddocument. - LLM Evaluation: The filtered document (
Dfiltered) and the original summary are then fed into an LLM (GPT-3.5 or GPT-4) using a carefully crafted prompt. The LLM is asked to determine if the summary contains any factual inconsistencies (omission, addition, or inaccuracy).
The paper also reevaluates the use of LLMs for this task by comparing the latest versions of GPT-3.5 Turbo and GPT-4 Turbo, and exploring the impact of benchmark-specific prompt templates and Chain-of-Thought (CoT) prompting.
Datasets
The primary dataset used for evaluation is SUMMAC (Laban et al., 2022). The authors also mention using the Polytope benchmark within SUMMAC, for which they developed a tailored prompt template. No new datasets were introduced.
Key Insights and Ideas
- GPT-4 is a Game Changer: The leap in performance from GPT-3.5 to GPT-4 for factual inconsistency detection is substantial, underscoring the importance of using the most capable LLMs for complex tasks.
- Context Matters for Prompts: A single, generic prompt template is insufficient for LLMs across different summarization benchmarks. Tailoring prompts to specific benchmark requirements significantly boosts performance.
- SIFiD's Efficiency Gain: By filtering out irrelevant sentences, SIFiD dramatically reduces the input size for LLMs, leading to more efficient and potentially more accurate evaluations without sacrificing crucial information.
- CoT is Not a Universal Fix: While Chain-of-Thought prompting can help some models (like GPT-3.5), it can actually degrade the performance of more capable models like GPT-4, suggesting that GPT-4's inherent reasoning abilities are already sufficient or that CoT can introduce unwanted biases.
Key Results
- GPT-4 Dominance: GPT-4 Turbo significantly outperforms GPT-3.5 Turbo, achieving a balanced accuracy of up to 78.8% on SUMMAC, compared to GPT-3.5's 69.7%.
- SIFiD Boosts GPT-4: Integrating SIFiD with GPT-4 further improved performance to 79.9% (SIFiD-Entailment +CoT).
- Sentence Reduction: SIFiD achieved an average sentence removal rate of 61.3% for entailment-based filtering and 67% for semantic similarity-based filtering across benchmarks.
- Benchmark-Specific Prompts: Tailoring prompts for the Polytope benchmark improved GPT-4's performance from 60.9% to 66.0%.
Limitations and Future Work
The primary limitation acknowledged is the cost associated with using powerful LLMs like GPT-4, even with SIFiD's sentence reduction. While SIFiD reduces input size, the financial implications of using GPT-4 remain considerable. The authors anticipate that advancements in LLM technology will lead to cost reductions, making such models more feasible for widespread real-world applications. Future work could explore further cost optimization and the application of SIFiD to a broader range of summarization tasks and datasets.
Summary updated