Molecular Facts: Desiderata for Decontextualization in LLM Fact Verification (2406.20079v1)

Abstract: Automatic factuality verification of LLM generations is becoming more and more widely used to combat hallucinations. A major point of tension in the literature is the granularity of this fact-checking: larger chunks of text are hard to fact-check, but more atomic facts like propositions may lack context to interpret correctly. In this work, we assess the role of context in these atomic facts. We argue that fully atomic facts are not the right representation, and define two criteria for molecular facts: decontextuality, or how well they can stand alone, and minimality, or how little extra information is added to achieve decontexuality. We quantify the impact of decontextualization on minimality, then present a baseline methodology for generating molecular facts automatically, aiming to add the right amount of information. We compare against various methods of decontextualization and find that molecular facts balance minimality with fact verification accuracy in ambiguous settings.

• The paper introduces molecular facts that balance decontextuality with minimal contextualization for improved fact verification.
• The methodology employs a two-step LLM process to identify ambiguities and generate minimal, contextually sufficient molecular facts.
• Controlled experiments demonstrate that molecular facts reduce error rates and enhance verification accuracy in ambiguous biography cases.

Molecular Facts: Desiderata for Decontextualization in LLM Fact Verification

Introduction

The paper "Molecular Facts: Desiderata for Decontextualization in LLM Fact Verification" by Anisha Gunjal and Greg Durrett addresses a critical issue in the domain of LLMs: the verification of generated facts to mitigate hallucinations. The focus is on the granularity of fact-checking—specifically, the balance between atomic and more contextually enriched "molecular" facts. This essay discusses the proposed framework and its implications.

Overview and Key Contributions

The authors identify a tension in the literature concerning the granularity of fact-checking. Atomic facts, although easier to verify, often lack the necessary context, leading to ambiguity. Conversely, larger chunks of text, while context-rich, pose a challenge in terms of verification. To address this, the authors introduce the concept of "molecular facts," designed to balance decontextuality and minimality. The main contributions include:

1. Definition of Molecular Facts: The paper introduces the concept of molecular facts, which are atomic facts enhanced with just enough context to make them unambiguous but minimal enough to avoid unnecessary complexity.
2. Criteria for Molecular Facts: Two essential criteria are defined—decontextuality (the ability of a fact to stand alone) and minimality (adding the least amount of information necessary for context).
3. Methodology for Generating Molecular Facts: A two-step process using LLMs to identify and generate molecular facts is presented.
4. Controlled Experiments: The authors devise a controlled experiment to illustrate the problem of non-minimality and evaluate the effectiveness of molecular facts.
5. Evaluation on Ambiguous Biographies Dataset: The methodology is tested on a dataset with ambiguous entity references, demonstrating that molecular facts improve verification accuracy while mitigating issues of minimality and ambiguity.

Methodology

The authors propose a systematic approach to generating molecular facts:

1. Stage One - Identifying Ambiguity: The primary subject of a claim is identified, and potential ambiguities are assessed based on the LLM's parametric knowledge.
2. Stage Two - Molecular Facts Generation: The LLM uses identified ambiguities and context to generate a decontextualized molecular fact.

The experiments employ two baselines, simple and safe decontextualization methods, and the proposed molecular fact generation method.

Key Findings

Controlled Experiment on Minimality

The controlled experiment illustrates the impact of minimality on error localization. The paper uses a synthetic dataset to show that non-minimal decontextualizations lead to higher error rates in fact verification. The results indicate that between 1.7% to 9.6% of decontextualized claims could cause error localization issues due to excessive information. The molecular approach demonstrates a better balance, with fewer instances of non-minimal claims.

Ambiguous Biographies Experiment

Evaluating the ambiguous biographies dataset reveals that molecular facts improve accuracy in ambiguous settings. Comparing various decontextualization methods, molecular facts show a significant advantage, striking a balance between specificity and minimality. The results highlight a noteworthy improvement in fact verification accuracy, especially for claims involving ambiguous entities.

Implications and Future Work

The implications of this work are both practical and theoretical:

1. Enhanced Fact-Checking Pipelines: The methodology can be integrated into existing fact-checking pipelines to improve the accuracy and reliability of LLM outputs.
2. Improved Disambiguation: Molecular facts offer a structured way to handle ambiguities in facts, potentially informing advancements in other NLP tasks such as entity linking and coreference resolution.
3. Broader NLP Applications: The principles of decontextuality and minimality could be applied to various NLP applications, such as summarization and information retrieval.
4. Future Research: Future studies could explore the extension of this framework to other domains and languages, as well as the integration with retrieval-augmented generation techniques.

Conclusion

The paper by Gunjal and Durrett offers a significant contribution to the field of LLM fact verification by introducing molecular facts. By balancing decontextuality and minimality, the proposed framework addresses key challenges in fact verification, particularly in handling ambiguities and maintaining conciseness. While there is room for improvement, especially in evaluating end-to-end LLM pipelines, the findings set a robust foundation for future research and practical applications in NLP.