RP29 - Lay Summaries: Verifying Truthfulness and Completeness

Continuing the lay summaries for patients project (RP15) will focus on assessing the degree of truthfulness and completeness, which is crucial when using LLMs. A significant challenge in text generation is the problem of so-called hallucinations (confabulations), where models generate plausible-sounding but incorrect or misleading information [1]. Addressing this issue is essential to improving the trustworthiness of generated summaries. One method to mitigate hallucinations involves metric-based matching techniques that compare the original content with the generated lay summaries. Several metrics have been proposed for fact-checking [2], but none are specific to producing lay summaries, which will be an objective of this research project. This process helps ensure that the key facts and messages are accurately conveyed, maintaining the integrity of the original information. To support this, an enhanced dictionary of lay-terms for melanoma will be developed. Prompt-Engineering is another technique to mitigate the problem, which is a standard approach. Another promising method is using knowledge graphs to verify the accuracy of the generated text. Knowledge graphs represent information in a structured form, capturing relationships between entities and facts. Cross-referencing the generated summaries with these graphs makes it possible to identify discrepancies and ensure that the content aligns with established knowledge. This method helps validate the factual accuracy and assess the completeness of the summaries, ensuring that all relevant information is included and correctly represented. We aim to enhance the reliability of lay summaries produced by LLMs by combining these techniques. Metric-based matching combined with pre-translation of lay-terms provides a straightforward way to spot-check the fidelity of summaries, while knowledge graphs offer a robust framework for systematic verification. This dual approach can help build greater confidence in using LLMs for generating lay summaries, making them a more dependable tool for disseminating information to non-experts. New presentation techniques with tool-tips or link-outs might help to avoid complex generated language. Ultimately, improving the truthfulness and completeness of these summaries will lead to better-informed audiences and foster greater trust in the use of advanced AI technologies for communication and education.

[1] Z. Ji et al., “Survey of Hallucination in Natural Language Generation,” ACM Comput Surv, vol. 55, no. 12, p. 248:1-248:38, Mar. 2023, doi: 10.1145/3571730.

[2] D. Tam, A. Mascarenhas, S. Zhang, S. Kwan, M. Bansal, and C. Raffel, “Evaluating the Factual Consistency of Large Language Models Through News Summarization,” in Findings of the Association for Computational Linguistics: ACL 2023, A. Rogers, J. Boyd-Graber, and N. Okazaki, Eds., Toronto, Canada: Association for Computational Linguistics, Jul. 2023, pp. 5220–5255. doi: 10.18653/v1/2023.findings-acl.322.