Explanatory Summarization with Discourse-Driven Planning (2504.19339v3)

Abstract: Lay summaries for scientific documents typically include explanations to help readers grasp sophisticated concepts or arguments. However, current automatic summarization methods do not explicitly model explanations, which makes it difficult to align the proportion of explanatory content with human-written summaries. In this paper, we present a plan-based approach that leverages discourse frameworks to organize summary generation and guide explanatory sentences by prompting responses to the plan. Specifically, we propose two discourse-driven planning strategies, where the plan is conditioned as part of the input or part of the output prefix, respectively. Empirical experiments on three lay summarization datasets show that our approach outperforms existing state-of-the-art methods in terms of summary quality, and it enhances model robustness, controllability, and mitigates hallucination.

Explanatory Summarization with Discourse-Driven Planning

The paper "Explanatory Summarization with Discourse-Driven Planning" introduces a novel method for generating lay summaries of scientific documents that specifically aims to incorporate explanatory content to enhance accessibility. Traditional summarization models often fail to explicitly model explanatory elements, which are necessary for translating complex scientific information into more readily understandable language. This paper proposes a plan-based approach leveraging discourse frameworks to structure the generation of summaries, guiding the integration of explanatory sentences through a thoughtful planning mechanism.

Methodology Overview

The authors introduce two discourse-driven planning strategies for summary generation: the first strategy treats the plan as part of the input, while the second strategy treats it as part of the output prefix. By using these strategies, the model can steer the generation process towards incorporating explanations. These plans are conceptualized as sequences of questions oriented towards producing explanations, which are then used to condition the generation process.

A discourse parser based on Rhetorical Structure Theory (RST) is employed to identify and characterize explanations within lay summarization datasets. This parser is used to ascertain the proportion and function of explanatory content within documents and lay summaries and informs model training by providing silver-standard plan annotations.

Experimental Findings

Empirical experiments across three datasets—SciNews, eLife, and PLOS—demonstrated that the proposed approaches outperform existing state-of-the-art summarization models. Notably, the discourse-driven approaches improve summary quality while enhancing model robustness and controllability. Quantitative metrics such as Rouge, BERTScore, and D-SARI, alongside qualitative assessments of accessibility and factual consistency, indicate that the plan-based models effectively balance the inclusion of explanatory content with readability and factuality.

Furthermore, the paper explores various plan generation strategies, revealing the importance of structured content planning in generating coherent and factually accurate summaries. The discourse-driven planning not only improves the generation quality but also provides a mechanism for controlling the nature and proportion of explanatory content.

Implications and Future Directions

The findings presented in this paper have significant implications for both theoretical and practical aspects of AI-driven summarization. The ability to explicitly control explanation generation within lay summaries opens up new opportunities for enhancing public access to scientific knowledge, particularly when bridging gaps between specialized content and non-expert audiences.

Theoretically, this work advocates for the integration of discourse frameworks with neural models, suggesting potential pathways for refining text generation processes to better align with human communication strategies. Practically, the discourse-driven planning techniques can be adapted for various narrative generation tasks, extending beyond summarization to encompass broader domains such as education, media, and policy communication.

Future work could explore the integration of the proposed methodologies with other narrative planning systems or extend the approach to multimodal contexts where textual summaries need to be aligned with graphical or audio-visual data. Additionally, adapting the planning framework to accommodate dynamic user interactions or personalization could further improve the efficacy of lay summarization systems.

In conclusion, the paper presents a comprehensive approach for enhancing explanatory summarization and contributes valuable insights into the integration of discourse analysis with neural planning techniques, offering promising directions for future research and development in AI text generation.