Abstract
Text summarization is a user-preference based task, i.e., for one document, users often have different priorities for summary. As a key aspect of customization in summarization, granularity is used to measure the semantic coverage between the summary and source document. However, developing systems that can generate summaries with customizable semantic coverage is still an under-explored topic. In this paper, we propose the first unsupervised multi-granularity summarization framework, GranuSum. We take events as the basic semantic units of the source documents and propose to rank these events by their salience. We also develop a model to summarize input documents with given events as anchors and hints. By inputting different numbers of events, GranuSum is capable of producing multi-granular summaries in an unsupervised manner. Meanwhile, we annotate a new benchmark GranuDUC that contains multiple summaries at different granularities for each document cluster. Experimental results confirm the substantial superiority of GranuSum on multi-granularity summarization over strong baselines. Further, by exploiting the event information, GranuSum also exhibits state-of-the-art performance under the conventional unsupervised abstractive setting.
Introduction
Standard text summarization systems typically adopt a "one size fits all" approach, producing a single output regardless of varying user needs. In practice, requirements range from high-level overviews (coarse-grained) to detailed reports (fine-grained). A critical challenge in this field is that summary length does not inherently equate to granularity; simply truncating a summary often sacrifices essential context. Furthermore, most existing controllable summarization models rely on extensive supervised datasets with human-labeled summaries at different granularities, which are expensive to produce.
To address these issues, we introduce GRANUSUM, the first unsupervised framework for multi-granularity summarization. By defining granularity as semantic coverage rather than just length, and using "events" (verb-centric phrases) as semantic anchors, GRANUSUM enables customizable control without requiring human-labeled training data.
The GRANUSUM Framework
The GRANUSUM framework is composed of two primary components: an Event-aware Summarizer and an Event Selector.
- Event-aware Summarizer: Built on a Longformer-Encoder-Decoder (LED) backbone, this component is pre-trained on a novel unsupervised task. Given a document with masked sentences, the model reconstructs the missing text using the remaining context and a list of extracted "events" as hints. These events are identified via 76 syntactic patterns (e.g., nsubj-verb-dobj). This process trains the model to generate coherent summaries anchored by specific semantic units.
- Event Selector: To determine event importance, the system employs a Salience Score based on perturbation. It generates a candidate summary with a set of events and then measures the impact of removing individual events. If an event's removal significantly alters the summary (as measured by ROUGE), it is deemed highly salient.
By adjusting the number of top-ranked salient events fed into the summarizer, the system can precisely scale the granularity of the output.
Datasets and the GranuDUC Benchmark
The model's performance was evaluated on several large-scale datasets, including Multi-News (multi-document news), arXiv (scientific papers), and DUC2004 (standard benchmark).
A significant contribution of this work is the introduction of GranuDUC, a new benchmark created by re-annotating DUC2004. GranuDUC provides human-written summaries at three distinct levels:
- Coarse: Focuses solely on the main event.
- Medium: Includes important conditions or results.
- Fine: Incorporates specific details such as time and location.
Experimental Results
GRANUSUM demonstrates substantial improvements over strong baselines like PEGASUS and PRIMERA across various settings.
On the GranuDUC benchmark, GRANUSUM achieved significant gains in ROUGE-1 scores, outperforming the strongest baselines by 1.78 points at the coarse level and up to 4.53 points at the fine-grained level. Human evaluations further validated these results, with judges rating GRANUSUM higher in fluency, relevance, and faithfulness compared to other controllable models. Notably, even in standard single-granularity unsupervised settings, GRANUSUM established a new state-of-the-art on Multi-News, arXiv, and DUC2004, indicating that event-based guidance effectively reduces hallucinations.
Discussion and Future Work
The success of GRANUSUM highlights the utility of verb-centric events as semantic units for controlling summary density. Unlike keywords or full sentences, events provide a flexible yet meaningful middle ground for guiding abstractive models.
While the current rule-based event extraction is effective, future iterations could benefit from neural event extractors to capture more complex semantic relations. Additionally, while the current model focuses on three discrete levels, exploring a continuous "granularity slider" could offer even more granular user control. Finally, increasing the diversity of event rephrasing could further enhance the abstractness of the generated summaries.
Summary updated