QUERY-FOCUSED EVENT SUMMARIZATION: A DATASET AND BENCHMARK

Research Question and Approach

Traditional Multi-Document Event Summarization (MDES) often produces generic summaries that fail to address the specific, multi-dimensional interests of different users. While Query-Focused Summarization (QFS) attempts to tailor outputs to user needs, existing datasets and methods are not optimized for large-scale event corpora. To bridge this gap, the authors propose the Query-Focused Event Summarization (QFES) task, which requires systems to first retrieve query-relevant documents from a large thematic corpus and then generate a summary focused on that specific aspect.

To support this, the authors constructed QFESum, a dataset containing 8 thematic events, 16,684 documents, and 104 queries. They also developed a two-stage framework:

1. Query-Focused Retrieval with Adaptive Thresholding (RAT): Uses dense retrieval combined with an LLM-based dynamic thresholding mechanism to filter relevant documents from the large corpus.
2. Query-Focused Summarization based on Hierarchical Clustering (SHC): Extracts events from the retrieved documents, clusters them hierarchically, selects key events based on intra-cluster centrality, and performs coreference resolution to produce a coherent summary.

Experimental results demonstrate that the proposed framework consistently outperforms current state-of-the-art baselines, including TopicTLS, GraphRAG, and FG-RAG. On the QFESum dataset, the SHC method achieved superior scores across standard lexical metrics (ROUGE, BLEU, METEOR) and semantic metrics (BERTScore). Human evaluation further confirmed that SHC produces more accurate and relevant event summaries compared to existing approaches, which often struggle with redundancy or fail to maintain global consistency when processing large-scale document sets.