Unfolding the Headline: Iterative Self-Questioning for News Retrieval and Timeline Summarization

Weiqi Wu, Shen Huang, Yong Jiang, Pengjun Xie, Fei Huang, Hai Zhao

Researchers from Shanghai Jiao Tong University and Alibaba Group have introduced CHRONOS, a novel framework designed to transform a single news headline into a comprehensive, chronological timeline. By mimicking the investigative behavior of a journalist—searching, learning, and asking follow-up questions—CHRONOS addresses the complexities of open-domain news retrieval with unprecedented efficiency.

The Challenge of Open-Domain Timelines

Timeline Summarization (TLS) involves extracting and ordering pivotal events to provide a structured narrative of a story's development. Traditionally, this has been treated as a "closed-domain" task where models process a pre-defined set of documents. However, real-world scenarios are "open-domain": a user starts with a headline and must navigate the vast, noisy landscape of the internet.

This transition introduces three primary hurdles:

  1. Information Overload: The sheer volume of search results creates significant noise.
  2. Missing Links: Initial searches often leave causal gaps in the narrative.
  3. Computational Cost: Large Language Models (LLMs) are too resource-intensive to process every potential document in a search corpus.

The CHRONOS Framework

CHRONOS (Causal Headline Retrieval for Open-domain News Timeline SummarizatiOn via Iterative Self-Questioning) addresses these issues through a surgical retrieval process. Rather than attempting to read everything, the framework uses an iterative loop to identify and fill information gaps.

The CHRONOS workflow: from context survey to iterative self-questioning and final timeline merging.

The pipeline consists of four distinct stages:

  1. Context Survey: An initial search establishes a baseline "News Context."
  2. Iterative Self-Questioning: The model reflects on the retrieved context to generate "5W1H" questions (Who, What, Where, When, Why, How), identifying what is missing to make the timeline complete.
  3. Question Rewriting: A dedicated "Rewriter" module decomposes conversational LLM questions into focused, search-friendly queries to maximize retrieval precision.
  4. Divide-and-Conquer Generation: The system generates mini-timelines for each retrieval round and merges them, resolving date conflicts and selecting milestone events for the final output.

Mimicking Human Inquiry

The core innovation of CHRONOS is its "Human-in-the-Loop" mimicry. Effective retrieval is treated not just as a vector matching problem, but as a process of active inquiry. By allowing the model to ask follow-up questions, it can discover causal links that a single keyword search would overlook.

An illustration of iterative self-questioning rounds applied to the 2023 Banking Crisis, showing how specific questions lead to the discovery of key chronological events.

To optimize this process, the authors introduced Chrono-Informativeness (CI), a metric that measures how effectively a generated question retrieves documents aligned with a ground-truth timeline.

Evaluation and Results

The study utilized classic benchmarks like T17 and Crisis, alongside a new resource: Open-TLS. This dataset features 50 timelines on recent news with significantly longer durations (averaging over 4,000 days) sourced from professional outlets like the Associated Press.

Key Findings:

  • Superior Efficiency: On the "Crisis" dataset, CHRONOS achieved state-of-the-art (SOTA) results with an 18x speedup, reducing processing time from over 7 hours to just 24 minutes.
  • Surgical Precision: The framework proved that retrieving ~30 high-quality documents is more effective and cost-efficient than brute-force processing of thousands of noisy results.
  • Model Agnostic: Performance gains were consistent across various LLMs, including GPT-3.5-Turbo, GPT-4o, and Qwen2.5-72B.

Conclusion

CHRONOS demonstrates that strategic, iterative inquiry is the key to mastering open-domain timeline summarization. By prioritizing "better inquiry" over "more data," the framework provides a scalable and efficient solution for turning headlines into structured history. Future work will focus on handling weakly linked temporal events and further refining the merging process for even larger data volumes.

Summary updated

BibTeX
@inproceedings{wu2025unfolding,
  author = {Weiqi Wu and Shen Huang and Yong Jiang and Pengjun Xie and Fei Huang and Hai Zhao},
  title = {{Unfolding the Headline: Iterative Self-Questioning for News Retrieval and Timeline Summarization}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2025},
  doi = {10.18653/v1/2025.findings-naacl.248},
}
Figure