Summarizing Scientific Papers Through Citation Context

179 concepts96 papers

When we want to understand a scientific paper, our default instinct is to read its abstract. The abstract provides a concise overview of the paper's goals, methods, and results. However, as highlighted by Mei et al. in Generating Impact-Based Summaries for Scientific Literature, an abstract only captures the author's intended significance at the time of writing. It suffers from an inherent author-centric bias and remains entirely static.

In reality, the true value and long-term impact of a scientific paper are dynamic properties defined by the research community over time. A community might find a minor methodological detail or an unexpected dataset application far more valuable than the primary conclusion the authors originally emphasized. To capture this evolving consensus, a prominent research direction shifts the focus of scientific document processing from what the authors wrote to how the community cites them.

Citances as Crowdsourced Micro-Summaries

The core mechanism behind this paradigm is the extraction and analysis of "citances"—the specific sentences or textual contexts surrounding a citation within an external, citing paper. Instead of treating a citation as a mere metadata link in a bibliography, this approach treats the surrounding text as a peer-provided micro-summary.

As noted in the overview of the CL-SciSumm shared tasks by Jaidka et al., these collective descriptions function as a form of community-driven summarization. When dozens of papers cite a target document, their combined citation contexts act as a virtual query against the original text. This query highlights the exact components of the work that proved useful in practice. For instance, one citing paper might describe how the target paper's dataset was constructed, while another might critique its experimental baseline. Aggregating these perspectives provides a multi-faceted, functional profile of the document that an abstract simply cannot capture.

The Challenge of Epistemic Drift and Citation Grounding

While relying on the community's voice provides rich, real-world context, it introduces a significant vulnerability: epistemic value drift. Citing authors frequently oversimplify, misinterpret, or strip away the original nuances and degrees of certainty present in the source material. If a summarization tool relies exclusively on external citances, it risks propagating these inaccuracies.

To prevent this, modern citation-based summarization frameworks do not just collect external sentences; they ground them back in the source text. As demonstrated by Cohan et al. in Scientific Article Summarization Using Citation-Context and Article's Discourse Structure, effective systems use external citations as semantic queries to map back to the exact "cited text spans" within the original document. By identifying where the external description aligns with the internal discourse structure (such as the specific paragraphs within the Methods or Results sections), the system ensures that the final summary reflects community impact while remaining strictly anchored in the original author's technical language.

From Summaries to Functional Representations

This paradigm extends far beyond generating text summaries; it fundamentally changes how documents are represented for downstream tasks like indexing, keyphrase extraction, and recommendation.

Traditional information retrieval relies heavily on internal document text and author-provided keywords. However, Caragea et al. demonstrated in Citation-Enhanced Keyphrase Extraction from Research Papers: A Supervised Approach that incorporating features from external citation networks drastically improves supervised keyphrase extraction. By tracking how a paper is described by its peers, models can index documents using the actual vocabulary of the active research community.

Similarly, in the domain of document embeddings, Han developed the hyperdoc2vec framework, which learns dual representations for hypertext documents based on their roles as both "linkers" (citing documents) and "linked" entities (cited documents). This approach explicitly leverages the textual context surrounding a link to capture the semantic reasoning behind inter-document connections, outperforming traditional text-only embedding models.

Personalizing the Literature Map

Ultimately, these functional, community-sourced representations enable interactive systems that help researchers navigate information overload without losing context. For example, Chang et al. developed CiteSee, a personalized reading environment that surfaces citing sentences from a user's own reading history to contextualize new citations.

By leveraging the relational synthesis already provided by the broader scientific community, these tools transform isolated references into an interconnected, personalized map of the scientific landscape, allowing scholars to understand not just what a paper is, but why it matters to their specific field.

Key Foundational Work

Towards Multi-paper Summarization Reference Information (1999) — contributed 5 concept elements to this direction.
Creating a Test Collection for Citation-based IR Experiments (2006) — contributed 2 concept elements to this direction.
History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 102 papers
  2. Version 1 · 2026-10-04 · 72 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_citation_based_summarization_f8a110dd,
  title = {citation-based summarization},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/citation-based-summarization/rdr_f8a110dd5d794ba4a4ea736fa9988738},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 2, frozen text):

@misc{lacuna_citation_based_summarization_f8a110dd_v2,
  title = {citation-based summarization},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/citation-based-summarization/rdr_f8a110dd5d794ba4a4ea736fa9988738/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-21},
}

See something wrong? Report a mistake