sanaa-11/citation_context

Citation Contexts for Scientific Evidence Retrieval

Dataset Description

This dataset contains 9,920 citation occurrences extracted from English-language scientific papers. Each row represents one occurrence of a citation in a source paper and links it to the cited paper. It provides four increasingly broad representations of the citation context:

1. 1.the sentence containing the citation (context_c1_sentence);
2. 2.the complete source paragraph (context_c2_paragraph);
3. 3.a local window containing the source body item and available neighboring body items (context_c3_window); and
4. 4.the same wider context with section information (context_c4_section_aware).

In the processed context fields, the citation represented by the row is replaced with <TARGET_CIT>, while other citations are replaced with <OTHER_CIT>. The dataset is intended for experiments that compare citation-context scopes in passage retrieval, citation grounding, scientific claim verification, and related information-retrieval tasks.

Dataset summary

Dataset Structure

Data instance

{
    "citation_id": "2101.00098__body_000013__cite_0013__start_1576__ref_...",
    "source_paper_id": "2101.00098",
    "cited_paper_id": "2108.13004",
    "source_section": "Single-View 3D Reconstruction",
    "citation_marker": "{{cite:...}}",
    "paragraph_with_target_marker": "... <OTHER_CIT>, <TARGET_CIT> ...",
    "context_c1_sentence": "... <TARGET_CIT> ...",
    "context_c2_paragraph": "...",
    "context_c3_window": "...",
    "context_c4_section_aware": "[SECTION] Single-View 3D Reconstruction\n\n[CONTEXT]\n...",
    "candidate_atomic_passage_count": 21,
    "candidate_contextual_passage_count": 4
}

The example is abbreviated; the CSV contains all fields listed below.

Fields

Context representations

Very long contexts reflect unusually long source body items in the extracted corpus and have not been truncated to a model-specific token limit.

Dataset Creation

Each input citation occurrence was matched to its source body item and bibliography reference. The four contexts were then constructed at sentence, paragraph, neighboring-body-item, and section-aware scopes. In processed contexts, the selected citation was replaced by <TARGET_CIT> and co-occurring citations by <OTHER_CIT>.

Candidate passage counts describe the retrieval pool linked to the cited paper. The dataset itself does not include passage text, relevance labels, or a gold supporting-passage identifier, so supervised evidence-ranking experiments require an accompanying candidate-passage resource or separately constructed labels.

Limitations and Biases

• —The dataset reflects the coverage and extraction quality of its source scientific-paper corpus; it should not be treated as representative of all disciplines, venues, languages, or publication periods.
• —Context text can contain extraction artifacts such as {{formula:...}}, {{figure:...}}, and normalized citation markers.
• —A citation context does not necessarily state a single factual claim, and a cited paper may support only part of the surrounding text.
• —The dataset does not provide citation-intent labels, entailment labels, or manually verified evidence annotations.
• —Multiple rows can come from the same source or cited paper. Evaluation without paper-level grouping may overestimate generalization.

Contact

sanaa.abril@gmail.com