MDContextCite: A Large-Scale Multi-disciplinary Citation Context Database

MDContextCite is a large-scale, multi-disciplinary citation context database designed to support research in citation-aware scholarly information retrieval, citation intent analysis, and citation-based retrieval benchmarking. The resource is built around citation contexts—local textual spans surrounding in-text citations—as the primary unit of analysis, enabling fine-grained study of how prior work is used, positioned, and retrieved in scientific discourse.

MDContextCite includes the original MDCite release as its core citation-context dataset component, together with an additional retrieval-oriented variant called EdgeCite, created for downstream large-scale retrieval experiments. In this sense, MDContextCite should be understood as a broader database and dissemination framework, while MDCite remains the primary citation-context dataset release inside it.

This Zenodo record contains the data artifacts required to reproduce both:

1. the original MDCite citation context dataset
2. a new retrieval-oriented version called EdgeCite

The retrieval dataset should be viewed as an additional dissemination variant of MDContextCite: created to support leakage-controlled large-scale citation retrieval experiments.

Data Sources

MDContextCite is constructed by integrating multiple large-scale scholarly data sources:

🔹Scopus bibliographic records (2000–2024)

Scopus bibliographic records are collected via the Scopus API, which provides programmatic access to journal articles, citation counts, and publication metadata. These records form the foundation of the dataset and are used to identify influential papers based on citation statistics, enabling reproducible large-scale data collection over a long temporal span.

🔹Web of Science (WoS) subject categories

Web of Science subject categories are used to group journals by scientific field and to select Top-5 Q1 journals per field. This field-aware categorization enables journal-stratified selection and controls for substantial differences in citation practices across disciplines.

🔹OpenAlex API

The OpenAlex API is used for DOI/title resolution and structured metadata support, especially when primary resolution through Scopus is unavailable. OpenAlex provides large-scale open scholarly indexing infrastructure that improves identifier coverage and metadata interoperability.

🔹Semantic Scholar Graph API

Citation contexts, citation-linked metadata, and related paper-level information are collected through Semantic Scholar infrastructure. For each citing paper, the pipeline retrieves structured citation metadata and local textual spans surrounding in-text citation markers, which form the core citation context units in MDContextCite.

🔹Automatic citation intent classification

Each citation context is automatically assigned one or more citation intent using the SynIntent model proposed in Understanding citation intents by generative intent model based on heterogeneous graph neural network (Phan et al., Information Processing & Management, 2026, 63(6):104743). For benchmark construction, one primary functional citation intent label is retained for each usable citation event.

Dataset Construction Pipeline

The MDContextCite construction process follows a transparent and reproducible pipeline:

1. Journal selection

Journals are grouped by WoS subject categories, and Top-5 Q1 journals are selected per scientific field.

2. Citation-based filtering

Within each selected journal, the Top 5% most-cited papers are identified independently. This journal-stratified strategy avoids overrepresentation of citation-intensive fields and preserves disciplinary diversity.

3. Citation context extraction

Citation contexts are extracted from papers citing the selected influential anchor papers. Each context corresponds to a local textual span surrounding an in-text citation marker.

4. Citation intent taxonomy

Each citation context is automatically assigned one of seven functional citation intents:

• background
• uses
• similarities
• differences
• motivation
• extends
• future_work

5. Dataset variants

MDContextCite contains both:

• an original citation-context dataset release (MDCite)
• an additional retrieval-oriented benchmark release (EdgeCite) constructed for citation retrieval experiments.

The latter should be interpreted as a new dissemination version of the database for downstream retrieval research.

File Overview

MDCite Dataset Files

dataset_context_intent_single.csv (1.45 GB)
Single-intent benchmark version of MDCite in CSV format.
Each row corresponds to one citation context.

dataset_context_intent_single.parquet (620 MB)
Same content as above, stored in Parquet format.
Recommended for large-scale indexing, retrieval, and analytics workflows.

Construction and Provenance Files

Scopus_(Year_2000–2024).zip (120 MB)
Raw bibliographic metadata collected via the Scopus API.

WOS_2024_Subject_categories_(Top-5_Q1_Journals_per_field)_data.zip (51 MB)
Journal lists grouped by scientific field.

Top_5%_cited_papers_per_journal_dataset.zip (8.05 MB)
Lists of Top-5% cited papers selected independently within each journal.

citation_context_intent_data.zip (2.29 GB)
Full citation context data prior to single-label filtering.
Includes intermediate JSON artifacts containing probabilistic/multi-intent model outputs and structured citation context records used during dataset construction. These files represent the complete multi-intent version of MDCite before benchmark conversion.

EdgeCite Dataset Files

retrieval_dataset_citing_disjoint_with_year.parquet (316.4 MB)
This file is a new retrieval-oriented variant of MDContextCite. It should be regarded as an additional dissemination version of the original MDCite resource—effectively a version 2 retrieval benchmark designed for downstream citation retrieval experiments.

Compared with the original MDCite single-intent release, this new file reorganizes the citation-context resource into a retrieval-oriented citation-event dataset with:

• citation-event level organization
• citing–cited paper linkage
• primary intent labels
• field metadata
• year metadata
• citing-disjoint train/validation/test protocol for leakage control
• fixed candidate-anchor construction for reproducible retrieval experiments

This file is intended for:

• intent-conditioned citation retrieval
• candidate generation analysis
• reranking and selective reranking
• leakage-controlled scholarly retrieval evaluation

Schema of MDCite Dataset

Each citation context instance in the original MDCite release includes the following fields:

• text — citation context text
• label — functional citation intent
• field — scientific field
• group_id — citing paper identifier (for document-level evaluation splits)
• paperId — Semantic Scholar work identifier
• doi — Digital Object Identifier
• venue — publication venue
• year — publication year
• source_file — provenance identifier

Schema of EdgeCite Dataset

Each row in retrieval_dataset_citing_disjoint_with_year.parquet (316.4 MB) corresponds to a single usable citation event and is organized for retrieval experiments. The released file contains the following columns:

• citing_doi — DOI of the citing paper
• cited_doi_norm — normalized DOI of the cited anchor paper
• cited_title — title of the cited anchor paper
• field — broad scientific field
• context — local citation context text
• primary_label — primary citation intent label
• split — data split indicator (train, validation, or test)
• year — publication year metadata associated with the citation event or paper record

This retrieval benchmark variant is reorganized specifically for retrieval evaluation. In particular, it provides a citation-event-level benchmark with normalized cited-anchor identifiers, title-based candidate representations, explicit split assignments, and year metadata, making it directly usable for leakage-controlled intent-conditioned citation retrieval experiments.

Dataset Statistics

MDCite release

• Citation contexts: 2,057,196
• Citing papers: ~100,000
• Scientific fields: 21
• Publication venues: >1,000
• Citation intent classes: 7

The intent distribution is highly imbalanced, reflecting real-world citation behavior rather than artificially balanced labels.

EdgeCite release

As reported in the accompanying paper, the retrieval-oriented benchmark contains:

• usable citation edges: 2,659,249
• unique citing DOIs: 1,196,415
• unique cited anchors (normalized DOI): 12,529
• anchor candidates with valid titles: 12,518
• train: 2,127,915
• validation: 264,886
• test: 266,448

Intended Use Cases

MDContextCite is intended for:

• citation-aware information retrieval
• intent-aware ranking and reranking
• candidate generation analysis
• large-scale citation intent classification
• scholarly search and citation analysis
• large-scale retrieval benchmark construction

The newly added retrieval benchmark dissemination file is particularly intended for reproducible experiments on:

• intent-conditioned citation retrieval
• selective reranking
• effectiveness–efficiency trade-offs in scholarly retrieval
• document-level leakage-controlled evaluation

Relation to the Accompanying Papers

MDContextCite includes artifacts associated with two complementary outputs:

• MDCite: the original large-scale citing context dataset release
• EdgeCite: the downstream retrieval study built on a new retrieval-oriented variant 

Reproducibility and Licensing

All files are released for research and educational use.

Reproducing the full CitationHub construction pipeline may require:

• access to the Scopus API (institutional entitlement may be required)
• access to the OpenAlex API (publicly accessible)
• access to the Semantic Scholar Graph API (publicly accessible)

API keys are not included in this repository and must be supplied externally.

The resource is designed to support end-to-end reproducible experimentation through clear provenance tracking, intermediate artifacts, and benchmark-specific dissemination files.