Semantic Scholar API

• Semantic Scholar API is a REST-style service that provides structured scholarly records from a vast academic graph with over 200 million papers.
• It integrates state-of-the-art models to offer advanced features such as TLDR summaries, citation intent classification, and SPECTER embeddings.
• The API supports flexible graph traversal and high-volume, tiered access for bibliometric workflows, combining public and proprietary data sources.

Semantic Scholar provides a unified, REST-style HTTP interface to its large, continuously updated academic knowledge graph (S2AG) and its associated semantic models. The Semantic Scholar Academic Graph encompasses over 200 million papers, 80 million authors, 550 million paper-authorship edges, and 2.4 billion citation edges, incorporating both public and proprietary data sources. Through state-of-the-art PDF content extraction and automated knowledge graph construction, the Semantic Scholar API enables programmatic access to structured scholarly records, advanced semantic features, and flexible graph traversal, supporting a variety of bibliometric and analytical workflows (Kinney et al., 2023).

1. High-Level System Architecture

The Semantic Scholar API (referred to as the "Graph API") is architected as a REST façade layered on top of the S2 data-processing pipeline and a set of semantic-feature models. The underlying data stack consists of:

• The S2 data-processing pipeline, which ingests raw PDFs and associated metadata, performs structured text and citation extraction, normalizes and disambiguates entities (authors, venues, institutions), and assembles the Academic Graph comprising papers, authors, venues, and citation relationships.
• Semantic-feature models that perform paper summarization, citation-intent classification, fields-of-study assignment, SPECTER embedding computation, and recommendation generation.

At runtime, API servers directly query the materialized graph and pre-computed feature tables, apply per-request filters and transformations, and return results as JSON via HTTPS. This architectural separation—ingestion/models, graph database, and API layer—permits independent evolution of each substrate while maintaining stable APIs (Kinney et al., 2023).

The principal goals of the Graph API are to (1) provide programmatic lookup of any paper, author, venue, or affiliation in the S2AG, (2) support basic graph traversal queries (citations, references, coauthors, papers by venue or field), and (3) expose advanced semantic signals such as TLDR summaries, citation intent, SPECTER embeddings, recommendations, and peer-review support.

2. Core API Endpoints and Data Schema

Graph API endpoints are organized by resource type: paper, author, venue, recommendation, and peer-review. Most endpoints follow the HTTP GET method, with recommendations and peer-review using POST. All APIs communicate via HTTPS and encode responses as application/json.

Major Endpoints

Endpoint Type	Example URL Pattern	Key Features
Paper Lookup	/graph/v1/paper/{paperId}	Metadata, identifiers, authorship, citations, references, TLDR, SPECTER embedding, fields of study
Author Lookup	/graph/v1/author/{authorId}	Author disambiguation, publication statistics, publication list
Keyword Search	/graph/v1/paper/search?query=...	Full-text title/abstract search with configurable pagination and output fields
Citation Graph	/graph/v1/paper/{paperId}/references, /citations	Outgoing references, incoming citations, citation context, and intent classification
Recommendations	/graph/v1/recommendations (POST)	Positive and negative paper inputs, personalized fresh-paper ranking
Peer-Review	/graph/v1/peer-review (POST)	Reviewer–submission matching, conflict-of-interest detection, match scoring
Datasets	/api-docs/datasets	Monthly gzipped JSON snapshots, embeddings, TLDR, bulk full-text S2ORC

Data Model

The S2AG may be formalized as:

$$G = (P \cup A \cup V \cup I, E_{pa} \cup E_{cc} \cup E_{pv} \cup E_{ai})$$

with

• $P$ = paper nodes, $A$ = author nodes, $V$ = venues, $I$ = institutions,
• $E_{pa} \subset P \times A$ (authorship), $E_{cc} \subset P \times P$ (citation edges),
• $E_{pv} \subset P \times V$ (publication venue), $E_{ai} \subset A \times I$ (affiliation).

Paper records return fields such as internal and external identifiers, authorship, citation lists, reference lists, TLDR one-sentence summaries, SPECTER vector embeddings, and multi-label field-of-study assignments.

3. Advanced Semantic Feature Integration

The Graph API enriches core bibliographic records with advanced semantic and vector-based signals:

• TLDR: One-sentence summarizations generated via a BART+CATTs model, available inline or as a separate TLDRs dataset.
• Citation Intent: Each reference is classified by the SciCite model as background, method, or result, surfaced in the citation "intent" field.
• Influential Citations: References flagged as “highly influential” using heuristic criteria.
• Fields-of-Study: Generated via multi-label SVM classifiers; each paper includes one or more high-level topical assignments.
• SPECTER Embeddings: Returned on demand as a 768-dimensional vector ("embedding" field) by a model trained with triplet margin loss over citation links:

$$L(\theta) = \sum_{(p, p^+, p^-) \in T} \max(0, d(f_\theta(p), f_\theta(p^+)) - d(f_\theta(p), f_\theta(p^-)) + \alpha)$$

where $P$0 are triplets (anchor, cited, uncited papers) and $P$1 is a distance metric (Kinney et al., 2023).

• Recommendations: Generated by an online SVM+FAISS pipeline over SPECTER embeddings for near-real-time fresh-paper suggestions.
• Peer-Review Matching: Computes reviewer–submission compatibility via SPECTER cosine similarity and detects co-authorship conflicts.

4. Access Control, Rate Limiting, and Bulk Data

The API is designed with tiered access and usage throttling mechanisms:

• Unauthenticated access permits low-volume usage (typically < 100 requests/day) with "lite" record versions.
• Authenticated users (free partner registration, API key via HTTP header or query param) receive substantially higher limits (hundreds of thousands of requests/day). Active abuse detection and throttling are enforced, and custom SLAs are available for high-throughput partners.

For bulk offline analysis and embeddings indexing, monthly gzipped JSON snapshots are available, including paper metadata, abstracts, authors, citations (with intent/context), SPECTER embeddings, TLDRs, venue metadata, paper identifier mappings, and the S2ORC full text corpus. An API key is recommended for high-volume downloads (Kinney et al., 2023).

5. Usage Examples and Query Patterns

Typical API usage patterns include:

• Metadata Retrieval: Direct lookup by DOI, arXiv ID, or internal identifier with selectable fields.
• Keyword Search: Full-text search over title and abstract with pagination and output field selection.
• Citation Traversal: Retrieval of incoming and outgoing citation lists, each with potential context and intent. Citation edges can be filtered based on context or classification.
• Author Analytics: Querying author records for publication and citation statistics (paper count, citation count, h-index) and enumerating authored papers.
• Recommendation: Supplying sets of positive/negative paper IDs to obtain ranked lists of recently published relevant papers.
• Peer Review Support: Submission of candidate reviewer and submission lists to receive suitability scores and conflict indicators.

6. Limitations and Prospective Extensions

Several current limitations and planned enhancements are documented:

• Full-text PDF structural and semantic annotations (sections, figures, tables) are presently supplied only in bulk S2ORC; not accessible via the Graph API.
• The API does not yet provide arbitrary vector-similarity search over embeddings. Users must download embedding datasets for local indexing.
• Semantic search capabilities over full text are in development.
• Forthcoming endpoints will expose additional granularity, such as paragraph boundaries, inline citation locations, figure/table bounding boxes, entity-and-relation extractions, semantic parsing services, and personalized library features.
• Author-affiliation normalization (linkage to ROR) is slated for integration into author records.

The architectural decoupling of the API from upstream extraction, normalization, and modeling pipelines enables the Semantic Scholar platform to independently advance its underlying technologies while maintaining stable, field-tested HTTP endpoints, supporting the world's largest open scholarly knowledge graph (Kinney et al., 2023).