SurGE

Welcome to the official Hugging Face repository for SurGE, a benchmark and dataset for end-to-end scientific survey generation in the computer science domain.

SurGE provides a comprehensive resource for evaluating automated survey generation systems through both a large-scale dataset and a fully automated evaluation framework.

More information at: https://github.com/oneal2000/SurGE

Overview

SurGE is designed to push the boundaries of automated survey generation by tackling the complex task of creating coherent, in-depth survey articles from a vast academic literature collection. Unlike traditional IR tasks focused solely on document retrieval, SurGE requires systems to:

• Retrieve: Identify relevant academic articles from a corpus of over 1 million papers.
• Organize: Construct a structured and hierarchical survey outline.
• Synthesize: Generate a coherent narrative with proper citations, reflecting expert-authored surveys.

The benchmark includes 205 carefully curated ground truth surveys, each accompanied by detailed metadata and a corresponding hierarchical structure, along with an extensive literature knowledge base sourced primarily from arXiv.

Data Release

This repository contains all necessary components for working with the SurGE dataset:

Dataset Files & Formats:

• Ground Truth Surveys: Each survey includes metadata fields such as title, authors, publication year, abstract, hierarchical structure, and citation lists.
• Literature Knowledge Base: A corpus of 1,086,992 academic papers with key fields (e.g., title, authors, abstract, publication date, and category).
• Auxiliary Mappings: Topic-to-publication mappings to support systematic survey generation.

The complete dataset can be downloaded from this Google Drive folder.

Then you will get the folder data .

Ground Truth Survey

A ground truth survey contains the full content of a survey and its citation information. However, due to space constraints, we cannot display it in its entirety here.

All ground truth surveys are available in data/surveys.json

A survey consists of the following fields:

Field	Description
authors	List of contributing researchers.
survey_title	The title of the survey paper.
year	The publication year of the survey.
date	The exact timestamp of publication.
category	Subject classification following the arXiv taxonomy.
abstract	The abstract of the survey paper.
structure	Hierarchical representation of the survey’s organization.
survey_id	A unique identifier for the survey.
all_cites	List of document IDs cited in the survey.
Bertopic_CD	A diversity measure computed using BERTopic.

Literature Knowledge Base

The corpus containing all literature articles is available in: data/corpus.json

Example : Here, we present how articles are organized in the knowledge base. Overly long abstract has been appropriately shortened.

{
        "Title": "Information Geometry of Evolution of Neural Network Parameters While   Training",
        "Authors": [
            "Abhiram Anand Thiruthummal",
            "Eun-jin Kim",
            "Sergiy Shelyag"
        ],
        "Year": "2024",
        "Date": "2024-06-07T23:42:54Z",
        "Abstract": "Artificial neural networks (ANNs) are powerful tools capable of approximating any arbitrary mathematical function, but their interpretability remains limited...",
        "Category": "cs.LG",
        "doc_id": 1086990
    }

The following are explanations of each field:

Key	Description
Title	The title of the research paper.
Authors	A list of contributing researchers.
Year	The publication year of the paper.
Date	The exact timestamp of the paper’s release.
Abstract	The abstract of the paper.
Category	The subject classification following the arXiv taxonomy.
doc_id	A unique identifier assigned for reference and retrieval.

Auxiliary Mappings:

The mapping containing all queries and their corresponding articles is available in: data/queries.json

Each query in data/queries.json corresponds to a section or paragraph from the ground truth surveys with high citation extraction quality. The associated articles are the references cited in that part of the survey.

Below is an example, Overly long content has been appropriately shortened.

  {
        "original_id": "23870233-7f5b-4ef1-9d38-e6f3adb0fa48",
        "query_id": 486,
        "date": "2020-07-16T09:23:13Z",
        "year": "2020",
        "category": "cs.LG",
        "content": "}\n{\nMachine learning classifiers can perpetuate and amplify the existing systemic injustices in society . Hence, fairness is becoming another important topic. Traditionally...",
        "prefix_titles": [
            [
                "title",
                "Learning from Noisy Labels with Deep Neural Networks: A Survey"
            ],
            [
                "section",
                "Future Research Directions"
            ],
            [
                "subsection",
                "{Robust and Fair Training"
            ]
        ],
        "prefix_titles_query": "What are the future research directions for robust and fair training in the context of learning from noisy labels with deep neural networks?",
        "cites": [
            7771,
            4163,
            3899,
            8740,
            8739
        ],
        "cite_extract_rate": 0.8333333333333334,
        "origin_cites_number": 6
    }

The following are explanations of each field:

Key	Description
original_id	The identifier for the section where this query is from.
query_id	The ID associated with the specific query.
content	The content of the section.
prefix_titles	A hierarchical list of titles of the section/subsection/paragraph
prefix_titles_query	The question this passage is relevant to. The goal of the question is to retrieve relevant documents.
cites	A list of document IDs that are cited within this section.
cite_extract_rate	The ratio of extracted citations to the total number of citations in the original document.
origin_cites_number	The total number of citations originally present in the section.

Note: Not every section in the ground truth surveys has a corresponding entry in queries.json. Only sections with a high citation extraction rate are included.
These entries can be used to train retrieval models, where prefix_titles_query serves as the query and cites contains the relevant document IDs.
The data in queries.json has not been pre-split into training and development sets—you may divide it manually as needed.

License

This project is licensed under MIT License. Please review the LICENSE file for more details.

---

license: mit task_categories: - text-generation language: - en