---
dataset_info:
  features:
  - name: id
    dtype: uint64
  - name: s2orc_id
    dtype: uint64
  - name: mag_id
    dtype: uint64
  - name: doi
    dtype: string
  - name: title
    dtype: string
  - name: abstract
    list:
      list:
      - name: title_path
        list: string
      - name: text
        dtype: string
      - name: citations
        list:
        - name: index
          dtype: uint16
        - name: start
          dtype: uint32
        - name: end
          dtype: uint32
      - name: references
        list:
        - name: index
          dtype: uint16
        - name: start
          dtype: uint32
        - name: end
          dtype: uint32
  - name: related_work
    dtype: string
  - name: hierarchy
    dtype: string
  - name: authors
    list: string
  - name: year
    dtype: uint16
  - name: fields_of_study
    list: string
  - name: referenced
    list:
    - name: id
      dtype: uint64
    - name: s2orc_id
      dtype: uint64
    - name: mag_id
      dtype: uint64
    - name: doi
      dtype: string
    - name: title
      dtype: string
    - name: hierarchy
      dtype: string
    - name: authors
      list: string
    - name: year
      dtype: uint16
    - name: fields_of_study
      list: string
    - name: citations
      list: uint64
    - name: bibliography
      list:
      - name: id
        dtype: uint64
      - name: title
        dtype: string
      - name: year
        dtype: uint16
      - name: authors
        list: string
    - name: non_plaintext_content
      list:
      - name: type
        dtype: string
      - name: description
        dtype: string
  - name: bibliography
    list:
    - name: id
      dtype: uint64
    - name: title
      dtype: string
    - name: year
      dtype: uint16
    - name: authors
      list: string
  - name: non_plaintext_content
    list:
    - name: type
      dtype: string
    - name: description
      dtype: string
  splits:
  - name: train
    num_bytes: 39235598318
    num_examples: 91445
  - name: validation
    num_bytes: 581643389
    num_examples: 1127
  - name: test
    num_bytes: 965353630
    num_examples: 1878
  download_size: 15174246190
  dataset_size: 40782595337
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: validation
    path: data/validation-*
  - split: test
    path: data/test-*
---

# OARelatedWork
OARelatedWork is a large-scale multi-document summarization dataset for related work generation containing whole related work sections and full-texts of cited papers. The dataset includes 94 450 papers and 5 824 689 unique referenced papers.

|          Split         | Samples |
|------------------------|---------|
|        Train           |   91,445|
|     Validation         |    1,127|
|         Test           |    1,878|

## Fields

* **id** - id from our corpus
* **s2orc_id** - SemanticScholar id
* **mag_id** - Microsoft Academic Graph id
* **DOI** - Might be DOI for another version of document than the one used for processing.
* **title** - title of publication
* **abstract** - list of paragraphs in an abstract, every paragraph is a list of sentences
* **related_work** - The target related work section. The format differs according to used configuration.
* **hierarchy** - Document body, but the abstract and related work section. The format differs according to used configuration.
* **authors** - authors of publication
* **year** - year of publication
* **fields_of_study** - list of fields of study
* **referenced** - List of referenced document. Each referenced document has the same fields, but the abstract, related_work, and referenced field are missing. All references have the abstract section as a first section in hierarchy.
* **bibliography** - document bibliography
* **non_plaintext_content** - tables and figures

## Structure
We provide multiple dataset configurations to make working with this dataset as simple as possible. Also, by the time this dataset is released, it is not possible to use hierarchical structures, which we use to represent document content. Thus, we used several workarounds, such as flattening the hierarchy or using a JSON representation of hierarchy.

We divide a document content into sections, subsections, paragraphs, and sentences. Not all documents have full text and subsections.

### Flattened hierarchy
The hierarchy is flattened on section level. meaning that it is a list of (sub)sections. Each(sub)section is represented by list of titles on tree path to given section and list of paragraphs in given (sub)section. Each paragraph is represented as a list of sentences. Every sentence also contains metadata such as citation spans.

### Configurations
* **oa_related_work**

   uses JSON format to represent hierarchy

* **abstracts**

   provides just abstracts of cited papers, hierarchy of target paper is flattened

* **flattened_sections**

   hierarchy is flattened, see the Flattened hierarchy section [above](#flattened-hierarchy)

* **greedy oracle based configurations**

   These configurations provide filtered content using greedy oracle. Since the greedy oracle is a cheating baseline, use these with care.

   * **greedy_oracle_sentences**

      Each referenced document is represented by sentences that are in greedy extractive oracle summary. It is using same format as flattened_sections.

   * **greedy_oracle_paragraphs**

      Each referenced document is represented by paragraphs that contain sentences that are in greedy extractive oracle summary. It is using same format as flattened_sections.

   * **greedy_oracle_per_input_doc_sentences**

      Each referenced document is represented by sentences that are in greedy extractive oracle summary done on each document separately. It is using same format as flattened_sections.

   * **greedy_oracle_per_input_doc_paragraphs**

      Each referenced document is represented by paragraphs that contain sentences that are in greedy extractive oracle summary done on each document separately. It is using same format as flattened_sections.

   * **abstracts_with_greedy_oracle_target_sentences**

      Same as abstracts, but target is greedy oracle summary of target document. Target document is the one for which the related work is generated for.

## I don't want to use Hugging Face loader
We also provide our custom loader that is available at [https://github.com/KNOT-FIT-BUT/OAPapersLoader](https://github.com/KNOT-FIT-BUT/OAPapersLoader).

## TUI Viewer
We provide a TUI viewer with the dataset ([https://github.com/KNOT-FIT-BUT/OAPapersViewer](https://github.com/KNOT-FIT-BUT/OAPapersViewer)), as it is difficult to navigate data of this kind, especially when one wants to investigate the content of cited papers.

![TUI Viewer](tui_viewer.png)

## Sources
The dataset contains open access papers obtained from **CORE** and **SemanticScholar** corpora. These corpora contain third party content and materials, such as open access works from publicly available sources. In addition to the licenses of those organizations (ODC-By, CC BY-NC), any underlying Third Party Content may be subject to separate license terms by the respective third party owner. We made the best effort to provide identifiers (title, authors, year, DOI, or SemanticScholar ID) of collected papers to allow the user of this dataset to check the license.