README.md

metadata

dataset_info:
  - config_name: full
    features:
      - name: doc_key
        dtype: string
      - name: gutenberg_key
        dtype: string
      - name: sentences
        sequence:
          sequence: string
      - name: clusters
        sequence:
          sequence:
            sequence: int64
      - name: characters
        list:
          - name: name
            dtype: string
          - name: mentions
            sequence:
              sequence: int64
    splits:
      - name: train
        num_bytes: 118643409
        num_examples: 45
      - name: validation
        num_bytes: 5893208
        num_examples: 5
      - name: test
        num_bytes: 2732407
        num_examples: 3
    download_size: 317560335
    dataset_size: 127269024
  - config_name: splitted
    features:
      - name: doc_key
        dtype: string
      - name: gutenberg_key
        dtype: string
      - name: sentences
        sequence:
          sequence: string
      - name: clusters
        sequence:
          sequence:
            sequence: int64
      - name: characters
        list:
          - name: name
            dtype: string
          - name: mentions
            sequence:
              sequence: int64
    splits:
      - name: train
        num_bytes: 118849212
        num_examples: 7544
      - name: validation
        num_bytes: 5905814
        num_examples: 398
      - name: test
        num_bytes: 2758250
        num_examples: 152
    download_size: 317560335
    dataset_size: 127513276
language:
  - en
pretty_name: BOOKCOREF
size_categories:
  - 10M<n<100M
tags:
  - coreference-resolution
license: cc-by-sa-4.0

This repository contains the BookCoref dataset, introduced in the paper "BookCoref: Coreference Resolution at Book Scale" by [Martinelli et al., 2025]().

We release both the manually-annotated test split (BookCoref_gold) and the pipeline-generated train and validation splits (BookCoref_silver). In order to enable the replication of our results, we also release the splitted version of each split, available by adding the suffix _splitted to each split name. As specified in the paper, this version is obtained through chunking the text into contiguous windows of 1500 tokens, retaining the coreference clusters of each window.

📚 Quickstart

Simply load the dataset through Huggingface's datasets library:

from datasets import load_dataset

bookcoref = load_dataset("sapienzanlp/bookcoref")

ℹ️ Data format

BookCoref is a collection of annotated books. Each item contains the annotations of one book following the structure of OntoNotes:

{
  doc_id: "pride_and_prejudice_142", # (str) i.e., id of document 
  sentences: [["Pride", "and", "Prejudice", "."], ["Begin", ...], ...], # list[list[str]] i.e., list of word-tokenized sentences
  clusters: [[[0,0], [3,5]], [[4,9]...], ...], # list[list[list[int]]] i.e., list of clusters' mention offsets
  characters: [
    {
      name: "Mr.Bennet", 
      cluster: [[0,0], ...],
    },
    {
      name: "Mr. Darcy",
      cluster: [[5,7], ...],
    }
  ] # list[character], list of characters objects with name and his mentions offsets, i,e., dict(name: str, cluster: list[list[int]])
}

We also include information on character names, which is not exploited in traditional coreference settings, but can be useful in future works.

📊 Dataset statistics

BookCoref has distinctly book-scale characteristics, as summarized in the following table:

🖋️ Cite this work

This work has been published at ACL 2025 (main conference). If you use any artifact of this dataset, please consider citing our paper as follows:

@inproceedings{martinelli-etal-2025-bookcoref,
    title = "{BookCoref}: Coreference Resolution at Book Scale",
    author = "Martinelli, Giuliano  and
      Bonomo, Tommaso  and
      Huguet Cabot, Pere-Llu{\'\i}s  and
      Navigli, Roberto",
    booktitle = "Proceedings of the 63nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
    month = jul,
    year = "2025",
    address = "Vienna, Austria",
    publisher = "Association for Computational Linguistics",
}

©️ License information

All the annotations provided by this repository are licensed under the Creative Commons Attribution Share Alike 4.0 license. The tokenized text of books is a modification of books from Project Gutenberg, following their license.