Code

• pke is an open source python-based keyphrase extraction toolkit. It provides an end-to-end keyphrase extraction pipeline in which each component can be easily modified or extended to develop new approaches. This toolkit also allows for easy benchmarking of state-of-the-art keyphrase extraction approaches, and ships with supervised models trained on the SemEval-2010 dataset.

• takahe is a multi-sentence compression module. Given a set of redundant sentences, a word-graph is constructed by iteratively adding sentences to it. The best compression is obtained by finding the shortest path in the word graph. A keyphrase-based reranking method can be applied to generate more informative compressions.

Models

• ContriBERT-ACL is a SciBERT model fine-tuned on the ARRContributions dataset to identify paper contributions.
  • Loading…
• bart-large-kp20k is a bart-large model for scientific keyphrase generation fine-tuned on the KP20k dataset.
  • Loading…
• bart-base-kp20k is a bart-base model for scientific keyphrase generation fine-tuned on the KP20k dataset.
  • Loading…

Datasets

• ARRContributions: A dataset of contribution types from ACL Rolling Review papers
  A dataset of more than 2,000 ACL Rolling Review papers labelled with contribution types by their authors, with a test set also annotated by domain experts.
  • Loading…
• EarlySciRev: A dataset of early-stage scientific revisions
  A dataset of revisions made by researchers while drafting their papers, extracted from the LaTeX writing traces of scientific articles, capturing how scientific text evolves in its early stages rather than between published versions.
  • Loading…
• SciClaimEval: A dataset for cross-modal claim verification in scientific Papers
  A dataset of authentic scientific claims paired with figure and table evidence. Refuted examples are built by perturbing the evidence rather than generating synthetic claims, resulting in realistic and controlled instances for evaluating cross-modal scientific claim verification systems.
  • Loading…
• ACL-rlg: A dataset for reading list generation
  The largest dataset of expert-crafted reading lists, made of 85 reading lists manually extracted from tutorial papers presented at ACL-related conferences between 2020 and 2024, enriched with Semantic Scholar metadata for evaluating reading list generation systems.
  • Loading…
• pararev: A dataset of revised scientific paragraphs with revision instructions
  A dataset of revised scientific paragraphs, providing pairs of original and revised versions enriched with detailed revision instructions and a manually annotated evaluation subset, enabling research on context-aware and instruction-guided scientific text revision.
  • Loading…
• parareval: Human evaluations of LLM-generated scientific revisions
  A dataset of human pairwise judgments of paragraph revisions generated by six LLMs from the pararev test set, annotated for instruction following, correctness and preference, for assessing how well automatic metrics evaluate scientific text revision.
  • Loading…
• CASIMIR: A Corpus of Scientific Articles enhanced with Multiple Author-Integrated Revisions
  A dataset of 15,646 scientific articles from OpenReview, providing sentence-aligned revisions across multiple versions along with peer reviews, automatically extracted edits, revision intentions, and paragraph-level metadata to support studies on scientific writing and revision.
  • Loading…
• silk: Silver-standard keyphrases from citation contexts for domain adaptation
  A dataset of synthetic samples for adapting keyphrase generation models to new domains. It contains synthetic samples for three specific domains: Natural Language Processing, Astrophysics and Paleontology, along with human-labeled test sets to evaluate keyphrase generation performance across these domains.
  • Loading…
• kp-biomed: A Large-Scale Dataset for Biomedical Keyphrase Generation
  A large-scale dataset of 5.6 million PubMed abstracts with author assigned keyphrases for training and evaluating neural keyphrase generation models on the biomedical domain.
  • Loading…
• KPNews: A Large-Scale Dataset for Keyphrase Generation on News Documents
  A large-scale dataset of 279,923 news texts paired with editor-curated keyphrases for training and evaluating neural keyphrase generation models on the news domain.
  • Loading…
• kp20k: KP20k benchmark dataset for keyphrase generation
  The KP20k benchmark (Meng et al., 2017) of 570,809 scientific abstracts with author-assigned keyphrases, preprocessed and with reference keyphrases categorized under the PRMU (Present-Reordered-Mixed-Unseen) scheme.
  • Loading…
• inspec: Inspec benchmark dataset for keyphrase generation
  The Inspec benchmark (Hulth, 2003) of 2,000 scientific abstracts with keyphrases assigned by professional indexers, preprocessed and with reference keyphrases categorized under the PRMU (Present-Reordered-Mixed-Unseen) scheme.
  • Loading…
• semeval-2010-pre: Preprocessed SemEval-2010 benchmark dataset
  The SemEval-2010 benchmark dataset (Kim et al., 2010) for automatic keyphrase extraction, already preprocessed at four increasingly sophisticated levels of linguistic preprocessing.
  • Loading…
• wikinews-fr-100: A Dataset of French Wikinews articles for keyphrase generation
  A dataset of 100 French Wikinews articles published between May and December 2012, annotated with a combination of three types of reader annotations.
  • Loading…
• termith-eval: A French benchmark dataset for keyphrase generation
  A dataset of 400 French scientific abstracts from the Inist FRANCIS and PASCAL databases, with keyphrases assigned by professional indexers, for evaluating keyphrase extraction and generation models in French.
  • Loading…
• TALN Archives: Digital archive of French research articles in Natural Language Processing
  TALN Archives is a digital repository of French research articles in Natural Language Processing, containing papers published at the TALN and RECITAL conferences from 1997 to 2015. This dataset originated from my initiative to create a central repository for French NLP papers and is continuously updated each year by the ATALA association at talnarchives.atala.org.
  • Loading…

• LINA-msc: LINA Multi-sentence Compression dataset
  LINA-msc is a dataset for evaluating Multi-sentence Compression in French. It is made of 40 sets of related sentences along with reference compressions composed by human assessors.

• ACM-CR: A manually annotated test collection for citation recommendation
  A test collection for context-aware citation recommendation built from open-access papers and bibliographic records of the ACM Digital Library, in which citation contexts are manually annotated with the papers they cite.

• CLIREC: CLinical Information Retrieval Evaluation Collection
  CLIREC is a test collection for clinical IR. From a set of systematic reviews, we have generated 423 queries with relevance data. Relevance judgments were manually collected from the References section containing the citations from which the synthesized results of the review were extracted.