ADS has been building and curating datasets to train deep learning models, both to provide a richer user experience and to enable automated data enrichment in our internal pipelines. The datasets described here are open to researchers and are publicly available and easy to access. The models are licensed under an MIT license and the datasets are licensed under a CC-BY 4.0 license. Briefly, these licenses allow researchers to use, share, modify or build upon these works as long as appropriate attribution is given. We discuss two of these datasets here.

The Detecting Entities in the Astrophysical Literature (DEAL) dataset is a curated dataset for Named Entity Recognition (NER). This task involves identifying predetermined entities in text, such as Organization or Location. The dataset consists of text fragments obtained from the fulltext or acknowledgements sections of the astrophysical literature (Astrophysical Journal, Astronomy & Astrophysics, and the Monthly Notices of the Royal Astronomical Society). Roughly 6000 text snippets containing over 147,000 labeled entities were manually labeled by a domain expert with 33 different entity types. The DEAL dataset was used as part of a shared task in the First Workshop on Information Extraction from the Scientific Literature (WIESP 2022) as part of the AACL-IJCNLP 2022 conference. The proceedings of this workshop are part of the ACL Anthology.

The Function Of Citation in Astrophysics Literature (FOCAL) dataset is a curated dataset for citation context analysis, which “facilitates the syntactic and semantic analysis of the contents of the citation context to understand how and why authors discuss others research work” (Kunnath et al. 2021), including as background or as motivation. The snippets that contain the citations are obtained from over 25,000 astronomy articles, from the same journals and publication years as the DEAL dataset. From this set of articles, over 2 million citations and their context are harvested. Further, only citations with context sizes between 2,000 and 10,000 characters are selected. This is to allow the determination of what portions of the context are most relevant to understanding the citation’s function. A domain area expert manually examined these text snippets to determine the citation function as well as label the relevant context. In total there are 6023 instances of annotated citations. The FOCAL dataset will be used for the Second Workshop on Information Extraction from the Scientific Literature (WIESP 2023) part of IJCNLP-AACL 2023.

To learn more about our datasets and astroBERT, our astrophysics-specific language model, see our blog post.

The automatic identification of planetary features, such as craters or mares, in astronomy and planetary science publications presents numerous challenges. 

Many feature names overlap with places or people’s names that they are named after, such as Tempe or Sagan. Some feature names have been used in many contexts, e.g. Apollo, or can appear in the text as adjectives, like the lunar craters Black, Green, and White. Additionally, some features share identical names across different celestial bodies, requiring disambiguation, such as the Adams crater, which exists on both the Moon and Mars.

We have developed a multi-step pipeline combining rule-based filtering, statistical relevance analysis, part-of-speech (POS) tagging, a named entity recognition (NER) model, hybrid keyword harvesting, knowledge graph (KG) matching, and inference with a locally installed large language model (LLM) to reliably identify planetary names despite these challenges. When evaluated on a dataset of astronomy papers from the Astrophysics Data System (ADS), this methodology achieves an F1-score over 0.97 in disambiguating planetary feature names.

With this pipeline, we have tagged over 6000 papers with features from over three dozen solar system bodies. Searching on these tags is now available in ADS using the gpn search tag, e.g. gpn:Mars. Browsing and filtering on these search tags is available in the new SciX interface.

To read more about the work behind the planetary features function, see our pre-print.