TheoremKB is a research project and a corresponding collection of tools to extract semantic information from (mathematical) research articles. This is an ongoing project, with preliminary code available from this repository.
For a high-level overview of the project, see this set of slides.
For a more in-depth look at some of the aspects of the project interms of publications/internship reports (chronologically sorted), see:
- Lucas Pluvinage's Master Thesis on using style-based information for Extracting scientific results from research articles.
- Théo Delemazure's Master Thesis on first steps towards A Knowledge Base of Mathematical Results.
- Shrey Mishra's paper on Towards Extraction of Theorems and Proofs in Scholarly Articles comparing various techniques evaluated individually at a single line level.
- Yacine Brihmouche's Master's thesis on TheoremKB: a knowledge base of Mathematical results connecting proofs and theorems from different papers.
- Antoine Gauquier's Master's thesis on Impact of the document class in the automatic extraction of mathematical environments in the scientific literature
- Antoine Gauquier's paper on Automatically inferring the document class used in a scientific article.
- Shrey Mishra's paper on Multimodal Machine Learning for Extraction of Theorems and Proofs in the Scientific Literature.
- Shufan JIANG's paper on Extracting Definienda in Mathematical Scholarly Articles with Transformers.
- Shrey Mishra's paper on First Steps in Building a Knowledge Base of Mathematical Results .
- Shrey Mishra's PhD thesis on Multimodal Extraction of Proofs and Theorems from the Scientific Literature
One of our dataset of reference is formed of 4400 articles extracted from arXiv, see arXiv Bulk Data Access for bulk access to the data. For licensing reasons, this datasets cannot be reshared, but we provide in Dataset/links.csv the reference to all articles of the dataset.
We are currently experimenting with the extraction of mathematical results based upon 3 approaches:
- Using style-based information
- Using Computer Vision based object detection to identify mathematical results
- Using NLP based techniques such as transformers and LSTM networks for sequence prediction
For Computer Vision and NLP based extractions (Please follow the jupyter notebooks) in the directory
/Computer_Vision
and NLP
- Computer Vision notebooks
/Computer_Vision/1.1 Computer vision preprocessing.ipynb
contains the preprocessing step and preparing the data into YOLO format
/Computer_Vision/obj.data
, /Computer_Vision/obj.names
, /Computer_Vision/yolov4-obj.cfg
contains the image annotations directory path, class labels and configuration file of the YOLO network trained
- NLP notebooks
/2.1 NLP text data preprocessing.ipynb
contains the preprocessing step and preparing of the xml files
/transformers_tkb.ipynb
contains application of several AutoEncoding Transformers all base models (SciBert, Bert, DistilBert)
/lstm_tkb_full.ipynb
contains LSTM implementation on Full data
/lstm_trimmed.ipynb
contains LSTM implementation on imbalanced data
- Style based
See the instructions within the Styling directory.
The project is led by Pierre Senellart, within the Valda research group joint between ENS, PSL University, CNRS and Inria.
The project has also involved:
- Théo Delemazure, Master's student, ENS
- Lucas Pluvinage, Master's student, ENS
- Shrey Mishra, PhD candidate, ENS
- Antoine Gauquier, PhD candidate, ENS
- Shufan JIANG, Postdoctoral Candidate, ENS
Contact Pierre Senellart for further information.
If you find our work useful and would like to cite it, please use the following BibTeX entry:
@inproceedings{mishra-etal-2024-first,
title = "First Steps in Building a Knowledge Base of Mathematical Results",
author = "Mishra, Shrey and
Brihmouche, Yacine and
Delemazure, Th{\'e}o and
Gauquier, Antoine and
Senellart, Pierre",
editor = "Ghosal, Tirthankar and
Singh, Amanpreet and
Waard, Anita and
Mayr, Philipp and
Naik, Aakanksha and
Weller, Orion and
Lee, Yoonjoo and
Shen, Shannon and
Qin, Yanxia",
booktitle = "Proceedings of the Fourth Workshop on Scholarly Document Processing (SDP 2024)",
month = aug,
year = "2024",
address = "Bangkok, Thailand",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2024.sdp-1.16",
pages = "165--174",
abstract = "This paper explores the initial steps towards extracting information about theorems and proofs from scholarly documents to build a knowledge base of interlinked results. Specifically, we consider two main tasks: extracting results and their proofs from the PDFs of scientific articles and establishing which results are used in the proofs of others across the scientific literature. We discuss the problem statement, methodologies, and preliminary findings employed in both phases of our approach, highlighting the challenges faced.",
}
@inproceedings{DBLP:conf/doceng/MishraPS21,
author = {Shrey Mishra and
Lucas Pluvinage and
Pierre Senellart},
editor = {Patrick Healy and
Mihai Bilauca and
Alexandra Bonnici},
title = {Towards extraction of theorems and proofs in scholarly articles},
booktitle = {DocEng '21: {ACM} Symposium on Document Engineering 2021, Limerick,
Ireland, August 24-27, 2021},
pages = {25:1--25:4},
publisher = {{ACM}},
year = {2021},
url = {https://doi.org/10.1145/3469096.3475059},
doi = {10.1145/3469096.3475059},
timestamp = {Fri, 20 Aug 2021 15:13:08 +0200},
biburl = {https://dblp.org/rec/conf/doceng/MishraPS21.bib},
bibsource = {dblp computer science bibliography, https://dblp.org}
}
All code is provided as open-source software under the MIT License. See LICENSE.
This work has been funded by the French government under management of Agence Nationale de la Recherche as part of the “Investissements d’avenir” program, reference ANR-19-P3IA-0001 (PRAIRIE 3IA Institute).
Pierre Senellart's work is also supported by his secondment to Institut Universitaire de France.