SemEval 2022 Task 12: Symlink - Linking Mathematical Symbols to their Descriptions

Proceedings of the 16th International Workshop on Semantic Evaluation (SemEval-2022), pages 1671 - 1678 July 14-15, 2022 ©2022 Association for Computational Linguistics SemEval 2022 Task 12: Symlink Linking Mathematical Symbols to their Descriptions 1Viet Dac Lai, 1Amir Pouran Ben Veyseh, 2Franck Dernoncourt, 1Thien Huu Nguyen 1Dept. of Computer and Information Science, University of Oregon, Eugene, Oregon, USA {vietl,apouranb,thien}@cs.uoregon.edu 2Adobe Research, Seattle, Washington, USA franck.dernoncourt@adobe.com Abstract We describe Symlink, a SemEval shared task of extracting mathematical symbols and their descriptions from LaTeX source of scientific documents. This is a new task in SemEval 2022, which attracted 180 individual registra- tions and 59 final submissions from 7 partic- ipant teams. We expect the data developed for this task and the findings reported to be valuable for the scientific knowledge extrac- tion and automated knowledge base construc- tion communities. The data used in this task is publicly accessible at https://github. com/nlp-uoregon/symlink. 1 Introduction The exponential growth of published articles may exceeds many readers’ ability to keep track of the development of their field of interest. Hence, au- tomatic reading comprehension of scientific doc- uments has attracted the attention of researchers across various domains such as Drug Discovery, Knowledge Base Construction, and Natural Lan- guage Processing. A crucial aspect of understand- ing scientific literature is understanding terminolo- gies and formulae because they offer an explicit and precise interface to present the relation between scientific concepts (Schubotz et al., 2018). As such, a reading comprehension machine needs to (i) iden- tify their descriptions and formulae, (ii) segment them into primitive terms and symbols, and (iii) link the associated terms and corresponding sym- bols. Working with mathematical formulae is arduous due to two fundamental reasons. First, common text encodings such as ASCII and Unicode do not fully support typing mathematical symbols. As a result, complex mathematical formulae are rarely written using either ASCII or Unicode. Rather, a higher level encoding (or typesetting) is often used to encode the content of scientific documents, in particular LaTeX. Second, most scientific docu- ments are stored in one of two forms: photos or Portable Document Format (PDF). Scientific doc- uments that were published prior to the graphical computer era are printed and now scanned and dis- tributed as photos. Nowadays, scientific documents are often composed in some text editors or word processing software, then exported and shared a PDF file. Unfortunately, analyzing textual infor- mation in photo images or PDF files is extremely difficult, and most of the natural language process- ing tools are not developed to handle this format. As such, to facilitate the understanding of scientific literature, documents should be stored using a uni- versal easy-to-process text-like encoding. In this paper, we use LaTeX as the typesetting to facili- tate document analysis. Thanks to recent advances in text processing and image recognition, a LaTeX document can often be restored to some extent from either a photo or a PDF file (Deng et al., 2017). This paper introduces the Symlink shared task for the extraction of mathematical symbols and …

Content truncated to 500 words