BIOfid Dataset: Publishing a German Gold Standard for Named Entity Recognition in Historical Biodiversity Literature

Proceedings of the 23rd Conference on Computational Natural Language Learning, pages 871–880 Hong Kong, China, November 3-4, 2019. c⃝2019 Association for Computational Linguistics 871 BIOfid Dataset: Publishing a German Gold Standard for Named Entity Recognition in Historical Biodiversity Literature Sajawel Ahmed1, Manuel Stoeckel1, Christine Driller2, Adrian Pachzelt3, Alexander Mehler1 1Goethe University Frankfurt 2Senckenberg Nature Research Society 3Frankfurt University Library {sahmed,mehler}@em.uni-frankfurt.de Abstract The Specialized Information Service Biodiver- sity Research (BIOfid) has been launched to mobilize valuable biological data from printed literature hidden in German libraries for over the past 250 years. In this project, we anno- tate German texts converted by OCR from his- torical scientific literature on the biodiversity of plants, birds, moths and butterflies. Our work enables the automatic extraction of bi- ological information previously buried in the mass of papers and volumes. For this pur- pose, we generated training data for the tasks of Named Entity Recognition (NER) and Taxa Recognition (TR) in biological documents. We use this data to train a number of leading ma- chine learning tools and create a gold standard for TR in biodiversity literature. More specif- ically, we perform a practical analysis of our newly generated BIOfid dataset through vari- ous downstream-task evaluations and establish a new state of the art for TR with 80.23% F- score. In this sense, our paper lays the founda- tions for future work in the field of information extraction in biology texts. 1 Introduction Data is the gold to any machine learning (ML). Most ML approaches to Natural Language Pro- cessing (NLP) address modern, high-resource lan- guages (such as English or Chinese) rather than historical, low-resource languages. As a conse- quence, feasible ML-tools for processing histori- cal documents are still rare. In this paper we con- sider corpora of historical German texts in order to extract useful information about biological sys- tems in the past (e.g. species, biotopes etc.). As a contribution to closing the gap between NLP of modern and of historical languages, we present the newly annotated BIOfid dataset for Named Entity Recognition (NER) and for Taxa Recognition (TR) in the domain of biology, the first of its kind concerning the German language. Our approach is especially designed to address the exploration of biodiversity data1 from historical documents. We perform a large-scale annotation of scanned texts converted by OCR from histori- cal scientific books on the biodiversity of plants, birds, moths and butterflies, thereby creating the necessary training data to accomplish the task of biological NER and TR using various ML algo- rithms. Our work facilitates an automatic extrac- tion of biological information so far buried in the bulk of papers and volumes (see Table 1). Over- Input sentence: Ahmed observes that Iris grows in Mai in Frankfurt. TR output: Ahmed observes that [Iris]TAXON grows in Mai in Frankfurt. Biological NER output: [Ahmed]PER observes that [Iris]TAXON grows in [Mai]TIME in [Frankfurt]LOC. Table 1: Example for our selected tasks. all, our newly generated dataset provides a gold standard and hereby lays the foundations for future work, …

Content truncated to 500 words