NEARSIDE: Structured kNowledge Extraction frAmework from SpecIes DEscriptions

Species descriptions are stored in textual form corpora such as floras and faunas, but this large amount of information cannot be used directly by algorithms, nor can it linked to other data sources. The production knowledge bases expressing structured benefit from collaborative easy-to-use platforms like Xper3 (Vignes-Lebbe et al. 2017, Kerner Vignes 2019, Saucède 2021) is very time-consuming at the human level. It therefore mandatory for task make contained species measurable compatible with computer techniques. One most structures on web deep learning community triplet structure. Each piece represented a set 3 elements (subject, predicate, object). first steps towards accessibility developing text-to-triplet model, also known text-to-graph, monograph descriptions. In work, we developed NEARSIDE, text-to-graph model adapted biology create normalized morphological characteristic Natural Language Processing, models have proven effective extracting open domain (Lample 2016, Sutskever 2014), especially since emergence attention-based (Devlin 2019b, Devlin 2019a). Several works been made biomedical (Fries 2017,Cho Lee 2019). our case, propose floras. Fully supervised require annotated training, nevertheless, annotation process implies an expensive intervention. Distant supervision technique that reduce cost. This paradigm uses small glossary project classes word level new complex longer text (see Fig. 1). Named Entity Recognition (NER) Processing (NLP) consists classifying words interest (Sutskever 2014, Lample 2016), while extraction compared Relation Extraction (RE) which semantic relations between pairs words. Distantly NER often studied subject literature comparison distantly RE (Liang 2020, Meng simply because subtask distant annotations generation less 2). Our contribution creating description dataset well-balanced test allows us bypass several biases induced observed datasets (Taillé 2021). dataset, each will classified into one 15 classes, class being specific kind organ or descriptor. second proposing trained fauna flora particularly long use technical vocabulary. We develop context-oriented pretraining language model. Thus encoder provides contextualized vectors extracted measure similarities different species. reaches 96% accuracy named entity classification set. third construction module applied model's outputs. based dependency rules inspired Xper3’s representation format 3). Finally, NEARSIDE end-to-end framework unstructured corpora, making easily linked, measured.