Handling Entity Normalization with no Annotated Corpus: Weakly Supervised Methods Based on Distributional Representation and Ontological Information

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 1959–1966 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 1959 Handling Entity Normalization with no Annotated Corpus: Weakly Supervised Methods Based on Distributional Representation and Ontological Information Arnaud Ferré1,2, Robert Bossy1, Mouhamadou Ba1, Louise Deléger1, Thomas Lavergne2, Pierre Zweigenbaum2, Claire Nédellec1 1Université Paris-Saclay, INRAE, MaIAGE, allée de Vilvert, 78350 Jouy-en-Josas, France 2Université Paris-Saclay, CNRS, LIMSI, rue John Von Neumann, 91400 Orsay, France {arnaud.ferre, robert.bossy, mouhamadou.ba, louise.deleger, claire.nedellec}@inrae.fr {thomas.lavergne, pierre.zweigenbaum}@limsi.fr Abstract Entity normalization (or entity linking) is an important subtask of information extraction that links entity mentions in text to categories or concepts in a reference vocabulary. Machine learning based normalization methods have good adaptability as long as they have enough training data per reference with a sufficient quality. Distributional representations are commonly used because of their capacity to handle different expressions with similar meanings. However, in specific technical and scientific domains, the small amount of training data and the relatively small size of specialized corpora remain major challenges. Recently, the machine learning-based CONTES method has addressed these challenges for reference vocabularies that are ontologies, as is often the case in life sciences and biomedical domains. Its performance is dependent on manually annotated corpus. Furthermore, like other machine learning based methods, parametrization remains tricky. We propose a new approach to address the scarcity of training data that extends the CONTES method by corpus selection, pre-processing and weak supervision strategies, which can yield high-performance results without any manually annotated examples. We also study which hyperparameters are most influential, with sometimes different patterns compared to previous work. The results show that our approach significantly improves accuracy and outperforms previous state-of-the-art algorithms. Keywords: Information Extraction, entity normalization, entity linking, supervised learning 1. Introduction Entity Normalization, also known as Entity Linking or Entity Grounding, is an Information Extraction subtask that consists in the assignment of categories or concepts from a reference vocabulary to entity mentions in the text (i.e. words or sequences of words). Ontologies and taxonomies have been frequently used as reference vocabularies since the late 1990s (Faure and Nédellec, 1998; Hwang, 1999). For instance, entity normalization with concepts of an ontology could consist in linking the textual entity mention “T-cell” to a concept referenced by a unique identifier and a label such as: <OBT:001342: lymphocyte>. Entity Normalization contributes to the reuse of the extracted information and to its integration with data in reference databases (Nédellec et al., 2009). Entity Normalization has recently gained traction in technical and scientific domains, especially in Life Sciences and Health Sciences. Several benchmark datasets have been proposed in these domains (Wei et al., 2015; Roberts et al., 2017; Deléger et al., 2016). Entity Normalization methods handle the problem as a classification problem. They are based either on pattern-matching rules (Aronson, 2001) or on Machine Learning (ML) algorithms (Leaman et al., 2013). Both are able to operate on the surface form of entities (i.e. their sequence of characters), on NLP analyses …

Content truncated to 500 words