Combining rule-based and embedding-based approaches to normalize textual entities with an ontology

Combining rule-based and embedding-based approaches to normalize textual entities with an ontology Arnaud Ferré1,2, Louise Deléger1, Pierre Zweigenbaum2, Claire Nédellec1 1MaIAGE, INRA, Université Paris-Saclay, 78350 Jouy-en-Josas, France 2LIMSI, CNRS, Université Paris-Saclay, 91405 Orsay, France arnaud.ferre@universite-paris-saclay.fr, louise.deleger@inra.fr, pierre.zweigenbaum@limsi.fr, claire.nedellec@inra.fr Abstract In this paper, we propose a two-step method to normalize multi-word terms with concepts from a domain-specific ontology. Normalization is a critical step of information extraction. The method uses vector representations of terms computed with word embedding information and hierarchical information among ontology concepts. A training dataset and a first result dataset with high precision and low recall are generated by using the ToMap unsupervised normalization method. It is based on the similarities between the form of the term to normalize and the form of concept labels. Then, a projection of the space of terms towards the space of concepts is learned by globally minimizing the distances between vectors of terms and vectors of concepts. It applies multivariate linear regression using the previously generated training dataset. Finally, a distance calculation is carried out between the projections of term vectors and the concept vectors, providing a prediction of normalization by a concept for each term. This method was evaluated through the categorization task of bacterial habitats of BioNLP Shared Task 2016. Our results largely outperform all existing systems on this task, opening up very encouraging prospects. Keywords: ontology-based entity normalization, vector-based model, distributional semantics, multivariate linear model 1. Introduction An important part of knowledge is expressed in textual form, such as in scientific articles. Specialized literature is characterized by the presence of terms of interest which are often complex nominal groups (e.g. "epithelial cells of the intestine") and display a high variability in their forms. At the same time, life sciences are a field with many structured, albeit incomplete, representations of knowledge: ontologies. These representations have the advantage of being machine interpretable, which can greatly improve the ability of programs to extract and reuse information from texts. The biomedical/biological field is therefore a good candidate for the development of more efficient generic methods that make use of these structured representations. To extract entity information from texts, two steps are commonly applied: recognition of named entities and normalization of these entities (also called entity linking in the general domain). The recognition step detects terms of interest (e.g. bacterial habitat references) while normalization identifies them precisely by linking them to specific concepts or categories of an ontology (e.g. "T- cells" is a bacterial habitat mention that can be identified by the "lymphocyte" labeled concept). We are focusing on this particular task. Today's best-performing normalization methods commonly rely on supervised learning from data that is manually annotated by experts of the field. However in specialized fields, these annotations are rare because they are difficult to obtain. Moreover, given the large number of target concepts in these fields (e.g. the biomedical metathesaurus UMLS contains more than 3 million concepts, the Ontobiotope ontology contains more than 2 thousand concepts, etc.), it seems unlikely to obtain sufficient data …

Content truncated to 500 words