Named Entity Recognition in the Medical Domain with Constrained CRF Models

Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 1, Long Papers, pages 839–849, Valencia, Spain, April 3-7, 2017. c⃝2017 Association for Computational Linguistics Named Entity Recognition in the Medical Domain with Constrained CRF Models Charles Jochim IBM Research - Ireland Dublin, Ireland charlesj@ie.ibm.com L´ea A. Deleris IBM Research - Ireland Dublin, Ireland lea.deleris@ie.ibm.com Abstract This paper investigates how to improve performance on information extraction tasks by constraining and sequencing CRF-based approaches. We consider two different relation extraction tasks, both from the medical literature: dependence relations and probability statements. We explore whether adding constraints can lead to an improvement over standard CRF decoding. Results on our relation extrac- tion tasks are promising, showing signif- icant increases in performance from both (i) adding constraints to post-process the output of a baseline CRF, which cap- tures “domain knowledge”, and (ii) fur- ther allowing flexibility in the application of those constraints by leveraging a binary classifier as a pre-processing step. 1 Introduction With the number of articles indexed by MED- LINE/PubMed exceeding one million articles per year1, manual consumption of published medical literature is no longer practical and researchers are increasingly turning to automated techniques to quickly identify and process relevant medical knowledge (e.g., literature-based discovery (Hris- tovski et al., 2006)). Our overall project’s objec- tive is to semi-automate the construction of de- cision support models by generating probabilistic graphical models of medical conditions and their associated risks based on the academic literature (Deleris et al., 2013). An essential task in this project consists of extracting from medical papers any relations mentioning (i) dependence or inde- pendence between variables and (ii) probability 1https://mbr.nlm.nih.gov/Background. shtml statements indicating the strength of a relation- ship. For both types of relations, we have ap- proached the entity extraction step as a sequence labeling problem. We then rely on a set of rules to construct relations from the entities extracted. This paper focuses specifically on the entity ex- traction step which we describe in more detail in the following section. We then proceed in Sec- tion 3 with details about our suggested constraint enforcement procedures. Section 4 reports details about the experiments, with numerical results re- ported and discussed in Section 5. 2 Dependence Relation and Probability Statement Extraction 2.1 Dependence Relation Extraction The first task concerns identifying dependence re- lations between pairs of potential variables men- tioned in text. Dependence and independence here are to be understood as defined by probability the- ory where A depends on B iff Pr(A) ̸= Pr(A|B). As independence statements seldom occur in the literature, we focus predominantly on dependence in this paper and define independence as the nega- tion of dependence. Our choice to use the term “dependence” to describe the task may lead to some ambiguity in the NLP world yet from a prob- ability perspective, it is a precise characterization. Thus we caution the reader that our use of the word “dependence” is exclusively related to its meaning in probability theory. We …

Content truncated to 500 words