UniPi: Recognition of Mentions of Disorders in Clinical Text

Proceedings of the 8th International Workshop on Semantic Evaluation (SemEval 2014), pages 754–760, Dublin, Ireland, August 23-24, 2014. UniPi: Recognition of Mentions of Disorders in Clinical Text Giuseppe Attardi, Vittoria Cozza, Daniele Sartiano Dipartimento di Informatica Università di Pisa Largo B. Pontecorvo, 3 I-56127 Pisa, Italy {attardi, cozza, sartiano}@di.unipi.it Abstract The paper describes our experiments ad- dressing the SemEval 2014 task on the Analysis of Clinical text. Our approach consists in extending the techniques of NE recognition, based on sequence label- ling, to address the special issues of this task, i.e. the presence of overlapping and discontiguous mentions and the require- ment to map the mentions to unique iden- tifiers. We explored using supervised methods in combination with word em- beddings generated from unannotated da- ta. 1 Introduction Clinical records provide detailed information on examination and findings of a patient consulta- tion expressed in a narrative style. Such records abound in mentions of clinical conditions, ana- tomical sites, medications, and procedures, whose accurate identification is crucial for any further activity of text mining. Many different surface forms are used to represent the same concept and the mentions are interleaved with modifiers, e.g. adjectives, verb or adverbs, or are abbreviated involving implicit terms. For example, in Abdomen is soft, nontender, nondistended, negative bruits the mention occurrences are “Abdomen nontender” and “Abdomen bruits”, which refer to the disorders: “nontender abdomen” and “abdomininal bruit”, with only the sec- ond having a corresponding UMLS Concept Unique Identifier (CUI). In this case the two mentions overlap and both are interleaved with other terms, not part of the mentions. Secondly, mentions can be nested, as in this example: left pleural and parenchymal calcifications where the mention calcifications is nested within pleural calcifications. Mentions of this kind are a considerable de- parture from those dealt in typical Named Entity recognition, which are contiguous and non- overlapping, and therefore they represents a new challenge for text analysis. The analysis of clinical records poses addi- tional difficulties with respect to other biomedi- cal NER tasks, which use corpora from the med- ical literature. Clinical records are entered by medical personnel on the fly and so they contain misspellings and inconsistent use of capitaliza- tion. The task 7 at SemEval 2014, Analysis of Clinical Text, addresses the problem of recogni- tion of mentions of disorders and is divided in two parts: A. recognition of mentions of bio-medical concepts that belong to the UMLS se- mantic group disorders; B. mapping of each disorder mention to a unique UMLS CUI (Concept Unique Identifiers). The challenge organizers provided the following resources:  A training corpus of clinical notes from MIMIC II database manually annotated for disorder mentions and normalized to an UMLS CUI, consisting of 9432 sen- tences, with 5816 annotations.  A collection of unannotated notes, consist- ing of 1,611,080 sentences. This work is licensed under a Creative Commons At- tribution 4.0 International Licence. Page numbers and pro- ceedings footer are added by the organisers. Licence details: http://creativecommons.org/licenses/by/4.0/ 754 We also had …

Content truncated to 500 words