Cimind: A phonetic-based tool for multilingual named entity recognition in biomedical texts.

BACKGROUND

Extracting concepts from biomedical texts is a key to support many advanced applications such as biomedical information retrieval. However, in clinical notes Named Entity Recognition (NER) has to deal with various types of errors such as spelling errors, grammatical errors, truncated sentences, and non-standard abbreviations. Moreover, in numerous countries, NER is challenged by the availability of many resources originally developed and only suitable for English texts. This paper presents the Cimind system, a multilingual system dedicated to named entity recognition in medical texts based on a phonetic similarity measure.

METHODS

Cimind performs entity recognition by combining phonetic recognition using the DM phonetic algorithm to deal with spelling errors and string similarity measures. Three main steps are processed to identify terms in a controlled vocabulary: normalization, candidate selection by phonetic similarity and candidate ranking.

RESULTS

Cimind was evaluated in the 2016 and 2017 editions of the CLEF eHealth challenge in the CépiDC/CDC tasks. In 2017, it obtained on each corpus the following results: English dataset: 83.9% P, 78.3% R, 81.0% F1; French raw dataset: 85.7% P, 68.9% R, 76.4% F1; French aligned dataset: 83.5% P, 77.5% R, 80.4% F1. It ranked first in French and fourth in English in officials runs.