Word embeddings for negation detection in health records written in Spanish.

This work focuses on the creation of a system to detect negated medical entities in electronic health records (EHRs) written in Spanish. The importance of this task rests on the influence that the negation can have in the automatic understanding of information given that it inverts the truth value of a clause. We explore a novel continuous characterization as an alternative to previous negation extraction approaches based on discrete characterizations. The aim is to increase the ability of the characterization to generalize over discrete features. We also included other features that could be useful for the negation detection task. In addition, the negation detection is approached as a named entity recognition task where we want to find only the negated entities. EHRs are represented by the corresponding embeddings. In addition, this approach is compared with a traditional discrete characterization based on words. These representations are employed by a supervised classifier such as conditional random fields to infer the predictive model. The approach is assessed on health records from different hospitals, namely IxaMed-GS and IULA. The best performance is achieved by virtue of the embedding-based characterization, leading to an f-measure of 75.3 and 81.6 for the IxaMed-GS and IULA corpus, respectively. With this work, we prove that the use of embedding-based representations can also be useful for the detection of negated medical entities. [ABSTRACT FROM AUTHOR]