JMIR Publications

ABSTRACT

Background:

Negation and speculation are critical elements in tasks related to Natural Language Processing, such as information extraction, as they change the truth-value of a proposition. In clinical narrative, these linguistic phenomena are used extensively with the objective of indicating hypothesis, impressions or negative findings. The previous state of the art approaches addressed negation and speculation detection tasks using rule-based methods but in the last few years, models based on machine learning and deep learning exploiting morphological, syntactic, semantic features represented as spare and dense vector, have emerged. However, although such methods of named entity recognition (NER) employ a broad set of features, they are limited to existing pre-trained models for a specific domain or language.

Objective:

A system for cross-lingual and cross-domain negation and speculation detection is introduced with special focus on biomedical scientific literature and clinical narrative. In this work, negation and speculation detection is considered as a sequence labeling task where cues and their scopes of both phenomena are recognized as a sequence of labels, recognized in an only phase.

Methods:

We propose two approaches: i) a Bidirectional Long Short-Term Memory (Bi-LSTM) and Conditional Random Field (CRF) using character, word and sense embeddings to deal with the extraction of semantic, syntactic and contextual patterns and ii) a Bidirectional Encoder Representations for Transformers (BERT) with fine-tuning for NER.

Results:

The approach was evaluated on English and Spanish in biomedical and reviews domains, particularly with the BioScope Corpus, IULA Corpus, and the SFU Spanish Review Corpus obtaining an F-measure of 86.6, 85.00, and 91.70, respectively.

Conclusions:

These results show that these architectures perform considerably better than the previous rule-based and machine learning-based systems. Moreover, our analysis results show that pre-training WordPieces and word embeddings on biomedical corpora help it to understand complexities inherent to biomedical texts.