An analysis of full-size Russian complexly NER labelled corpus of Internet user reviews on the drugs based on deep learning and language neural nets

TL;DR

This paper introduces a comprehensive Russian NER-labeled corpus of Internet drug reviews and evaluates deep learning models for extracting pharmacological entities, establishing state-of-the-art results for Russian text analysis.

Contribution

It provides a large, annotated Russian corpus for pharmacological entity recognition and analyzes the impact of different model modifications on extraction accuracy.

Findings

Achieved 61.1 F1 score for adverse drug reaction recognition.

Baseline coreference extraction precision is 71%.

Model modifications significantly influence extraction performance.

Abstract

We present the full-size Russian complexly NER-labeled corpus of Internet user reviews, along with an evaluation of accuracy levels reached on this corpus by a set of advanced deep learning neural networks to extract the pharmacologically meaningful entities from Russian texts. The corpus annotation includes mentions of the following entities: Medication (33005 mentions), Adverse Drug Reaction (1778), Disease (17403), and Note (4490). Two of them - Medication and Disease - comprise a set of attributes. A part of the corpus has the coreference annotation with 1560 coreference chains in 300 documents. Special multi-label model based on a language model and the set of features is developed, appropriate for presented corpus labeling. The influence of the choice of different modifications of the models: word vector representations, types of language models pre-trained for Russian, text normalization styles, and other preliminary processing are analyzed. The sufficient size of our corpus allows to study the effects of particularities of corpus labeling and balancing entities in the corpus. As a result, the state of the art for the pharmacological entity extraction problem for Russian is established on a full-size labeled corpus. In case of the adverse drug reaction (ADR) recognition, it is 61.1 by the F1-exact metric that, as our analysis shows, is on par with the accuracy level for other language corpora with similar characteristics and the ADR representativnes. The evaluated baseline precision of coreference relation extraction on the corpus is 71, that is higher the results reached on other Russian corpora.