Combining the tasks of entity linking and relation extraction using a unified neural network model

In this paper we describe methods for training neural network models for extracting pharmacologically significant entities from natural language texts with their further transformation into a formalized form of thesauruses and specialized dictionaries, as well as establishing relations between them. The task of extracting relevant pharmaceutical information from Internet texts is in demand by pharmacovigilance to monitor the effects and conditions of taking medicines. The analysis of texts from the Internet is complicated by the presence of informal speech and distorted terminology. Therefore, the analysis requires not only extracting pharmacologically relevant information, but also bringing it to a standardized form. The purpose of this work is to obtain an end-to-end neural network model that solves all three tasks – entity recognition, relation extraction, and entity disambiguation – in order to avoid sequential processing of one text by independent models. We consider approaches based on generative neural networks that create sequences of words according to a given input text and extractive ones that select and classify words and sequences within the source text. The results of the comparison showed the advantage of the extractive approach over the generative one on the considered set of tasks. The models of this approach outperform the generative model by 5% (f1-micro=85.9) in the task of extracting pharmaceutical entities, by 10% (f1-micro=72.8) in the task of extracting relations and by 4% (f1-micro=64.5) in the entity disambiguation. A joint extractive model was also obtained for three tasks with f1-micro accuracy: 83.4, 68.2, 57.4 for each of the tasks.