Key points are not available for this paper at this time.
Discovering more time-effective and a wider range of adverse drug reactions (ADRs) from social texts related to feelings concerning taking medication has recently received significant interest in pharmacovigilance research. Recognizing the posts that include ADRs is an important step for detecting ADRs from social texts. The existing systems show the unsatisfactory performance due to the insufficient expression of emotions and the inadequacy of information expression in short social texts. Although these systems exploit emotional features to improve the performance of their methods, the representation of word-level emotional scores is insufficient for emotional expression. Moreover, most of the systems make less use of medical knowledge to enhance the detection of the potential relationship between drugs and adverse reactions in posts. Therefore, enough expression of emotion and medical knowledge in sparse medical social texts may be explored to improve system performance. This paper proposed an effective method integrating sufficient emotional expression and medical knowledge to detect ADRs from medical tweets. First, the proposed method utilized sentence-level emotional context and word-level emotional score to learn sufficient emotional information for distinguishing between ADR and non-ADR tweets. Furthermore, a co-occurrence dictionary of each drug and its relevant ADRs was constructed by means of a medical resource (MedDRA) and drug site (https://www.drug.com) to help the proposed model focus on posts containing drugs and ADRs. Finally, a convolutional neural network (CNN) model on the basis of bidirectional encoder representations from transformers (BERT) performed the classification task. The proposed model achieved better overall performance than the other existing methods on two Twitter datasets (F1-scores of 72.64% and 64.98% on PSB2016 and SMM4H, respectively).
Li et al. (Wed,) studied this question.