Semisupervised neural biomedical sense disambiguation approach for aspect-based sentiment analysis on social networks

Patient narratives on social networks contain large amounts of objective information, such as the descriptions of examinations and interventions. Sentiment analysis (SA) models are mostly used to evaluate the conveyed sentiments by patients in these narratives to assess positive or negative clinical outcomes or to judge the impact of a drug or a medical condition. To date, many state-of-the-art SA models often result in false assessment coverage due to the natural medical entities recognition deficiency and ambiguity problem. In this work, we propose a semisupervised-based neural sense disambiguation approach that helps to substantially define ambiguities, their levels, and the relational mappings between biomedical targets and dependencies for accurate aspect-based sentiment prediction. Three main modules are proposed: (1) generate a sentiment value based on extracted concepts and their synsets, (2) encode the representations of the contextual senses and sentiment inputs, and (3) estimate an aspect-based sentiment weight based on the context-dependency sentiment units vs. the biomedical sense. Both intrinsic and extrinsic evaluation proved how the proposed method have succeeded in pruning contextual sense feature generation and showed a strong agreement for biomedical data property parameterization and ambiguity type extraction. Thus, the model offers a significant rate of discrimination of biomedical natural concept senses by critically analysing constraints from conjunctions of positive or negative contextual semantics. A total of 21% of the vocabulary is drug names, 11% is a multiword drug reaction expressions, 7% is disease symptoms, and 5% is disease-related concepts such as symptoms and related therapy terms. Furthermore, the experiments on a multisource data from Twitter and health-related forums have overshadowed sentiment assessment and achieved an accuracy of 0.91 regarding concepts-based biomedical aspects. These results provide fresh insights into how to investigate biomedical knowledge, e.g., Medical Subject Headings (MeSH) and PubMed, to clarify the correspondence of various biomedical descriptive entities, definitions, and data properties from shared medication-related content.