Identification and analysis of health states in twitter messages

• Autor(a) principal: Morais, Edgar Guilherme Silva
• Data de Publicação: 2022
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10773/36519
• topic: Social media; Health information; Natural language processing; Machine learning

Social media has become very widely used all over the world for its ability to connect people from different countries and create global communities. One of the most prominent social media platforms is Twitter. Twitter is a platform where users can share text segments with a maximum length of 280 characters. Due to the nature of the platform, it generates very large amounts of text data about its users’ lives. This data can be used to extract health information about a segment of the population for the purpose of public health surveillance. Social Media Mining for Health Shared Task is a challenge that encompasses many Natural Language Processing tasks related to the use of social media data for health research purposes. This dissertation describes the approach I used in my participation in the Social Media Mining for Health Shared Task. I participated in task 1 of the Shared Task. This task was divided into three subtasks. Subtask 1a consisted of the classification of Tweets regarding the presence of Adverse Drug Events. Subtask 1b was a Named Entity Recognition task that aimed at detecting Adverse Drug Effect spans in tweets. Subtask 1c was a normalization task that sought to match an Adverse Drug Event mention to a Medical Dictionary for Regulatory Activities preferred term ID. Toward discovering the best approach for each of the subtasks I made many experiments with different models and techniques to distinguish the ones that were more suited for each subtask. To solve these subtasks, I used transformer-based models as well as other techniques that aim at solving the challenges present in each of the subtasks. The best-performing approach for subtask 1a was a BERTweet large model trained with an augmented training set. As for subtask 1b, the best results were obtained through a RoBERTa large model with oversampled training data. Regarding subtask 1c, I used a RoBERTa base model trained with data from an additional dataset beyond the one made available by the shared task organizers. The systems used for subtasks 1a and 1b both achieved state-of-the-art performance, however, the approach for the third subtask was not able to achieve favorable results. The system used in subtask 1a achieved an F1 score of 0.698, the one used in subtask 1b achieved a relaxed F1 score of 0.661, and the one used in the final subtask achieved a relaxed F1 score of 0.116.