Sweeping the social minefield : detecting online opinions and emotions with NLP
Jens Lemmens · 2024
This thesis deals with opinion and emotion mining in social media messages, and addresses three key issues in these tasks.Part I of the thesis tackles the issue of domainspecific language.We focus on the domain of COVID-19 vaccines, and present a social media monitor named "Vaccinpraat".The monitor estimates the public stance towards COVID-19 vaccines in Dutch-language Facebook and Twitter messages by predicting which messages are vaccine-hesitant.Additionally, Vaccinpraat detects which fine-grained arguments are used for vaccine hesitancy.Combined with named entity recognition (based on existing tools) and hashtag extraction, Vaccinpraat provides a comprehensive overview of topical vaccine opinions.To improve predictions regarding stance and arguments in Vaccinpraat, we developed the first publicly available Dutch language model specialized in COVID-19 related social media content: CoNTACT.This RobBERT-based model is domain-adapted using 2.8 million Dutch tweets about COVID-19, and shows substantial improvements over the baseline model.After the development of CoNTACT, it was used together with the Vaccinpraat pipeline for a large-scale social media study on vaccine hesitancy in pregnant and lactating women.The results of this study suggest that approximately 40% of the studied messages were hesitant towards vaccines, with the main concern being the medical safety of both the mother and child.