Introducing the NLP task of negative attitudinal function identification

Nicolás José Fernández-Martínez


On social media, users often express emotions, judgments, and evaluations on various social and private topics, detectable through automated methods. While NLP tasks like emotion detection and dialogue act classification focus on identifying emotions and intentions in texts, little attention has been paid to the attitudinal function of a text, such as expressing dislike, disagreement, pessimism, disapproval, etc. Our main contribution introduces the NLP task of negative attitudinal function identification, going beyond emotion detection and dialogue classification by focusing on users’ intent and the expression of negative emotions, and negative ethical and aesthetic evaluations. We present a basic synthetic dataset for negative attitudinal functions built with foreign language teaching and learning resources. The dataset was used to develop negative attitudinal function models with supervised approaches, which were compared against other baseline models based on social media emotion detection datasets whose emotion categories were mapped to negative attitudinal functions. Our models, though not consistently outperforming baselines due to the qualitative differences of the tasks, use of out-of-domain data, and labeling issues found in the emotion detection datasets, exhibit promising capabilities with unseen data and in multilingual contexts.

Texto completo: