Towards Identifying Objectivity in Short Informal Text

Short informal texts are increasingly prevalent in modern communication, often containing fragmented grammar, personal opinions, and limited context. Traditional NLP tasks for the texts ordinarily focus on the subjective aspect learning, such as sentiment analysis and polarity classification. The study of learning objectivity from the texts is similarly significant, which can benefit many real-world applications including information filtering, content verification, etc. Unfortunately, this study is not being explored. This paper proposes a novel task that aims at identifying objectivity in short informal texts. Inspired by the characteristics of objective statements that normally need complete syntax structures for knowledge expression and delivery, we try to leverage the viewpoint of subjects (U), the tense of predicates (V), and the viewpoint of objects (O) as critical factors for objectivity learning. Upon that, we further propose a two-stage objectivity identification approach: (1) a UVO quantification module is implemented via a proposed OpenIE and large language model (LLM)-based triple feature quantification procedure; (2) an objectivity identification module employs pre-trained base models like BERT or RoBERTa that are constrained with the quantified UVO. The experimental result demonstrates our approach can outperform the base models up to 5.91% in objective-F1 and up to 6.97% in accuracy.