Classification of Constructiveness in Social Media Comments Using Large Language Models and Machine Learning
In today’s digital age, online platforms are a primary source of knowledge exchange and social interactions. However, the quality of interactions can vary widely and not all contributions are constructive and conducive to a productive discourse. Constructive feedback is crucial as it fosters learning, promotes understanding, and encourages a respectful exchange of ideas. Prior work has focused on classifying constructiveness of user comments on news websites, however is not transferable to social media websites since news websites are heavily moderated. In response, in this paper, we developed a model trained on reddit comments capable of effectively discerning the quality and constructiveness of a user's comments in an unfiltered environment. This model can be deployed on social media websites as a black box feature to filter out non-constructive comments that lack substance to encourage more insightful discussions and promote more meaningful exchanges. Our novel approach leverages Large Language Models to auto annotate constructiveness and then build a task-specific model that learns to detect constructiveness of comments based on a taxonomy of carefully engineered features. We were able to achieve X% accuracy with the X classifier.