When AI reads between the lines: Detecting hidden insincerity in text

In the current digital age, technology has become an integral part of every facet of life, often shaping human perceptions through the content shared on social media platforms. However, this digital transformation has also led to the emergence of significant challenges, particularly the proliferation of irrelevant and harmful content, such as "toxic" material. This paper addresses a key issue faced by many online platforms, particularly in the context of Quora, which is the prevalence of short, insincere questions that lack meaningful content. These insincere questions are typically characterized by exaggerated falsehoods, argumentative tones, and unethical language, contributing to a negative user experience. This study focuses on tackling these challenges, specifically within the framework of a Kaggle competition, by utilizing innovative techniques to detect and mitigate problematic content. We propose a solution that leverages a pre-trained deep bidirectional model, specifically fine-tuned with BERT-base, to classify textual questions as either sincere or insincere. The model is built upon a series of carefully engineered features, applied after basic data preprocessing. Our proposed approach enables the automated identification of toxic content, allowing for the removal of insincere questions and thus enhancing the overall quality of information on the platform. The effectiveness of our model is demonstrated through its superior performance compared to existing classification methods, achieving an impressive F1 score of 0.721. This result highlights the potential of our approach in addressing the challenges posed by insincere and toxic content in online discussions.