Text Characteristics Vector: Rethinking Human-Centered Sentiment Analysis with Emotion-Related Text Characteristics

To evaluate how emotional expression is related to valence, we built a sentimental classifier based only on quantitative information of the count of eight text characteristics on a dataset of fanfiction reviews. We also integrated the adjusted character count into the counting. To evaluate our text characteristics vector (TCV) model, we compared it with a model built based on the same data through a TF-IDF vectorizer. Our TCV model performed equally to the TF-IDF model with both F1 scores around 0.90. Our research is critical since it shows how emotional expression can be deterministic on the valence of text. Our result reveals the importance of a human-centered approach to NLP. The result is generalizable to all reviews on any social media. We release data and code in our research for reproducing purposes.