Sensory Features in Affective Analysis: A Study Based on Neural Network Models

This study proposes an ensemble model to incorporate sensory features of lexical items in English from external resources into neural affective analysis frameworks. This allows the models to take the combined effects of bi-directional feeling between the sensory lexicon and the writer to infer human affective knowledge. We evaluate our model on two affective analysis tasks. The ensemble model exhibits the best accuracy and the results with 1% F1-score improvement over the baseline LSTM model in the sentiment analysis task. The performance shows that perceptual information can contribute to the performance of sentiment classification tasks significantly. This study also provides a support for the linguistic finding that correlations exist between sensory features and sentiments in the language.