Leveraging Recursive Processing for Neural-Symbolic Affect-Target Associations

摘要: Explaining the outcome of deep learning decisions based on affect is challenging but necessary if we expect social companion robots to interact with users an emotional level. In this paper, present a commonsense approach that utilizes interpretable hybrid neural-symbolic system associate extracted targets, noun chunks determined be associated expressed emotion, affective labels from natural language expression. We leverage pre-trained neural network well adapted tree and sub-tree processing, Dependency Tree-LSTM, learn dynamic through symbolic rules, in language. find making use unique properties recursive provides higher accuracy interpretability when compared other unstructured sequential methods for determining target-affect associations aspect-based sentiment analysis task.