Multi-domain sentiment analysis with mimicked and polarized word embeddings for human-robot interaction

This paper presents a state-of-the-art approach for sentiment polarity classification. Our approach relies on an ensemble of Bidirectional Long Short-Term Memory networks equipped with a neural attention mechanism. The system makes use of pre-trained word embeddings, and is capable of predicting new vectors for out-of-vocabulary words, by learning distributional representations based on word spellings. Also, during the training process the recurrent neural network is used to perform a fine-tuning of the original word embeddings, taking into account information about sentiment polarity. This step can be particularly helpful for sentiment analysis, as word embeddings are usually built based on context information, while words with opposite sentiment polarity often occur in similar contexts. The system described in this paper is an improved version of an approach that competed in a recent challenge on semantic sentiment analysis. We evaluate the performance of the system on the same multi-domain test set used by the organizers of the challenge, showing that our approach allows reaching better results with respect to the previous top-scoring system. Last but not least, we embedded the proposed sentiment polarity approach on top of a humanoid robot to lively identify the sentiment of the speaking user