Deep learning models and datasets for aspect term sentiment classification: Implementing holistic recurrent attention on target-dependent memories

摘要

An essential challenge in aspect term sentiment classification using deep learning is modeling a tailor-made sentence representation towards given aspect terms to enhance the classification performance. To seek a solution to this, we have two main research questions: (1) Which factors are vital for a sentiment classifier? (2) How will these factors interact with dataset characteristics? Regarding the first question, harmonious combination of location attention and content attention may be crucial to alleviate semantic mismatch problem between aspect terms and opinion words. However, location attention does not reflect the fact that critical opinion words usually come left or right of corresponding aspect terms, as implied in the target-dependent method although not well elucidated before. Besides, content attention needs to be sophisticated to combine multiple attention outcomes nonlinearly and consider the entire context to address complicated sentences. We merge all these significant factors for the first time, and design two models differing a little in the implementation of a few factors. Concerning the second question, we suggest a new multifaceted view on the dataset beyond the current tendency to be somewhat indifferent to the dataset in pursuit of a universal best performer. We then observe the interaction between factors of model architecture and dimensions of dataset characteristics. Experimental results show that our models achieve state-of-the-art or comparable performances and that there exist some useful relationships such as superior performance of bi-directional LSTM over one-directional LSTM for sentences containing multiple aspects and vice versa for sentences containing only one aspect.