Aspect-Level Sentiment Classification

Overview | People | Collaborators | Sponsors | Publications | Tools


Traditional document-level sentiment classification tries to identify the general sentiment polarity of a given text as positive, negative, or neutral. Unlike document-level sentiment, aspect-level sentiment classification identifies the sentiment of one specific aspect in its context sentence. For example, given a sentence "great" food but the service was dreadful" the sentiment polarity for aspects "food" and "service" are positive and negative respectively.


In this project, we develop multiple methods to approach this problem. Our first attempt is based on the context sentiment, where we classify the sentiment using the context words of the target aspect. We further adopt deep neural networks for this problem. We proposed parameterized convolutional neural networks (PCNN) to take the aspect information into CNN. In our second model, we use attention-over attention neural networks to automatically focus on the aspect-related parts in the text.


We experiment our two neural methods on the aspect-level sentiemnt classification dataset released in SemEval 2014 competition. Our best model achieves accuracy of 0.812 and 0.745 on the restaurant and laptop datasets respectively.

Aspect Level weights for Sentences
Figure 1: Examples of final attention weights for sentences. The color depth denotes the importance degree of that word towards the aspect sentiment