Review aspect extraction based on character-enhanced embedding models

User reviews, in the form of short unstructured natural texts, often provide rich information to benefit product adoption or service improvement. Aspect can be extracted as the abstract meaning from the reviews. Traditional methods have employed either rule-based templates or bag-of-words features for aspect extraction from text. However, these models cannot effectively handle short texts, especially in Chinese reviews. In this paper, we address the issue by learning the character embeddings as the basic semantic unit and incorporating the compositional sentence-level representation into a neural network approach for review aspect classification. For that, the character embeddings from the reviews are learned in position-based and clustered-based fashions, and then combined into sentence vectors to yield better text representations. Extensive experiments on real world data set suggest that our proposed model highly outperforms the state-of-the-art methods for review aspect extraction task.