Opinion Target Extraction Based On Deep Learning Approach

Opinion target extraction, which aims to identify fine-grained opinion targets from opinion texts, is a subtask of fine-grained opinion mining. Existing work mainly focuses on unsupervised and shallow architecture approaches, but less work has been done based on deep architecture. Recent work has shown that unsupervised and shallow architecture approaches perform well. However, they not only rely on human crafted rules or opinion word lexicon, but also need heavy manual work and have poor generalization ability.Deep learning focuses on the process of automatically extracting hierarchical features, and is independent with human crafted features. Therefore, this thesis employs a model based on stacked auto-encoders to extract opinion targets. This model uses only word embeddings as input without human crafted extraction rules and opinion word lexicon. The main contents of this thesis are as follows.(1) Model construction. The employed model uses word embeddings as input, auto-encoder as hidden layer and softmax regression as output layer. The model training process consists of two phases, unsupervised pre-training and supervised parameter tuning.(2) Model improvement. A greedy algorithm is used to extend the stacked auto-encoders model. The original stacked auto-encoders model is a classification model. However, the opinion target extraction task is regarded as a sequential labeling task in this thesis. In order to adapt the original classification model to a sequential labeling model, a greedy algorithm is used to extend the original model.(3) Model selection. A set of experiments are designed to compare the performance of the model with different parameters (model depth, text window size, etc.) and methods that are used to prevent over-fitting. The methods used to prevent over-fitting include L2, sparse auto-encoder, denoising auto-encoder and dropout.Experimental results show that the proposed method is effective. We also compare our model with the state-of-the-art models in SemEval-2014. Two datasets (Laptop and Restaurant) are used in this thesis. For the Laptop dataset, our model (F1 is 72.42%) outperforms the best constrained model (F1 is 70.40%). For the Restaurant dataset, our model (F1 is 82.43%) is as good as the best constrained model (F1 is 83.98%).