With the increasing computing power and exploding volumes of data in the past few years, researchers have made continuous breakthroughs in deep learning technology including its application to natural language processing (NLP). However, deep learning models for Chinese language processing are still underdeveloped. Inspired by the potential of Chinese NLP with deep learning, this project aims at exploring and developing different deep learning models for sentiment analysis of Chinese text, classifying short Chinese sentences according to their underlying sentiments. Cooperating with Microsoft, this project has designed, implemented, and evaluated two deep learning models, namely LSTM model and CNN model, which leverage the neural networks to capture the high-level semantics of Chinese text. In order to improve the accuracy of the models, this project has conducted a series of experiments to optimize the dataset configuration and model’s hyperparameters. As an industry-based project, this project adopts Microsoft Cognitive Toolkit as the deep learning frameworks, and the experiments were powered by Microsoft Azure, the Microsoft could computing platform.

This project collected two datasets from Douban website. One dataset consists of movie reviews and another one is composed of literature book reviews. We performed a series of data preprocessing procedures, including simplification of Chinese, word segmentation, dataset splitting, filtering, and replacing. Different experiments have been conducted to compare the impacts of different methods of data preprocessing. Classification and regression have been applied to the tasks to investigate the performance of the models. Grid researches have been conducted to find the best combination of the hyperparameters for the LSTM model and the CNN model.

This project adopts CNTK as the deep learning frameworks, and Microsoft Azure provides the computing power. Our models are implemented using Python. Some third-party pachages have been introduced to help process the data and leverage our experiments.