Sentiment analysis is a well researched natural language processing field. It is a challenging machine learning task due to the recursive nature of sentences, different length of documents and sarcasm. Traditional approaches to sentiment analysis use count or frequency of words in the text which are assigned sentiment value by some expert. These approaches disregard the order of words and the complex meanings they can convey. Gated Recurrent Units are recent form of recurrent neural network which have the ability to store information of long term dependencies in sequential data. In this work we showed that GRU are suitable for processing long textual data and applied it to the task of sentiment analysis. We showed its effectiveness by comparing with tf-idf and word2vec models. We also showed that GRUs are faster in convergence than LSTM, another gating network. We applied a number of modifications to the standard GRU to make it train faster and yet less prone to over training. We found the better performimg hyperparameters of the GRU-net through extensive cross-validation testing. Finally we ensembled the best performing GRU models for even better performance.
展开▼
