An Effective Approach of Sentence Compression Based on "Re-read" Mechanism and Bayesian Combination Model

As the mass social media texts and the increasingly popular "small screen" interaction mode are producing a huge text compression requirement, this paper presents an approach for English sentences compression based on a "Re-read" Mechanism and Bayesian combination model. Firstly, we build an encoder-decoder consisting of Long Short-Term Memory (LSTM) model. In the encoding stage, the original sentence's semantics is modeled twice. The result from the first encoder, as global information, and the original sentence are input into the second encoder together, obtaining a more comprehensive semantic vector. In the decoding stage, we adopt a simple attention mechanism, focusing on the most relevant semantic information to improve the decoding efficiency. Then, a Bayesian combination model combines the explicit prior information and "Re-read" models to enhance the use of explicit training data features. Experimental results on Google Newswire sentence compression dataset show that the method proposed in this paper can greatly improve the compression accuracy, and the F1 score reaches 0.80.