Sentence similarity is an important problem in Natural language processing. Its techniques can be applied in a wide variety of tasks such as paraphrase identification, answer selection, and machine comprehension. For this problem, previous works have investigated a number of deep learning techniques, with different architectural configurations such as Convolutional neural network, Recurrent neural network, Denoising autoencoder, Attention mechanism, etc.
In this project, we developed an attentional model for sentence similarity matching with a goal of finding similar questions in community question answering systems. The model includes a bidirectional LSTM to encode contextual information followed by multi-level attention mechanism to search for parts in the source question that are relevant/irrelevant to the target question. Experiments conducted on Quora dataset show that the model achieves competitive results with the state-of-the-art methods on classifying and ranking similar questions.