Paraphrase Detection using Dependency Tree Recursive Autoencoder

摘要

An architecture is proposed based on recursive autoencoder for paraphrase detection. The proposed architecture embeds the semantic information by using word representations generated from the neural network language model and the syntactic information by implementing the dependency tree over the recursive autoencoder, where dependency tree reveals the syntactic information of the given sentence in recursive form. The proposed architecture is tested on the MSRP dataset for paraphrase detection and the results are above the baseline. The proposed system reached a moderate accuracy and F1 score for the paraphrase detection test on MSRP dataset.