Page 88 - Undergraduate Research Opportunities Program Proceedings 2017-18

School of Engineering
Department of Computer Science & Engineering

In this paper, we explore methods of determining semantic equivalence between pairs of questions, with a dataset
released by Quora. A deep learning approach that uses a Siamese GRU neural network connected to a distance measure
layer to determine the semantic equivalence is used in this paper. Variations of the network are explored to improve
network performance. Applying an adapted attention mechanism on the GRU network has been found to improve
model performance significantly. The effectiveness of several regularization techniques on the GRU cells are evaluated,
improving the model performance close to the state-of-art level.

Supervisor: YEUNG Dit Yan/ CSE
Student: ZHANG Lin / COSC Course: UROP1100, Summer

The goal of this project is to build an AI TA which can answer questions already having duplicate pairs and corresponding
answers in our database. Therefore, how to detect duplicate questions accurately has been a major problem. In this
report, we first introduce our baseline method, the vanilla GRU model, to detect duplicate questions. We then propose
and analyze several methods to improve the accuracy. They ca be divided into two categories: Improving the model,
including using deeper neural network, LSTM networks, Convolutional neural networks, adding POS tags; Or creating
augmented dataset, especially applying Variational Auto Encoder (VAE) to generating duplicate questions. Finally, we
will discuss the possible future development.

78