Paraphrase Detection with BiMPM

Description

This repository contains the PyTorch implementation of the Bilateral Multi-perspective Matching model BiMPM described in the paper by Wang et al. The model is used to perform a paraphrase detection task on the Quora Questions Pairs dataset. In order to maintain consistency in comparison, we adopted the train/dev/test partition by Wang et al. The program takes two phrases as inputs and predicts a value to indicate if the two phrases are paraphrases of each other or not.

Requirements

• python 3.5
• torch 0.1.12

Train

To train the model using the setting described in the paper, run

python trainer.py --embedding wordvec.txt --data quora_data/ --word-len 15 --seq-len 50 --perspectives 5 --batch-size 32 --cuda 

Test

To test a model on the test dataset, run

python test.py --embedding wordvec.txt --data quora_data/test.tsv --word-len 15 --seq-len 50 --perspectives 5 --batch-size 32 --model model.pth

Issues

Please report any issues to me juiting.hsu@nyu.edu.

Reference

• Zhiguo Wang, Wael Hamza, Radu Florian. Bilateral Multi-Perspective Matching for Natural Language Sentences. IJCAI (2017)
• Quora Question Pairs Dataset