Semantic Paraphrase Detection

By Pranav Goyal and Saransh Rajput

Task

Given two sentences, the task is to classify if the the sentences are paraphrase of each other. This is a binary classification. The task is a generalization of the semantic textual similarity which would be about finding the level of similarity among two sentences.

Datasets

• Quora Question Pairs (QQP) - 404300 sentence pairs
• Microsoft Paraphrase COrpus (MSRP) - 5800 sentence pairs

Models Implemented

Saved Model Checkpoints

1. Bidirectional Multi-Perspective Matching (BiMPM)

Details:

• Uses pretrained Glove embeddings (6B)
• Bidirectional LSTM to encode the sentences
• Three Distinct multi-perspective matching mechanisms to measure the relations across sentences
• Final LSTM to generate fixed-size vector and a feed-forward network to predict the output

2. Multiway Attention

Details:

• GRU for contextual representatoins.
• Four attention mechanism to get matching vectors representation from the other sentence.
• Further aggregated using GRU and attention applied based on inputs to make the final predictions.

3. Finetune BERT-Base

Details:

• Fine-tuned Bert Base Model (12 layers of stacked transformers with 768 hidden dimention) with a binary classification dense layer on top.
• Original Pretrained Model: bert_en_uncased_L-12_H-768_A-12_3

Results

No	Models	Validation Accuracy on QQP	Validation Accuracy on MSRP
1	BiMPM	85.28 %	72.52 %
2	Multiway Attention	83.47 %	72.56 %
3	Bert Finetuning	90.10 %	84.12 %

Report

Project_Report.pdf

Additional Links

Project Presentation

Repository Link

References

• “Fine-Tuning a BERT Model : TensorFlow Core.” TensorFlow, www.tensorflow.org/official_models/fine_tuning_bert.
• Devlin, Jacob, et al. "Bert: Pre-training of deep bidirectional transformers for language understanding." arXiv preprint arXiv:1810.04805 (2018).
• Wang, Zhiguo, Wael Hamza, and Radu Florian. "Bilateral multi-perspective matching for natural language sentences." arXiv preprint arXiv:1702.03814 (2017).