Character Embedding + ESIM + Focal Loss for Chinese Answer Sentence Selection

This is a course project for Web Data Mining. The task is to decide whether a sentence contains the answer to the questions. We use ESIM (Enhanced LSTM for Natural Language Inference) as our main model. Pretrained Chinese character embedding is adopted to faciliate character-level matching between questions and answers. We employ focal loss to address the unbalanced label. A PowerPoint slide is attached in which we further explain our method.

Requirement

• Python (>= 3.6)
• PyTorch (>= 1.0)
• torchtext

Dataset

The dataset for this project is NLPCC DBQA 2016.

Result

	MAP	MRR
All-0	25.30	25.81
BERT	93.73	93.83
Ours	90.33	90.48

Reference

• Enhanced LSTM for Natural Language Inference
• 神经网络不收敛的11个常见问题
• 自用Pytorch笔记(十五:防止过拟合的几种方法)
• 如何处理不平衡数据集的分类任务
• https://github.com/pengshuang/Text-Similarity/blob/master/models/ESIM.py
• https://github.com/marvis/pytorch-yolo2/blob/master/FocalLoss.py