PPDAI Magic Mirror Data Application Contest
Introduction
This is the repository for PPDAI contest, which is a natural language processing (NLP) model aims to detect duplicate questions in Chinese.
Data
Data was provided by PPDAI, which are pairs of questions labeled with 0 and 1 represents similar or not. The questions are represented by two sequences of integers which are the indices of corresponding embedding vectors (word and character).
Model
We proposed three models including a RNN based model, CNN based model and a RCNN based model. These models have the following characteristics:
- Bi-Directional GRU in RNN based models for semantic learning.
- 1-D Convolution in CNN and RCNN based models for local feature extraction.
- Co-Attention was used to learn the semantic correlations between two sequences.
- Self-Attention was used to enhance the feature representation.
- Word embedding and Character Embedding were used simultaneously.
Performance:
Our ensemble model achieved 0.203930 of loss in the semi-final, at the top 15% in ranking.
Reference
QANet: Combining Local Convolution with Global Self-Attention for Reading Comprehension[ICLR 2018]
Wenhui Wang et al. “Gated Self-Matching Networks for Reading Comprehension and Question Answering”