A simple Keras implementation of MatchPyramid model for semantic matching.
Please refer paper：Text Matching as Image Recognition
- Input Data Format
- Train/Valid set:
label |q1 |q2
1 |Q2119 |D18821
0 |Q2119 |D18822
- Test set:
q1 |q2
Q2241 |D19682
Q2241 |D19684
- Preprocessed Corpus:
qid |words
D9980 |47 0 1 2 3 4 5 6 7 8 9 10
D5796 |21 40 41 42 43 44 14 45
- Word Embedding:
word |embedding (50-dimension)
28137 |-0.54645991 2.28509140 ... -0.34052843 -2.01874685
8417 |-9.01635551 -3.80108356 ... 1.86873138 2.14706421
- Word Dictionary:
word |wid
preparing |0
to |1
rebuild |2
- Train the model
$ python match_pyramid.py
- Loss and Accuracy
| Best Valid loss | Best Valid accuracy | |
|---|---|---|
| WikiQA | 0.3973 | 0.8786 |
| QouraQP | 0.4525 | 0.7797 |
- Python 3.5
- TensorFlow 1.8.0
- Keras 2.1.6