Tensorflow implementation of Facebook #TagSpace
You can read more about #TagSpace from here
Special thanks to Facebook research team's Starspace project, it was really good reference.
Beside choosing 1000 random negative tag (for performance reason I guess), I choosed worst positive tag, best negative tag. It's not good for performance but since we don't have much tags(labels) as Facebook, it seems okay.
Download ag news dataset as below
$ tree ./data ./data └── ag_news_csv ├── classes.txt ├── readme.txt ├── test.csv ├── train.csv └── train_mini.csv
$ python train.py
Accuracy 0.89 (ag test data, compare 0.91 from StarSpace with same condition [5 epoch, 10 dim])
- Clean up messy code
- Better class structure
- improve Tokenizer
- support multiple dataset
- improve performance
- adopt WARP sampling
- add Tensorboard metrics
- add Korean