This repository is a keras implementation about the ESIM model proposed in paper Enhanced LSTM for Natural Language Inference. The ESIM model is a strong baseline used for text matching tasks, such as Natural Language Inference(NLI) and paraphrase detection.
- keras, tested on 2.3.1, with tensorflow 1.14.0 as its backend.
- python 3.5+
- NLTK
First, you should download the Quora Question Pairs dataset via http://qim.fs.quoracdn.net/quora_duplicate_questions.tsv. We also need GloVe, which is a pretrained word embeddings, you can download it here. Put the glove.840B.300d.txt file to ./.
Atfer that, you can train, test, and save the model via:
python run.py
The model will be saved to the ./saved_models.
Except the GloVe word embeddings, I also use character embeddings obtained from bi-lstm, you can remove it anyway if you do not need it.
| Model | Dataset | test loss | Accuracy |
|---|---|---|---|
| ESIM-Glove | QQP | 0.3737 | 0.8812 |