This is the code we used to establish a baseline for the Chinese Natural Language Inference (CNLI) corpus.
The CNLI dataset can be downloaded at here
Both the train and dev set are tab-separated format. Each line in the train (or dev) file corresponds to an instance, and it is arranged as:
sentence-id premise hypothesis label
This repository includes the baseline model for Chinese Natural Language Inference (CNLI) dataset. We choose the Decomposable Attention Model as our baseline model. More details about the model can be found in the original paper.
- python 3.5
- tensorflow '1.4.0'
- jieba 0.39
Data Preprocessing
We use jieba to tokenize the sentences. During trainging, we use the pre-trained SGNS embedding introduced in Analogical Reasoning on Chinese Morphological and Semantic Relations. You can download the sgns.merge.word from here.
Main Scripts
config.py:the parameter configuration.
decomposable_att.py: implementation of the Decomposable Attention Model.
data_reader.py: preparing data for the model.
train.py: training the Decomposable Attention Model.
Running Model
You can train the model by the following command line:
python3 train.py
We adopt early stopping on dev set. The best results are shown in the following table:
data | accuracy(%) |
---|---|
train | 64.88 |
dev | 58.70 |
Please let us know, if you encounter any problems.