pyTorch implementation of story cloze task on ROCStories Corpora

python==3.x

torch==0.4

nltk (with 'punkt' installed using nltk.download('punkt'))

pickle

tensorflow (optional, for tensorboard logging)

训练数据、验证数据与测试数据均在./data目录下

首先执行data.py预处理数据，生成vocab.freq.pkl、vocab.pkl、vocab_embed.pkl、train.pkl、test.pkl和val.pkl

使用InferSent的准备：

​ 下载fastText词向量并解压至./embedding/InferSent/dataset/fastText目录下

mkdir embedding/InferSent/dataset/fastText
curl -Lo embedding/InferSent/dataset/fastText/crawl-300d-2M.vec.zip https://s3-us-west-1.amazonaws.com/fasttext-vectors/crawl-300d-2M.vec.zip
unzip embedding/InferSent/dataset/fastText/crawl-300d-2M.vec.zip -d embedding/InferSent/dataset/fastText/

​ 下载预训练的InferSent（V2）模型至./embedding/InferSent/encoder目录下

curl -Lo embedding/InferSent/encoder/infersent2.pkl https://s3.amazonaws.com/senteval/infersent/infersent2.pkl

然后就可以直接执行train.py进行模型的训练

或者使用test.py加载预训练的模型进行故事结尾的预测

4500000次iteration的一个模型可以从https://pan.baidu.com/s/1o8XfXX9HGHnt88G81aTiyw下载

放入./checkpoint目录下即可