文件列表:

data.py (3658, 2018-05-18)
dataset (0, 2018-05-18)
dataset\get_data.bash (2151, 2018-05-18)
dataset\tokenizer.sed (2192, 2018-05-18)
encoder (0, 2018-05-18)
encoder\demo.ipynb (47349, 2018-05-18)
encoder\evaluate_model.py (1988, 2018-05-18)
encoder\models.py (11118, 2018-05-18)
encoder\samples.txt (580976, 2018-05-18)
encoder\xutils.py (175, 2018-05-18)
models.py (37344, 2018-05-18)
mutils.py (2517, 2018-05-18)
train_nli.py (13588, 2018-05-18)

# BiLSTM with BF-max pooling for Universal sentence representations *BiLSTM with BF-max pooling* is a *sentence embeddings* method that provides semantic sentence representations. It is trained on natural language inference data and generalizes well to many different tasks. *BiLSTM with BF-max pooling* is inspired by [InferSent](https://arxiv.org/abs/1705.023***),we use an imporved pooling mechanisim to boost the BiLSTM with max pooling. ## Dependencies This code is written in python. The dependencies are: * Python 2.7 (with recent versions of [NumPy](http://www.numpy.org/)/[SciPy](http://www.scipy.org/)) * [Pytorch](http://pytorch.org/) >= 0.12 * NLTK >= 3 ## Download datasets To get GloVe, SNLI and MultiNLI [2GB, 90MB, 216MB], run (in dataset/): ```bash ./get_data.bash ``` This will download GloVe and preprocess SNLI/MultiNLI datasets. For MacOS, you may have to use *p7zip* instead of *unzip*. ## Train model on Natural Language Inference (SNLI) To reproduce our results and train our models on [SNLI](https://nlp.stanford.edu/projects/snli/), set **GLOVE_PATH** in *train_nli.py*, then run: ```bash python train_nli.py --bf=1 use bf-balanced(1) or --bf=2 bf-combined(2) ``` You should obtain a dev accuracy of 85.1 and a test accuracy of 84.9 with the default setting. ## Reproduce our results on transfer tasks To reproduce our results on transfer tasks, clone [SentEval](https://github.com/facebookresearch/SentEval) and set **PATH_SENTEVAL**, **PATH_TRANSFER_TASKS** in *evaluate_model.py*, then run: ```bash python evaluate_model.py ``` Note that while BiLSTM with BF-max pooling provides good features for many different tasks, our approach also obtains strong results on STS tasks which evaluate the quality of the cosine metrics in the embedding space.

近期下载者：

相关文件：

评论：[我要评论] [举报此文件]

收藏者：