A Model for Natural Language Attack on Text Classification and Inference
使用时序评分对单词重要性排序，对关键词做同意替换，选取替换前后相似度较大且影响较大样本作为对抗样本。
本仓库仿照TextFooler，对平台进行了适配。相关环境安装请参考原仓库，或者参考build.sh。
在环境安装好后，请在myTextFooler目录下解压counter-fitted-vectors.txt.zip
先构建TextFooler类的实例：
textfooler = TextFooler(model=predictor, device='gpu0', IsTargeted=False)然后将输入数据与标签输入，来生成对抗样本：
adv_xs = textfooler.generate(texts, labels)具体运行样例参考text.py（注：需要原仓库中的一些代码依赖，具体看import。text.py仅为测试，具体使用不需要原仓库依赖）
本仓库代码仅在与原仓库相同配置下测试过。
perturb_ratio: Whether use random perturbation for ablation study. 默认为0
sim_score_threshold: Required minimum semantic similarity score. 默认为0.7
import_score_threshold: Required mininum importance score. 默认为-1
sim_score_window: Text length or token number to compute the semantic similarity score 默认为15
synonym_num: Number of synonyms to extract 默认为50
batch_size: Batch size to get prediction 默认为32
counter_fitting_embeddings_path: path to the counter-fitting embeddings we used to find synonyms 默认为myTextFooler/counter-fitted-vectors.txt
counter_fitting_cos_sim_path: pre-compute the cosine similarity scores based on the counter-fitting embeddings 默认为空
USE_model_path: Path to the USE model. 默认为空