george0828zhang/nlp

NLP

hw1

Required packages

• tensorflow-gpu 1.12
• jieba
• tqdm
• jupyter notebook
• tensorflow-gpu 1.13.1 (model_bert_basic)
• bert-as-service

Preprocessing

• Use Preprocessing.ipynb to tokenize sentences and turn them into sequences of int values.
• Use WordEmbedding.ipynb to pretrain word vectors.
• Use model_bert_basic/preprocessing.py to generate datasets.

Training

• Use RNN_basic.ipynb to train a simple model w/o attention mechanism.
• Use RNN_complex.ipynb to train a attentional model.
• Use model_bert_basic/predict.py to train a model based on BERT.

Experiment

• Add Convolution Layer
  • Use Var1_Simple_CNN_network.ipynb to train a CNN model w/o attention
  • Use Var2_Convolution_before_RNN.ipynb to train a RNN_basic model w/ additional Convolution Layer
  • Use Var3_Convolution_Layer_after_Attentional_bi-LSTM_network.ipynb to train a RNN_complex mode w/ additional Convolution Layer

Data & Submission

https://www.kaggle.com/c/fake-news-pair-classification-challenge/

hw2

Required packages

• pytorch 1.0.1
• wordninja (pip install wordninja)
• nltk
• sklearn (for metrics)
• tqdm
• jupyter notebook

How to execute

• To train biLSTM+POS models, open biLSTM-POS/Trainer.ipynb, set the datapaths, then run all cells.