Note: this code is no longer actively maintained.

introduction

• attention based summarization on tensorflow using seq2seq model
• my graduation project code
• do not provide data for the time

environment

• ubuntu 16.04 lts
• anaconda python 3.6
• recompiled tensorflow r1.7 gpu version
• CUDA 9.0
• cudnn 7.1.2
• rouge

run

• This work use Gigaword dataset which is not for public. You need fetch the data yourself.
• The SentiWordNet 3.0 dataset can be found here :SentiWordNet3.0
• The codes are written in an early version of tensorflow. I do not recommend run this code directly. Just for reference.
• run python main.py -help for help.
• run python main.py -w2v to train the wordvector from Gigaword dataset using Word2Vec，then run python main.py -train to train the model and python main.py -testto test the model(just get the output of testset).
• you need install ROUGE to test the output. All the results are collected in the original PERL version of ROUGE. Using PyRouge make cause the result a little bit higher.

progress

• finish word embedding matrix
• build seq2seq model
• test lstm and gru core
• test bidirectional core
• fix infer problem
• test multilayer with dropout core
• fix lazy loading
• fix pre-processing
• try training with non-mentor model
• secondary activation
• test attention decoder(luong attention)
• choose last batch in each epoch as the validation set
• learning rate decay:gradient descent,low init value,decay=0.995
• cut vocab size to 3000,replace unusual word to unk
• enlarge rnn hidden units size
• fix word embedding matrix and try to load model
• divide infer and train into two graphs
• use rouge to value model
• save each test result
• fix unk problems
• train sentiment classification svm
• add sentiment-blended word embeddings
• test sentiment classify
• use larger corpus
• collect ROUGE

current effect

• ROUGE files collected in the './ROUGE_ANSWER'