SEG

This repository contains code for the NLPCC 2018 paper From Plots to Endings: A Reinforced Pointer Generator for Story Ending Generation.

Description

Our code is based on the pointer-generator for the ACL 2017 paper Get To The Point: Summarization with Pointer-Generator Networks. We build three modes in the code: seq2seq_train, rl_train, and beam_search_decode.

Requirements

python 3.5
tensorflow 1.4.0
numpy 1.14.4
nlgeval

We use the evaluation package nlg-eval for automatic evaluation. For more detailed information, see the following paper: Relevance of Unsupervised Metrics in Task-Oriented Dialogue for Evaluating Natural Language Generation.

How to run

Pretrain Seq2seq/Pointer-generator+coverage Model

A directory will be created according to the super-parameter exp_name. Two subdirectory train_seq2seq and eval_seq2seq of exp_name will be produced to save the checkpoints during training and validation.

run seq2seq model

python3 run.py --mode seq2seq_train --data_path ./data/ROC_data --exp_name model_seq2seq --pointer_gen False --word_vocab_size None --coverage False --convert_to_coverage_model False >logs/seq2seq

run pointer-generator+coverage model

train pointer-generator for 10 epochs

python3 run.py --mode seq2seq_train --data_path ./data/ROC_data --exp_name model_pgen10e --epochs 10 --pointer_gen True --word_vocab_size 15000 --coverage False --convert_to_coverage_model False >logs/pgen10e

convert pointer-generator to a model for running coverage

python3 run.py --mode seq2seq_train --data_path ./data/ROC_data --exp_name model_pgen10e --pointer_gen True --word_vocab_size 15000 --coverage True --convert_to_coverage_model True

continue to train pointer-generator model with coverage

python3 run.py --exp_name model_pgen10e_10cov --mode seq2seq_train --epochs 10 --pointer_gen True --word_vocab_size 15000 --coverage True --convert_to_coverage_model False >logs/pgen10e_10cov

Restore Pretrained Model and Run RL Training

You need to give the exp_name for restoring the pretained model, e.g. model_seq2seq/model_pgen10e_10cov

python3 run.py --mode rl_train  --reward_type bleu_4 --data_path ./data/ROC_data --exp_name=model_seq2seq >logs/seq2seq_rl

python3 run.py --mode=rl_train  --reward_type bleu_4 --data_path ./data/ROC_data --exp_name model_pgen10e_10cov >logs/pgen_rl

Run Beam Search Decoding

python3 run.py --mode beam_search_decode --data_path ./data/ROC_data --exp_name model_seq2seq --decode_ckpt_dir eval_seq2seq

Run Automatic Evaluation Script

cd evaluation
python2 evaluation_py2.py --data_path ./data/ROC_data --exp_name model_seq2seq --result_dir eval_seq2seq12600

blcunlp/SEG