This repo contains code for Data-to-Text Generation with Content Selection and Planning (Puduppully, R., Dong, L., & Lapata, M.; AAAI 2019); this code is based on an earlier fork of OpenNMT-py. The Pytorch version is 0.3.1.
Update: For a model with better relation generation precision (RG P%) and other metrics, please see the macro planning repository and the corresponding TACL 2021 paper.
@inproceedings{DBLP:conf/aaai/Puduppully0L19,
author = {Ratish Puduppully and
Li Dong and
Mirella Lapata},
title = {Data-to-Text Generation with Content Selection and Planning},
booktitle = {The Thirty-Third {AAAI} Conference on Artificial Intelligence, {AAAI}
2019, The Thirty-First Innovative Applications of Artificial Intelligence
Conference, {IAAI} 2019, The Ninth {AAAI} Symposium on Educational
Advances in Artificial Intelligence, {EAAI} 2019, Honolulu, Hawaii,
USA, January 27 - February 1, 2019},
pages = {6908--6915},
publisher = {{AAAI} Press},
year = {2019},
url = {https://doi.org/10.1609/aaai.v33i01.33016908},
doi = {10.1609/aaai.v33i01.33016908},
timestamp = {Tue, 02 Feb 2021 08:00:48 +0100},
biburl = {https://dblp.org/rec/conf/aaai/Puduppully0L19.bib},
bibsource = {dblp computer science bibliography, https://dblp.org}
}
The test set output for the model can be found here
All dependencies can be installed via:
pip install -r requirements.txtNote that the Pytorch version is 0.3.1 and Python version is 2.7.
The path to Pytorch wheel in requirements.txt is configured with CUDA 8.0. You may change it to the desired CUDA version.
The boxscore-data json files can be downloaded from the boxscore-data repo.
The input dataset for data2text-plan-py can be created by running the script create_dataset.py in scripts folder.
The dataset so obtained is available at link https://drive.google.com/open?id=1R_82ifGiybHKuXnVnC8JhBTW8BAkdwek
Assuming the OpenNMT-py input files reside at ~/boxscore-data, the following command will preprocess the data
BASE=~/boxscore-data
IDENTIFIER=cc
mkdir $BASE/preprocess
python preprocess.py -train_src1 $BASE/rotowire/src_train.txt -train_tgt1 $BASE/rotowire/train_content_plan.txt -train_src2 $BASE/rotowire/inter/train_content_plan.txt -train_tgt2 $BASE/rotowire/tgt_train.txt -valid_src1 $BASE/rotowire/src_valid.txt -valid_tgt1 $BASE/rotowire/valid_content_plan.txt -valid_src2 $BASE/rotowire/inter/valid_content_plan.txt -valid_tgt2 $BASE/rotowire/tgt_valid.txt -save_data $BASE/preprocess/roto -src_seq_length 1000 -tgt_seq_length 1000 -dynamic_dict -train_ptr $BASE/rotowire/train-roto-ptrs.txt
The train-roto-ptrs.txt file is available along with the dataset and can also be created by the following command
python data_utils.py -mode ptrs -input_path $BASE/rotowire/train.json -train_content_plan $BASE/rotowire/inter/train_content_plan.txt -output_fi $BASE/rotowire/train-roto-ptrs.txt
The command for training the Neural Content Planning model with conditional copy NCP+CC is as follows:
BASE=~/boxscore-data
IDENTIFIER=cc
python train.py -data $BASE/preprocess/roto -save_model $BASE/gen_model/$IDENTIFIER/roto -encoder_type1 mean -decoder_type1 pointer -enc_layers1 1 -dec_layers1 1 -encoder_type2 brnn -decoder_type2 rnn -enc_layers2 2 -dec_layers2 2 -batch_size 5 -feat_merge mlp -feat_vec_size 600 -word_vec_size 600 -rnn_size 600 -seed 1234 -start_checkpoint_at 4 -epochs 25 -optim adagrad -learning_rate 0.15 -adagrad_accumulator_init 0.1 -report_every 100 -copy_attn -truncated_decoder 100 -gpuid $GPUID -attn_hidden 64 -reuse_copy_attn -start_decay_at 4 -learning_rate_decay 0.97 -valid_batch_size 5
The NCP+CC model can be downloaded from https://www.dropbox.com/sh/vo5wb2fuq7m0bk0/AABikW0KomOKIor24wD8VSFWa?dl=0
During inference, we first generate the content plan
MODEL_PATH=<path to model1>
python translate.py -model $MODEL_PATH -src1 $BASE/rotowire/inf_src_valid.txt -output $BASE/gen/roto_stage1_$IDENTIFIER-beam5_gens.txt -batch_size 10 -max_length 80 -gpu $GPUID -min_length 35 -stage1
This script generates the content plan with records from input of content plan with indices
python scripts/create_content_plan_from_index.py $BASE/rotowire/inf_src_valid.txt $BASE/gen/roto_stage1_$IDENTIFIER-beam5_gens.txt $BASE/transform_gen/roto_stage1_$IDENTIFIER-beam5_gens.h5-tuples.txt $BASE/gen/roto_stage1_inter_$IDENTIFIER-beam5_gens.txt
The accuracy of content plan in first stage can be evaluated using the following command
python non_rg_metrics.py $BASE/transform_gen/roto-gold-val-beam5_gens.h5-tuples.txt $BASE/transform_gen/roto_stage1_$IDENTIFIER-beam5_gens.h5-tuples.txt
The output summary is generated using the command
MODEL_PATH2=<path to model2>
python translate.py -model $MODEL_PATH -model2 $MODEL_PATH2 -src1 $BASE/rotowire/inf_src_valid.txt -tgt1 $BASE/gen/roto_stage1_$IDENTIFIER-beam5_gens.txt -src2 $BASE/gen/roto_stage1_inter_$IDENTIFIER-beam5_gens.txt -output $BASE/gen/roto_stage2_$IDENTIFIER-beam5_gens.txt -batch_size 10 -max_length 850 -min_length 150 -gpu $GPUID
Metrics of RG, CS, CO are computed using the below commands.
python data_utils.py -mode prep_gen_data -gen_fi $BASE/gen/roto_stage2_$IDENTIFIER-beam5_gens.txt -dict_pfx "roto-ie" -output_fi $BASE/transform_gen/roto_stage2_$IDENTIFIER-beam5_gens.h5 -input_path "/boxcore-json/rotowire"
th extractor.lua -gpuid $GPUID -datafile roto-ie.h5 -preddata $BASE/transform_gen/roto_stage2_$IDENTIFIER-beam5_gens.h5 -dict_pfx "roto-ie" -just_eval
python non_rg_metrics.py $BASE/transform_gen/roto-gold-val-beam5_gens.h5-tuples.txt $BASE/transform_gen/roto_stage2_$IDENTIFIER-beam5_gens.h5-tuples.txt
The BLEU perl script can be obtained from https://github.com/moses-smt/mosesdecoder/blob/master/scripts/generic/multi-bleu.perl Command to compute BLEU score:
~/multi-bleu.perl $BASE/rotowire/inf_tgt_valid.txt < $BASE/gen/roto_stage2_$IDENTIFIER-beam5_gens.txt
For training the IE models, follow the updated code in https://github.com/ratishsp/data2text-1 which contains bug fixes for number handling. The repo contains the downloadable links for IE models too.