KPCNet

Code for the WWW 2021 paper: Diverse and Specific Clarification Question Generation with Keywords

A Chinese blog with my reviews of clarification question generation, for readers interested in this topic.

阐明性问题生成(clarification question generation)概览

The system requirements are:

Linux OS
Python 3

To run the code, first install the required python packages.

pip install -r requirements.txt

Then download corresponding resources link to this directory and unzip it.

wget https://zenodo.org/record/4587504/files/KPCNet_resources.zip.001
wget https://zenodo.org/record/4587504/files/KPCNet_resources.zip.002
wget https://zenodo.org/record/4587504/files/KPCNet_resources.zip.003
wget https://zenodo.org/record/4587504/files/KPCNet_resources.zip.004

cat KPCNet_resources.zip.00* > KPCNet_resources.zip
unzip KPCNet_resources.zip

Quick start: Web demo with pretrained model

Play with pretrained model in a flask-based web demo.

The required model and data directory has been specified in the default arguments (trained on Amazon Home & Kitchen dataset). You should be able to start it by:

python web_demo.py

The visit 127.0.0.1:10100 if on a local environment, or server_ip:10100 if you are running the code in a remote server. You should see something like this:

After click "Question" button, you can some CQs being generated like this:

Results may slightly differ as there are random factors in the model like the clustering step.

1st stage: train and evaluate the keyword predictor

train the keyword predictor

export CUDA_VISIBLE_DEVICES=0
python -u main.py --train_context ./data/train_context.txt \
            --train_ques ./data/train_ques.txt \
            --train_ids ./data/train_asin.txt \
            --tune_context ./data/tune_context.txt \
            --tune_ques ./data/tune_ques.txt \
            --tune_ids ./data/tune_asin.txt \
            --test_context ./data/test_context.txt \
            --test_ques ./data/test_ques.txt \
            --test_ids ./data/test_asin.txt \
            --word_embeddings ./data/word_embeddings.p \
            --vocab ./data/vocab.p \
            --kwd_data_dir ./data \
            --kwd_vocab ./data/train_kwd_vocab.txt \
            --batch_size 256 \
            --patience 5 \
            --save_hparams_dir ./hparams \
            --n_epochs 100 \
            --seed 777 \
            --kwd_model_layer 1 \
            --kwd_predictor_type cnn \
            --save_dir ./ckpt \
            --no_neg_sample \
            --pretrain_kwd

predict with the trained model

remember to replace ##YOUR_MODEL## with the produced filename.

export CUDA_VISIBLE_DEVICES=0
python -u predict.py --test_context ./data/test_context.txt \
                  --test_ques ./data/test_ques.txt \
                  --test_ids ./data/test_asin.txt \
                  --word_embeddings ./data/word_embeddings.p \
                  --vocab ./data/vocab.p \
                  --kwd_vocab ./data/train_kwd_vocab.txt \
                  --kwd_model_dir ./ckpt/##YOUR_MODEL##.kwd_pred \
                  --load_hparams_dir ./hparams/##YOUR_MODEL##.json \
                  --out_dir ./output \
                  --batch_size 128 \
                  --eval_kwd

evaluate the keyword prediction

remember to replace ##YOUR_MODEL## with the produced filename.

python ./eval/kwd_eval_pre.py --truth ./data/test_ref.kwds \
                              --kwd ./output/##YOUR_MODEL##.kwd_pred.kwd_prob

python ./eval/kwd_eval_pre.py --truth ./data/test_ref.kwds \
                              --kwd ./output/kwd_D0.3_cnn_noneg.last.kwd_pred.kwd_prob

2nd stage: train and evaluate the whole model pipeline

train the pipeline

export CUDA_VISIBLE_DEVICES=0
python main.py --train_context ./data/train_context.txt \
            --train_ques ./data/train_ques.txt \
            --train_ids ./data/train_asin.txt \
            --tune_context ./data/tune_context.txt \
            --tune_ques ./data/tune_ques.txt \
            --tune_ids ./data/tune_asin.txt \
            --test_context ./data/test_context.txt \
            --test_ques ./data/test_ques.txt \
            --test_ids ./data/test_asin.txt \
            --save_dir ./ckpt \
            --word_embeddings ./data/word_embeddings.p \
            --vocab ./data/vocab.p \
            --kwd_data_dir ./data \
            --kwd_vocab ./data/train_kwd_vocab.txt \
            --batch_size 128 \
            --n_epochs 60 \
            --kwd_predictor_type cnn \
            --no_neg_sample \
            --freeze_kwd_model \
            --save_hparams_dir ./hparams \
            --kwd_model_dir ./ckpt/##YOUR_MODEL##.kwd_pred \
            --pretrain_ques

predict with pipeline

remember to replace ##YOUR_MODEL## with the produced filename.

Threshold

export CUDA_VISIBLE_DEVICES=0
python predict.py	--test_context ./data/test_context.txt \
                  --test_ques ./data/test_ques.txt \
                  --test_ids ./data/test_asin.txt \
                  --word_embeddings ./data/word_embeddings.p \
                  --vocab ./data/vocab.p \
                  --kwd_vocab ./data/train_kwd_vocab.txt \
                  --load_models_dir ./ckpt/##YOUR_MODEL##.models \
                  --load_hparams_dir ./hparams/##YOUR_MODEL##.json \
                  --out_dir ./output \
                  --threshold 0.075 \
                  --save_all_beam \
                  --batch_size 128

Can further apply keyword filtering with --load_filter_dir ./data/kwd_filter_dict.json

CLUSTER

evaluate the pipeline

export CUDA_VISIBLE_DEVICES=0
python predict.py	--test_context ./data/test_context.txt \
                  --test_ques ./data/test_ques.txt \
                  --test_ids ./data/test_asin.txt \
                  --word_embeddings ./data/word_embeddings.p \
                  --vocab ./data/vocab.p \
                  --kwd_vocab ./data/train_kwd_vocab.txt \
                  --load_models_dir ./ckpt/##YOUR_MODEL##.models \
                  --load_hparams_dir ./hparams/##YOUR_MODEL##.json \
                  --out_dir ./output \
                  --load_kwd_edge_dir ./data/kwd_edges.npz \
                  --load_filter_dir ./data/kwd_filter_dict.json \
                  --save_all_beam \
                  --cluster_kwd \
                  --sample_top_k 6 \
                  --kwd_clusters 2 \
                  --sample_times 2 \
                  --batch_size 128

Threshold + Diverse beam search

python predict.py	--test_context ./data/test_context.txt \
                  --test_ques ./data/test_ques.txt \
                  --test_ids ./data/test_asin.txt \
                  --word_embeddings ./data/word_embeddings.p \
                  --vocab ./data/vocab.p \
                  --kwd_vocab ./data/train_kwd_vocab.txt \
                  --load_models_dir ./ckpt/##YOUR_MODEL##.models \
                  --load_hparams_dir ./hparams/##YOUR_MODEL##.json \
                  --load_filter_dir ./data/kwd_filter_dict.json \
                  --out_dir ./output \
                  --threshold 0.075 \
                  --batch_size 128 \
                  --diverse_beam \
                  --save_all_beam \
                  --beam_size 6 \
                  --diverse_group 3 \
                  --diverse_lambda 0.4 \
                  --batch_size 128

SAMPLE

python predict.py	--test_context ./data/test_context.txt \
                  --test_ques ./data/test_ques.txt \
                  --test_ids ./data/test_asin.txt \
                  --word_embeddings ./data/word_embeddings.p \
                  --vocab ./data/vocab.p \
                  --kwd_vocab ./data/train_kwd_vocab.txt \
                  --load_models_dir ./ckpt/##YOUR_MODEL##.models \
                  --load_hparams_dir ./hparams/##YOUR_MODEL##.json \
                  --out_dir ./output \
                  --save_all_beam \
                  --sample_top_k 6 \
                  --sample_top_p 0.9 \
                  --sample_times 2 \
                  --batch_size 128

evaluate the generated sequence

remember to replace ##YOUR_OUTPUT##

BLEU, Distinct-3, response, P@5, length

sh eval_all.sh ##YOUR_OUTPUT##.beam0

METEOR

Need METEOR installed

METEOR=/home/xxx/meteor-1.5
java -Xmx2G -jar $METEOR/meteor-1.5.jar ./output/##YOUR_OUTPUT##.beam0_2 ./data/test_ref_combined \
										-l en -norm -r 10 \

Pairwise BLEU and Avg. BLEU

For threshold decoding (w/o diverse beam search)

python eval/pairwise_bleu.py --hyp_prefix ./output/##YOUR_MODEL##.beam \
                             --ref_prefix ./data/test_ref \
                             --deduplicate

For other selection methods, note to end with a with indicates the id of group

python eval/pairwise_bleu.py --hyp_prefix ./output/##YOUR_MODEL##.a \
                             --ref_prefix ./data/test_ref \
                             --deduplicate

inference

The the example provided in the jupyter notebook infer.ipynb

vhrehfdl/KPCNet

KPCNet

Quick start: Web demo with pretrained model

1st stage: train and evaluate the keyword predictor

train the keyword predictor

predict with the trained model

evaluate the keyword prediction

2nd stage: train and evaluate the whole model pipeline

train the pipeline

predict with pipeline

evaluate the generated sequence

inference