Seq2Seq Dialogue

datasets

• Natural Questions on github
  • wget https://storage.googleapis.com/natural_questions/v1.0-simplified/simplified-nq-train.jsonl.gz
• topical-chat
  • git clone https://github.com/Shikib/Response-Generation-Baselines.git
  • unzip Response-Generation-Baselines/data.zip -d topical-chat

  topicalchat-train: 179708
  topicalchat-val-freq: 11273
  topicalchat-val-rare: 11245
  topicalchat-test: 11206

• hotpotqa; hotpot-paper
• coqa
• SQUAD 2.0
• convai-challenge and persona-chat
  • https://gist.github.com/thomwolf/ecc52ea728d29c9724320b38619bd6a6
• quangaroo
  • multi-document QA, how to use? concat documents?? Context way to big to fit in transformer

number of seq2seq-examples

topicalchat-train: 182347
personachat-train: 262875
coqa-train: 108646
squad20-train: 86820
coqa-val: 7982
squad20-val: 20301

training

0. build corpus: python build_seq2seq_corpus.py
1. train: optionally continue from--checkpoint=$HOME/data/bart_seq2seq_dialogue_new/checkpointepoch=2.ckpt \
2. only once run train-command on frontend with OMP_NUM_THREADS=4 to downloads pretrained model

OMP_NUM_THREADS=2 wandb init # on frontend

export PYTHONPATH=~/transformers/examples
CUDA_VISIBLE_DEVICES=1 WANDB_MODE=dryrun python ../transformers/examples/seq2seq/finetune.py \
--data_dir=$HOME/data/seq2seq_dialogue_topicalchat \
--model_name_or_path=sshleifer/distilbart-xsum-12-1 \
--learning_rate=3e-5 \
--max_source_length=1024 \
--max_target_length=128 \
--train_batch_size=2 \
--eval_batch_size=2 \
--output_dir=coqa-distilbart-xsum-12-1 \
--num_train_epochs 3 \
--fp16 \
--gpus 1 \
--do_train \
--do_predict \
--n_val 1000 \
--val_check_interval 0.1 \
--sortish_sampler \
--logger wandb \
--wandb_project seq2seq-chatbot

evaluating

DSTC9-Baseline

• Transformer trained by Shikib Mehri
• no cheating!

export PYTHONPATH=~/code/NLP/transformers/examples
python evaluation.py --pred_file ~/gunther/data/transformer_trained/test_rare_epoch_20.pred --target_file ~/gunther/Response-Generation-Baselines/processed_output/test_rare.tgt
{'f1-scores': {'rouge-1': 0.18832909781578547,
               'rouge-2': 0.047195065416391825,
               'rouge-l': 0.18423218591116589},
 'huggingface-rouge': {'rouge1': 0.1902879866916785,
                       'rouge2': 0.049791688778764095,
                       'rougeL': 0.16477289347786828}}

• evaluated with cheating: cd Response-Generation-Baselines && python3 test.py --save_path /docker-share/data/transformer_trained --epoch 20 --cheating

{'f1-scores': {'rouge-1': 0.19093731158816457,
               'rouge-2': 0.04742658960617347,
               'rouge-l': 0.1866664146489359},
 'huggingface-rouge': {'rouge1': 0.19244537896234243,
                       'rouge2': 0.049771674296700374,
                       'rougeL': 0.16611361064019586}}

-> cheating makes no big difference!

• trained no cheating

python evaluation.py --pred_file ~/gunther/Response-Generation-Baselines/no_cheating/test_rare_epoch_4.pred
{'f1-scores': {'rouge-1': 0.20489976209624824,
               'rouge-2': 0.05517858678345437,
               'rouge-l': 0.20176257339875925},
 'huggingface-rouge': {'rouge1': 0.20666656629287844,
                       'rouge2': 0.05805848849887245,
                       'rougeL': 0.17717360382688946}}

• trained and evaluated WITH cheating!
  • train: python3 train.py --use_knowledge --transformer --batch_size=8 --save_path cheating/ --num_layers=1 --num_epochs=10 --cheating
  • evaluate: CUDA_VISIBLE_DEVICES=1 python3 test.py --save_path cheating/ --epoch 2 --cheating

python evaluation.py --pred_file ~/gunther/Response-Generation-Baselines/cheating/test_rare_epoch_9.pred
{'f1-scores': {'rouge-1': 0.6284242142836528,
               'rouge-2': 0.382027446998092,
               'rouge-l': 0.6279301955302672},
 'huggingface-rouge': {'rouge1': 0.6254876724740767,
                       'rouge2': 0.3868426341376391,
                       'rougeL': 0.5984314895063148}}

huggingface-evaluation

• coqa-distilbart-xsum-12-1

(ml_gpu) [tilo-himmelsbach@gpu010 seq2seq]$ python evaluate_transformer.py 
100%|████| 998/998 [14:02<00:00,  1.18it/s]
{'f1-scores': {'rouge-1': 0.2750953579003743,
               'rouge-2': 0.15929957668606087,
               'rouge-l': 0.28888274540104353},
 'huggingface-rouge': {'rouge1': 0.3280108921933519,
                       'rouge2': 0.19008676583777595,
                       'rougeL': 0.324871404293655}}

• topicalchat

python evaluate_transformer.py --model_path topicalchat-distilbart-xsum-12-1/best_tfmr --source_file ~/data/seq2seq_dialogue_topicalchat/test.source --target_file ~/data/seq2seq_dialogue_topicalchat/test.target
100%|████| 1401/1401 [27:21<00:00,  1.17s/it]
{'f1-scores': {'rouge-1': 0.2938204560963201,
               'rouge-2': 0.1319923371771301,
               'rouge-l': 0.2853012250506464},
 'huggingface-rouge': {'rouge1': 0.2980726989168491,
                       'rouge2': 0.14077689234989404,
                       'rougeL': 0.2667963920497738}}

coqa-baselines-evaluation

{'bart': {'em': 47.9, 'f1': 65.5, 'turns': 7983},
 'bart-danqi': {'em': 31.6, 'f1': 43.7, 'turns': 7983}, # TODO(tilo): why is danqi not working?
 'cheatbot': {'em': 94.7, 'f1': 97.3, 'turns': 7983}, # should be at 100 percent! but it is not!
 'echobot': {'em': 0.0, 'f1': 3.5, 'turns': 7983}}

dash-frontend

1. on gunther, get model from hpc: rsync -avz -e "ssh -o StrictHostKeyChecking=no -o UserKnownHostsFile=/dev/null" --exclude=.git tilo-himmelsbach@gateway.hpc.tu-berlin.de:/home/users/t/tilo-himmelsbach/data/bart_seq2seq_dialogue_continued/checkpointepoch=2.ckpt ~/data/bart_coqa_seq2seq/
2. dash-app in browser: gunther:8050

similar projects

• https://github.com/NTU-SQUAD/transformers-coqa -> extractive QA! not generative!

nlg-eval

pip install git+https://github.com/Maluuba/nlg-eval.git@master