beckdaniel / NLPerf

Performance Prediction for NLP Tasks

NLPerf

NLPerf is an exploratory framework to evaluate performance for NLP tasks by training predictors using past experimental records. We provide data, code, experiment logs for 9 tasks in this repository. For more technical details, please refer to our paper:

Predicting Performance for Natural Language Processing Tasks

Mengzhou Xia, Antonios Anastasopoulos, Ruochen Xu, Yiming Yang, Graham Neubig

Scripts

Predicting Performance

K-fold evaluation for TED-MT over 10 random runs. You'll have to modify the parameters in the code.

python3 main_code.py 

...
The average rmse for 10 runs is 6.177170443398633
The average k_split baseline rmse for 10 runs is 12.653809064828215
The average Source Language baseline rmse for 10 runs is 12.653809064828215
The average Target Language baseline rmse for 10 runs is 12.653809064828215

For single model tasks:

k_fold_evaluation("wiki",
                  shuffle=True,
                  selected_feats=None,
                  combine_models=False,
                  regressor="xgboost",
                  k=5,
                  num_running=10)

For multi model tasks:

k_fold_evaluation("ud",
                  shuffle=True,
                  selected_feats=None,
                  combine_models=True,
                  regressor="xgboost",
                  k=5,
                  num_running=10)

You can specify a set of features for training by passing a list of feature names to selected_feats.

Representativeness

Search types include best_search, worst_search and random_search. n denotes the maximum size of the final set and beam_size denotes the number of expanded nodes in each step. Note that monomt denotes the TED-MT task.

python3 representativeness.py --task monomt \ 
                              --log logs/monomt_bs.log \
                              --n 5 \
                              --beam_size 100 \
                              --type best_search

New Models

Using existing experimental records for other models and n experimental records of 50% of the experimental records from a new coming model to make predictions on the rest of the 50%.

n = 5
python3 new_model.py --task ud \
                     --log logs/ud_nm_${n}.log \
                     --n 5 \
                     --portion 0.5 \
                     --test_id_options_num 100 \
                     --sample_options_num 100