TARexp
is an opensource Python framework for conducting TAR experiments with various
reference implementation to algorithms and methods that are commonly-used.
The experiments are fully reproducible and easy to conduct ablation studies.
For studying components that do not change the selection process of the review documents,
TARexp
supports replying TAR runs and experimenting these components offline.
Helper functions to support results analysis are also avaiable.
Please visit our Google Colab Demo to check out the full running example
Please refer to the documentation for more detail: https://eugene.zone/tarexp.
You can install TARexp
from PyPi by running
pip install tarexp
Or install it with the lastest version from GitHub
pip install git+https://github.com/eugene-yang/tarexp.git
If you like to build it from source, please use
git clone https://github.com/eugene-yang/tarexp.git
cd tarexp
python setup.py bdist_wheel
pip install dist/*.whl
In Python, please use the following command to import both the main package and the components
import tarexp
from tarexp import component
The following snippet is an example of creating a dataset
instance for TARexp
.
For scikit-learn
rankers, the structure of the dataset is bascially a sparse scipy
matrix
for the vectorized dataset and a list or an array of binary labels with the same length of the matrix.
from sklearn import datasets
import pandas as pd
rcv1 = datasets.fetch_rcv1()
X = rcv1['data']
rel_info = pd.DataFrame(rcv1['target'].todense().astype(bool), columns=rcv1['target_names'])
ds = tarexp.SparseVectorDataset.from_sparse(X)
The following snippet defines a set of componets to use for a workflow,
setting = component.combine(component.SklearnRanker(LogisticRegression, solver='liblinear'),
component.PerfectLabeler(),
component.RelevanceSampler(),
component.FixedRoundStoppingRule(max_round=20))()
And to declare a workflow, simply put in your dataset, setting, and other parameters to the workflow.
workflow = tarexp.OnePhaseTARWorkflow(
ds.set_label(rel_info['GPRO']),
setting,
seed_doc=[1023],
batch_size=200,
random_seed=123
)
And finally, you can start executing the workflow by running it as an iterator.
We also support everything from ir-measures
as evaluation metrics.
recording_metrics = [ir_measures.RPrec, tarexp.OptimisticCost(target_recall=0.8, cost_structure=(25,5,5,1))]
for ledger in workflow:
print("Round {}: found {} positives in total".format(ledger.n_rounds, ledger.n_pos_annotated))
print("metric:", workflow.getMetrics(recording_metrics))
Besides standard IR evaluation metrics, we also implement OptimisticCost
as cost-based evaluation metrics in TARexp
. Please refer to this paper for more information and consider citing it if you use this measurement.
tarexp.TARExperiment
is a wrapper and dispatcher for running TAR experiments with different settings.
It construct all combinations of the input settings and dispath each TAR run to execute.
The following command defines a set of 6 TAR runs which consists of 3 topics and each has 2 runs with batch size 200 and 100.
exp = tarexp.TARExperiment('./my_tar_exp/', random_seed=123, max_round_exec=20,
metrics=[RPrec, P@10, tarexp.OptimisticCost(target_recall=0.8, cost_structure=(1,10,1,10))],
tasks=tarexp.TaskFeeder(ds, rel_info[['GPRO', 'GOBIT', 'E141']]),
components=setting,
workflow=tarexp.OnePhaseTARWorkflow, batch_size=[200, 100])
To start running the experiment, please use the following command which will execute with single processor and resume from any crash runs if exist in the output directory.
results = exp.run(n_processes=1, resume=True, dump_frequency=10)
TARexp
also encourages experiments on stopping rules.
We have built-in a number of stopping rules in the package and continuing to update them.
The following snippet is an exmaple for running a replay experiment based on a set of existing
TAR runs with a list of stopping rules defined in stopping_rules
arguments.
replay_exp = tarexp.StoppingExperimentOnReplay(
'./test_stopping_rules', random_seed=123,
tasks=tarexp.TaskFeeder(ds, rel_info[['GPRO','GOBIT', 'E141']]),
replay=tarexp.OnePhaseTARWorkflowReplay,
saved_exp_path='./my_tar_exp',
metrics=[tarexp.OptimisticCost(target_recall=0.8, cost_structure=(1,1,1,1)),
tarexp.OptimisticCost(target_recall=0.9, cost_structure=(1,1,1,1))],
stopping_rules=[
component.KneeStoppingRule(),
component.BudgetStoppingRule(),
component.BatchPrecStoppingRule(),
component.ReviewHalfStoppingRule(),
component.Rule2399StoppingRule(),
component.QuantStoppingRule(0.4, 0),
component.QuantStoppingRule(0.2, 0),
component.QuantStoppingRule(0.8, 0),
component.CHMHeuristicsStoppingRule(0.8),
component.CHMHeuristicsStoppingRule(0.4),
component.CHMHeuristicsStoppingRule(0.2),
]
)
stopping_results = replay_exp.run(resume=True, dump_frequency=10)
TARexp
also provide visualization tools for TAR runs.
createDFfromResults
creates a pandas DataFrame from either the result variable
df = tarexp.helper.createDFfromResults(results, remove_redundant_level=True)
Or the output directory
df = tarexp.helper.createDFfromResults('./my_tar_exp', remove_redundant_level=True)
And the following command provides you the cost dynamic graph introduced in this paper.
tarexp.helper.cost_dynamic(
df.loc[:, 'GOBIT', :].groupby(level='dataset'),
recall_targets=[0.8], cost_structures=[(1,1,1,1), (10, 10, 1, 1), (25, 5, 5, 1)],
with_hatches=True
)
Alternatively, you can also create this graph by using a command line interface
python -m tarexp.helper.plotting \
--runs GPRO=./my_tar_exp/GPRO.61b1f31a0a29de634939db77c0dde383/ \
GOBIT=./my_tar_exp/GOBIT.ae86e0b37809cb139dfa1f4cf914fb9b/ \
--cost_structures 1-1-1-1 25-5-5-1 --y_thousands --with_hatches
Any feedback is welcome! You can reach out to us either by emailing the author or rasing an issue!
The demo paper of TARexp
is currently under review.
If you use the cost measure or the cost dynamic graphs, pleas consider citing this paper
@inproceedings{cost-structure,
author = {Eugene Yang and David D. Lewis and Ophir Frieder},
title = {On Minimizing Cost in Legal Document Review Workflows},
booktitle = {Proceedings of the ACM Symposium on Document Engineering (DocEng)},
year = {2021},
url = {https://arxiv.org/abs/2106.09866}
}