Jupyter Notebook
Beyond Accuracy: Behavioral Testing of NLP models with CheckList

CheckList

This repository contains code for testing NLP Models as described in the following paper:

Beyond Accuracy: Behavioral Testing of NLP models with CheckList
Marco Tulio Ribeiro, Tongshuang Wu, Carlos Guestrin, Sameer Singh Association for Computational Linguistics (ACL), 2020

Bibtex for citations:

@inproceedings{checklist:acl20,
author = {Marco Tulio Ribeiro and Tongshuang Wu and Carlos Guestrin and Sameer Singh},
title = { Beyond Accuracy: Behavioral Testing of NLP models with CheckList},
booktitle = {Association for Computational Linguistics (ACL)},
year = {2020}
}

Table of Contents

• Installation
• Tutorials
• Paper tests
  • Notebooks: how we created the tests in the paper
  • Replicating paper tests, or running them with new models
    • Sentiment Analysis
    • QQP
    • SQuAD
    • Testing huggingface transformer pipelines
• Code snippets
  • Templates
  • RoBERTa suggestions
  • Perturbing data for INVs and DIRs
  • Creating and running tests
  • Custom expectation functions
  • Test Suites
• API reference
• Code of Conduct

Installation

From pypi:

pip install checklist
jupyter nbextension install --py --user checklist.viewer
jupyter nbextension enable --py --user checklist.viewer

From source:

git clone git@github.com:marcotcr/checklist.git
cd checklist
pip install -e .

Either way, you need to install pytorch or tensorflow if you want to use masked language model suggestions:

pip install torch

For most tutorials, you also need to download a spacy model:

python -m spacy download en_core_web_sm

Tutorials

1. Generating data
2. Perturbing data
3. Test types, expectation functions, running tests
4. The CheckList process

Paper tests

Notebooks: how we created the tests in the paper

1. Sentiment analysis
2. QQP
3. SQuAD

Replicating paper tests, or running them with new models

For all of these, you need to unpack the release data (in the main repo folder after cloning):

tar xvzf release_data.tar.gz

Sentiment Analysis

Loading the suite:

import checklist
from checklist.test_suite import TestSuite
suite_path = 'release_data/sentiment/sentiment_suite.pkl'
suite = TestSuite.from_file(suite_path)

Running tests with precomputed bert predictions (replace bert on pred_path with amazon, google, microsoft, or roberta for others):

pred_path = 'release_data/sentiment/predictions/bert'
suite.run_from_file(pred_path, overwrite=True)
suite.summary() # or suite.visual_summary_table()

To test your own model, get predictions for the texts in release_data/sentiment/tests_n500 and save them in a file where each line has 4 numbers: the prediction (0 for negative, 1 for neutral, 2 for positive) and the prediction probabilities for (negative, neutral, positive).
Then, update pred_path with this file and run the lines above.

QQP

import checklist
from checklist.test_suite import TestSuite
suite_path = 'release_data/qqp/qqp_suite.pkl'
suite = TestSuite.from_file(suite_path)

Running tests with precomputed bert predictions (replace bert on pred_path with roberta if you want):

pred_path = 'release_data/qqp/predictions/bert'
suite.run_from_file(pred_path, overwrite=True, file_format='binary_conf')
suite.visual_summary_table()

To test your own model, get predictions for pairs in release_data/qqp/tests_n500 (format: tsv) and output them in a file where each line has a single number: the probability that the pair is a duplicate.

SQuAD

import checklist
from checklist.test_suite import TestSuite
suite_path = 'release_data/squad/squad_suite.pkl'
suite = TestSuite.from_file(suite_path)

Running tests with precomputed bert predictions:

pred_path = 'release_data/squad/predictions/bert'
suite.run_from_file(pred_path, overwrite=True, file_format='pred_only')
suite.visual_summary_table()

To test your own model, get predictions for pairs in release_data/squad/squad.jsonl (format: jsonl) or release_data/squad/squad.json (format: json, like SQuAD dev) and output them in a file where each line has a single string: the prediction span.

Testing huggingface transformer pipelines

See this notebook.

Code snippets

Templates

See 1. Generating data for more details.

import checklist
from checklist.editor import Editor
import numpy as np
editor = Editor()
ret = editor.template('{first_name} is {a:profession} from {country}.',
                       profession=['lawyer', 'doctor', 'accountant'])
np.random.choice(ret.data, 3)

['Mary is a doctor from Afghanistan.',
'Jordan is an accountant from Indonesia.',
'Kayla is a lawyer from Sierra Leone.']

RoBERTa suggestions

See 1. Generating data for more details.
In template:

ret = editor.template('This is {a:adj} {mask}.',  
                      adj=['good', 'bad', 'great', 'terrible'])
ret.data[:3]

['This is a good idea.',
'This is a good sign.',
'This is a good thing.']

Multiple masks:

ret = editor.template('This is {a:adj} {mask} {mask}.',
                      adj=['good', 'bad', 'great', 'terrible'])
ret.data[:3]

['This is a good history lesson.',
'This is a good chess move.',
'This is a good news story.']

Getting suggestions rather than filling out templates:

editor.suggest('This is {a:adj} {mask}.',
               adj=['good', 'bad', 'great', 'terrible'])[:5]

['idea', 'sign', 'thing', 'example', 'start']

Getting suggestions through jupyter visualization:

editor.visual_suggest('This is {a:mask} movie.')

Perturbing data for INVs and DIRs

See 2.Perturbing data for more details.
Custom perturbation function:

import re
import checklist
from checklist.perturb import Perturb
def replace_john_with_others(x, *args, **kwargs):
    # Returns empty (if John is not present) or list of strings with John replaced by Luke and Mark
    if not re.search(r'\bJohn\b', x):
        return None
    return [re.sub(r'\bJohn\b', n, x) for n in ['Luke', 'Mark']]

dataset = ['John is a man', 'Mary is a woman', 'John is an apostle']
ret = Perturb.perturb(dataset, replace_john_with_others)
ret.data

[['John is a man', 'Luke is a man', 'Mark is a man'],
['John is an apostle', 'Luke is an apostle', 'Mark is an apostle']]

General purpose perturbations (see tutorial for more):

import spacy
nlp = spacy.load('en_core_web_sm')
pdataset = list(nlp.pipe(dataset))
ret = Perturb.perturb(pdataset, Perturb.change_names, n=2)
ret.data

[['John is a man', 'Ian is a man', 'Robert is a man'],
['Mary is a woman', 'Katherine is a woman', 'Alexandra is a woman'],
['John is an apostle', 'Paul is an apostle', 'Gabriel is an apostle']]

ret = Perturb.perturb(pdataset, Perturb.add_negation)
ret.data

[['John is a man', 'John is not a man'],
['Mary is a woman', 'Mary is not a woman'],
['John is an apostle', 'John is not an apostle']]

Creating and running tests

See 3. Test types, expectation functions, running tests for more details.

MFT:

import checklist
from checklist.editor import Editor
from checklist.perturb import Perturb
from checklist.test_types import MFT, INV, DIR
editor = Editor()

t = editor.template('This is {a:adj} {mask}.',  
                      adj=['good', 'great', 'excellent', 'awesome'])
test1 = MFT(t.data, labels=1, name='Simple positives',
           capability='Vocabulary', description='')

INV:

dataset = ['This was a very nice movie directed by John Smith.',
           'Mary Keen was brilliant.',
          'I hated everything about this.',
          'This movie was very bad.',
          'I really liked this movie.',
          'just bad.',
          'amazing.',
          ]
t = Perturb.perturb(dataset, Perturb.add_typos)
test2 = INV(**t)

DIR:

from checklist.expect import Expect
def add_negative(x):
    phrases = ['Anyway, I thought it was bad.', 'Having said this, I hated it', 'The director should be fired.']
    return ['%s %s' % (x, p) for p in phrases]

t = Perturb.perturb(dataset, add_negative)
monotonic_decreasing = Expect.monotonic(label=1, increasing=False, tolerance=0.1)
test3 = DIR(**t, expect=monotonic_decreasing)

Running tests directly:

from checklist.pred_wrapper import PredictorWrapper
# wrapped_pp returns a tuple with (predictions, softmax confidences)
wrapped_pp = PredictorWrapper.wrap_softmax(model.predict_proba)
test.run(wrapped_pp)

Running from a file:

# One line per example
test.to_raw_file('/tmp/raw_file.txt')
# each line has prediction probabilities (softmax)
test.run_from_file('/tmp/softmax_preds.txt', file_format='softmax', overwrite=True)

Summary of results:

test.summary(n=1)

Test cases: 400
Fails (rate): 200 (50.0%)

Example fails:
0.2 This is a good idea

Visual summary:

test.visual_summary()

Saving and loading individual tests:

# save
test.save(path)
# load
test = MFT.from_file(path)

Custom expectation functions

See 3. Test types, expectation functions, running tests for more details.

If you are writing a custom expectation functions, it must return a float or bool for each example such that:

• > 0 (or True) means passed,
• <= 0 or False means fail, and (optionally) the magnitude of the failure, indicated by distance from 0, e.g. -10 is worse than -1
• None means the test does not apply, and this should not be counted

Expectation on a single example:

def high_confidence(x, pred, conf, label=None, meta=None):
    return conf.max() > 0.95
expect_fn = Expect.single(high_confidence)

Expectation on pairs of (orig, new) examples (for INV and DIR):

def changed_pred(orig_pred, pred, orig_conf, conf, labels=None, meta=None):
    return pred != orig_pred
expect_fn = Expect.pairwise(changed_pred)

There's also Expect.testcase and Expect.test, amongst many others.
Check out expect.py for more details.

Test Suites

See 4. The CheckList process for more details.

Adding tests:

from checklist.test_suite import TestSuite
# assuming test exists:
suite.add(test)

Running a suite is the same as running an individual test, either directly or through a file:

from checklist.pred_wrapper import PredictorWrapper
# wrapped_pp returns a tuple with (predictions, softmax confidences)
wrapped_pp = PredictorWrapper.wrap_softmax(model.predict_proba)
suite.run(wrapped_pp)
# or suite.run_from_file, see examples above

To visualize results, you can call suite.summary() (same as test.summary), or suite.visual_summary_table(). This is what the latter looks like for BERT on sentiment analysis:

suite.visual_summary_table()

Finally, it's easy to save, load, and share a suite:

# save
suite.save(path)
# load
suite = TestSuite.from_file(path)

API reference

On readthedocs

Code of Conduct

Microsoft Open Source Code of Conduct