Skip to content

Behavioral Testing

What is it?

Performing behavioral testing on ML models was first introduced in the checklist paper (Ribeiro, Marco Tulio, et al., 20201). Behavioral tests provide an assessment of the model robustness to small modifications to the input. Proper behavioral testing can help in detecting bias or other potential harmful aspects of the model that may not be otherwise obvious.

Where is this used in Azimuth?

In Azimuth, behavioral tests are automatically executed when launching the tool, using the provided dataset and model.

  • The details of all the tests that were computed for a given utterance are shown in the Utterance Details.
  • A summary of each test for all utterances in both dataset splits (training and evaluation) is available in the Behavioral Testing Summary.
  • Finally, a Smart Tag is generated for each utterance for which at least one test of each family has failed.

How is it computed?

The tests are deterministic for reproducibility purposes.

Test Failing Criteria

The tests can fail for two reasons.

  • The test will fail if the predicted class for the modified utterance is different from the predicted class of the original utterance.

    Failing examples
    Original Utterance Predicted Class
    Azimuth is the best tool positive
    Modified Utterance Predicted Class Test fails?
    Hello Azimuth is the best tool positive NO
    Azimuth is the best tool!!! negative YES

  • The test will fail if the confidence associated with the predicted class of the modified utterance is too different (based on a threshold) from the confidence of the original utterance. By default, the threshold is set to 1, meaning the tests will never fail due to a change in confidence for the same predicted class.

    Failing examples
    Original Utterance Predicted Class Confidence
    Azimuth is the best tool positive 95%

    Threshold set to 0.1:

    Modified Utterance Predicted Class Confidence Test fails?
    Hello Azimuth is the best tool positive 82% YES

    Threshold set to 1:

    Modified Utterance Predicted Class Confidence Test fails?
    Hello Azimuth is the best tool positive 82% NO

Available Tests

All tests are invariant (the modification should not change the predicted class) and assess the robustness of the model.

  • The tool currently has 2 families of tests: Fuzzy Matching and Punctuation.
  • For each test, different modification types can be applied (Insertion, Deletion, etc.)
    • For certain tests, all modification types are applied to each utterance (e.g., Typos and Neutral Token).
    • For others, only one modification type is applied based on the presence of a certain pattern in the utterance (e.g., Punctuation and Contractions tests).

Fuzzy Matching

  • Typos: For this test, we simulate common typos that might happen when typing an utterance. By default, the test creates one typo per utterance. The different types of simulated typos ( modification types) are:

    • Swap: Random swap of two adjacent characters in a word.
    • Deletion: Deletion of random characters in a word.
    • Replacement: Keyboard proximity-based typos inserted in a word.

  • Neutral Token: Default neutral tokens are added to the utterance.

    • PreInsertion: One of ["pls", "please", "hello", "greetings"] is added at the beginning of the utterance.
    • PostInsertion: One of ["pls", "please", "thank you", "appreciated"] is added at the end of an utterance.

  • Contractions: This test is applied only when the utterance contains a relevant expression that can be contracted or expanded. The list is taken from NL-Augmenter .

    • Contraction: Contract relevant expressions, if present.
    • Expansion: Expand relevant expressions, if present.

Punctuation

  • Question Mark: Adds/Deletes/Replaces question marks.

    • Deletion: Removes the ending question mark, if present.
    • Replacement: Replaces the ending punctuation sign ('.', '!', ','), if present, by a question mark.
    • PostInsertion: Adds an ending question mark when the utterance does not end with a punctuation sign.

  • Ending period: Same logic as the Question Mark test, with a period.

    • Deletion: Removes the ending period, if present.
    • Replacement: Replaces the ending punctuation sign ('?', '!', ','), if present, by a period.
    • PostInsertion: Adds an ending period when the utterance does not end with a punctuation sign.

  • Inner Comma: Adds/Deletes comma inside the utterance (not at the end).

    • Deletion: Removes all commas inside the utterance, if present.
    • Insertion: Adds a comma near the middle of the utterance.

  • Inner Period: Same logic as the Inner Comma test, with a period.

    • Deletion: Removes all periods inside the utterance, if present.
    • Insertion: Adds a period near the middle of the utterance.

Configuration

Behavioral Testing Configuration details how to change some parameters, such as the lists of neutral tokens, the number of typos and the threshold confidence delta above which the tests should fail.

  1. Ribeiro, Marco Tulio, et al. "Beyond accuracy: Behavioral testing of NLP models with CheckList." Association for Computational Linguistics (ACL), 2020. ↩