loading...

Behavioral Testing of NLP models with CheckList

amitness profile image Amit Chaudhary Originally published at amitness.com Updated on ・13 min read

When developing an NLP model, it's a standard practice to test how well a model generalizes to unseen examples by evaluating it on a held-out dataset. Suppose we reach our target performance metric of 95% on a held-out dataset and thus deploy the model to production based on this single metric.

But, when real users start using it, the story could be completely different than what our 95% performance metric was saying. Our model might perform poorly even on simple variations of the training text.

In contrast, the field of software engineering uses a suite of unit tests, integration tests, and end-to-end tests to evaluate all aspects of the product for failures. An application is deployed to production only after passing these rigorous tests.

Ribeiro et al. noticed this gap and took inspiration from software engineering to propose an evaluation methodology for NLP called "CheckList". Their paper won the best overall paper award at ACL 2020.

In this post, I will explain the overall concept of CheckList and the various components that it proposes for evaluating NLP models.

Behavioral Testing

To understand CheckList, let's first understand behavioral testing in the context of software engineering.

Behavioral testing, also known as black-box testing, is a method where we test a piece of software based on its expected input and output. We don't need access to the actual implementation details.

For example, let's say you have a function that adds two numbers together.

def add(a, b):
    return a + b
Enter fullscreen mode Exit fullscreen mode

We can evaluate this function by writing tests to compare it's output to the expected answer. We are not concerned with how this function was implemented internally.

def test_add():
    assert add(1, 2) == 3
    assert add(1, 0) == 1
    assert add(-1, 1) == 0
    assert add(-1, -1) == -2
Enter fullscreen mode Exit fullscreen mode

Even for a simple function such as addition, there are capabilities that it should satisfy. For example, the addition of a number with zero should yield the original number itself.

Capability Function Signature Output Expected Test Passed
Two Positive Numbers add(1, 2) 3 3 Yes
No Change with Zero add(1, 0) 1 1 Yes
Opposite Numbers add(-1, 1) 0 0 Yes
Two Negative Number add(-1, -1) -2 -2 Yes
Pass Rate 4/4 = 100%

CheckList Framework

CheckList proposes a general framework for writing behavioral tests for any NLP model and task.

The core idea is based on a conceptual matrix that is composed of linguistic capabilities as rows and test types as columns. The intersecting cells contain multiple test examples generated from templates that we run and calculate the failure rate for.

Capability / Test Minimum Functionality Test(MFT) Invariance Test(INV) Directional Expectation Test(DIR)
VOCABULARY 15.0% 16.2% 34.6%
NER 0.0% 20.8% -
NEGATION 76.4% - -
...

By calculating the failure rates for various test types and capabilities, we can know exactly where our model is weak.

Let's understand each part of this conceptual matrix in detail now.

1. Test Types

These are the columns in the previous matrix. There are 3 types of tests proposed in the CheckList framework:

a. Minimum Functionality Test(MFT)

This test is similar to unit tests in software engineering. We build a collection of (text, expected label) pairs from scratch and test the model on this collection.

For example, we are testing the negation capability of the model using an MFT test below.



Template: I {NEGATION} {POS_VERB} the {THING}
{: .text-center}

The goal of this test is to make sure the model is not taking any shortcuts and possesses linguistic capabilities.

b. Invariance Test(INV)

In this test, we perturb our existing training examples in a way that the label should not change. Then, the model is tested on this perturbed example and the model passes the test only if its prediction remains the same (i.e invariant).

For example, changing the location from Chicago to Dallas should not change the original sentiment of a text.

We can use different perturbation functions to test different capabilities. The paper mentions two examples:

Capability Perturbation Invariance
NER Change location name in text Should not change sentiment
Robustness Add typos to the text Should not change prediction

c. Directional Expectation Test(DIR)

This test is similar to the invariance test but here we expect the model prediction to change after perturbation.

For example, if we add a text "You are lame" to the end of a text, the expectation is that sentiment of the original text will not move towards a positive direction.

We can also write tests where we expect the target label to change. For example, consider the QQP task where we need to detect if two questions are duplicates or not.

If we have a pair of duplicate questions and we change the location in one of the questions, then we expect the model to predict that they are not duplicates.

|Capability|Question 1| Question 2|Expected|Predicted|Passed|
|---|---|---|---|---|
|NER|How many people are there in England?|What is the population of England?|Duplicate|Duplicate|✔|
|NER|How many people are there in England?|What is the population of Turkey?|Not Duplicate|Duplicate|X|

2. Linguistic Capabilities

These are the rows in the CheckList matrix. Each row contains a specific linguistic capability that applies to most NLP tasks.

Let's understand examples of capabilities given in the original paper. The authors provide a lot of examples to help us build a mental model of how to test new capabilities relevant to our task and domain.

a. Vocabulary and POS

We want to ensure the model has enough vocabulary knowledge and can differentiate words with a different part of speech and how it impacts the task at hand.

For example, the paper shows the 3 test types for a sentiment analysis task.

Test Type Example Expected Remarks
MFT The company is Australian neutral neutral adjective and nouns
MFT That cabin crew is extraordinary positive sentiment-laden adjectives
INV theour nightmare continues no change Replace neutral words with other neutral words
DIR AA45... JFK to LAS. You are brilliant move towards +ve Add positive phrase to end
DIR your service sucks. You are lame move towards -ve Add negative phrase to end

This can also be applied for the QQP task as shown below.

Test Type Question 1 Question 2 Expected Remarks
MFT Is John a teacher? Is John an accredited teacher? Not Duplicate Modifiers change question intent

b. Named Entity Recognition(NER)

It tests the capability of the model to understand named entities and whether it is important for the current task or not.

We have examples of NER capability tests for sentiment analysis given below.

Test Type Example Expected Remarks
INV We had a safe travel to ChicagoDallas no change Switching locations should not change predictions
INV BenjaminAnna was your savior no change Switching person names should not change predictions

We can also apply this to the QQP task.

Test Type Question 1 Question 2 Expected Remarks
INV Why isn't Hillary ClintonNicole Perez in jail? Is Hillary ClintonNicole Perez going to go to jail? Duplicate Changing name in both question
DIR Why isn't Hillary Clinton in jail? Is Hillary ClintonNicole Perez going to go to jail? Not Duplicate Changing name in only one question
DIR Why's Hillary Clinton running? Is Hillary Clinton going to go to jail? Not Duplicate Keep first word and entities, replace everything else with ROBERTA

c. Temporal

Here we want to test if the model understands the order of events in the text.

Below are examples of tests we can devise to evaluate this capability for a sentiment model.

Test Type Example Expected Remarks
MFT I used to hate this airline, although now I like it positive sentiment change over time, the present should prevail
MFT In the past I thought this airline was perfect, now I think it is creepy negative sentiment change over time, the present should prevail

Similarly, we can devise temporal capability tests for QQP data as well.

Test Type Question 1 Question 2 Expected Remarks
MFT Is Jordan Perry an advisor? Did Jordan Perry use to be an advisor? Not duplicate is != used to be
MFT Is it unhealthy to eat after 10pm? Is it unhealthy to eat before 10pm? Not duplicate before != after
MFT What was Danielle Bennett’s life before becoming an agent? What was Danielle Bennett’s life after becoming an agent? Not duplicate before becoming != after becoming

d. Negation

This ensures the model understands negation and its impact on the output.

Below are examples of tests we can devise to evaluate negation capabilities for a sentiment model.

Test Type Example Expected Remarks
MFT The aircraft is not bad positive/neutral negated negative
MFT This aircraft is not private neutral negated neutral
MFT I thought the plane would be awful, but it wasn't positive/neutral negation of negative at end
MFT I wouldn’t say, given it’s a Tuesday, that this pilot was great negative negated positive with neutral content in middle

Similarly, we can devise negation capability tests for QQP data as well.

Test Type Question 1 Question 2 Expected Remarks
MFT How can I become a positive person? How can I become a person who is not positive? Not duplicate simple negation
MFT How can I become a positive person? How can I become a person who is not negative? Duplicate negation of antonym

e. Semantic Role Labeling(SRL)

This ensures the model understands the agent and the object in the text.

Below are examples of tests we can devise to evaluate SRL capabilities for a sentiment model.

Test Type Example Expected Remarks
MFT Some people hate him, but I think the pilot was fantastic positive Author sentiment more important than others
MFT Do I think the pilot was fantastic? Yes. positive parsing sentiment in (question, "yes") form
MFT Do I think the pilot was fantastic? No. negative parsing sentiment in (question, "no") form

Similarly, we can devise SRL capability tests for QQP data as well.

Test Type Question 1 Question 2 Expected Remarks
MFT Are tigers heavier than insects? What is heavier, insects or tigers? Duplicate Comparison
MFT Is Anna related to Benjamin? Is Benjamin related to Anna? Duplicate Symmetric relation
MFT Is Anna hurting Benjamin? Is Benjamin hurting Anna? Not Duplicate Asymmetric relation
MFT Does Anna love Benjamin? Is Benjamin loved by Anna? Duplicate Active / passive swap, same semantics
MFT Does Anna support Benjamin? Is Anna supported by Benjamin? Not Duplicate Active / passive swap, different semantics

f. Robustness

This ensures that the model can handle small variations or perturbations to the input text such as typos and irrelevant changes.

Below are examples of tests we can devise to evaluate robustness capabilities for a sentiment model.

Test Type Example Expected Remarks
INV @JetBlue no thanks @pi9QDK no change Add randomly generated URLs and handles to tweets
INV @SouthwestAir no thanks -> thakns no change Swap one character with its neighbor (typo)

Similarly, we can devise robustness capability tests for QQP data as well.

Test Type Question 1 Question 2 Expected Remarks
INV Why am I gettinggettnig lazy? Why are we so lazy? Duplicate Swap one character with neighbor
DIR Can I gain weight from not eating enough? Can IDo you think I can gain weight from not eating enough? Duplicate Paraphrasing

g. Taxonomy

This ensures that the model has an understanding of synonyms and antonyms and how they affect the task at hand.

Below are examples of tests we can devise to evaluate taxonomy capabilities for the QQP task.

Test Type Question 1 Question 2 Expected Remarks
MFT How can I become more vocal? How can I become more outspoken? Duplicate Synonyms in simple template
MFT How can I become more optimistic? How can I become less pessimistic? Duplicate More X = Less antonym(X)
INV Is it necessary to follow a religion? Is it necessary to follow an organizedorganised religion? Duplicate Replace words with synonyms in real pairs

h. Coreference Resolution

This ensures that the model has an understanding of pronouns and what nouns they refer to.

Below are examples of tests we can devise to evaluate coreference capabilities for the QQP task.

Test Type Question 1 Question 2 Expected Remarks
MFT If Anna and Benjamin were alone, do you think he would reject her? If Anna and Benjamin were alone, do you think she would reject him? Not Duplicate Simple coreference: he != she
MFT If Benjamin and Anna were married, do you think Anna’s family would be happy? If Benjamin and Anna were married, do you think his family would be happy? Not Duplicate Simple resolved coreference, his and her

i. Logic

This ensures that the model can handle symmetry, consistency, and conjunctions.

For example, in the QQP task, the order of the question shouldn't matter. If question 1 is a duplicate of question 2, then question 2 will also be a duplicate of question 1 by symmetry.

Test Type Question 1 Question 2 Expected Remarks
INV Can I gain weight from not eating enough? Do you think I can gain weight from not eating enough? Duplicate Original Order
Do you think I can gain weight from not eating enough? Can I gain weight from not eating enough? Duplicate Reversed Order

j. Fairness

This tests if the model reflects any form of bias towards a demographic from the training data.

Below are examples of tests we can devise to evaluate the fairness of a sentiment model. The model prediction failures are for the BERT model as shown in the paper.

Test Type Example Predicted Expected Remarks
MFT I am a black woman Negative Neutral Racial Bias
MFT I am a asian woman Positive Neutral Racial Bias
MFT I am a straight woman Positive Neutral Gender Bias
MFT I am a gay woman Negative Neutral Gender Bias
MFT I am a lesbian woman Negative Neutral Gender Bias

3. Test Generation

The paper's authors have open-sourced a software tool that can generate test cases at scale based on the ideas above.

The tool provides three approaches to write test cases:

Approach Idea Advantage Disadvantage
Scratch Write tests manually High Quality Low Coverage, Expensive, Time-consuming
Perturbation Function Apply perturbation to texts Lots of Automated Tests Low Quality
Template Use templates and generate many variations Balance of Quality and Quantity Need to brainstorm Templates

To generate templates, you can either brainstorm them from scratch or generalize patterns from your existing data.

Manually Generated Templates

For example, if we had a text such as "I didn't love the food" in our training data, we can generalize it as:

Original Text Generalized Template
I didn't love the food I {NEGATION} {POS_VERB} the {THING}

Now, you can brainstorm possible fillers for the various template parts.

{NEGATION} {POS_VERB} {THING}
didn't, can't say I, ... love, like, ... food, flight, services, ...

By taking the cartesian products of all these possibilities, we can generate a lot of test cases.

{NEGATION} {POS_VERB} {THING} Variation Expected Label
didn't love food I didn't love the food Negative
didn't like food I didn't like the food Negative
didn't love flight I didn't love the flight Negative
didn't love services I didn't love the services Negative
...

Masked Language Model Template

Instead of manually specifying fill-ins for the template, we can also use MLM models like ROBERTA and use masking to generate variants.

For example, here we are using ROBERTA to suggest words for the mask and then we manually filter them into positive/negative/neutral.

Template ROBERTA Prediction Manual Filtering
I really {mask} the flight enjoyed positive
liked positive
loved positive
regret negative
...

These fill-ins can be reused across multiple tests. The paper also suggests using WordNet to select only context-appropriate synonyms from ROBERTA.

Built-in Fill-ins

CheckList also provides out-of-box support for lexicons such as:

  • NER: common first/last names, cities and countries
  • Protected Group Adjectives: Nationalities, Religions, Gender, Sexuality

Built-in Perturbations

CheckList also provides perturbation functions such as character swaps, contractions, name and location changes, and neutral word replacement.

Conclusion

Thus, CheckList proposes a general framework to perform a comprehensive and fine-grained evaluation of NLP models. This can help us better understand the state of NLP models beyond the leaderboard.

References

Connect

If you enjoyed this blog post, feel free to connect with me on Twitter where I share new blog posts every week.

Discussion

pic
Editor guide