Custom pipeline components in NLP - Model Metrics & Evaluation

Metrics & Evaluation - Custom pipeline components

Which metric matters for Custom pipeline components and WHY

When building custom pipeline components in NLP, the key metrics depend on the task the component performs. For example, if the component classifies text, accuracy, precision, and recall matter to measure how well it predicts correct labels. If it extracts information, metrics like F1 score balance precision and recall to show overall quality. These metrics help us know if the component improves the pipeline or not.

Confusion matrix example for a classification component

      | Predicted Positive | Predicted Negative |
      |--------------------|--------------------|
      | True Positive (TP) = 50  | False Negative (FN) = 10 |
      | False Positive (FP) = 5  | True Negative (TN) = 35  |

      Total samples = 50 + 10 + 5 + 35 = 100

      Precision = TP / (TP + FP) = 50 / (50 + 5) = 0.91
      Recall = TP / (TP + FN) = 50 / (50 + 10) = 0.83
      F1 Score = 2 * (Precision * Recall) / (Precision + Recall) ≈ 0.87
    

Precision vs Recall tradeoff with examples

In custom NLP components, sometimes you want to catch as many correct cases as possible (high recall), even if some are wrong. For example, a component detecting sensitive info should find all instances (high recall) to avoid leaks.

Other times, you want to be very sure when the component says "yes" (high precision). For example, a spam detector should not mark good emails as spam, so precision is key.

Balancing precision and recall depends on the use case. The F1 score helps find a good middle ground.

What good vs bad metric values look like for custom pipeline components

• Good: Precision and recall above 0.8, showing the component finds most correct cases and makes few mistakes.
• Bad: Precision or recall below 0.5, meaning many wrong predictions or many missed cases.
• Accuracy: Can be misleading if classes are imbalanced. For example, 90% accuracy might be bad if the component misses all rare but important cases.

Common pitfalls in metrics for custom pipeline components

• Accuracy paradox: High accuracy but poor recall on rare classes.
• Data leakage: Training data accidentally includes test info, inflating metrics.
• Overfitting: Great metrics on training data but poor on new data.
• Ignoring class imbalance: Not using precision/recall or F1 when classes are uneven.

Self-check question

Your custom NLP component has 98% accuracy but only 12% recall on the important class. Is it good for production? Why or why not?

Answer: No, it is not good. The low recall means it misses most important cases, even though accuracy is high. This can cause serious problems if those cases matter. You should improve recall before using it in production.

Key Result

Precision, recall, and F1 score are key to evaluate custom NLP pipeline components because they show how well the component finds correct cases and avoids mistakes.

Practice

1. What is the main purpose of a custom pipeline component in an NLP pipeline?

easy

A. To store the processed documents in a database

B. To replace the entire NLP model with a new one

C. To visualize the text data in charts

D. To add your own processing steps that modify the document

Solution

1. Step 1: Understand the role of pipeline components

   Pipeline components process text step-by-step, modifying or analyzing it.

2. Step 2: Identify what custom components do

   Custom components let you add your own processing steps that change the document or add data.

3. Final Answer:

   To add your own processing steps that modify the document -> Option D

4. Quick Check:

   Custom pipeline components = add processing steps [OK]

Hint: Custom components add steps that change the document [OK]

Common Mistakes:

• Thinking custom components replace the whole model
• Confusing visualization with processing
• Assuming storage is part of pipeline components

2. Which of the following is the correct way to define a custom pipeline component function in Python?

easy

A. def custom_component(text): return text

B. def custom_component(doc): print(doc)

C. def custom_component(doc): return doc

D. def custom_component(): return None

Solution

1. Step 1: Recall the function signature for custom components

   Custom components take a doc object and return it after processing.

2. Step 2: Check each option

   def custom_component(doc): return doc matches the signature and returns the doc. Others either take wrong input or don't return doc.

3. Final Answer:

   def custom_component(doc): return doc -> Option C

4. Quick Check:

   Function takes doc and returns doc [OK]

Hint: Custom component functions take and return doc objects [OK]

Common Mistakes:

• Using text instead of doc as input
• Not returning the doc object
• Missing the doc parameter

3. Given this custom component code:

def add_custom_attr(doc):
    for token in doc:
        token._.is_custom = token.text.isalpha()
    return doc

nlp.add_pipe(add_custom_attr, last=True)

text = 'Hello 123!'
doc = nlp(text)
print([token._.is_custom for token in doc])

What will be the printed output?

medium

A. [True, True, False]

B. [True, False, False]

C. [True, False, True]

D. [False, False, False]

Solution

1. Step 1: Analyze the tokens in the text

   The text 'Hello 123!' splits into tokens: 'Hello', '123', '!'.

2. Step 2: Check the custom attribute logic

   For each token, isalpha() returns True if all characters are letters. 'Hello' is True, '123' and '!' are False.

3. Final Answer:

   [True, False, False] -> Option B

4. Quick Check:

   isalpha() per token = [True, False, False] [OK]

Hint: Check token text with isalpha() for True/False [OK]

Common Mistakes:

• Assuming punctuation is alpha
• Counting tokens incorrectly
• Forgetting to return doc

4. What is wrong with this custom pipeline component code?

def faulty_component(doc):
    for token in doc:
        token._.is_custom = token.text.isdigit()
    # Missing return statement

nlp.add_pipe(faulty_component, last=True)

medium

A. It does not return the doc object

B. It uses an invalid attribute name

C. It modifies tokens outside the loop

D. It should not be added to the pipeline

Solution

1. Step 1: Check the function structure

   The function loops over tokens and sets a custom attribute but does not return the doc.

2. Step 2: Recall pipeline component requirements

   Custom components must return the doc object to continue the pipeline correctly.

3. Final Answer:

   It does not return the doc object -> Option A

4. Quick Check:

   Missing return doc causes pipeline failure [OK]

Hint: Always return doc at end of custom component [OK]

Common Mistakes:

• Forgetting to return doc
• Using wrong attribute names without registration
• Adding component incorrectly

5. You want to create a custom pipeline component that counts how many tokens in a document are uppercase and stores this count as doc._.uppercase_count. Which of the following is the correct approach?

hard

A. Register a doc extension for 'uppercase_count', define a component that counts uppercase tokens, assign the count to doc._.uppercase_count, and return doc

B. Add a token extension for 'uppercase_count' and count uppercase tokens per token

C. Modify tokens in place without registering any extension and return doc

D. Create a new NLP model that outputs uppercase counts directly

Solution

1. Step 1: Understand extension registration

   To add a new attribute to doc._, you must register a doc extension first.

2. Step 2: Implement counting and assignment

   Count uppercase tokens in the component, assign the count to doc._.uppercase_count, then return doc.

3. Final Answer:

   Register a doc extension for 'uppercase_count', define a component that counts uppercase tokens, assign the count to doc._.uppercase_count, and return doc -> Option A

4. Quick Check:

   Doc extension + count + assign + return doc [OK]

Hint: Register doc extension before assigning custom doc attributes [OK]

Common Mistakes:

• Not registering the doc extension before use
• Using token extension for doc-level data
• Not returning doc at the end