Custom pipeline components in NLP - Practice Problems & Coding Challenges

Challenge - 5 Problems

❓ Predict Output

intermediate

Output of a simple custom pipeline component

What will be the output of the following code that adds a custom component to a spaCy pipeline which counts tokens?

NLP

import spacy
from spacy.language import Language

@Language.component('token_counter')
def token_counter(doc):
    doc._.token_count = len(doc)
    return doc

nlp = spacy.blank('en')

# Register extension attribute
from spacy.tokens import Doc
Doc.set_extension('token_count', default=0)

nlp.add_pipe('token_counter')
doc = nlp('Hello world! This is a test.')
print(doc._.token_count)

A SyntaxError

B 5

C 6

D 8

💡 Hint

Count the number of tokens in the sentence including punctuation.

The sentence 'Hello world! This is a test.' has 8 tokens: 'Hello', 'world', '!', 'This', 'is', 'a', 'test', '.'. The code counts len(doc) which is 8 tokens, so the correct output is 8.

❓ Model Choice

intermediate

Choosing the right custom pipeline component for sentiment analysis

You want to add a custom pipeline component that assigns a sentiment score to each document. Which component design is best?

A A component that deletes tokens from the Doc to keep only positive words.

B A component that modifies the Doc object by adding a custom attribute with the sentiment score.

C A component that returns None instead of a Doc object.

D A component that replaces the Doc object with a string containing the sentiment score.

💡 Hint

Remember that pipeline components must return a Doc object.

Custom components should add information to the Doc without breaking the pipeline. Returning None or replacing the Doc with a string breaks the pipeline. Deleting tokens is destructive and not recommended for sentiment scoring.

❓ Hyperparameter

advanced

Setting hyperparameters in a custom spaCy pipeline component

You want to create a custom pipeline component that filters tokens by a minimum length parameter. How should you pass this parameter to the component?

A Define the component as a factory function that accepts the parameter and returns the actual component function.

B Hardcode the minimum length inside the component function without parameters.

C Pass the parameter as a global variable outside the component.

D Set the parameter inside the Doc object before processing.

💡 Hint

Think about how spaCy components are registered and initialized.

spaCy components can be created as factory functions that accept parameters and return the component function. This allows flexible configuration when adding the component to the pipeline.

🔧 Debug

advanced

Debugging a custom pipeline component that raises an error

Consider this custom component code snippet: @Language.component('uppercase_tokens') def uppercase_tokens(doc): for token in doc: token.text = token.text.upper() return doc Why does this code raise an error when added to the pipeline?

A Because token.text is a read-only property and cannot be assigned to.

B Because the function does not return a Doc object.

C Because the pipeline does not support loops over tokens.

D Because the component name is invalid.

💡 Hint

Check if token.text can be changed directly.

In spaCy, token.text is read-only. You cannot modify it directly. To change token text, you must create a new Doc or use other methods.

🧠 Conceptual

expert

Understanding the order of custom pipeline components

You have two custom components: one that lemmatizes tokens and another that filters out stop words. Which order should you add them to the pipeline for best results?

A Order does not matter; both can be added in any sequence.

B Add the stop word filter first, then the lemmatizer.

C Add the lemmatizer first, then the stop word filter.

D Add both components simultaneously using add_pipe with the same name.

💡 Hint

Think about how lemmatization affects token forms before filtering.

Lemmatization should happen before filtering stop words because filtering usually depends on the lemma or normalized form. Filtering first might remove tokens that could be lemmatized differently.