Introduction

One CountVectorizer instance can use only one analyzer at a time. If you want both word-level and character-level features from the same text, the normal solution is to build two vectorizers and combine their outputs.

Why One Vectorizer Is Not Enough

CountVectorizer has an analyzer parameter, but it accepts one mode per instance:

• 'word'
• 'char'
• 'char_wb'

That means this is not possible in one object:

CountVectorizer(analyzer=["word", "char"])

Instead, create one vectorizer for words and another for character n-grams.

Combine Two Vectorizers with FeatureUnion

Scikit-learn's FeatureUnion is the standard tool for combining multiple feature extractors that operate on the same input.

1from sklearn.feature_extraction.text import CountVectorizer
2from sklearn.pipeline import FeatureUnion
3
4texts = [
5    "The cat sat on the mat",
6    "The dog chased the cat",
7    "Cats and dogs are pets",
8]
9
10features = FeatureUnion([
11    ("word", CountVectorizer(analyzer="word", ngram_range=(1, 2))),
12    ("char", CountVectorizer(analyzer="char", ngram_range=(3, 5))),
13])
14
15X = features.fit_transform(texts)
16print(X.shape)

The result is one sparse matrix whose columns include both word n-grams and character n-grams.

Why This Is Useful

Word features capture meaning at the token level, while character features capture spelling patterns, prefixes, suffixes, and resilience to small typos.

That combination is often helpful when:

• text contains misspellings or noisy user input
• morphology matters
• you want both semantic and subword information

It is a common pattern in text classification tasks.

Add the Combined Features to a Classifier Pipeline

You can feed the union directly into a classifier:

1from sklearn.feature_extraction.text import CountVectorizer
2from sklearn.linear_model import LogisticRegression
3from sklearn.pipeline import FeatureUnion, Pipeline
4
5pipeline = Pipeline([
6    ("features", FeatureUnion([
7        ("word", CountVectorizer(analyzer="word", ngram_range=(1, 2))),
8        ("char", CountVectorizer(analyzer="char", ngram_range=(3, 5))),
9    ])),
10    ("clf", LogisticRegression(max_iter=1000)),
11])
12
13labels = [0, 1, 1]
14pipeline.fit(texts, labels)
15print(pipeline.predict(["dog on mat"]))

This keeps the preprocessing and model together in one reusable object.

Consider char_wb as a Variation

If plain character n-grams feel too noisy, char_wb can be a good compromise. It creates character n-grams inside word boundaries, which often reduces cross-word artifacts.

You can swap the character branch easily:

CountVectorizer(analyzer="char_wb", ngram_range=(3, 5))

That small change can matter in short-text classification or language-identification tasks.

It is especially useful when you want character features that stay inside individual words instead of creating n-grams that cross whitespace boundaries.

That often produces a cleaner feature space for short texts, product names, and typo-prone user input.

Watch the Feature Space Size

Combining word and character analyzers increases dimensionality quickly. That is not automatically bad, but it does mean:

• more memory usage
• larger sparse matrices
• potentially slower model training

Practical controls include min_df, max_features, and narrower n-gram ranges.

Common Pitfalls

The biggest pitfall is trying to force both analyzers into one CountVectorizer. The API is not designed that way.

Another issue is converting the combined sparse matrix to dense form with .toarray() on large datasets. That can explode memory usage immediately.

People also underestimate how large character n-gram vocabularies become. Start with a narrow range and tune from there instead of enabling everything at once.

Summary

• One CountVectorizer can use only one analyzer setting at a time.
• To use both word and character features, build two vectorizers and combine them with FeatureUnion.
• The combined output is a single sparse matrix suitable for normal scikit-learn pipelines.
• 'char_wb is worth considering when raw character n-grams are too noisy.'
• Control vocabulary growth with sensible n-gram ranges and feature limits.