คุณลักษณะ N-Gram ใน scikit-learn

You could hand-roll n-grams, but scikit-learn builds them for you. The trick is one small argument on your vectorizer.

Every text vectorizer accepts ngram_range, a tuple of min and max sizes. It tells the vectorizer which n-grams to count.

Leave it alone and ngram_range is (1, 1), pure single words. That is the plain bag-of-words you already know.

Set ngram_range to (1, 2) and the vectorizer keeps single words and every adjacent pair, all in one feature space.

Want pairs alone? Use (2, 2). Now single words vanish and only bigrams survive as features.

After fitting, peek at the learned features with get_feature_names_out. You will see both words and joined phrases listed.

scikit-learn names each bigram by joining its words with a single space, like not good. That string becomes one column.

The same ngram_range works on TfidfVectorizer too. You get n-gram phrases that are also weighted by how distinctive they are.

With bigrams on, not good lands in its own column. Your classifier can finally learn that this pair signals a negative review.

Adding bigrams can multiply your columns dramatically. The matrix stays sparse, but the vocabulary grows fast.

To tame the blow-up, combine ngram_range with min_df. Dropping rare phrases keeps only the n-grams that repeat usefully.

Which ngram_range gives you single words plus adjacent pairs in one vectorizer?

One ngram_range tuple turns any vectorizer into an n-gram machine. Inspect features, then trim with min_df. Next you will choose the right range. 🎯