견고성을 위한 문자 N-그램

Word-level features break on typos and rare spellings. The model sees runnning as a brand-new word, so a single slip can lose all its signal.

Instead of whole words, we slice text into short chunks of letters. These overlapping pieces are called character n-grams.

For the word cat, character 3-grams are simple letter triples. They slide across the text one position at a time to form overlapping chunks.

The words run and runn share most of their character grams. So a typo changes only a few features instead of erasing the whole word.

play, playing, and player all share the gram pla. Character grams quietly link related forms without any stemming step.

Social media, usernames, and product codes are full of odd spellings. Character grams stay robust where word features fall apart.

scikit-learn makes this one parameter. Set analyzer to char and pick an ngram_range to extract character grams instead of words.

The char_wb analyzer only makes grams inside word boundaries, padding each word. It keeps the robustness while avoiding grams that span two words.

Languages that glue words together or have rich endings are tough for word tokenizers. Character grams sidestep that and stay language-friendly.

Character grams create many more features than words do. That extra dimensionality needs more memory, so keep the range tight.

Character grams shine when combined with word features, not as a replacement. Together they cover both meaning and robustness. 💪

Why do character n-grams handle typos so well?

Character n-grams slice text into letter chunks, survive typos, and link word families. Use analyzer char in scikit-learn alongside word features. ✅