nlp-domain-wordpiece-tfidf

Trains a WordPiece tokenizer on in-domain text, then feeds its subword token IDs into TF-IDF vectorization for domain-adapted sparse features.

SKILL.md

GitHub

Domain WordPiece TF-IDF

Overview

Standard TF-IDF uses whitespace or regex tokenization, which misses subword patterns important in domain-specific text (misspellings, slang, coded language in toxic text). Training a WordPiece tokenizer on the domain corpus learns meaningful subword units, then feeding these token IDs into TF-IDF creates sparse features that capture domain-specific vocabulary at the subword level. This bridges the gap between neural tokenizers and classical ML — you get BPE-quality tokenization with TF-IDF + Ridge regression speed.

Quick Start

from tokenizers import Tokenizer, models, normalizers, pre_tokenizers, trainers
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.linear_model import Ridge

# Train WordPiece tokenizer on domain text
tokenizer = Tokenizer(models.WordPiece(unk_token="[UNK]"))
tokenizer.normalizer = normalizers.BertNormalizer(lowercase=True)
tokenizer.pre_tokenizer = pre_tokenizers.BertPreTokenizer()
trainer = trainers.WordPieceTrainer(vocab_size=25000)

def corpus_iter():
    for text in df['text']:
        yield text

tokenizer.train_from_iterator(corpus_iter(), trainer=trainer)

# Tokenize all texts
tokenized = [tokenizer.encode(t).ids for t in df['text']]

# Feed into TF-IDF with identity tokenizer (already tokenized)
identity = lambda x: x
vectorizer = TfidfVectorizer(
    analyzer='word', tokenizer=identity,
    preprocessor=identity, token_pattern=None)
X = vectorizer.fit_transform(tokenized)

model = Ridge(alpha=0.8)
model.fit(X, df['score'])

Workflow

  1. Train WordPiece tokenizer on all available text (train + test)
  2. Tokenize texts to get lists of token IDs
  3. Pass token ID lists to TfidfVectorizer with identity tokenizer/preprocessor
  4. Train a linear model (Ridge, SVM) on the sparse TF-IDF matrix

Key Decisions

  • vocab_size: 20k-30k; larger captures more subwords but increases sparsity
  • vs BPE: WordPiece and BPE produce similar subwords; WordPiece is simpler to train via HuggingFace tokenizers
  • Why not just transformers: 100x faster training, no GPU needed, competitive for simple ranking tasks
  • Train on all text: Include test text in tokenizer training (not labels) for better vocabulary coverage

References