Classic Text Processing and TF-IDF · AI Engineering

Tokenize, stem, lemmatize, count, weight with TF-IDF, classify sentiment and find topics with LDA or BERTopic.

Classic text pipeline

Tokenize split the string into word tokens. Keep punctuation as its own tokens.
Stem or lemmatize stem for speed. Lemmatize with a POS tag when meaning matters.
Negation scope prefix tokens after a negation word with NOT_ up to punctuation.
Count bag of words. One row per document, one column per vocabulary word.
Weight TF-IDF scales down common words. L2 rows make cosine a dot product.
Classify Naive Bayes for a baseline, logistic regression for production.
Evaluate per-class precision and recall, macro-F1, confusion matrix.

Sources

Tasks in code

Task	Code
Tokenize, stem and lemmatize	`preprocess(text, pos_tagger=demo_pos_tagger)`
Porter step 1a	`stem_step_1a(word)`
Table lemmatizer with fallback	`lemmatize(word, pos)`
Count matrix	`bag_of_words(docs, build_vocab(docs))`
TF-IDF rows on the unit sphere	`l2_normalize(tfidf(bow))`
Scope negation	`apply_negation(tokens)`
Naive Bayes baseline	`train_nb(docs_by_class, vocab, alpha=1.0)`
LDA by collapsed Gibbs sampling	`collapsed_gibbs_lda(docs, n_topics=3, n_iters=300, seed=42)`

Sources

NLTK: tokens, stems, lemmas

nltk.download("punkt_tab")

nltk.download("wordnet")

nltk.download("averaged_perceptron_tagger_eng")

tokens = word_tokenize(text)contractions, Unicode

stems = [PorterStemmer().stem(t) for t in tokens] # all five phases

tagged = pos_tag(tokens)Penn Treebank tags

lemmatizer = WordNetLemmatizer()

lemmas = [lemmatizer.lemmatize(t, nltk_pos_to_wordnet(tag)) for t, tag in tagged]V>v J>a R>r else n

Sources

phases/05-nlp-foundations-to-advanced/01-text-processing/docs/en.md

Sentiment baseline in scikit-learn

pipe = Pipeline([

("tfidf", TfidfVectorizer(

ngram_range=(1, 2),bigrams, so not_good is a feature

min_df=2,

sublinear_tf=True,dampens repeated words

stop_words=None)),keeps negations

("clf", LogisticRegression(C=1.0, max_iter=1000)),

])

pipe.fit(X_train, y_train)

print(pipe.score(X_test, y_test))75% to 85% on SST-2 with the 3 flags

Sources

phases/05-nlp-foundations-to-advanced/05-sentiment-analysis/docs/en.md

Vectorizer knobs

Command	What it does
`CountVectorizer()`	Tokenization, vocabulary and BoW in one call
`TfidfVectorizer()`	Adds IDF weighting and L2 normalization
`ngram_range=(1, 2)`	Includes bigrams. Usually boosts classification
`min_df=2`	Drops words in fewer than 2 docs
`max_df=0.95`	Drops words in more than 95% of docs
`stop_words="english"`	Builtin stopword list. Not for sentiment
`sublinear_tf=True`	Uses `1 + log(tf)` instead of raw `tf`
`get_feature_names_out()`	Lists the vocabulary in column order

Sources

phases/05-nlp-foundations-to-advanced/02-bag-of-words-tfidf/docs/en.md

Topic models

cv = CountVectorizer(max_features=max_features, stop_words="english",

min_df=2, max_df=0.9)LDA needs raw counts

X = cv.fit_transform(documents)

lda = LatentDirichletAllocation(n_components=n_topics, random_state=42,

max_iter=50, learning_method="online")

doc_topic = lda.fit_transform(X)(n_docs, n_topics), rows sum to 1

topic_model = BERTopic(

embedding_model="sentence-transformers/all-MiniLM-L6-v2",

min_topic_size=15)library default is 10

topics, probs = topic_model.fit_transform(documents)

info = topic_model.get_topic_info()Topic -1 is the outlier bucket

Sources

phases/05-nlp-foundations-to-advanced/15-topic-modeling/docs/en.md

Formulas

TF-IDF: \frac{\mathrm{count}(w \text{ in } d)}{|d|} \cdot \log\frac{N}{\mathrm{df}(w)}
Smoothed IDF: \log\frac{n + 1}{d + 1} + 1. A word in every doc keeps IDF 1.
L2 row: divide each row by its norm. Long docs stop dominating.
Naive Bayes: log prior plus the sum of log P(\text{word} \mid \text{class}). Take the max.
Laplace: (\mathrm{count} + \alpha) / (\mathrm{total} + \alpha |V|). No zero probabilities.
Logistic regression: sigmoid of Xw + b. Predict 1 at 0.5 or more.
Macro-F1: mean of per-class F1, equally weighted.
c_v coherence: NPMI of top-word pairs, compared by cosine. Higher is better.

Sources

Defaults

Setting	Default	Note
NB `alpha`	1.0	Teaching default. 0.01 is common in practice
LR `l2`	0.01	Start here and tune. Sparse features overfit
LR `epochs`, `lr`	500	Learning rate 0.05
LR `C`	1.0	With `max_iter=1000` in scikit-learn
BERTopic `min_topic_size`	10	Use 50 or 100 over 10,000 docs
Gibbs `alpha`, `beta`	0.1	Beta is 0.01. Both must be positive
Gibbs `n_iters`	200	The demo runs 300
Topic count	`max(5, round(sqrt(n_docs)))`	Clamp to 200 under 40,000 docs

Sources

Troubleshooting

Symptom	Cause	Fix
Accuracy drops after deploy	Training and inference preprocess differ	Ship one preprocess function in the model
Tokens change after an upgrade	Library version drift	Pin versions. Freeze 20 sample sentences
`ponies` stems to `poni`	Step 1a `ies` rule only	Use full Porter with step 1b
`watched` stays `watched`	Lemma table has no `ed` rule	Use WordNet or spaCy
"not good" reads positive	Stopword list drops negations	`stop_words=None` plus bigrams
High score, bad minority class	Micro-F1 or accuracy on imbalance	Report macro-F1, AUPRC over 95-5
LDA junk topics	Stopwords and frequent terms	Set `min_df` and `max_df`
Many docs in topic -1	HDBSCAN outlier bucket	Check `min_topic_size`, read the -1 docs

Sources

Use X when

Stem or lemmatize?: Stem for search indexing. Lemmatize for text the user reads.
NLTK or spaCy?: NLTK to teach and swap parts. spaCy for production speed.
TF-IDF or embeddings?: TF-IDF for spam, logs, low data and explainable models.
When a transformer?: Sarcasm, aspects, sentiment shifts and low-resource languages.
LDA or BERTopic?: LDA for long docs with topic mixtures. BERTopic for short text.
No GPU?: Use LDA or NMF.

Sources

Tips

Stay sparse. Dense TF-IDF for 100k documents does not fit in memory.
Read 5 wrong predictions per class before you trust a metric.
Regenerate the baseline on your own pipeline, not the paper's number.

Sources

Original source

---
id: text-processing-classic
title: Classic Text Processing and TF-IDF
code: CT
subtitle: Tokenize, stem, lemmatize, count, weight with TF-IDF, classify sentiment and find topics with LDA or BERTopic.
hue: 30
sources: phases/05-nlp-foundations-to-advanced/01-text-processing, phases/05-nlp-foundations-to-advanced/02-bag-of-words-tfidf, phases/05-nlp-foundations-to-advanced/05-sentiment-analysis, phases/05-nlp-foundations-to-advanced/15-topic-modeling
source_hash: 8993666a55a0debd
---

## Classic text pipeline
type: flow
from: phases/05-nlp-foundations-to-advanced/01-text-processing/docs/en.md, phases/05-nlp-foundations-to-advanced/02-bag-of-words-tfidf/docs/en.md, phases/05-nlp-foundations-to-advanced/05-sentiment-analysis/docs/en.md

1. **Tokenize**: split the string into word tokens. Keep punctuation as its own tokens.
2. **Stem or lemmatize**: stem for speed. Lemmatize with a POS tag when meaning matters.
3. **Negation scope**: prefix tokens after a negation word with `NOT_` up to punctuation.
4. **Count**: bag of words. One row per document, one column per vocabulary word.
5. **Weight**: TF-IDF scales down common words. L2 rows make cosine a dot product.
6. **Classify**: Naive Bayes for a baseline, logistic regression for production.
7. **Evaluate**: per-class precision and recall, macro-F1, confusion matrix.

## Tasks in code
type: task-table
span: 2
from: phases/05-nlp-foundations-to-advanced/01-text-processing/code/main.py, phases/05-nlp-foundations-to-advanced/02-bag-of-words-tfidf/code/main.py, phases/05-nlp-foundations-to-advanced/05-sentiment-analysis/docs/en.md, phases/05-nlp-foundations-to-advanced/15-topic-modeling/code/main.py

| Task | Code |
|---|---|
| Tokenize, stem and lemmatize | `preprocess(text, pos_tagger=demo_pos_tagger)` |
| Porter step 1a | `stem_step_1a(word)` |
| Table lemmatizer with fallback | `lemmatize(word, pos)` |
| Count matrix | `bag_of_words(docs, build_vocab(docs))` |
| TF-IDF rows on the unit sphere | `l2_normalize(tfidf(bow))` |
| Scope negation | `apply_negation(tokens)` |
| Naive Bayes baseline | `train_nb(docs_by_class, vocab, alpha=1.0)` |
| LDA by collapsed Gibbs sampling | `collapsed_gibbs_lda(docs, n_topics=3, n_iters=300, seed=42)` |

## NLTK: tokens, stems, lemmas
type: annotated-code
lang: python
from: phases/05-nlp-foundations-to-advanced/01-text-processing/docs/en.md

```python
nltk.download("punkt_tab")
nltk.download("wordnet")
nltk.download("averaged_perceptron_tagger_eng")
tokens = word_tokenize(text)                      # contractions, Unicode
stems = [PorterStemmer().stem(t) for t in tokens] # all five phases
tagged = pos_tag(tokens)                          # Penn Treebank tags
lemmatizer = WordNetLemmatizer()
lemmas = [lemmatizer.lemmatize(t, nltk_pos_to_wordnet(tag)) for t, tag in tagged]  # V>v J>a R>r else n
```

## Sentiment baseline in scikit-learn
type: annotated-code
lang: python
from: phases/05-nlp-foundations-to-advanced/05-sentiment-analysis/docs/en.md

```python
pipe = Pipeline([
    ("tfidf", TfidfVectorizer(
        ngram_range=(1, 2),       # bigrams, so not_good is a feature
        min_df=2,
        sublinear_tf=True,        # dampens repeated words
        stop_words=None)),        # keeps negations
    ("clf", LogisticRegression(C=1.0, max_iter=1000)),
])
pipe.fit(X_train, y_train)
print(pipe.score(X_test, y_test))  # 75% to 85% on SST-2 with the 3 flags
```

## Vectorizer knobs
type: command-table
from: phases/05-nlp-foundations-to-advanced/02-bag-of-words-tfidf/docs/en.md

| Command | What it does |
|---|---|
| `CountVectorizer()` | Tokenization, vocabulary and BoW in one call |
| `TfidfVectorizer()` | Adds IDF weighting and L2 normalization |
| `ngram_range=(1, 2)` | Includes bigrams. Usually boosts classification |
| `min_df=2` | Drops words in fewer than 2 docs |
| `max_df=0.95` | Drops words in more than 95% of docs |
| `stop_words="english"` | Builtin stopword list. Not for sentiment |
| `sublinear_tf=True` | Uses `1 + log(tf)` instead of raw `tf` |
| `get_feature_names_out()` | Lists the vocabulary in column order |

## Topic models
type: annotated-code
lang: python
from: phases/05-nlp-foundations-to-advanced/15-topic-modeling/docs/en.md

```python
cv = CountVectorizer(max_features=max_features, stop_words="english",
                     min_df=2, max_df=0.9)   # LDA needs raw counts
X = cv.fit_transform(documents)
lda = LatentDirichletAllocation(n_components=n_topics, random_state=42,
                                max_iter=50, learning_method="online")
doc_topic = lda.fit_transform(X)           # (n_docs, n_topics), rows sum to 1
topic_model = BERTopic(
    embedding_model="sentence-transformers/all-MiniLM-L6-v2",
    min_topic_size=15)                     # library default is 10
topics, probs = topic_model.fit_transform(documents)
info = topic_model.get_topic_info()       # Topic -1 is the outlier bucket
```

## Formulas
type: concepts
from: phases/05-nlp-foundations-to-advanced/02-bag-of-words-tfidf/docs/en.md, phases/05-nlp-foundations-to-advanced/05-sentiment-analysis/docs/en.md, phases/05-nlp-foundations-to-advanced/15-topic-modeling/docs/en.md

- **TF-IDF**: $\frac{\mathrm{count}(w \text{ in } d)}{|d|} \cdot \log\frac{N}{\mathrm{df}(w)}$
- **Smoothed IDF**: $\log\frac{n + 1}{d + 1} + 1$. A word in every doc keeps IDF 1.
- **L2 row**: divide each row by its norm. Long docs stop dominating.
- **Naive Bayes**: log prior plus the sum of log $P(\text{word} \mid \text{class})$. Take the max.
- **Laplace**: $(\mathrm{count} + \alpha) / (\mathrm{total} + \alpha |V|)$. No zero probabilities.
- **Logistic regression**: sigmoid of $Xw + b$. Predict 1 at 0.5 or more.
- **Macro-F1**: mean of per-class F1, equally weighted.
- **c_v coherence**: NPMI of top-word pairs, compared by cosine. Higher is better.

## Defaults
type: defaults
from: phases/05-nlp-foundations-to-advanced/05-sentiment-analysis/docs/en.md, phases/05-nlp-foundations-to-advanced/15-topic-modeling/docs/en.md, phases/05-nlp-foundations-to-advanced/15-topic-modeling/code/main.py

| Setting | Default | Note |
|---|---|---|
| NB `alpha` | 1.0 | Teaching default. 0.01 is common in practice |
| LR `l2` | 0.01 | Start here and tune. Sparse features overfit |
| LR `epochs`, `lr` | 500 | Learning rate 0.05 |
| LR `C` | 1.0 | With `max_iter=1000` in scikit-learn |
| BERTopic `min_topic_size` | 10 | Use 50 or 100 over 10,000 docs |
| Gibbs `alpha`, `beta` | 0.1 | Beta is 0.01. Both must be positive |
| Gibbs `n_iters` | 200 | The demo runs 300 |
| Topic count | `max(5, round(sqrt(n_docs)))` | Clamp to 200 under 40,000 docs |

## Troubleshooting
type: troubleshooting
from: phases/05-nlp-foundations-to-advanced/01-text-processing/docs/en.md, phases/05-nlp-foundations-to-advanced/02-bag-of-words-tfidf/docs/en.md, phases/05-nlp-foundations-to-advanced/05-sentiment-analysis/docs/en.md, phases/05-nlp-foundations-to-advanced/15-topic-modeling/docs/en.md

| Symptom | Cause | Fix |
|---|---|---|
| Accuracy drops after deploy | Training and inference preprocess differ | Ship one preprocess function in the model |
| Tokens change after an upgrade | Library version drift | Pin versions. Freeze 20 sample sentences |
| `ponies` stems to `poni` | Step 1a `ies` rule only | Use full Porter with step 1b |
| `watched` stays `watched` | Lemma table has no `ed` rule | Use WordNet or spaCy |
| "not good" reads positive | Stopword list drops negations | `stop_words=None` plus bigrams |
| High score, bad minority class | Micro-F1 or accuracy on imbalance | Report macro-F1, AUPRC over 95-5 |
| LDA junk topics | Stopwords and frequent terms | Set `min_df` and `max_df` |
| Many docs in topic -1 | HDBSCAN outlier bucket | Check `min_topic_size`, read the -1 docs |

## Use X when
type: qa
from: phases/05-nlp-foundations-to-advanced/01-text-processing/docs/en.md, phases/05-nlp-foundations-to-advanced/02-bag-of-words-tfidf/docs/en.md, phases/05-nlp-foundations-to-advanced/05-sentiment-analysis/docs/en.md, phases/05-nlp-foundations-to-advanced/15-topic-modeling/docs/en.md

- **Stem or lemmatize?** Stem for search indexing. Lemmatize for text the user reads.
- **NLTK or spaCy?** NLTK to teach and swap parts. spaCy for production speed.
- **TF-IDF or embeddings?** TF-IDF for spam, logs, low data and explainable models.
- **When a transformer?** Sarcasm, aspects, sentiment shifts and low-resource languages.
- **LDA or BERTopic?** LDA for long docs with topic mixtures. BERTopic for short text.
- **No GPU?** Use LDA or NMF.

## Tips
type: tips
from: phases/05-nlp-foundations-to-advanced/02-bag-of-words-tfidf/docs/en.md, phases/05-nlp-foundations-to-advanced/05-sentiment-analysis/docs/en.md, phases/05-nlp-foundations-to-advanced/15-topic-modeling/docs/en.md

- Stay sparse. Dense TF-IDF for 100k documents does not fit in memory.
- Read 5 wrong predictions per class before you trust a metric.
- Regenerate the baseline on your own pipeline, not the paper's number.