---
license: mit
language:
- pt
tags:
- portuguese
- pt-PT
- european-portuguese
- lexicon
- ngram
- phrase-bank
- hospitality
- nlp
pretty_name: booteek pt-PT lexicon
size_categories:
- 10K<n<100K
task_categories:
- text-generation
- feature-extraction
---

# booteek pt-PT lexicon

A SQLite lexicon of European Portuguese (**pt-PT**) unigrams, bigrams (PMI-ranked),
phrases (3/4/5-grams), and diagnostic tokens (pt-PT vs pt-BR variants), derived from
the [`booteek-ai/fineweb2-bagaco2`](https://huggingface.co/datasets/booteek-ai/fineweb2-bagaco2)
corpus (mirror of [`duarteocarmo/fineweb2-bagaco2`](https://huggingface.co/datasets/duarteocarmo/fineweb2-bagaco2)).

Built for grounding LLM outputs in **European Portuguese** (not Brazilian) — Haiku,
GPT-4, Gemini all default to pt-BR phrasing when asked for "Portuguese" without aggressive
prompting. This lexicon provides a phrase bank + collocation index + diagnostic tokens that
applications can query at runtime to:

- Score the pt-PT confidence of generated text
- Detect pt-BR-isms (`você` vs `tu`, `ônibus` vs `autocarro`, `acreditar` vs `crer`, …)
- Inject collocations and phrases into prompts as few-shot grounding
- Surface under-served pt-PT search vocabulary for SEO/AEO

## Provenance

| Field | Value |
|---|---|
| Source corpus | [`booteek-ai/fineweb2-bagaco2`](https://huggingface.co/datasets/booteek-ai/fineweb2-bagaco2) (`all` config) |
| Upstream | [`duarteocarmo/fineweb2-bagaco2`](https://huggingface.co/datasets/duarteocarmo/fineweb2-bagaco2) → [`uonlp/CulturaX`](https://huggingface.co/datasets/uonlp/CulturaX) (pt split) filtered by [`duarteocarmo/fasttext-euptvid`](https://huggingface.co/duarteocarmo/fasttext-euptvid) |
| Docs processed | 1,000,000 |
| Tokens processed | ~390 million |
| Extraction date | 2026-05-13 |
| Extractor | `extract-pt-pt-corpus.py` v2026-05-13-v2-all-1M |
| Output size | ~17 MB |

## Schema

Five tables. Inspect provenance with `SELECT key, value FROM meta`.

### `unigrams`
```sql
CREATE TABLE unigrams (
  token TEXT PRIMARY KEY,   -- surface form, accents preserved
  count INTEGER NOT NULL,   -- corpus frequency
  doc_freq INTEGER NOT NULL,
  pt_pt_score REAL          -- avg ptpt_score from upstream
);
```

### `bigrams`
```sql
CREATE TABLE bigrams (
  token1 TEXT, token2 TEXT,
  count INTEGER NOT NULL,
  pmi REAL NOT NULL,        -- pointwise mutual information, ≥ 3.0
  PRIMARY KEY (token1, token2)
);
```

### `phrases`
```sql
CREATE TABLE phrases (
  phrase TEXT PRIMARY KEY,
  length INTEGER NOT NULL,  -- 3, 4, or 5
  count INTEGER NOT NULL,
  domain_hint TEXT          -- optional: food | service | team | NULL
);
```

### `diagnostic_tokens`
```sql
CREATE TABLE diagnostic_tokens (
  token TEXT PRIMARY KEY,
  variant TEXT NOT NULL,    -- 'pt-PT' | 'pt-BR'
  notes TEXT
);
```

Used at query-time to compute pt-PT confidence and detect pt-BR-isms.

### `meta`
Extraction provenance (source dataset, doc count, token count, extracted_at,
extractor_version, license, attribution chain).

## Usage

### Python
```python
import sqlite3
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="booteek-ai/pt-pt-lexicon",
    filename="pt-pt-corpus.sqlite",
    repo_type="dataset",
)
db = sqlite3.connect(path)

# Top collocations for "vinho"
rows = db.execute(
    "SELECT token2, count, pmi FROM bigrams WHERE token1=? ORDER BY pmi DESC LIMIT 5",
    ("vinho",),
).fetchall()
# → [('tinto', ...), ('branco', ...), ('verde', ...), ('porto', ...), ...]
```

### TypeScript (via better-sqlite3)
booteek consumes this lexicon via `src/lib/lexicon/pt-pt.ts`. Surface API:
```ts
import {
  getTopCollocations,
  getCommonPhrases,
  ptPtConfidence,
  detectPtBrIsms,
} from '@/lib/lexicon/pt-pt';

getTopCollocations('vinho', 5);
// → ['vinho tinto', 'vinho branco', 'vinho do porto', ...]

ptPtConfidence('Apanho o autocarro e ponho a comida no frigorífico.');
// → 0.95   (high pt-PT confidence)

detectPtBrIsms('Vou pegar o ônibus.');
// → [{ token: 'ônibus', variant: 'pt-BR', ptPtAlternative: 'autocarro' }]
```

## Re-extraction

The corpus is built from `booteek-ai/fineweb2-bagaco2`. See the extraction pipeline in
the [booteek monorepo](https://github.com/anthonyporto/booteek):
`scripts/lexicon/extract-pt-pt-corpus.py`.

```bash
# Full corpus extraction (1M docs, ~90 minutes against pre-cached parquet shards)
.venv/bin/python scripts/lexicon/extract-pt-pt-corpus.py \
    --config all \
    --max-docs 1000000 \
    --min-bigram-pmi 4.0
```

## License

**MIT** — preserved from the upstream
[`duarteocarmo/fineweb2-bagaco2`](https://huggingface.co/datasets/duarteocarmo/fineweb2-bagaco2)
dataset card.

## Attribution

Derived from the FineWeb2 Bagaço2 pt-PT slice maintained by
[**Duarte O. Carmo**](https://huggingface.co/duarteocarmo) (FOSDEM speaker, pt-PT NLP).
Upstream corpus: [CulturaX](https://huggingface.co/datasets/uonlp/CulturaX).
pt-PT classifier: [fasttext-euptvid](https://huggingface.co/duarteocarmo/fasttext-euptvid).

## Maintainer

[`booteek-ai`](https://huggingface.co/booteek-ai) on HuggingFace.
Booteek is Europe's first AI Visibility platform for independent restaurants and bars
([booteek.ai](https://booteek.ai)). The lexicon is open under MIT for any pt-PT
NLP work — not specific to hospitality.