README.md

---
license: cc-by-4.0
pretty_name: 'Leipzig Corpora Frequency Data'
tags:
  - language
  - frequency
  - corpus
  - linguistics
  - nlp
configs:
  - config_name: default
    data_files: 'base/**/*.parquet'
    default: true
---

# Leipzig Corpora Frequency Data

Word frequency lists and co-occurrence data from the Leipzig Corpora
Collection, converted to Parquet.

Covers hundreds of languages across news, web, Wikipedia, and mixed
sources. Each corpus includes token frequencies, source provenance,
and statistical co-occurrence pairs.

## Contents

```
base/
  <language>/
    <source>-<date>-<size>/
      metadata.json
      string.0001.parquet
      source.0001.parquet
      cooccurrence.sentence.0001.parquet
      cooccurrence.neighbor.0001.parquet
```

Shards are split at ~200-400 MB each. Small corpora may have a single
shard. Large corpora have multiple (`0001.parquet`, `0002.parquet`,
etc.).

Example: `base/afr/news-2020-30K/string.0001.parquet`

Languages use ISO 639-3 codes, sometimes with region suffixes (e.g.
`ara-eg` for Egyptian Arabic).

## Files per corpus

| File                                 | Format  | Description                                          |
| ------------------------------------ | ------- | ---------------------------------------------------- |
| `metadata.json`                      | JSON    | Language, source type, date, size, original filename |
| `string.NNNN.parquet`                | Parquet | Token frequency list (words and punctuation)         |
| `source.NNNN.parquet`                | Parquet | Source article URLs and dates                        |
| `cooccurrence.sentence.NNNN.parquet` | Parquet | Word pairs appearing in the same sentence            |
| `cooccurrence.neighbor.NNNN.parquet` | Parquet | Word pairs appearing adjacent to each other          |

"Strings" instead of "words" because the list includes punctuation,
special characters, and other non-word tokens alongside actual words.

Parquet files use ZSTD compression for ~4x smaller size than equivalent
JSONL, with column-wise reads for fast filtering.

## Usage

```python
from datasets import load_dataset

ds = load_dataset("cluesurf/leipzig-frequency")
```

Or query directly with DuckDB:

```sql
SELECT text, frequency
FROM 'base/afr/news-2020-30K/string.*.parquet'
ORDER BY frequency DESC
LIMIT 20;
```

## Record schemas

### metadata.json

```json
{
  "language": "afr",
  "source": "news",
  "date": "2020",
  "size": "30K",
  "file": "afr_news_2020_30K"
}
```

### string.NNNN.parquet

| Column    | Type   |
| --------- | ------ |
| id        | int32  |
| text      | string |
| frequency | int64  |

Example row: `{ id: 101, text: "die", frequency: 30994 }`

### source.NNNN.parquet

| Column | Type   |
| ------ | ------ |
| id     | int32  |
| url    | string |
| date   | string |

Example row: `{ id: 1, url: "https://carletonvilleherald.com/...", date: "2020-05-17" }`

### cooccurrence.sentence.NNNN.parquet

| Column       | Type    |
| ------------ | ------- |
| string_1_id  | int32   |
| string_2_id  | int32   |
| frequency    | int64   |
| significance | float64 |

Example row: `{ string_1_id: 116, string_2_id: 4688, frequency: 5, significance: 7.61 }`

### cooccurrence.neighbor.NNNN.parquet

Same schema as `cooccurrence.sentence`, but for words appearing
adjacent to each other rather than in the same sentence.

## Source

Downloaded from the
[Leipzig Corpora Collection](https://wortschatz.uni-leipzig.de/en/download)
at the University of Leipzig.

Original archives are `.tar.gz` files containing tab-delimited `.txt`
data following the Wortschatz database schema.

## Sources

- [Leipzig Corpora Collection](https://wortschatz.uni-leipzig.de/en/download)
- [Wortschatz project](https://wortschatz.uni-leipzig.de)
- D. Goldhahn, T. Eckart, U. Quasthoff: Building Large Monolingual
  Dictionaries at the Leipzig Corpora Collection: From 100 to 200
  Languages. In: Proceedings of LREC, 2012.

## License

CC-BY-4.0, as specified by the Leipzig Corpora Collection.