README.md

---
language:
- ar
pretty_name: TypePrediction
task_categories:
- token-classification
tags:
- named-entity-recognition
- arabic
- wojood
- entity-typing
size_categories:
- 100K<n<1M
---

# TypePrediction

Unified mention-level type prediction data built from all available local Wojood NER datasets except `WojoodRelations`.

## Main file

- `type_predictor_data.jsonl`
- `type_predictor_train.jsonl`
- `type_predictor_val.jsonl`
- `type_predictor_test.jsonl`
- `summary.json`
- `split_summary.json`
- `build_type_prediction_dataset.py`
- `push_to_hf.py`
- `split_data.py`
- `artifacts/ambiguous_span_type_conflicts.jsonl`
- `artifacts/split_assignments.jsonl`

Each row contains:

```json
{
  "id": "tp_000001",
  "entity": "النص",
  "type": "ORG",
  "sentence": "الجملة الكاملة",
  "start_token": 0,
  "end_token": 1,
  "start_char": 0,
  "end_char": 10,
  "raw_types": ["ORG", "NONGOV"],
  "sources": [
    {"dataset": "WojoodFine", "split": "train"},
    {"dataset": "Wojood1_1_nested", "split": "train"}
  ]
}
```

## Stored row examples

Simple flat/coarse example:

```json
{
  "id": "tp_000001",
  "entity": "عبد الحميد عبد العاطي",
  "type": "PERS",
  "sentence": "! ! ! تم النشر قبل بواسطة عبد الحميد عبد العاطي ازهقنا .",
  "start_token": 7,
  "end_token": 10,
  "start_char": 26,
  "end_char": 47,
  "raw_types": ["PERS"],
  "sources": [{"dataset": "WojoodFine", "split": "train"}]
}
```

Example where a fine label is mapped to a coarse label:

```json
{
  "id": "tp_001001",
  "entity": "للعراقيين",
  "type": "NORP",
  "sentence": ", أثارت محاكمات 11 مشتبها بهم فرنسيين في العراق في مايو / أيار هذه المخاوف ، ليس فقط للمشتبه بهم الأجانب بل للعراقيين أيضا .",
  "start_token": 22,
  "end_token": 22,
  "start_char": 108,
  "end_char": 117,
  "raw_types": ["NORP"],
  "sources": [{"dataset": "WojoodFine", "split": "train"}]
}
```

Example merged across multiple source datasets:

```json
{
  "id": "tp_050001",
  "entity": "مصر",
  "type": "GPE",
  "sentence": "بعدما أطاح به آية الله الخميني وبنظامه ، حيث توجه بداية إلى مصر قبل أن يستقر به الأمر في المغرب ، ,",
  "start_token": 12,
  "end_token": 12,
  "start_char": 60,
  "end_char": 63,
  "raw_types": ["GPE", "COUNTRY"],
  "sources": [
    {"dataset": "Wojood1_1_flat", "split": "test"},
    {"dataset": "Wojood1_1_nested", "split": "test"},
    {"dataset": "WojoodFine-Flat", "split": "split20"}
  ]
}
```

## Construction

1. Parse all sentence-separated CoNLL-style files from:
   - `Wojood1_1_flat`
   - `Wojood1_1_nested`
   - `WojoodFine`
   - `WojoodFine-Flat`
2. Extract every entity span, including nested spans when multiple BIO tags are present on the same token sequence.
3. Map fine labels to coarse labels using `wojood_ontology.json`.
4. Reconstruct full sentence text and both token and character spans.
5. Deduplicate by `(sentence, entity, type, start_token, end_token)` and merge provenance.
6. Remove exact span/type conflicts from the main file:
   - conflict key: `(sentence, entity, start_token, end_token, start_char, end_char)`
   - if the same exact span has more than one coarse type, all candidate rows for that span are written to `artifacts/ambiguous_span_type_conflicts.jsonl`
   - the main `type_predictor_data.jsonl` keeps only unambiguous rows
7. Assign deterministic ids in final sorted order: `tp_000001`, `tp_000002`, ...

## Parsing details and edge cases

- `Wojood1_1_flat` uses simple space-separated `token tag` rows.
- `Wojood1_1_nested` uses space-separated rows with multiple BIO tags on the same token when nested annotations exist.
- `WojoodFine` and `WojoodFine-Flat` use tab-separated rows.
- Some `WojoodFine` rows place multiple tags inside a single tab field, for example:
  - `نبيه\tI-OCC B-PERS\t\t\t`
  The builder normalizes this by splitting tag cells on internal whitespace.
- Some fine labels appear with alias spellings that differ from the ontology ids, for example:
  - `GPE-ORG` -> `GPE_ORG`
  - `Cluster` -> `CLUSTER`
  - `Path` -> `PATH`
  The builder canonicalizes these before mapping to coarse parents.
- Sentence boundaries are read from blank lines in the CoNLL files.
- Character spans are reconstructed after joining sentence tokens with single spaces.
- The main file is single-label by construction. Ambiguous multi-type spans are excluded rather than guessed.

## Summary

- total rows: `125996`
- unique sentences: `33933`
- raw extracted mentions before dedupe: `408035`
- rows merged by dedupe: `275272`
- ambiguous span groups excluded: `3383`
- candidate rows excluded by ambiguity filtering: `6767`

## Source breakdown

- `Wojood1_1_flat/test`: 14573 mentions from 6606 sentences
- `Wojood1_1_flat/train`: 49959 mentions from 23125 sentences
- `Wojood1_1_flat/val`: 7143 mentions from 3304 sentences
- `Wojood1_1_nested/test`: 18045 mentions from 6606 sentences
- `Wojood1_1_nested/train`: 62377 mentions from 23125 sentences
- `Wojood1_1_nested/val`: 8944 mentions from 3304 sentences
- `WojoodFine/test`: 27848 mentions from 5748 sentences
- `WojoodFine/train`: 96188 mentions from 19484 sentences
- `WojoodFine/val`: 13800 mentions from 2828 sentences
- `WojoodFine-Flat/split10`: 10859 mentions from 3304 sentences
- `WojoodFine-Flat/split20`: 22208 mentions from 6606 sentences
- `WojoodFine-Flat/split70`: 76091 mentions from 23125 sentences

## Top coarse types

- `GPE`: 26902
- `ORG`: 26593
- `DATE`: 19762
- `PERS`: 10712
- `NORP`: 10217
- `ORDINAL`: 7807
- `OCC`: 7696
- `EVENT`: 3769
- `CARDINAL`: 3724
- `LOC`: 2388
- `WEBSITE`: 1478
- `FAC`: 1296
- `LAW`: 904
- `TIME`: 871
- `MONEY`: 420
- `CURR`: 411
- `LANGUAGE`: 332
- `PERCENT`: 312
- `PRODUCT`: 190
- `QUANTITY`: 106

## Validation performed

- verified that all four non-relation Wojood sources were included
- verified nested extraction is enabled when multiple tags occur on one token
- verified fine-to-coarse mapping uses the repo's `wojood_ontology.json`
- verified `sentence[start_char:end_char] == entity` across the built file
- verified deduped rows merge provenance rather than silently dropping it
- verified no exact span in the main file has more than one coarse type
- verified ids are unique and deterministic

## Split files

The repo also includes a mention-level stratified split built from `type_predictor_data.jsonl`:

- `type_predictor_train.jsonl`
- `type_predictor_val.jsonl`
- `type_predictor_test.jsonl`
- `split_summary.json`
- `artifacts/split_assignments.jsonl`
- `split_data.py`

Split policy:

1. Start from the clean `type_predictor_data.jsonl`.
2. Perform a stratified `80/20` split by the `type` column with `random_state = 42`.
3. Split the holdout pool again, stratified by `type`, into equal halves.
4. Keep sentence overlap allowed across splits.
5. Preserve all original fields.
6. Add `evaluation_category` only to validation and test rows:
   - `unseen_sentence`
   - `seen_sentence_new_entity`

Current split sizes:

- train: `100,796`
- validation: `12,600`
- test: `12,600`

Current evaluation-category counts:

- validation `seen_sentence_new_entity`: `11,661`
- validation `unseen_sentence`: `939`
- test `seen_sentence_new_entity`: `11,606`
- test `unseen_sentence`: `994`