README.md

---
language:
- en
license: apache-2.0
task_categories:
- text-generation
tags:
- crossword
- rebus
- nested-clues
- word-puzzles
pretty_name: Nested Crossword Clues
size_categories:
- 10K<n<100K
source_datasets:
- albertxu/CrosswordQA
- wikimedia/wikipedia
---

# Nested Crossword Clues

A dataset of English sentences with recursively nested crossword clue definitions. Each example starts from a natural sentence sampled from English Wikipedia, where content words are replaced by crossword-style definitions from [CrosswordQA](https://huggingface.co/datasets/albertxu/CrosswordQA). The definitions themselves are then recursively nested — words within clues are replaced by further clues — up to 10 levels deep.

## Example

**Original sentence (level 0):**

> He developed the theory of relativity.

**Level 1** (words replaced with crossword clues):

> He [Brought to maturity] the [Hypothesis] of [Family connection].

**Level 2** (words within clues replaced further):

> He [Brought to [Full growth]] the [[Educated guess]] of [Family [Link]].

Each example records the sentence at every nesting level, from the plain original to the maximally nested version.

## Dataset Fields

| Field | Type | Description |
|-------|------|-------------|
| `example_id` | `int` | Unique example identifier |
| `source_article_title` | `string` | Title of the Wikipedia article the sentence was sampled from |
| `original_sentence` | `string` | The original unmodified sentence (level 0) |
| `levels` | `list[string]` | The sentence at each nesting depth. `levels[0]` is the original, `levels[k]` has `k` layers of `[bracketed clue]` nesting |
| `max_nesting_depth` | `int` | The deepest nesting level achieved for this example |
| `num_replacements_per_level` | `list[int]` | Number of word-to-clue replacements made at each depth level |

## Construction

The dataset is built in four phases:

1. **Clue Index**: A lookup table is built from the [CrosswordQA](https://huggingface.co/datasets/albertxu/CrosswordQA) train split (~6.4M clue-answer pairs), mapping lowercase answers (up to 2 words) to their corresponding clue strings.

2. **Sentence Sampling**: Sentences are sampled from [English Wikipedia](https://huggingface.co/datasets/wikimedia/wikipedia) (November 2023 snapshot) via streaming and reservoir sampling. Sentences are filtered to 5–25 words, must end with sentence-ending punctuation, and must be free of wiki markup artifacts.

3. **Recursive Nesting**: For each sentence, content words are replaced with randomly chosen crossword clues from the index. The replacement is then applied recursively to words within the clues themselves, building a tree of nested definitions up to a maximum depth of 10. A max-gap constraint ensures definitions are densely distributed — if too many consecutive words go unreplaced, replacement is forced or the example is dropped.

4. **Anti-cycle Protection**: A set of already-replaced words is tracked across all nesting levels. Clues containing previously replaced words are rejected to prevent circular definitions.

### Filters and Design Choices

- **Stopwords skipped**: Common English function words (articles, prepositions, pronouns, auxiliaries) are left unreplaced to maintain readability.
- **Bigram matching**: Two-word answers from CrosswordQA are matched greedily before single words.
- **Max-gap constraint**: At the sentence level (depth 1), if 3+ consecutive content words go unreplaced, the next replacement is forced. If no valid clue exists, the entire example is dropped and a new sentence is sampled.
- **Min answer length**: Answers shorter than 3 characters are excluded from the index.
- **Max clue length**: Clues longer than 80 characters are excluded.
- **No brackets in clues**: Clues containing `[`, `]`, `(`, `)`, `{`, `}` are excluded to avoid ambiguity with the nesting notation.

## Source Datasets

- **[CrosswordQA](https://huggingface.co/datasets/albertxu/CrosswordQA)** — 6.4M crossword clue-answer pairs (train split)
- **[English Wikipedia](https://huggingface.co/datasets/wikimedia/wikipedia)** — November 2023 snapshot (`20231101.en` config)

## Generation Script

The dataset was generated using [`scripts/build_nested_clues.py`](https://github.com/gsarti/rebus-reasoning/blob/main/scripts/build_nested_clues.py) from the [rebus-reasoning](https://github.com/gsarti/rebus-reasoning) repository.

```bash
python scripts/build_nested_clues.py \
    --num-examples 100000 \
    --max-depth 10 \
    --seed 42 \
    --max-answer-words 2 \
    --max-gap 3 \
    --replacement-prob 0.8 \
    --articles-to-scan 50000 \
    --save-index output/clue_index.json \
    --output output/nested_clues.json
```

## License

Apache 2.0