OpenVoiceOS/yes_no_answers

---
language:
- en
- de
- fr
- es
- it
- pt
- ru
- uk
- pl
- nl
- sv
- da
- fi
- nb
- nn
- cs
- sk
- ro
- hr
- sl
- hu
- bg
- el
- ca
- lt
- lv
- et
- eu
- gl
- is
- an
- ja
- ko
- zh
- ar
- he
- fa
- tr
- id
- ms
- fil
- vi
- th
license: apache-2.0
task_categories:
- text-classification
task_ids:
- intent-classification
tags:
- yes-no
- dialogue
- multilingual
- agreement
- intent
size_categories:
- 10K<n<100K
---

# Yes/No Multilingual Answers Dataset

A dataset of **10,709** conversational utterances for classifying yes/no/ambiguous responses across **43 languages**.

## Dataset Description

Each sample is a natural language utterance a person might say in response to a yes/no question. The dataset covers three classes:

| Label | Description |
|-------|-------------|
| `yes` | Affirmation, agreement, or confirmation |
| `no` | Negation, refusal, or disagreement |
| `None` | Genuinely ambiguous — cannot be resolved without context |

### Schema

```
utterance,agreement,subtype,language
"ja","yes","Y1","de"
"absolument pas","no","N2","fr"
"peut-être","None","C1","fr"
```

## Statistics

| Metric | Value |
|--------|-------|
| Total samples | 10,709 |
| Languages | 43 |
| Samples per language | 224–290 (avg 249) |
| Label: yes | 3,873 (36.2%) |
| Label: no | 3,826 (35.7%) |
| Label: None | 3,010 (28.1%) |
| Semantic subtypes | 28 |
| Min samples per subtype per language | 8 |

## Languages

**European:** English · German · French · Spanish · Italian · Portuguese · Russian · Ukrainian · Polish · Dutch · Swedish · Danish · Finnish · Norwegian Bokmål · Norwegian Nynorsk · Czech · Slovak · Romanian · Croatian · Slovenian · Hungarian · Bulgarian · Greek · Catalan · Lithuanian · Latvian · Estonian · Basque · Galician · Icelandic · Aragonese

**Asian & Middle Eastern:** Japanese · Korean · Chinese · Arabic · Hebrew · Persian · Turkish · Indonesian · Malay · Filipino · Vietnamese · Thai

## Semantic Subtypes

### YES (Y1–Y10)
| ID | Description | English Examples |
|----|-------------|-----------------|
| Y1 | Direct affirmation | yes, yeah, yep, aye |
| Y2 | Emphatic affirmation | absolutely, definitely, without a doubt |
| Y3 | Polite/soft affirmation | of course, gladly, with pleasure |
| Y4 | Colloquial/slang affirmation | you bet, totally, hell yeah |
| Y5 | Agreement with proposition | I agree, exactly, spot on |
| Y6 | Preference/willingness | I'd love to, I'm in, sounds good |
| Y7 | Paradox resolving to yes | I can't say no, I don't disagree |
| Y8 | Rhetorical confirmation | is the sky blue?, does a bear live in the woods? |
| Y9 | Non-verbal/gestural description | *nods*, *thumbs up* |
| Y10 | Contextual indirect yes | let's do it, that works for me |

### NO (N1–N10)
| ID | Description | English Examples |
|----|-------------|-----------------|
| N1 | Direct negation | no, nope, nay, nah |
| N2 | Emphatic negation | absolutely not, never, no way |
| N3 | Polite/soft negation | I'd rather not, I'm afraid not |
| N4 | Colloquial/slang negation | hard pass, not happening, fat chance |
| N5 | Disagreement with proposition | I disagree, you're wrong, that's incorrect |
| N6 | Refusal/aversion | I refuse, count me out, I won't |
| N7 | Paradox resolving to no | yes but actually no, yes yes yes but no |
| N8 | Rhetorical denial | when pigs fly, not in a million years |
| N9 | Non-verbal/gestural description | *shakes head*, *thumbs down* |
| N10 | Contextual indirect no | I'll pass, no thank you, I'm good |

### NONE / Ambiguous (C1–C8)
| ID | Description | English Examples |
|----|-------------|-----------------|
| C1 | Pure uncertainty | maybe, perhaps, I'm not sure |
| C2 | Conditional yes | only if, depends on the price |
| C3 | Conditional no | unless you can prove it, not if it costs money |
| C4 | Deferral / time-based | later, not now, ask me again |
| C5 | Processing / thinking | let me think, I'm considering it |
| C6 | Ambiguous both-sides | it depends, I have mixed feelings |
| C7 | Redirection / clarification | why do you ask?, what do you mean? |
| C8 | Partial agreement | sort of, kind of, more or less |

## Files

| File | Description |
|------|-------------|
| `yesno_multilingual.csv` | Main dataset (10,709 rows) |
| `taxonomy.md` | Full taxonomy, subtype definitions, and golden rules |

## Usage

```python
from datasets import load_dataset

ds = load_dataset("Jarbas/yes-no-multilingual")
```

### Filter by language

```python
en = ds["train"].filter(lambda x: x["language"] == "en")
```

### Filter by label

```python
yes_only = ds["train"].filter(lambda x: x["agreement"] == "yes")
```

## How the Data Was Generated

All utterances were generated directly by a large language model (Claude) acting as a multilingual conversational AI. No machine translation was used — each utterance was composed idiomatically in its target language from scratch.

The generation process followed a strict per-language protocol:

1. **Taxonomy-first**: Each language block was generated by iterating over all 28 semantic subtypes (Y1–Y10, N1–N10, C1–C8) and producing multiple idiomatic examples per subtype.
2. **Register coverage**: Examples span formal, neutral, and casual registers. Languages with formal/informal T–V distinction (German du/Sie, French tu/vous, Spanish tú/usted, Japanese plain/polite forms, Korean formal/informal, etc.) include both.
3. **Golden rules enforcement**: Each utterance was checked against validation rules covering label integrity, no label leaking, length ≤ 75 characters, naturalism, and uniqueness.
4. **Cultural authenticity**: Rhetorical forms (Y8, N8) use idioms native to each language's culture rather than translated English expressions.
5. **Deduplication**: A global deduplication pass ensures no utterance appears twice across the entire dataset.

The language set aligns with the [OVOS localize](https://github.com/OpenVoiceOS/ovos-localize) classification dataset, covering European, Middle Eastern, and Asian languages including minority and regional languages (Basque, Catalan, Galician, Aragonese, Norwegian Nynorsk, Icelandic).

## Quality Guarantees

- **No machine translation** — all utterances are idiomatically authentic per language
- **≥ 8 samples per subtype per language** — every (language × subtype) cell is covered
- **Zero duplicates** — global case-insensitive deduplication across all 43 languages
- **Zero overlength entries** — all utterances ≤ 75 characters
- **Register diversity** — formal, neutral, and casual speech per language
- **Paradox handling** — utterances like "yes but actually no" are labeled by final resolution

## License

Apache 2.0