ambiguous-questions-clarifying

A synthetic dataset of 119 ambiguous natural-language questions paired with:

• Multiple valid interpretations
• Targeted clarifying questions that resolve the ambiguity
• A minimal resolution context and a fully grounded answer

The dataset spans 6 structural ambiguity types and 14 domains, making it suitable for training and evaluating models that handle ambiguity in dialogue, question answering, and natural language understanding.

Technical report: Taxonomy and Construction of Synthetic Clarification Datasets for Natural Language Ambiguity — covers generation methodology, analysis, evaluation protocols, and limitations in depth.

---

Dataset at a Glance

Split	Examples
train	95
validation	11
test	13
Total	119

Ambiguity Types

Type	Count	Description
lexical	33	A word with multiple meanings (e.g., "bank", "table", "fork")
referential	23	Unclear pronoun or noun-phrase antecedent
pragmatic	22	Literal vs. intended meaning; indirect speech acts
scope	14	Quantifier or negation scope ambiguity
presupposition	14	Question embeds a possibly-false hidden assumption
attachment	13	Syntactic attachment of a phrase to one of two heads

Difficulty Distribution

Level	Count
easy	35
medium	63
hard	21

Pattern: Lexical ambiguity is always easy or medium (polysemy resolves with a single direct question). Attachment and presupposition are always medium or hard (require syntactic or pragmatic inference).

Domains Covered

General (41), workplace (23), technical (18), education (8), business (4), science (4), medical (4), legal (4), cooking (4), finance (3), hospitality (2), travel (2), photography (1), design (1).

---

Schema

Each record is a JSON object with the following fields:

Field	Type	Description
id	string	MD5-derived 12-character unique identifier
ambiguous_question	string	The original ambiguous question or statement
ambiguity_type	string	One of: lexical, referential, scope, presupposition, pragmatic, attachment
domain	string	Topic domain (e.g., workplace, technical, cooking)
interpretations	list[string]	Two or more valid interpretations of the question
clarifying_questions	list[string]	One or more questions that would resolve the ambiguity
resolution_context	string	Minimal additional context that disambiguates the question
resolved_answer	string	The correct answer given the resolved context
difficulty	string	One of: easy, medium, hard

---

Sample Records

Lexical — easy

{
  "ambiguous_question": "I want to book a table.",
  "ambiguity_type": "lexical",
  "domain": "hospitality",
  "interpretations": [
    "Make a restaurant reservation.",
    "Schedule a meeting slot in a calendar (colloquially 'book a slot on the table').",
    "Purchase a table (furniture) from a shop."
  ],
  "clarifying_questions": [
    "Are you looking to make a restaurant reservation, schedule a meeting, or buy furniture?",
    "What kind of table — dining, meeting, or furniture?"
  ],
  "resolution_context": "User mentions a restaurant name and a time.",
  "resolved_answer": "You can book a table at most restaurants via their website, OpenTable, Resy, or by calling directly. What date, time, and party size do you need?",
  "difficulty": "easy",
  "id": "e651507f65d0"
}

Referential — medium

{
  "ambiguous_question": "The CEO told the CFO that she reviewed the financial report.",
  "ambiguity_type": "referential",
  "domain": "workplace",
  "interpretations": [
    "The CEO reviewed the financial report.",
    "The CFO reviewed the financial report."
  ],
  "clarifying_questions": [
    "Who reviewed the financial report — the CEO or the CFO?",
    "Does 'she' refer to the CEO or the CFO?"
  ],
  "resolution_context": "CFO prepared and reviewed the quarterly earnings.",
  "resolved_answer": "The CFO reviewed the financial report — it was part of her quarterly preparation duties.",
  "difficulty": "medium",
  "id": "60ceeb2a1e68"
}

Pragmatic — hard

{
  "ambiguous_question": "That's an interesting idea.",
  "ambiguity_type": "pragmatic",
  "domain": "workplace",
  "interpretations": [
    "Genuine enthusiasm — the idea is considered interesting and worth pursuing.",
    "Polite dismissal — 'interesting' used to avoid saying the idea is bad."
  ],
  "clarifying_questions": [
    "Do you mean interesting in a positive way, or is it a polite no?",
    "Would you like to explore this idea further, or are you moving on?"
  ],
  "resolution_context": "Said by a senior executive without further follow-up.",
  "resolved_answer": "Without further follow-up, 'interesting' from a senior executive often signals polite disinterest. Explicitly ask: 'Should we take this forward, or would you like to see more data first?'",
  "difficulty": "hard",
  "id": "905324ee0467"
}

Scope — hard

{
  "ambiguous_question": "The drug works on most patients.",
  "ambiguity_type": "scope",
  "domain": "medical",
  "interpretations": [
    "The majority (>50%) of patients respond.",
    "A high proportion (e.g., >80%) respond — 'most' could imply near-universal."
  ],
  "clarifying_questions": [
    "What percentage does 'most' correspond to in the clinical data?",
    "Are we talking 55% or 90%? The clinical implications differ greatly."
  ],
  "resolution_context": "Drug shows 73% response rate in Phase III trial.",
  "resolved_answer": "The trial shows a 73% response rate — a majority, but roughly one in four patients won't respond. Discuss alternative treatment pathways for non-responders.",
  "difficulty": "hard",
  "id": "093f93bd3d83"
}

Presupposition — hard

{
  "ambiguous_question": "Have you told your partner about the affair?",
  "ambiguity_type": "presupposition",
  "domain": "general",
  "interpretations": [
    "Presupposes there was an affair; asks about disclosure.",
    "No affair occurred — the presupposition is false."
  ],
  "clarifying_questions": [
    "Are you assuming I had an affair, or asking hypothetically?",
    "What makes you think there was an affair?"
  ],
  "resolution_context": "Question asked based on a misunderstanding of a rumor.",
  "resolved_answer": "There was no affair — the questioner can be corrected: 'That's a misunderstanding. There is nothing to disclose.'",
  "difficulty": "hard",
  "id": "f3a1eccb36b1"
}

Attachment — hard

{
  "ambiguous_question": "Old men and women should get priority seating.",
  "ambiguity_type": "attachment",
  "domain": "general",
  "interpretations": [
    "Old men and old women should get priority seating.",
    "Old men, and all women (regardless of age) should get priority seating."
  ],
  "clarifying_questions": [
    "Does 'old' modify both men and women, or only men?",
    "Should all women get priority, or only elderly women?"
  ],
  "resolution_context": "A transit policy for elderly and disabled passengers.",
  "resolved_answer": "In transit policy context, 'old' typically modifies both: elderly men and elderly women both qualify for priority seating.",
  "difficulty": "hard",
  "id": "fc82ace50c1b"
}

---

Usage

Loading the Dataset

from datasets import load_dataset

ds = load_dataset("SalutefaceAAC/ambiguous-questions-clarifying")
print(ds)
# DatasetDict({
#     train: Dataset({features: ['id', 'ambiguous_question', 'ambiguity_type', ...], num_rows: 95})
#     validation: Dataset({features: [...], num_rows: 11})
#     test: Dataset({features: [...], num_rows: 13})
# })

Task A: Ambiguity Classification

Predict the ambiguity type (6-class) given only the question text.

from datasets import load_dataset

ds = load_dataset("SalutefaceAAC/ambiguous-questions-clarifying")

# Build classification examples
for split in ["train", "validation", "test"]:
    for ex in ds[split]:
        x = ex["ambiguous_question"]
        y = ex["ambiguity_type"]
        # e.g. train a text classifier or few-shot prompt

Baseline: Random = 16.7% accuracy. Bag-of-words = ~45%. Strong LLM (zero-shot) expected > 75%.

Task B: Clarification Generation

Given an ambiguous question, generate targeted clarifying questions.

from datasets import load_dataset

ds = load_dataset("SalutefaceAAC/ambiguous-questions-clarifying")

for ex in ds["train"]:
    input_text = ex["ambiguous_question"]
    target_clarifiers = ex["clarifying_questions"]  # list[str]
    # Fine-tune seq2seq or use as few-shot prompt examples
    # Evaluate with BERTScore or LLM-as-judge (ROUGE-L undercounts valid alternatives)

Task C: Disambiguation (Grounded QA)

Given the ambiguous question and a resolution context, produce the correct answer.

from datasets import load_dataset

ds = load_dataset("SalutefaceAAC/ambiguous-questions-clarifying")

for ex in ds["train"]:
    question = ex["ambiguous_question"]
    context = ex["resolution_context"]
    answer = ex["resolved_answer"]
    # Frame as: "Q: {question}\nContext: {context}\nA:"

Filtering by Type or Difficulty

from datasets import load_dataset

ds = load_dataset("SalutefaceAAC/ambiguous-questions-clarifying")

# Only hard presupposition examples
hard_presupposition = ds["train"].filter(
    lambda x: x["ambiguity_type"] == "presupposition" and x["difficulty"] == "hard"
)

# Only workplace domain
workplace = ds["train"].filter(lambda x: x["domain"] == "workplace")

---

Intended Uses

• Clarification generation: Train models to ask targeted clarifying questions on ambiguous input.
• Ambiguity classification: Classify which structural ambiguity type is present.
• Disambiguation: Given a question + resolution context, produce the correct grounded answer.
• Dialogue systems: Improve chatbot handling of ambiguous user inputs.
• NLU benchmarking: Evaluate models on fine-grained pragmatic and structural understanding.

---

Generation Methodology

Records were generated in two phases:

1. Hand-authored base examples (40 records): Each record was individually crafted to represent a linguistically distinct ambiguity pattern. Examples are drawn from real-world communication scenarios (restaurant booking, workplace meeting, medical consultation, transit policy).

2. Parametric augmentation (79 records): Templates parameterized over roles, words, and domains, each with ≥3 independent fill slots. A cap of 15% per ambiguity type enforces balance. MD5 deduplication on question + interpretations removed 12 collisions (13% collision rate).

See REPORT.md for full methodology, taxonomy definitions, and evaluation protocol details.

---

Limitations

• English only: No cross-lingual coverage.
• Small scale: 119 examples — use for evaluation or as seed data, not solo fine-tuning.
• No unambiguous controls: Every record is ambiguous by design. Production systems also need unambiguous examples to calibrate when not to ask.
• Cultural specificity: Pragmatic examples assume Anglo-American conversational norms (e.g., "interesting" as polite dismissal). May not generalize cross-culturally.
• Synthetic distribution: Type proportions (lexical 28%, referential 19%, ...) reflect editorial choices, not natural corpus statistics.
• Incomplete interpretation enumeration: Scope ambiguity can have more readings than listed; the dataset lists the two most clinically/practically significant.

---

License

Apache 2.0

---

Citation

@dataset{salutefaceaac2026ambiguous,
  title     = {ambiguous-questions-clarifying},
  author    = {SalutefaceAAC},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/SalutefaceAAC/ambiguous-questions-clarifying}
}