DATA_STATEMENT.md

# Data Statement: STAQ (Synthetic Technology Assistance Queries)

This data statement follows the schema proposed by Bender and Friedman (2018),
"Data Statements for Natural Language Processing." Because STAQ is a *synthetic*
dataset, the text was produced by a large language model rather than by human
speakers; the sections below distinguish the model that generated the text from
the real population whose queries it was grounded in.

For dataset contents, schema, and usage, see the [README](README.md). For full
generation details, see the [datasheet](DATASHEET.md).

## A. Curation rationale

STAQ was created to support the development of age-inclusive artificial
intelligence (AI) systems. Training data for foundation models tends to
underrepresent older adults, and asking older adults to produce the volume of
data needed to train or fine-tune such models is both impractical and ethically
burdensome. STAQ is intended as a scalable bridge for model training and
evaluation — not a replacement for participatory design or human evaluation.

The queries were generated to amplify empirically observed communication
patterns rather than assumed stereotypes. Generation was anchored in real,
unstructured technology-support queries collected during a formative diary study
(see the paper, Section 3), and organized around four communication
characteristics observed in that study: verbosity, over-specification,
under-specification, and incompleteness.

## B. Language variety

The dataset is in English as used in the United States (BCP-47 tag: `en-US`).
The text is model-generated and written to imitate the informal, conversational
register of older adults seeking technology help — including hesitations,
expressions of uncertainty, and everyday phrasing — paired with a clearer,
standard-register expert paraphrase.

## C. Speaker demographic

No real speakers produced the text in STAQ. The literal producer of every query
is the language model GPT-4o. The queries were, however, grounded in and
designed to reflect a specific real population: the participants of the
formative diary study.

Diary-study eligibility required participants to be aged 60 or older, to use a
mobile device (smartphone or tablet) at least once per week, and to be
proficient in English. All participants were community-dwelling adults residing 
in the United States. Full participant demographics are reported in the paper's
Section 3. STAQ does not represent older adults as a monolithic group, and the
synthetic queries should not be read as evidence about any real individual or
subgroup.

## D. Annotator / expert demographic

Domain experts in aging and technology wrote the paraphrases used as few-shot 
examples in the generation prompt; the paraphrases accompanying the released 
queries are model-generated (GPT-4o). Face-validity ratings were produced by 
two such experts, who independently rated a random sample of 50 queries; their
ratings agreed strongly (Cohen's kappa = 0.83). Expert identities and detailed
demographics are not released; they are described here only at the level of
their relevant expertise.

## E. Speech situation

The queries are written text, not transcribed speech, and were produced
asynchronously by a language model rather than in a live interaction. They
simulate a one-to-one, informal, help-seeking situation — an older adult
describing a technology problem to someone who might help — and are intended to
resemble spontaneous, unscripted requests rather than edited or formal writing.
The data was generated in 2025.

## F. Text characteristics

Each instance is a short, technology-related help-seeking query paired with an
AI-generated paraphrase that clarifies the intended meaning. Every query is labeled
with one of four communication characteristics. The distribution across the 514
instances is shown below:

| Characteristic | Count |
| --- | --- |
| under-specification | 154 |
| verbosity | 120 |
| over-specification | 120 |
| incompleteness | 120 |

The queries are longer and less lexically varied than their paraphrases,
mirroring the real data: synthetic queries averaged 58.89 tokens versus 23.75
for paraphrases (range 13–174), with a type-token ratio of 0.858 for queries
versus 0.937 for paraphrases — close to the real-world values of 0.842 and
0.890.

## G. Recording quality

Not applicable. STAQ contains no audio; all content is text.

## H. Other considerations

STAQ is synthetic. Although grounded in and validated against real diary-study
data, it may not capture the full range of how older adults actually
communicate, and it should complement rather than replace human evaluation and
participatory research. Generation was deliberately anchored in real data to
reduce ageist stereotyping, but residual bias from the underlying model cannot
be ruled out. Note also that the `expert_rating` field is populated only for the
50-query face-validity sample; the remaining rows are unrated.

## I. Provenance appendix

Queries were generated with GPT-4o using few-shot prompting. Each prompt
included three example pairs — an original older-adult-style query and its
expert paraphrase — and instructed the model to produce one new example in the
same conversational style, output as JSON. The model was directed to vary the
technical issues represented and was explicitly not asked to emulate a generic
"older adult," instead mirroring the phrasing, uncertainty, and tone of the real
diary-study queries.

Two evaluations support the dataset. For fidelity, sentence embeddings (from a
Sentence-BERT model) of the synthetic and real queries were compared using
Principal Component Analysis; the distributions overlapped strongly. For face
validity, the expert review described in Section D rated 40 of 50 sampled
queries as "likely" to have been said by an older adult (1 "possibly," 9
"unlikely").

The full generation prompt and example pairs appear in the paper's appendix.

## About this data statement

Schema: Bender, E. M., & Friedman, B. (2018). Data Statements for Natural
Language Processing: Toward Mitigating System Bias and Enabling Better Science.
*Transactions of the Association for Computational Linguistics*, 6, 587–604.

Version [v1] — [July 12, 2026]. Maintained alongside the STAQ dataset; see the
[README](README.md) for citation and contact details.