README.md

# STAQ: Synthetic Technology Assistance Queries

STAQ (Synthetic Technology Assistance Queries, pronounced "stack") is a dataset
of technology help-seeking queries written in the communication style of older
adults, each paired with a AI-generated paraphrase that clarifies the intended meaning. 
**Synthetic — generated by a large language model and grounded
in real diary-study data, not collected directly from real users.**

## Key facts about the dataset

| Field | Value |
| --- | --- |
| Number of instances | 514 |
| File format | CSV |
| Fields | `id`, `query`, `rephrased_query`, `category`, `expert_rating` |
| Language | English |
| Generation model | GPT-4o (few-shot prompting) |
| License | CC BY 4.0 |
| Canonical copy | Zenodo (see Archived at) |
| Mirror | Hugging Face: `hhshomee2/STAQ` |
| Archived at | Zenodo — DOI [10.5281/zenodo.21323107] |
| Associated paper | Sharifi et al., ASSETS '26 — DOI 10.1145/3797867.3829031 |
| Version [v1] — [July 12, 2026] |

## Contents

- `staq.csv` — the dataset (514 rows).
- `README.md` — this file.
- `DATASHEET.md` — generation details, example pairs, provenance, and evaluation.
- `DATA_STATEMENT.md` — language variety and assumed population.


## What's in the data

Each row is one older-adult-style query, its paraphrase, the communication
characteristic it exhibits, and (for a sampled subset) an expert plausibility
rating. The columns are:

| Column | Description |
| --- | --- |
| `id` | Unique instance identifier. |
| `query` | Technology help-seeking query in the communication style of older adults. |
| `rephrased_query` | AI-generated (GPT-4o) paraphrase clarifying the intended meaning. |
| `category` | Communication characteristic. Exactly one of: `verbosity`, `over-specification`, `under-specification`, `incompleteness`. |
| expert_rating | Expert plausibility rating of how likely an older adult would phrase the query, coded 0 = unlikely, 1 = likely, and 2 = maybe/possibly. **Populated only for the 50-query face-validity sample; unrated rows are marked N/A.** See Limitations. |


The 514 instances are distributed across four categories, shown below:

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

Two illustrative pairs (`query` → `rephrased_query`):

- **Under-specification:** "I want to send pictures to my grandson but my Gmail
  is just not doing it." → "How can I successfully send images through email
  using Gmail?"
- **Incompleteness:** "Hey, my Facebook is not working like before. How do I make
  it normal?" → "How can I revert to the previous version or settings of my
  Facebook application as it is not behaving as expected currently?"

## How it was generated

Queries were generated with GPT-4o using few-shot prompting. Each prompt included
three example pairs — an original query exhibiting one of four communication
characteristics (verbosity, over-specification, under-specification, or
incompleteness) and a corresponding expert-written paraphrase — and instructed
the model to produce one new example in the same casual, conversational style,
output as JSON. To avoid ageist stereotypes, the model was **not** asked to
emulate a generic "older adult"; instead, generation was anchored in real,
unstructured queries from a formative diary study, requiring the model to mirror
the specific phrasing, expressions of uncertainty, and informal tone observed in
that data. The full prompt and example pairs are in the paper's appendix.

## Linguistic properties and evaluation

Synthetic queries averaged 58.89 tokens; paraphrases averaged 23.75 tokens
(range 13–174). Type-token ratio was 0.858 for the older-adult-style queries
versus 0.937 for paraphrases — closely matching the real-world data
(0.842 vs 0.890).

The dataset was evaluated on two dimensions. For **fidelity**, sentence
embeddings (SBERT) of synthetic and real queries were compared with Principal
Component Analysis (PCA); the distributions overlapped strongly, indicating the
synthetic queries capture the semantic structure of the real data. For **face
validity**, two experts in aging and technology independently rated a random
sample of 50 queries; 40 were rated "likely," 1 "possibly," and 9 "unlikely,"
with high inter-rater reliability (Cohen's kappa = 0.83).

## Quick start

```python
from datasets import load_dataset
ds = load_dataset("hhshomee2/STAQ")

# or, from the CSV directly:
import pandas as pd
df = pd.read_csv("staq.csv")
print(len(df))                         # expected: 514
print(df["category"].value_counts())   # under-specification 154, others 120 each
print(df["expert_rating"].notna().sum())  # expected: 50 (rated sample)
```

## Intended use

STAQ is intended as a scalable, training-oriented resource for building
age-inclusive AI systems — for example, fine-tuning or evaluating models that
help articulate or clarify technology-support requests from older adults. It is
meant to complement, never replace, participatory design and human evaluation
with older adults.

### Out of scope

- Do not treat these queries as a substitute for real older adults' behavior or
  as evidence about any individual or group.
- Do not use STAQ to replace human evaluation or participatory research.


## Limitations

- **Synthetic, not real.** Queries are model-generated. Although grounded in and
  validated against real diary-study data, they may not capture the full range of
  how older adults actually communicate.
- **Population framing.** The underlying diary study recruited  participants 
  aged 60 or older who used a mobile device (smartphone or tablet) at least 
  weekly and were proficient in English. All participants were community-dwelling 
  adults residing in the United States. The data does not represent older adults 
  as a monolith.
- **Language and cultural scope.** English only; phrasing and examples may not
  transfer to other languages or cultural contexts.
- **Model-encoded bias.** Generation was anchored in real data specifically to
  reduce ageist stereotyping, but residual bias from GPT-4o cannot be ruled out.
- **Partial expert ratings.** `expert_rating` is populated only for the random
  50-query face-validity sample described above; the remaining rows are unrated.
  Do not treat missing ratings as a low rating.

## License

Released under the Creative Commons Attribution 4.0 International license
(CC BY 4.0). You are free to share and adapt the data
with attribution.

## How to cite

If you use STAQ, please cite the paper. The dataset is also archived on Zenodo
with its own DOI (see the "Archived at" row above) if you need to cite the data
directly.


```bibtex
@inproceedings{sharifi2026staq,
author = {Sharifi, Hasti and Shomee, Homaira Huda and Lamar, Melissa and Medya, Sourav and Chattopadhyay, Debaleena},
title = {Helping the Helper: {LLM}-Assisted Problem Articulation for Older Adults Seeking Technology Support},
year = {2026},
publisher = {Association for Computing Machinery},
address = {New York, NY, USA},
url = {https://doi.org/10.1145/3797867.3829031},
doi = {10.1145/3797867.3829031},
booktitle = {Proceedings of the 28th International ACM SIGACCESS Conference on Computers and Accessibility},
articleno = {61},
numpages = {22},
keywords = {Older adults, Large language models, Synthetic data, Technology support},
location = {Vila Nova de Gaia, Portugal},
series = {ASSETS '26}
}
```

## Versioning and changelog

- **[1.0.0]** — [date] — Initial release accompanying the ASSETS '26 paper.


## Related documentation

- [Datasheet](DATASHEET.md) — generation details, example pairs, provenance, and evaluation.
- [Data statement](DATA_STATEMENT.md) — language variety and assumed population.

## Contact

Debaleena Chattopadhyay — debchatt@uic.edu — [ORCID](https://orcid.org/0000-0002-8197-9905)