# Datasheet — ESCI LLM Aspect Mining

Following the *Datasheets for Datasets* framework (Gebru et al., 2021).

## Motivation

**Why was it created?** To provide a product-intrinsic, polarity-labeled **aspect-annotation
layer** over the reviews in the Amazon ESCI benchmark, and to serve as a worked reference for
*calibrated LLM-as-ETL*: extraction whose quality and cost are measured before and during a
multi-million-record run. It supports aspect-based sentiment analysis, review-aspect retrieval,
and research on LLM extraction reliability at scale.

## Composition

**What do the instances represent?** Each row of `review_aspects.parquet` is one *(review,
extracted aspect)* pair over the **`us` locale of ESCI-S**. `product_aspects.parquet` aggregates
these to *(product, facet)* polarity counts.

**Scope and size.** US locale only. 447,924 catalog products; 412,739 carry reviews (≤13 each).
**9,507,120** review rows (**8,850,361** with an aspect; the rest examined with no
product-intrinsic aspect, kept as null-facet rows), over **3,950,828 reviews** (99.96% of the
3,952,486 in scope). **5,372,779** product–facet pairs over **408,099** products.

**What data does each instance contain?** Derived facet phrases (English, 1–4-word lowercase noun
phrases naming a product *dimension*), a polarity label, and join keys (`asin`, `review_no`,
`review_md5`). **No original review text is included.**

**Is any information missing?** By design: (a) `evidence` — the verbatim quote grounding each
aspect — is computed internally but **stripped from the release** (Amazon text is not
redistributed); (b) 1,658 reviews (0.04%) failed extraction (over-length, exceeding the 4096-token
context) and are absent; (c) reviews yielding no product-intrinsic aspect appear as null-facet
rows, distinguishing "examined, nothing found" from "not processed".

**Labels / ground truth.** Facets and polarities are **model-generated**, not human gold, except
the 598-review gold set in [`gold/`](gold/) (dev-248 / test-350) used for calibration. Held-out
facet F1 is **0.604** (semantic θ=0.80, 95% CI [0.566, 0.638]); polarity accuracy **0.9455**.

**Errors, noise, redundancies.** The annotations carry the extractor's measured error rate. The
corpus-scale §4.3 star cross-check confirms polarity is sound in aggregate (monotone in stars
across all slices) but makes **no per-review guarantee** — a low-star review can carry a true
positive facet ("comfortable, but fell apart").

## Collection & preprocessing

**Source.** Reviews come from [ESCI-S](https://github.com/shuttie/esci-s), an unofficial
January-2023 scrape of Amazon, joined by ASIN to the
[Amazon ESCI benchmark](https://github.com/amazon-science/esci-data). This project does **not**
scrape Amazon and does **not** mirror ESCI-S.

**Extraction.** One pass per review through `google/gemma-4-12B-it` (BF16, served by vLLM with
guided JSON decoding, temperature 0), prompt `gi9` selected by a GEPA-style loop against the gold
set. Facets are constrained to product-intrinsic dimensions; delivery/shipping/price/seller are
excluded by the prompt. The full engineering record is in
[`docs/review-aspect-extraction.md`](docs/review-aspect-extraction.md).

## Uses

**Suitable for:** aspect-based sentiment analysis, aspect-aware product retrieval/reranking,
faceted-search prototypes, and studies of LLM extraction quality, cost, and determinism at scale.

**Not suitable for:** claims about individual reviews (the signal is aggregate); a locale other
than US; or as a source of Amazon review text (there is none here).

**PII.** The released columns are abstracted facet phrases and hashes — no free text, no user
identifiers. Facet phrases are product dimensions and carry negligible PII risk. (The stripped
`evidence` column could in principle have contained incidental PII fragments — another reason it is
not distributed.)

## Distribution & licensing

- **Derived annotations** (parquets, gold set): **CC-BY-4.0**. The license applies to the facet
  phrases and polarity labels — the derivative annotation layer, which is the author's
  contribution. Attribution is required (see the repository `CITATION.cff`). The license does
  **not**, and legally cannot, extend to the underlying review *text*: that content is the
  property of Amazon and its users, is **not redistributed** here, and no rights over it are
  granted or implied. The reviews were obtained via ESCI-S, an unofficial scrape that this project
  neither mirrors nor endorses; anyone reconstructing quotes (by joining ESCI-S on
  `asin`+`review_no`) is responsible for their own lawful access to and use of that source text.
- **Code**: MIT.
- **ESCI benchmark labels**: Apache-2.0 (`amazon-science/esci-data`).

## Maintenance

Maintained by the author. Corrections and re-runs (new prompt/model versions) would be published
as new dataset versions; the extractor and prompt are pinned by revision so any version is
re-derivable. `review_md5` lets consumers verify they joined the correct source review.