Datasets:

ThiagoFNJ
/

esci-llm-aspect-mining

license: cc-by-4.0
language:
  - en
size_categories:
  - 1M<n<10M
task_categories:
  - text-classification
  - feature-extraction
task_ids:
  - sentiment-classification
source_datasets:
  - extended|other-amazon-esci
tags:
  - aspect-based-sentiment-analysis
  - aspect-extraction
  - e-commerce
  - product-reviews
  - llm-as-etl
  - esci
pretty_name: ESCI LLM Aspect Mining (US)
configs:
  - config_name: review_aspects
    data_files: review_aspects.parquet
  - config_name: product_aspects
    data_files: product_aspects.parquet

ESCI LLM Aspect Mining (US)

A product-intrinsic, polarity-labeled aspect-annotation layer over the us locale of ESCI-S, extracted with a calibrated local LLM (google/gemma-4-12B-it, guided JSON, temperature 0). 8.85M aspects over 3.95M reviews; held-out facet F1 0.604, polarity accuracy 0.9455.

⚠️ No Amazon review text is included. Rows carry derived facets + join keys (asin, review_no, review_md5). Reconstruct the original quote by joining your own copy of ESCI-S on asin+review_no and verifying review_md5.

Full methodology, cost model, and corpus-scale QA (the star cross-check) live in the code repository; this dataset's provenance and licensing are in its DATASHEET.

Configs

review_aspects (9.5M rows): asin, review_no, facet, polarity, review_md5.
product_aspects (5.4M rows): asin, facet, pos, neg, neu, total.

Scope & limitations

US locale only. Model-generated labels (calibrated, not human gold except the 598-review gold split used for tuning). Single-annotator gold set (no inter-annotator agreement). Aggregate polarity is validated (monotone in stars across all slices); no per-review guarantee.

License & provenance

The cc-by-4.0 license applies to the derived annotation layer (the facet phrases and polarity labels) — the author's contribution. Use it freely, including commercially, with attribution (see the citation below).

It does not and cannot grant rights over the underlying review text, which is the property of Amazon and its users and is not included in this dataset. The reviews were obtained via ESCI-S, an unofficial January-2023 scrape, which this dataset neither mirrors nor endorses. If you reconstruct original quotes by joining ESCI-S, you are responsible for your own lawful access to and use of that source text. Full statement: DATASHEET.