TrEcomQA: Turkish E-Commerce Product Question Answering

TrEcomQA is a dataset of real customer questions and seller answers collected from the public Q&A sections of product pages on Turkish e-commerce platforms (Trendyol, Hepsiburada). It covers 63 product categories in 8 domains (white goods, small kitchen appliances, electronics, personal care, power tools, cookware, household consumables) and targets product question answering, answer generation, and question classification in Turkish.

Status: 🚧 work in progress — schema is fixed; statistics below are filled in at release.

Türkçe Özet

TrEcomQA, Türk e-ticaret sitelerinin ürün sayfalarındaki herkese açık soru-cevap bölümlerinden toplanmış gerçek müşteri soruları ve satıcı cevaplarından oluşan bir veri setidir. 8 alan altında 63 ürün kategorisini kapsar. Kişisel veriler alınmamış (kullanıcı adları yok, satıcı adları hash'lenmiş) ve metin içindeki telefon, e-posta, IBAN, URL, sipariş numarası ve TCKN maskelenmiştir. Eğitim/doğrulama/test bölmeleri ürün bazında yapılmıştır; aynı ürün iki farklı bölmede yer almaz.

Dataset Structure

Three configurations, joinable on id (question) and (source, product_id) (product):

Config	Granularity	Splits	Description
qa (default)	one row per question	train / validation / test	Question + its primary answer + derived labels
answers	one row per answer	train / validation / test	All answers to each question (several sellers may answer the same question)
products	one row per product	train	Product metadata; gtin allows matching the same product across platforms

Loading

from datasets import load_dataset

qa = load_dataset("mustafaoksuz/TrEcomQA", "qa")
answers = load_dataset("mustafaoksuz/TrEcomQA", "answers")
products = load_dataset("mustafaoksuz/TrEcomQA", "products", split="train")

Data Instance (qa)

{
  "id": "hepsiburada:HBV00000AI8YX:10610bbe-017c-40b2-86bc-96d6c932057c",
  "source": "hepsiburada",
  "product_id": "HBV00000AI8YX",
  "question_id": "10610bbe-017c-40b2-86bc-96d6c932057c",
  "question": "Izgara olarak kullanılır mı?",
  "question_date": "2022-07-12T17:02:24",
  "question_type": "yes_no",
  "question_scope": "product",
  "question_topics": null,
  "topics_source": null,
  "topics_model": null,
  "primary_answer": "Izgara özelliği bulunmamaktadır.",
  "primary_answer_date": "2022-07-13T06:50:36",
  "primary_answer_latency_min": 828,
  "primary_seller_id": "3f1c…",
  "primary_seller_is_brand_store": false,
  "num_answers": 3,
  "category": "sandwich_maker",
  "domain": "kitchen_appliance",
  "split": "train"
}

Data Fields (qa)

Field	Type	Nullable	Description
id	string	no	Global unique id: {source}:{product_id}:{question_id}
source	string	no	Platform: trendyol, hepsiburada
product_id	string	no	Platform-native product id (string, never cast to int)
question_id	string	no	Platform-native question id
question	string	no	Customer question (cleaned, PII-masked)
question_date	timestamp	yes	When the question was posted (UTC)
question_type	string	no	yes_no or open_ended; rule-based on the Turkish question particle (-mı/-mi/-mu/-mü)
question_scope	string	no	product (about the item itself), seller_service (authenticity, warranty, dealer, service), transaction (stock, shipping, price, order), unknown
question_topics	list[string]	yes	Multi-label topic set (≤ 4) from the taxonomy below. Empty in this release; see Annotations
topics_source	string	yes	llm or human — who produced question_topics
topics_model	string	yes	Model name when topics_source = llm
primary_answer	string	no	First answer to the question (earliest; cleaned, PII-masked)
primary_answer_date	timestamp	yes	Date of the primary answer. Null for Trendyol (platform shows only relative time)
primary_answer_latency_min	int32	yes	Minutes between question and primary answer. Hepsiburada: exact; Trendyol: parsed from "1 saat içinde cevaplandı"
primary_seller_id	string	no	Salted hash of the answering seller's store name (store names are not released)
primary_seller_is_brand_store	bool	no	Seller name contains the product brand (official/brand store heuristic)
num_answers	int16	no	Number of answers in answers for this question
category	string	no	Normalized product category (63 values, see below)
domain	string	no	Category group (8 values)
split	string	no	train / validation / test

Data Fields (answers)

Field	Type	Nullable	Description
id	string	no	Question id, joins to qa.id
answer_id	string	no	Unique answer id
source, product_id, question_id	string	no	As in qa
answer	string	no	Answer text (cleaned, PII-masked)
answer_date	timestamp	yes	Null for Trendyol
answer_latency_min	int32	yes	As primary_answer_latency_min
answer_rank	int16	no	1 = earliest answer (the one used as primary_answer)
seller_id	string	no	Hashed seller
seller_is_brand_store	bool	no	Brand-store heuristic
is_ai_reviewed	bool	no	Platform flag on the answer, as crawled
answer_polarity	string	yes	(planned) Y / N / ? for yes_no questions
split	string	no	Same as the question's split

Data Fields (products)

Field	Type	Nullable	Description
source, product_id	string	no	Product key
title	string	no	Product title
brand	string	yes	Brand
gtin	string	yes	EAN/GTIN barcode — the same physical product on both platforms shares a gtin
category_path	string	yes	Raw platform category (Hepsiburada: leaf name; Trendyol: full path)
category	string	no	Normalized category
category_tr	string	yes	Turkish display name of category
domain	string	no	Category group
color	string	yes	Color variant, if listed
url	string	yes	Product page URL
num_questions	int32	no	Questions in qa for this product
num_answers	int32	no	Answers in answers for this product
split	string	no	Split all of this product's questions belong to

Categories

63 normalized categories in 8 domains. The mapping from raw platform category names is in docs/category_map.csv.

Domain	Categories
electronics	smartphone, tablet, television, monitor, smartwatch, tws_earbuds, bluetooth_speaker, soundbar, powerbank
white_goods	refrigerator, freezer, washing_machine, dryer, dishwasher, oven, cooktop, built_in_set, range_hood, microwave, air_conditioner, water_dispenser
home_appliance	robot_vacuum, stick_vacuum, bagless_vacuum, bagged_vacuum, steam_iron, steam_generator_iron, air_purifier, fan
kitchen_appliance	espresso_machine, turkish_coffee_machine, filter_coffee_machine, tea_maker, kettle, blender, mixer, chopper, sandwich_maker, air_fryer, toaster
personal_care	shaver_trimmer, hair_styler, hair_dryer, epilator, ipl_device, body_lotion
power_tools	cordless_drill_driver, rotary_hammer, drill, angle_grinder, laser_measure
cookware	pan, pot, pressure_cooker, thermos
household_consumable	dishwasher_detergent, laundry_detergent, fabric_softener, stain_remover, coffee_capsule, filter_coffee, led_bulb, smart_bulb

Question Topic Taxonomy

question_topics is multi-label (at most 4 per question):

Label	Türkçe	Scope
product_specifications	Teknik özellik (watt, kapasite, boyut)	product
product_features	Özellik / fonksiyon	product
usage_and_maintenance	Kullanım ve bakım	product
usage_suitability	Belirli bir amaca uygunluk	product
product_comparison	Model karşılaştırma	product
package_contents	Kutu içeriği	product
battery_and_charging	Pil / şarj	product
product_listing_accuracy	İlan doğruluğu, orijinallik	seller_service
warranty_and_service	Garanti, servis, bayi	seller_service
shipping_and_delivery	Kargo / teslimat	transaction

question_scope is derived from these labels when present, otherwise from a keyword rule. Annotation guidelines: docs/annotation_guidelines.md.

Data Splits

Splits are made at the product level so that no product appears in two splits:

bucket = int(sha1(f"{source}:{product_id}").hexdigest(), 16) % 100
train: bucket < 80   validation: 80 ≤ bucket < 90   test: bucket ≥ 90

Split	# products	# questions	# answers
train	TBD	TBD	TBD
validation	TBD	TBD	TBD
test	TBD	TBD	TBD

Per-source counts, question_type and question_scope distributions: TBD at release.

Dataset Creation

Source Data

Public Q&A sections of product pages on Trendyol and Hepsiburada, restricted to the 63 categories above. Questions are written by customers; answers are written by the sellers of the listing (on Hepsiburada several sellers of the same product may each answer the same question — these are all kept in answers). Collection period: TBD.

Preprocessing

1. Cleaning: HTML entity decoding and tag removal, Unicode NFC normalization, line-break and whitespace normalization. Turkish characters are preserved; no lowercasing.
2. Anonymization: customer usernames are never collected. Seller store names are replaced by a salted hash (seller_id). Inside question/answer text the following are replaced by placeholder tokens: [TELEFON], [EPOSTA], [IBAN], [URL], [SIPARIS_NO], [TCKN].
3. Question merging: on Hepsiburada each seller's answer is stored as a separate Q&A record with the same question text; these are merged into one question with several answers. Exact duplicate (seller, answer) pairs from repeated crawls are dropped.
4. Filtering: rows with an empty question or answer, questions shorter than 3 characters or longer than 2000, answers longer than 5000 characters.
5. Derived fields: question_type (question-particle rule), question_scope (topic labels or keyword rule), answer latency (timestamp difference, or parsed from Trendyol's relative "answered within …" text).
6. Splitting: product-level hash split (see above).

Annotations

The text data is "found" (no manual annotation). question_topics was produced internally with an LLM for quality analysis but is not part of the public release; the columns are kept in the schema (all null) so that a manually annotated gold subset (topics_source = "human") can be added in a later version. answer_polarity is likewise planned.

Considerations

Personal and Sensitive Information

The data comes from publicly visible pages. Following KVKK (Turkish Personal Data Protection Law) and GDPR principles, customer identities are not included, seller names are hashed, and identifying strings inside texts are masked. If you find remaining personal data, please open a discussion on the dataset page and it will be removed.

Biases and Limitations

• Answers are written by sellers, not by other customers; they can be promotional, templated ("Merhaba, … iyi günler dileriz") or uninformative ("sorunuzu mağazaya sorunuz").
• Trendyol provides no absolute answer timestamp; answer_date is null and latency is approximate.
• question_type, question_scope and *_is_brand_store are heuristics, not gold labels.
• Popular products and categories are over-represented; category coverage reflects the selected 63 categories, not the whole platform.
• Answers are not verified for factual correctness.

Intended Use

Academic research on Turkish product QA, answer generation, question classification and retrieval. Not intended for re-identifying customers or sellers, or for commercial re-publication of platform content.

Repository Structure

TrEcomQA/
├── README.md                     # this dataset card
├── data/
│   ├── qa/        {train,validation,test}-*.parquet
│   ├── answers/   {train,validation,test}-*.parquet
│   └── products/  products-*.parquet
└── docs/
    ├── category_map.csv          # raw platform category -> category / category_tr / domain
    └── annotation_guidelines.md  # question_topics taxonomy and rules

Crawler and processing code are kept in a separate repository.

Citation

@misc{oksuz2026trecomqa,
  title        = {TrEcomQA: A Turkish E-Commerce Product Question Answering Dataset},
  author       = {Öksüz, Mustafa},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/mustafaoksuz/TrEcomQA}}
}

License

Dataset card and derived fields: CC BY 4.0. The underlying texts originate from the respective platforms and their users; please also respect the platforms' terms of use.