Datasets:

KeyvanMahmoudi
/

PerSHOP

license: cc-by-nc-4.0
language:
  - fa
pretty_name: PerSHOP
size_categories:
  - 10K<n<100K
task_categories:
  - text-classification
  - token-classification
  - text-generation
task_ids:
  - intent-classification
  - named-entity-recognition
  - dialogue-modeling
tags:
  - task-oriented-dialogue
  - shopping
  - persian
  - dialogue-state-tracking
  - slot-filling
configs:
  - config_name: dialogues
    default: true
    data_files:
      - split: train
        path: data/dialogues/train.jsonl
      - split: validation
        path: data/dialogues/validation.jsonl
      - split: test
        path: data/dialogues/test.jsonl
  - config_name: products
    data_files:
      - split: database
        path: data/products/products.jsonl
  - config_name: slots
    data_files:
      - split: descriptions
        path: data/products/slots.jsonl

PerSHOP: A Multi-Domain Persian Dataset for Shopping Dialogue Systems

PerSHOP is a Persian task-oriented dialogue dataset in the shopping domain. Each dialogue is a conversation between a customer and a seller who fill a shopping basket from a fixed product database. Utterances are annotated with domain, intent/dialogue act, product category and inline slot spans.

This is the official release maintained by the authors, with fixed dialogue-level train/validation/test splits and the product database used to collect the dialogues.


Dialogues	1,061
Utterances	21,923
Domains	14 product domains + chit-chat
Intents / dialogue acts	20
Slot types	27
Products in the database	750
Mean utterances per dialogue	20.7

Collection

The dialogues were collected at the University of Tehran through a web platform built for this purpose. About 120 volunteer students (out of roughly 340 enrolled in the Data Structures and Natural Language Processing courses over three semesters) took the customer and seller roles and conducted shopping conversations over a shared product database, following a written guideline. The dialogues were then annotated by the first author.

The conversations are role-played; the released data contains only the speaker role (user or system) and no identifiers of the participants.

Splits

Split	Dialogues	Utterances
train	849	17,482
validation	106	2,242
test	106	2,199

The split is made at the dialogue level: all utterances of a dialogue are in the same split, so no dialogue context leaks into evaluation data.
80/10/10, stratified by each dialogue's primary domain (its most frequent product domain), with a fixed seed (42).
The dialogue IDs of each split are listed in splits/, and scripts/make_splits.py rebuilds the release exactly from the original spreadsheets.

Please report results on the test split and use validation for model selection.

Data fields

`dialogues` (one row per utterance)

Field	Description
`utterance_id`	Utterance ID from the original release (unique; IDs are not contiguous)
`dialogue_id`	Dialogue ID (1–1061)
`turn_index`	Position of the utterance within its dialogue, from 0
`speaker`	`user` (customer) or `system` (seller)
`domain`	Domain in Persian, as annotated (`Chats` for chit-chat turns)
`domain_en`	English name of the domain
`intent`	Intent or dialogue act, e.g. `Purchase`, `Inform`, `Ask_slot`, `Ask_product`, `Add_product`
`super_product`	Product category the turn is about, e.g. `milk`, `water` (null for chit-chat)
`text`	Utterance text with annotation markup removed
`annotated_text`	Annotated utterance; slots are marked inline as `[span](slot)` (see Corrections)
`raw_text`	`RawText` column of the original release, kept unchanged
`entities`	Slot spans parsed from `annotated_text`: `value`, `slot`, and `start`/`end` character offsets in `text`

`products`

The product database (750 items). Each record has id, full_name, domain, domain_en, super_product, product, price and the attribute slots that apply to that product (e.g. brand, flavor, volume, weight, fat).

`slots`

Names and English descriptions of the product attribute slots.

Usage

from datasets import load_dataset

dialogues = load_dataset("KeyvanMahmoudi/PerSHOP", "dialogues")
products = load_dataset("KeyvanMahmoudi/PerSHOP", "products", split="database")

Corrections

Compared with the original spreadsheet, this release applies a small set of documented corrections, listed in scripts/corrections.json:

13 utterances had slot markup that could not be parsed (e.g. (volume)[1.5], a missing bracket, or a missing slot label). Their syntax was fixed so the annotated spans are recoverable; in two of them (utterance_id 7835 and
1. the product and bag_number labels had been attached to the wrong spans and were swapped back.
2 empty user utterances (utterance_id 12536 and 13069) were removed.
The slot super-product was renamed super_product, matching the naming of all other slots.

raw_text is unchanged. In 242 cases an utterance follows another utterance by the same speaker within a dialogue; these are kept as collected.

License

The dataset (dialogues, annotations and product database) is released under the Creative Commons Attribution-NonCommercial 4.0 International license. Research and other non-commercial use is permitted with citation of the paper below.

Commercial use — including training, fine-tuning or deploying models for commercial products or services — is not covered by this license and requires a separate written agreement with the authors.

Copies of PerSHOP published by third parties are not maintained by the authors; this license applies to all copies.

Citation

@inproceedings{mahmoudi2026pershop,
  title     = {PerSHOP: A Multi-Domain Persian Dataset for Shopping Dialogue Systems},
  author    = {Mahmoudi, Keyvan and Faili, Heshaam},
  booktitle = {2026 International Interdisciplinary Conference on Artificial Intelligence: Engineering, Health, Finance and Humanities (IICAI)},
  pages     = {1--6},
  year      = {2026},
  address   = {Tehran, Iran},
  doi       = {10.1109/IICAI70155.2026.11620935}
}

@article{mahmoudi2024pershop,
  title   = {PerSHOP -- A Persian dataset for shopping dialogue systems modeling},
  author  = {Mahmoudi, Keyvan and Faili, Heshaam},
  journal = {arXiv preprint arXiv:2401.00811},
  year    = {2024}
}