blueharu/persian-response-ranking

PerSHOP Response Ranking

A Persian response-ranking benchmark derived from PerSHOP and used to evaluate lexical, semantic-embedding and LLM-based ranking methods. It includes the existing Random and Same-Domain configurations alongside a newly developed Domain-Lexical hard-negative configuration.

Each instance pairs a customer query with 5 candidate responses: 1 correct (gold) and 4 negatives. The task is to rank the gold response first.

The same 2,116 query–gold-response instances and dialogue-level train, validation and test partitions are used across all three configurations. The negative candidates vary according to the sampling strategy.

Config	How negatives are chosen	Difficulty
random	Random responses from anywhere in the dataset	Easy
same_domain	Responses from the same product domain (e.g. dairy)	Medium
domain_lexical	Same domain and high word overlap with the query (new in this work)	Hard

Splits

Split	Instances (per config)
train	1,483
validation	303
test	330
total	2,116

Splits are made at the dialogue level. No dialogue, and no query–response pair, appears in more than one split.

Usage

from datasets import load_dataset

ds = load_dataset("blueharu/persian-response-ranking", "domain_lexical")
print(ds["test"][0])

Fields

Field	Description
id	Instance ID
dialogue_number	Source dialogue in PerSHOP
domain	Product domain (in Persian, e.g. لبنیات = dairy)
super_product	Product category (e.g. cheese)
query	Customer utterance
query_intents	Intent label(s) of the query, |-separated
positive_response	The gold response
response_intents	Intent label(s) of the gold response, |-separated
candidates	List of 5 {text, label} objects; label is 1 for gold, 0 for negatives. Gold position is shuffled.
negative_strategy	random, domain (= same_domain), or domain_lexical
split	train, val (= validation), or test

Note: in some instances, especially in domain_lexical, two negatives have identical text. The gold response never appears as a negative.

Baseline results (test, Hits@1)

Method	random	same_domain	domain_lexical
BGE-m3 (semantic)	0.894	–	0.518
Lean prompt + PSC (LLM)	0.96	0.88	0.52
RAG-augmented fine-tuning (Gemma-2-9B)	0.94	0.827	0.594

From the author's bachelor's thesis on response ranking for Persian social-commerce dialogue.

Source and license

This dataset is derived from PerSHOP (Mahmoudi & Faili, 2024). All rights in the original dialogue text remain with the PerSHOP authors. Permission to publicly redistribute the underlying dialogue text as part of this derived response-ranking benchmark has been obtained from the PerSHOP authors. See LICENSE for the applicable terms and citation requirements.

Citation

If you use this dataset, please cite both the original PerSHOP paper and this dataset.

@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},
  url     = {https://arxiv.org/abs/2401.00811}
}

@misc{shafieian2026pershopranking,
  title        = {PerSHOP Response Ranking: A Persian Response-Ranking Benchmark with Domain-Lexical Hard Negatives},
  author       = {Shafieian, Bahar},
  year         = {2026},
  howpublished = {Hugging Face dataset, \url{https://huggingface.co/datasets/blueharu/persian-response-ranking}},
  note         = {Bachelor's thesis dataset, University of Cassino and Southern Lazio}
}

Author

Bahar Shafieian, University of Cassino and Southern Lazio (UniCas)