jangedoo/nepali-ecommerce-retrieval

Nepali E-commerce Retrieval

This is a synthetic e-commerce retrieval dataset containing 1,990 query-positive pairs and three supplied negatives per query. Queries cover English, Devanagari Nepali, Romanized Nepali, and mixed modes; the documents are almost entirely English. The most accurate description is therefore cross-lingual query-to-English retrieval, not a fully multilingual document corpus.

Dataset structure

Field	Description
row_id	Original integer row identifier; IDs are not contiguous.
query	E-commerce retrieval query.
document	Labeled positive document.
negative1–negative3	Supplied negative documents.
product_category	Product-domain label.
language_mode	Intended query-language style.
query_intent	Intended shopping or support intent.
doc_type	Document-format label.

split	rows	product_categories	language_modes	query_intents	document_types	Recall@1	Recall@10	MRR@10	median_rank
train	1367	12	7	10	7	0.200	0.565	0.309	7
test	425	12	7	10	7	0.259	0.666	0.379	5
valid	198	12	7	10	7	0.182	0.566	0.293	6

Composition

dimension	values	minimum rows/value	maximum rows/value
product_category	12	164	167
language_mode	7	283	285
query_intent	10	197	200
doc_type	7	283	285

Text lengths

field	min_words	median_words	mean_words	max_words	median_characters	mean_characters
query	4	12	12.9	33	71	73.9
document	38	68	69.6	129	414	420.8
negative1	25	47	47.8	86	278	282.5
negative2	19	42	43	84	247	252.2
negative3	9	40	40.5	73	235	237.7

E5 retrieval baseline

The baseline uses intfloat/multilingual-e5-small with the required query: and passage: prefixes and L2-normalized embeddings. Exact normalized duplicate queries treat all associated positive documents as relevant. The invalid four-document row is excluded only from the local four-way metric.

• Four-way positive-at-rank-1 accuracy: 91.75% over 1,989 valid rows.
• Four-way MRR: 0.9571.
• Full-corpus Recall@1 / Recall@5 / Recall@10: 0.2106 / 0.4719 / 0.5864.
• Full-corpus MRR@10: 0.3222; median positive rank: 6.

Full-corpus results by labeled language mode

language_mode	rows	Recall@1	Recall@5	Recall@10	MRR@10	median_rank
Romanized Nepali with English terms	284	0.303	0.585	0.739	0.427	4
English only	285	0.263	0.572	0.681	0.389	4
Romanized Nepali	284	0.268	0.556	0.673	0.388	4.5
English-Romanized Nepali	283	0.166	0.473	0.583	0.300	7
English-Nepali mixed	285	0.232	0.477	0.575	0.336	6
Trilingual mixed	284	0.151	0.391	0.504	0.249	10
Nepali (Devanagari)	285	0.091	0.249	0.351	0.167	26

Because language_mode and doc_type are one-to-one confounded, this table must not be used to claim a purely linguistic performance difference.

Split methodology

The split is designed so valid remains broadly representative while test is a semantic out-of-distribution holdout:

1. Queries are Unicode-NFKC normalized, case-folded, stripped of punctuation, and whitespace-normalized. Each duplicate-query group stays in one split.
2. Each row is represented by the equal-weight concatenation of its normalized E5 query and positive-document embeddings.
3. Query groups are clustered independently inside every product category with five-cluster K-means (random_state=42, n_init=20).
4. One 10–30%-sized, maximally isolated cluster per category is reserved for test. Consequently every product domain occurs in test, but its selected E5 clusters do not occur in train or valid.
5. valid targets 10% of all rows and is selected from the remainder by grouped candidate search, minimizing size and marginal-distribution error across category, language mode, intent, and document type. All other rows become train.

There is zero normalized-query overlap between splits. Every split contains all 12 categories, 7 language modes, 10 intents, and 7 document types.

The selected test clusters have mean within-category centroid cosine distance 0.037927 from the remainder, compared with 0.002932 for 2,000 size-matched random holdouts. The shift is 12.94× larger (p=0.000500, one-sided permutation test), confirming that test is materially different in E5 representation space.

Intended uses

• Contrastive retrieval or reranking experiments using supplied negatives.
• Cross-lingual Nepali/Romanized-Nepali query-to-English retrieval.
• Domain-generalization evaluation with the E5-clustered test split.
• Error analysis across shopping intents and product categories.

Limitations and release recommendations

• The examples are synthetic/AI-generated and may contain unrealistic details or relevance errors.
• The dataset is small and has no completed human relevance assessment.
• The provided negatives are much easier than retrieval over the full corpus, so four-way accuracy alone overstates retrieval quality.
• Duplicate queries can have multiple linked positives and must be evaluated with multi-positive relevance.
• language_mode is inseparable from doc_type in the present design.
• Product and price claims should not be treated as current factual information.
• Confirm and add an appropriate license before public upload.

This dataset is best treated as a synthetic training and evaluation seed, not as production ground truth or a definitive Nepali retrieval benchmark without cleanup and human validation.