Datasetpublic

k-chirkunov/jeeran_semeval_2016_arabic_adaptation_coarse

Jeeran — SemEval-2016 Task 5 (Arabic) adaptation, coarse categories The k-chirkunov/jeeran_semeval_2016_arabic_adaptation dataset with its 48 fine-grained Arabic aspect labels collapsed onto 14 coarse families. Everything else — targets, character offsets, polarity, the human-annotated spans, the train/test split — is carried over unchanged. Contents split reviews sentences opinions train 43,365 69,524 181,406 test 10,838 17,594 45,225 The… See the full description on the dataset page: https://huggingface.co/datasets/k-chirkunov/jeeran_semeval_2016_arabic_adaptation_coarse.

sourceHugging Faceotherupdated 1mo agoView on Hugging Face

0likes28downloads

Dataset Card

Jeeran — SemEval-2016 Task 5 (Arabic) adaptation, coarse categories

The `k-chirkunov/jeeran_semeval_2016_arabic_adaptation` dataset with its 48 fine-grained Arabic aspect labels collapsed onto 14 coarse families. Everything else — targets, character offsets, polarity, the human-annotated spans, the train/test split — is carried over unchanged.

split	reviews	sentences	opinions
train	43,365	69,524	181,406
test	10,838	17,594	45,225

The datasets view is flat: one row per opinion, with the sentence text repeated across the opinions it contains. SemEval SB1 XML is under semeval_xml/.

Fields

field	description
`review_id`, `sentence_id`	`sentence_id` is `"{review_id}:{n}"`, consecutive from 0
`text`	the sentence
`target`	opinion target expression — verbatim substring of `text`
`from` / `to`	character offsets of `target` in `text`
`category`	coarse family, one of the 14 below
`category_fine`	the original fine label this was mapped from (48 values)
`polarity`	`positive` / `negative` / `neutral`
`span`, `span_from`, `span_to`	the human-annotated evaluative span the target came from

Coarse taxonomy

#	family	covers
1	تقييم عام والجودة	general or holistic praise/criticism, a personal opinion or impression, a rating/comparison, overall quality, product/goods quality, device/equipment quality, advertising/promotion, notices/announcements, or collected opinions/surveys.
2	توفر وتنوع البضاعة	variety or diversity of products/goods/models, product availability, stock, shortage, or abundance.
3	الاحترافية والكفاءة	professionalism, skill, expertise or experience, dedication and commitment to work, responsibility and accountability, honesty and integrity, or cheating/fraud.
4	المعاملة والاهتمام	how staff or management TREAT and ATTEND TO customers, visitors, or patients — politeness, respect, helpfulness, reception, bedside manner — and the establishment's attentiveness, responsiveness, effort, and concern to meet people's needs, comfort, and wishes (including whether customers end up satisfied or pleased).
5	السعر	price, cost, being expensive or cheap, or value for money.
6	النظافة	cleanliness, hygiene, tidiness, or care for the cleanliness of the place.
7	المكان	location, view, accessibility, surroundings, or the size/area/space of the place.
8	الوقت	punctuality, appointments, delays, waiting time, or working/opening hours.
9	الازدحام وعدد الموظفين	crowding or congestion, or the number of employees/staff.
10	النظام والحماية	rules, organization, order, procedures, discipline, safety, security, protection, or guards.
11	التوصية والتحذير	advice, guidance, a recommendation or suggestion, or a warning about danger, risk, or harm.
12	تهنئة	congratulations, thanks, appreciation, or blessing.
13	تخصص	something intended for a specific group, a specialization, or a dedicated section.
14	توتر	stress, anxiety, tension, confusion, or discomfort.

Distribution

family	opinions	share
تقييم عام والجودة	93,784	41.38%
السعر	23,822	10.51%
المعاملة والاهتمام	22,181	9.79%
توفر وتنوع البضاعة	17,944	7.92%
المكان	17,921	7.91%
الاحترافية والكفاءة	14,770	6.52%
النظافة	11,496	5.07%
الوقت	6,777	2.99%
النظام والحماية	5,578	2.46%
التوصية والتحذير	4,052	1.79%
الازدحام وعدد الموظفين	3,014	1.33%
تخصص	2,543	1.12%
تهنئة	2,262	1.00%
توتر	487	0.21%

Mapping provenance

The fine → coarse mapping is coarsen_aspect_category from the project's evaluation_coarse/prompts_override.py (verified identical to the ASPECT_CATEGORY_MERGE in evaluation/eval_utils.py), applied verbatim — no relabelling by hand or by model.

Two details worth knowing:

—The fine dataset contains 80 distinct category strings, not 48: 30 are whitespace variants ("السعر ", " الخبرة"). The mapping strips before lookup, so they normalize onto their 48 canonical labels; 365 rows were affected.
—1,677 opinions (0.73%) were dropped — fine label unknown or empty, which belong to no coarse family. Sentences and reviews left empty by that were dropped too, and sentence ids renumbered to stay gap-free.

Verification

Every released opinion satisfies text[from:to] == target, and the target lies within its own span. Checked on the CSVs, on the XML after reparsing, and after the Hub round-trip — 0 exceptions. No review or sentence id appears in both splits.

Caveats inherited from the fine dataset

target and category_fine are model-generated (Gemma); span boundaries and polarity are human. Opinions with no explicit target noun phrase were already dropped upstream, which removed speech-act labels preferentially — so نصيحة/ذم-derived families (التوصية والتحذير) are thinner here than in the source corpus.

Licensing

The underlying reviews were collected from Jeeran; this repository does not assert a license over them. Consult the source terms before redistribution or commercial use.