---
language:
- ar
license: other
license_name: see-provenance
task_categories:
- token-classification
- text-classification
tags:
- aspect-based-sentiment-analysis
- absa
- arabic
- dialectal-arabic
- semeval-2016
---

# Jeeran — SemEval-2016 Task 5 (Arabic) adaptation

Aspect-based sentiment annotations over Jeeran reviews (Jordanian/Levantine
dialectal Arabic, 29 business domains), rendered in the **SemEval-2016 Task 5
subtask 1** format so that tooling written for `SemEval2016_arabic` runs unchanged.

## Contents

| split | reviews | sentences | opinions |
|---|--:|--:|--:|
| train | 43420 | 69795 | 182788 |
| test | 10851 | 17644 | 45520 |

Raw SemEval XML lives under `semeval_xml/`; the `datasets` view has one row per
sentence with a nested `opinions` list.

## Fields

Per opinion:

| field | description |
|---|---|
| `target` | opinion target expression — verbatim substring of `text` |
| `from` / `to` | character offsets of `target` in `text` |
| `category` | Jeeran aspect category (flat Arabic label, 80 values) |
| `polarity` | `positive` / `negative` / `neutral` |
| `span` / `span_from` / `span_to` | the original human-annotated evaluative span the target was extracted from |

## How it differs from SemEval-2016 Arabic hotels

* **`category` is not `ENTITY#ATTRIBUTE`.** SemEval's 34 `E#A` labels are
  hotel-specific; Jeeran spans many business domains, so the original flat Arabic
  aspect label is kept verbatim rather than forced into an invented entity scheme.
* **No implicit (`NULL`) targets.** SemEval marks implicit aspects with
  `target="NULL"`; here opinions without an explicit target noun phrase are dropped.
* **Sentences are opinion-bearing only.** Segments carrying no annotated span are
  not emitted, and segments are merged where a span straddles a boundary, so no
  span is ever split across sentences.

## Provenance and annotation quality

| layer | source |
|---|---|
| `span`, `span_from`/`span_to`, `polarity` | **human** annotation (inline `[[…]]` / `{…}` markup in the Jeeran corpus) |
| `category` | **model-generated** (Gemma) — known to be noisy |
| `target`, `from`/`to` | **model-extracted** from the human span (Gemma), accepted only when a verbatim contiguous substring of it |

### Target extraction outcome (286,957 candidate opinions)

| status | share | kept? |
|---|--:|---|
| target extracted and verified | 79.56% | yes |
| no explicit target noun phrase | 19.41% | dropped |
| model would not copy verbatim | 1.03% | dropped |

Extracted targets average **1.53 words** (SemEval-2016 Arabic gold: 1.20;
the Jeeran spans they came from: 5.18), so the target slot is comparable in shape
to SemEval rather than clause-like.

Every retained opinion is offset-verified: `text[from:to] == target` and the target
lies inside its own `span`. Verified on the released files, 0 exceptions.

Two caveats for anyone using this:

* Dropping targetless opinions is **not uniform across categories** — it removes
  speech-act labels preferentially (نصيحة 66%, ذم 36%, مدح عام 23%, vs الموقع 5%),
  so the category distribution differs from the source corpus.
* ~4.7% of targets are enumerations (brand or name lists) kept as a single target
  where SemEval convention would emit one opinion per item.

## 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.