update dataset card for ATE v2 schema

README.md

@@ -2,7 +2,6 @@
 license: cc-by-sa-4.0
 task_categories:
   - token-classification
-  - text-generation
 language:
   - vi
 size_categories:
@@ -10,86 +9,93 @@ size_categories:
 tags:
   - vietnamese
   - reviews
+  - aspect-term-extraction
   - aspect-based-sentiment
-  - cause-extraction
   - e-commerce
-pretty_name: CausaSent
+  - absa
+pretty_name: CausaSent ATE v2
 ---
 
-# CausaSent — Vietnamese Review Causal Sentiment Dataset
+# CausaSent ATE v2 — Vietnamese E-commerce Aspect Term Extraction + ABSA
 
-Per-review tuples of `(aspect, sentiment, cause_span, action)` for Vietnamese e-commerce reviews.
+Vietnamese e-commerce reviews labeled with **aspect terms (span)**, **aspect category**, and **binary sentiment** — combined from a small manually-annotated gold pool (hotel + restaurant + smartphone reviews) and a larger LLM-labeled silver pool (Tiki product reviews).
 
 ## Schema
 
-Each row in `gold/{train,val,test}.json` is:
+Each row in `train.json` / `val.json` / `test.json` is:
 
 ```json
 {
-  "id": "string",
-  "review": "Ship lâu nhưng đóng gói đẹp",
+  "id": "tiki-14075703",
+  "source": "tiki",
+  "review": "Hàng đóng gói không cẩn thận, hộp méo mó, rách nhiều.",
   "annotations": [
     {
-      "aspect": "delivery",
-      "sentiment": "negative",
-      "cause_text": "Ship lâu",
-      "cause_span": [0, 8],
-      "action": "Rút ngắn thời gian giao hàng"
+      "aspect_term": "Hàng",
+      "aspect_term_span": [0, 4],
+      "aspect_category": "appearance",
+      "sentiment": "negative"
     },
     {
-      "aspect": "packaging",
-      "sentiment": "positive",
-      "cause_text": "đóng gói đẹp",
-      "cause_span": [15, 27],
-      "action": "Duy trì cách đóng gói"
+      "aspect_term": "đóng gói",
+      "aspect_term_span": [5, 13],
+      "aspect_category": "packaging",
+      "sentiment": "negative"
     }
   ]
 }
 ```
 
-Hard invariants on every annotation (validated at load time):
+**Hard invariants:**
 
-- `review[cause_span[0]:cause_span[1]] == cause_text` (exact substring; Python codepoint indexing).
-- `aspect ∈ {delivery, packaging, product_quality, price, customer_service, usability, appearance}`.
-- `sentiment ∈ {positive, negative, neutral}`.
-- Within a review, no two `cause_span`s overlap.
-- `action` is imperative, verb-first, ≤10 Vietnamese words.
+- `review[aspect_term_span[0]:aspect_term_span[1]] == aspect_term` byte-for-byte.
+- `aspect_category` ∈ {`delivery`, `packaging`, `product_quality`, `price`, `customer_service`, `usability`, `appearance`} — closed 7-class taxonomy.
+- `sentiment` ∈ {`positive`, `negative`} — neutral was dropped (only 2.7% of gold, too few to train).
+- `source` ∈ {`gold`, `tiki`}.
 
 ## Splits
 
 | Split | Reviews | Annotations |
-|-------|---------|-------------|
-| train | TBD     | TBD         |
-| val   | TBD     | TBD         |
-| test  | TBD     | TBD         |
+|---|---|---|
+| train | 5,647 | 8,251 |
+| val | 700 | 1,033 |
+| test | 719 | 1,023 |
+| **total** | **7,066** | **10,307** |
 
-Numbers populated after `scripts/split_dataset.py` runs.
+Stratified 80/10/10 by dominant `(aspect_category, sentiment)` per review.
 
-## Annotation process
+Aspect coverage in train (annotation count):
 
-1. Raw reviews collected from public Vietnamese review datasets (UIT-VSFC, ViSFD, Foody) plus a small manual export of Shopee / Google Maps reviews (ToS-safe, copy-paste only).
-2. Weak-labeled with Gemini 2.5 Flash via `src/data/weak_label.py` (closed-taxonomy prompt, JSON mode, exact-substring constraint).
-3. Validated by `src/data/validate.py` (drops hallucinated spans, bad aspects, duplicates, overlaps).
-4. Manually refined to ≥1000 gold samples per [`docs/annotation-guide.md`](https://github.com/tamir39/causa-sent/blob/develop/docs/annotation-guide.md) via the Gradio review UI in `src/tools/review_ui.py`.
-5. 80/10/10 splits by review id with seed 42.
+| aspect_category | count |
+|---|---|
+| appearance | 2,789 |
+| product_quality | 2,133 |
+| delivery | 982 |
+| packaging | 896 |
+| price | 922 |
+| customer_service | 424 |
+| usability | 105 |
 
-## Intended use
+Sentiment leans negative for `appearance` / `packaging` (Tiki complaints) and positive for `price` / `customer_service`.
 
-- Train aspect-sentiment + cause extraction models (e.g., PhoBERT-large two-head tagger).
-- Train action generation models (e.g., mT5-base) conditioned on extracted tuples.
-- Benchmark fine-grained ABSA + causal-rationale tasks on Vietnamese.
+## Data sources
 
-## License
+- **Gold (1,132 reviews, 2,205 annotations):** subset of CausaSent original manual annotations (drop `neutral`, drop `cause_span`/`action` fields). Schema converted via Claude-assisted aspect-term span extraction; 3 NFD-encoding edge cases dropped. See [Tamir39/causasent](https://huggingface.co/datasets/Tamir39/causasent) for the pre-pivot 4-tuple version.
+- **Tiki silver (5,934 reviews, 8,102 annotations):** Vietnamese Tiki product reviews originally labeled with `aspect_category` + `sentiment` by an LLM (mid-term project); we added `aspect_term_span` via the same Claude pipeline. 14 records dropped on span verification.
 
-Released under **CC-BY-SA-4.0**. Derivative works must share-alike. Underlying public review corpora retain their original licenses; respect those when redistributing.
+## Annotation pipeline
 
-## Citation
+Both sources passed through the same Claude-based annotation pipeline (`scripts/annotate_aspect_terms.py` orchestrator → `scripts/SUBAGENT_PROMPT.md` per-batch worker → `scripts/verify_annotation_batch.py` byte-match check → `scripts/merge_annotation_batches.py`). 99.86% of gold and 99.83% of Tiki survived span verification.
 
-```bibtex
-@misc{causasent2026,
-  title  = {CausaSent: Causal and Actionable Sentiment Analysis for Vietnamese Reviews},
-  author = {Phí Vương Tường Tâm},
-  year   = {2026},
-  howpublished = {\url{https://huggingface.co/datasets/Tamir39/causasent}}
-}
-```
+A stratified sample of 300 gold records was independently re-validated; inter-annotator agreement ≈ 91% (judged term matches LLM's pick).
+
+## Intended use
+
+Joint aspect-term extraction + aspect-category-aware sentiment classification on Vietnamese e-commerce text. Suitable for:
+- Training PhoBERT/XLM-R BIO span taggers with auxiliary sentiment head.
+- Comparing fine-tuned models vs. zero-shot LLM baselines on Vietnamese ATE.
+- ABSA-informed action recommendation pipelines (per the paper this dataset accompanies).
+
+## License & ToS
+
+Released under CC-BY-SA-4.0. The Tiki subset comes from publicly-displayed product reviews; we redistribute only the text and our derived annotations. If you use this dataset, please cite the original CausaSent project.