README.md

# TRABL: Travel-Domain Aspect-Based Sentiment Analysis Dataset

This repository contains the **TRABL dataset**, released in support of our paper accepted to **The ACM Web Conference 2026 (WWW 2026)**:

> **TRABL: A Unified Framework for Travel Domain Aspect-Based Sentiment Analysis Applications of Large Language Models**

The dataset is designed to support research on **Aspect-Based Sentiment Analysis (ABSA)** in the travel domain, with a particular focus on *joint extraction* of structured sentiment tuples and *supporting evidence snippets*.

---

## 1. Overview

TRABL extends the standard ABSA *quad prediction* task by additionally requiring the extraction of **textual evidence snippets** that justify each extracted sentiment tuple. The dataset enables training and evaluation of models for both fine-grained opinion mining and downstream travel applications such as recommendation, review summarization, and property type detection.

The data consists of **English user reviews** from three travel-related domains:

* **Hotels**
* **Attractions**
* **Destinations**

Each review is annotated by **two annotators**, following carefully designed guidelines and quality control procedures.

---

## 2. Task Definition

Given a review text (T), the task is to extract a set of tuples of the form:

```
(aspect_term, aspect_category, opinion_span, sentiment, snippet)
```

Where:

* **Aspect Term**: Explicit mention of the target entity (e.g., *"room"*)
* **Aspect Category**: Semantic category from a **closed set of 112 travel-domain categories**
* **Opinion Span**: Evaluative expression (e.g., *"spacious"*)
* **Sentiment**: One of `{positive, negative, neutral}`
* **Snippet**: A *contiguous span* of text from the review that provides sufficient evidence for the other fields

Some fields may be `null` if not explicitly mentioned (e.g., implicit aspects).

This formulation generalizes multiple ABSA subtasks, including:

* Category extraction `(c)`
* Category–sentiment extraction `(c, p)`
* Category–sentiment with evidence `(c, p, s)`
* Standard quad prediction `(a, o, c, p)`
* Quad prediction with snippet extraction `(a, o, c, p, s)`

---

## 3. Dataset Splits

The dataset is split into **train**, **validation**, and **test** sets as follows:

| Split      | #Rows |
| ---------- | ----- |
| Train      | 5,768 |
| Validation | 898   |
| Test       | 900   |

> Note: A *row* corresponds to a single annotated review instance (including annotator-specific annotations).

---

## 4. Data Format

The data is released in **JSONL** format (one JSON object per line).

Each example contains the following top-level fields:

* **review_id**: Unique identifier for the review
* **annotator**: Identifier of the annotator (`1` or `2`)
* **text**: Full review text

  * Hotel reviews include title, positive, and negative sections
  * Attraction and destination reviews contain free text only
* **labels**: List of labeled tuples

Each entry in `annotations` has the structure:

```json
{
  "aspect_term": "value for money",
  "aspect_category": "value for money",
  "opinion_span": "good",
  "sentiment": "positive",
  "snippet": "The good value for money"
}
```

---

## 5. Example

**Input Review (excerpt):**

```
The good value for money, the helpful & polite staff, the level of cleanliness.
```

**Annotations (excerpt):**

```json
[
  {
    "aspect_term": "value for money",
    "aspect_category": "value for money",
    "opinion_span": "good",
    "sentiment": "positive",
    "snippet": "The good value for money"
  },
  {
    "aspect_term": "staff",
    "aspect_category": "property staff and service support",
    "opinion_span": "helpful & polite",
    "sentiment": "positive",
    "snippet": "the helpful & polite staff"
  }
]
```

---

## 6. Annotation Process and Quality

* Each review was annotated by **two annotators**
* Annotation was supported by LLM-generated initial labels, followed by human correction
* Disagreements were normalized using post-processing rules
* **Overall annotator agreement** (F1): ~67%, consistent with prior ABSA benchmarks
* Agreement exceeds **80%** when evaluated on partial field subsets (e.g., category or sentiment only)

---

## 7. Statistics

* **Domains**: Hotels, Attractions, Destinations
* **Total number of reviews**: 3,783
* **Average review length**: ~42 words
* **Average number of extracted tuples per review**:

  * Annotator 1: ~6.96
  * Annotator 2: ~6.77
* **Sentiment distribution**: Skewed positive, reflecting real-world travel review behavior

---

## 8. Intended Use

This dataset is intended for **research purposes only**, including but not limited to:

* Aspect-Based Sentiment Analysis
* Opinion mining and evidence extraction
* Travel-domain NLP
* Evaluation of LLMs and lightweight fine-tuned models
* Structured information extraction with explainability

---

## 9. Citation

If you use this dataset, please cite:

```
@inproceedings{madmon2026trabl,
  title     = {TRABL: A Unified Framework for Travel Domain Aspect-Based Sentiment Analysis Applications of Large Language Models},
  author    = {Madmon, Omer and Golan, Shiran and Fainman, Eran and Kleinfeld, Ofri and Beladev, Moran},
  booktitle = {Proceedings of The ACM Web Conference},
  year      = {2026}
}
```

---

## 10. License and Privacy

* All reviews were processed under strict internal privacy and PII guidelines
* The dataset is released for **non-commercial research use**, cc-by-sa-4.0 license
* No personally identifiable information is included

---

For questions or issues related to the dataset, please refer to the paper or open an issue on the dataset repository.