---
license: apache-2.0
language:
- en
tags:
- sentiment
- wikipedia
- vader
- textblob
- distilbert
- roberta
- cvader
size_categories:
- 1M<n<10M
---

This dataset provides sentiment annotations for English Wikipedia articles using five different sentiment analysis models, as described in the study *“[Cross-Topic Sentiment Analysis of Wikipedia Articles: A Comparative Study of AI Models](https://doi.org/10.1007/978-3-032-18920-2_34)”* (see also the [supplementary materials](https://data.lewoniewski.info/sentiment/) for the study).

## Dataset Structure

Each record corresponds to a single Wikipedia article and contains sentiment labels produced by multiple models.

| Field     | Description                                                         |
| --------- | ------------------------------------------------------------------- |
| `page_id` | Unique identifier of the Wikipedia article                          |
| `title`   | Title of the Wikipedia article                                      |
| `model1`  | Sentiment label from cVADER (lexicon-based, compound score variant) |
| `model2`  | Sentiment label from DistilBERT (transformer-based)                 |
| `model3`  | Sentiment label from RoBERTa (transformer-based)                    |
| `model4`  | Sentiment label from TextBlob (lexicon-based)                       |
| `model5`  | Sentiment label from VADER (lexicon-based)                          |

Each model assigns one of the following labels:
* `pos` – positive
* `neu` – neutral
* `neg` – negative

## Data Collection and Processing

### Source Data
* English Wikipedia dump (March 2025)
* Only main namespace articles (no redirects)

### Preprocessing
* Articles were cleaned and extracted using [WikiExtractor](https://github.com/attardi/wikiextractor)
* Long articles were split into smaller text segments due to model input limits

### Sentiment Inference
* Each segment was analyzed independently by all models
* Transformer models output class probabilities
* Lexicon-based models output polarity scores mapped to classes

### Aggregation Method
Similarly to our [previous work](https://doi.org/10.1007/978-3-031-94193-1_8), we used the advanced aggregation method to obtain the overall sentiment assessment for whole Wikipedia article. Segment-level predictions were aggregated into article-level sentiment using a length-weighted averaging scheme. The final label for each model corresponds to the dominant sentiment class.

---

## Citation

If you use this dataset, please cite:

```bibtex
@InProceedings{10.1007/978-3-032-18920-2_34,
author="Lewoniewski, W{\l}odzimierz and Str{\'o}{\.{z}}yna, Milena and Czuma{\l}owska, Izabela and Wojewoda, Aleksandra and W{\k{e}}cel, Krzysztof",
title="Cross-Topic Sentiment Analysis of Wikipedia Articles: A Comparative Study of AI Models",
booktitle="Artificial Intelligence for Knowledge Acquisition and Management",
year="2026",
publisher="Springer Nature Switzerland",
address="Cham",
pages="364--370",
isbn="978-3-032-18920-2"
}
```