xtinge/turkish-extractive-summarization-dataset

---
dataset_info:
- config_name: mlsum_tr_ext
- config_name: xtinge-sum_tr_ext
- config_name: tes
configs:
- config_name: mlsum_tr_ext
data_files:
- split: train
path: MLSUM_TR_EXT/train*
- split: test
path: MLSUM_TR_EXT/test*
- split: val
path: MLSUM_TR_EXT/val*
- config_name: xtinge-sum_tr_ext
data_files:
- split: test
path: XTINGE-SUM_TR_EXT/XTINGE-SUM_TR_EXT*
- config_name: tes
data_files:
- split: test
path: TES/tes*
task_categories:
- summarization
license: gpl-3.0
---
# XTINGE Turkish Extractive Summarization Datasets

This repository hosts three datasets created for advancing Turkish extractive text summarization research: MLSUM_TR_EXT, TES, and XTINGE-SUM_TR_EXT. These datasets are designed to support the development of models capable of generating concise and relevant extractive summaries of Turkish texts.

Below is a Python example showcasing how to download and use these datasets:

```python
from datasets import load_dataset

# Load the MLSUM_TR_EXT dataset
mlsum_tr_ext = load_dataset("xtinge/turkish-extractive-summarization-dataset", "mlsum_tr_ext")
# Load the TES dataset
tes = load_dataset("xtinge/turkish-extractive-summarization-dataset", "tes")
# Load the xtinge-sum_tr_ext dataset
xtinge_sum_tr_ext = load_dataset("xtinge/turkish-extractive-summarization-dataset", "xtinge-sum_tr_ext")

```

## Dataset Details

### Dataset Description

The datasets, having a focus on Turkish text summarization, aim to advance research in this area by providing structured, annotated resources for extractive summarization tasks. These datasets are:

1. **MLSUM_TR_EXT**:
- Originates as an extension of the Turkish subset from the [MLSUM dataset](https://huggingface.co/datasets/mlsum), focusing on extractive summarization.
- Comprises articles from internethaber.com, with summaries derived from existing headlines for creating contextually rich extractive summaries.
- Sentences within these articles were selected based on their SBERT Similarity and ROUGE Scores compared to the original summaries, ensuring relevance and conciseness.

2. **TES**:
- Represents a unique collection found on [Hugging Face](https://huggingface.co/erturkerdagi/turkishExtractiveSummarization/tree/main) tailored for Turkish extractive summarization.
- Contains a variety of news articles annotated by three distinct annotators, each providing different perspectives and lengths, thus contributing to a rich set of summarization examples.

3. **XTINGE-SUM_TR_EXT**:
- Specifically developed to supplement existing resources by providing detailed sentence importance rankings within lengthy Wikipedia documents.
- Features annotations by three different annotators who meticulously ranked all sentences by importance, contributing to a comprehensive resource for studying extractive summarization.
- The annotation process considered Inter Annotator Agreement, specifically employing Krippendorff's alpha to ensure consistency and reliability in sentence importance assessments.

- **Language(s) (NLP):** Turkish
- **License:** [gpl-3.0]

## Dataset Structure

### Generic Structure Across Datasets

All three datasets share a generic structure tailored for extractive summarization tasks, comprising the following elements:

- **Title**: The title of the document or article, serving as a concise representation of the content.
- **Sentences**: The body of the text, split into sentences. This segmentation facilitates the identification of individual sentences that contribute to the summary.
- **Annotations**: This section includes annotations for selecting summary sentences. It is subdivided into:
- **Indexes**: Indices of sentences that have been selected for the summary. This field varies across datasets based on the number of annotators.
- **Ranking**: Rankings assigned to sentences based on their perceived importance for the summary. This feature is more prominent in datasets focusing on sentence importance ranking.

```python
{
'Title': '<title_of_document>',
'Sentences': ['<sentence_1>', '<sentence_2>', ..., '<sentence_n>'],
'Annotations': {
'Indexes': {

'Annotator1': [<index_of_selected_sentence_1>, ..., <index_of_selected_sentence_m>],
# If there are more than one annotator
'Annotator2': [...],
# etc.
},
'Ranking': {
'Annotator1': [<ranking_of_first_sentence>,<ranking_of_second_sentence>,..., <ranking_of_mth_sentence>],
# If there are more than one annotator
'Annotator2': [...],
# etc.
}
}
}
```

## Cite XTINGE Turkish Extractive Summarization Dataset
```
@inproceedings{xtinge_turkish_extractive,
title = {Extractive Summarization Data Sets Generated with Measurable Analyses},
author = {Demir, İrem and Küpçü, Emel and Küpçü, Alptekin},
booktitle = {Proceedings of the 32nd IEEE Conference on Signal Processing and Communications Applications},
year = {2024}
}
```

原始信息汇总

XTINGE Turkish Extractive Summarization Datasets

数据集概述

该数据集包含三个用于土耳其文本抽取式摘要研究的子数据集：MLSUM_TR_EXT、TES 和 XTINGE-SUM_TR_EXT。这些数据集旨在支持开发能够生成简洁且相关摘要的模型。

数据集详情

数据集描述

1. MLSUM_TR_EXT:

   • 源自 MLSUM 数据集 的土耳其子集扩展，专注于抽取式摘要。
   • 包含来自 internethaber.com 的文章，摘要由现有标题生成，以创建上下文丰富的抽取式摘要。
   • 文章中的句子根据与原始摘要的 SBERT 相似度和 ROUGE 分数进行选择，确保相关性和简洁性。

2. TES:

   • 代表一个独特的集合，可在 Hugging Face 上找到，专为土耳其抽取式摘要设计。
   • 包含由三位不同标注者标注的各种新闻文章，每个标注者提供不同的视角和长度，从而提供丰富的摘要示例。

3. XTINGE-SUM_TR_EXT:

   • 专门开发以补充现有资源，提供详细的句子重要性排序，适用于长篇维基百科文档。
   • 由三位不同的标注者仔细对所有句子按重要性进行排序，为研究抽取式摘要提供全面的资源。
   • 标注过程中考虑了标注者间的一致性，特别是使用 Krippendorffs alpha 来确保句子重要性评估的一致性和可靠性。

数据集结构

所有三个数据集共享一个适用于抽取式摘要任务的通用结构，包括以下元素：

• Title: 文档或文章的标题，作为内容的简洁表示。
• Sentences: 文本的主体，分割成句子。这种分割有助于识别对摘要有贡献的单个句子。
• Annotations: 包括选择摘要句子的标注。分为：
  • Indexes: 被选为摘要的句子索引。该字段因标注者数量而异。
  • Ranking: 根据句子对摘要的重要性分配的排名。这一特征在专注于句子重要性排序的数据集中更为突出。

python { Title: <title_of_document>, Sentences: [<sentence_1>, <sentence_2>, ..., <sentence_n>], Annotations: { Indexes: { Annotator1: [<index_of_selected_sentence_1>, ..., <index_of_selected_sentence_m>], # 如果有多个标注者 Annotator2: [...], # 等等 }, Ranking: { Annotator1: [<ranking_of_first_sentence>,<ranking_of_second_sentence>,..., <ranking_of_mth_sentence>], # 如果有多个标注者 Annotator2: [...], # 等等 } } }

引用

@inproceedings{xtinge_turkish_extractive, title = {Extractive Summarization Data Sets Generated with Measurable Analyses}, author = {Demir, İrem and Küpçü, Emel and Küpçü, Alptekin}, booktitle = {Proceedings of the 32nd IEEE Conference on Signal Processing and Communications Applications}, year = {2024} }