Datasets:

MBZUAI
/

dialseg-ar

language:
  - ar
license: cc-by-nc-4.0
task_categories:
  - text-classification
pretty_name: DialSeg-Ar
size_categories:
  - 1K<n<10K
tags:
  - linear_text_segmentation
configs:
  - config_name: default
    data_files:
      - split: eval
        path: data/eval-*
dataset_info:
  features:
    - name: subset
      dtype: string
    - name: arabic_dialect
      dtype: string
    - name: genre
      dtype: string
    - name: file_name
      dtype: string
    - name: lines
      dtype: string
    - name: segmentation
      dtype: string
  splits:
    - name: eval
      num_bytes: 6781199
      num_examples: 1010
  download_size: 1975170
  dataset_size: 6781199

DialSeg-Ar

Dataset Summary

DialSeg-Ar is a multi-genre benchmark for linear semantic segmentation in Arabic, with a focus on dialectal conversational and transcribed speech. The dataset is designed to evaluate how well models can split a sequence of utterances into contiguous topic-coherent segments.

The benchmark covers diverse and underrepresented Arabic genres, including:

dialectal telephone conversations
code-switched Gulf Arabic–English podcasts
dialectal emotional dialogue from novels
Modern Standard Arabic (MSA) news commentary

DialSeg-Ar is intended for research on:

linear text segmentation
discourse analysis
conversational structure modeling
segmentation for retrieval, summarization, and downstream discourse tasks
low-resource and dialectal Arabic NLP

This dataset was introduced in:

Chirkunov, K., Samih, Y., Freihat, A. A., & Aldarmaki, H. (2026). Linear Semantic Segmentation for Low-Resource Spoken Dialects. In Proceedings of ACL 2026.

Supported Tasks and Leaderboards

DialSeg-Ar supports the task of linear semantic segmentation, also known as linear text segmentation.

Given an ordered sequence of utterances, the goal is to predict the positions where a new topic segment begins.

Typical evaluation metrics include:

Boundary F1
Pk
WindowDiff

Languages

The dataset primarily contains Arabic, including:

Modern Standard Arabic (MSA)
Moroccan Arabic
Gulf Arabic
Levantine Arabic
Iraqi Arabic

Some subsets also include English code-switching, especially in the podcast portion.

Dataset Structure

Each row corresponds to one JSONL filename and contains:

subset (string): subset name (for example ldc_gulf, mgb-5, rewayat).
arabic_dialect (string): language label.
genre (string): genre label (for example phone conversations, spontaneous speech, shows, drama).
file_name (string): matched JSONL file name.
lines (string): input text content in jsonl format.
segmentation (string): text segments in jsonl format.