language:
- it
license: cc-by-nc-sa-4.0
dataset_info:
features:
- name: id
dtype: string
- name: text
dtype: string
- name: score
dtype: float64
splits:
- name: train
num_bytes: 486202
num_examples: 864
- name: test
num_bytes: 117482
num_examples: 200
download_size: 410135
dataset_size: 603684
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
- split: test
path: data/test-*
DisCoTEX - Coherence
Disclaimer: This dataset is not the official DisCoTEX repository from EVALITA. For the official repository and more information, please visit the EVALITA DiscoTEX page or the DiscoTEX repository.
Overview
The DisCoTEX dataset is part of the DisCoTEX shared task at EVALITA 2023, aimed at assessing discourse coherence in Italian texts. This dataset focuses on Italian real-world texts and provides resources to model coherence in natural language. The dataset comprises texts from two distinct sources: Italian Wikipedia and the Multilingual TEDx corpus (mTEDx), representing different language varieties and genres.
Dataset Description
Structure:
The dataset consists of text passages containing four consecutive sentences, used as the unit of analysis. For Wikipedia, paragraphs were used to select four-sentence segments. For TEDx, transcripts were automatically split into four-sentence passages. The dataset is balanced between original and artificially modified texts. The artificially modified dataset includes passages modified by either shuffling the sentence order or substituting one sentence with another from the same document. Human ratings of coherence were collected for each passage via a crowdsourcing platform, with each passage rated by at least 10 native Italian speakers on a 1-5 Likert scale.
Format:
The dataset is composed of:
id: Identifier for the entry.text: A snippet of text (4 sentences).score: Mean coherence score based on human judgments.