language: it
license: cc-by-sa-4.0
multilinguality: monolingual
task_categories:
- token-classification
tags:
- Frame Parsing
- Event Extraction
EventNet-ITA - Dataset
Dataset Description
Dataset Summary
EventNet-ITA is a textual dataset annotated full-text with semantic frames in Italian. It can be used to train multi-label models for Frame Parsing or Event Extraction. The schema consists of 205 semantic frames (and relative frame elements) and covers different macro-domains, like conflictual, social, communication, legal, geopolitical, economic and biographical events, among others. The dataset counts 53,843 annotated sentences and over 1,583,000 tokens. For more details, please refer to the paper. If you want to requests the full documentation of the resource (guidelines, detailed frame-level description, lexical units and annotation examples), please fill out this form or email the author.
Annotation Process
EventNet-ITA has been annotated at token level, adopting the IOB2 style. The annotation is full-text, i.e., for each sentence any frame mention and all relative frame elements (provided in the schema) are annotated. Example:
La O B-EVENT O
costruzione B-BUILDING I-EVENT O
della B-CREATED_ENTITY I-EVENT O
fortificazione I-CREATED_ENTITY I-EVENT O
alvitana I-CREATED_ENTITY I-EVENT O
risale O B-TEMPORAL_ORIGIN O
dunque O O O
all' O B-ORIGIN O
epoca O I-ORIGIN O
dell' O I-ORIGIN O
invasione O I-ORIGIN B-INVADING
normanna O I-ORIGIN B-INVADER
. O O O
By convention, in the dataset frame elements are represented as a concatenation of their label name with the name of the corresponding frame. For example, the CREATED_ENTITY frame element, associated to the BUILDING frame, will be represented as CREATED_ENTITY*BUILDING.
Data format
The dataset is formatted as a two-column tsv. The first column contains the token, the second column contains all corresponding labels (both frames and frame elements), separated by |. This format makes the dataset ready-to-train with the MaChAmp multi-sequence task type. Please see the model page for more details about training.
Data Split
For the sake of reproducibility, the three folds used in the paper are provided. The data split follows a 80/10/10 ratio and has been created in a stratified way. This means each train/dev/test set contains the same relative distribution of (frame) classes.
Additional Information
Licensing Information
The EventNet-ITA dataset is released under the CC-BY-SA-4.0 License.
Citation Information
If you use EventNet-ITA, please cite the following paper:
@inproceedings{rovera-2024-eventnet,
title = "{E}vent{N}et-{ITA}: {I}talian Frame Parsing for Events",
author = "Rovera, Marco",
editor = "Bizzoni, Yuri and
Degaetano-Ortlieb, Stefania and
Kazantseva, Anna and
Szpakowicz, Stan",
booktitle = "Proceedings of the 8th Joint SIGHUM Workshop on Computational Linguistics for Cultural Heritage, Social Sciences, Humanities and Literature (LaTeCH-CLfL 2024)",
year = "2024",
publisher = "Association for Computational Linguistics",
pages = "77--90",
}