NEO-NDA: Neo Natural Language Data Augmentation

Data augmentation generates synthetic data by making changes in data already obtained. It is applied to distinct data types like images, audio, and text. For textual data augmentation, many works propose restrictive transformations, for instance, they only work with one language (monolingual) or cre...

Celý popis

Uloženo v:
Podrobná bibliografie
Vydáno v:2022 IEEE 16th International Conference on Semantic Computing (ICSC) s. 99 - 102
Hlavní autoři: Ladeira, Lucas Z., Santos, Frances, Cleopas, Lucas, Buteneers, Pieter, Villas, Leandro
Médium: Konferenční příspěvek
Jazyk:angličtina
Vydáno: IEEE 01.01.2022
Témata:
Conferences
data augmentation
Data models
hybrid-approach
multilingual
Natural languages
Semantics
Switches
text classification
On-line přístup:Získat plný text
Tagy: Přidat tag
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
Popis
Shrnutí:Data augmentation generates synthetic data by making changes in data already obtained. It is applied to distinct data types like images, audio, and text. For textual data augmentation, many works propose restrictive transformations, for instance, they only work with one language (monolingual) or create samples with a fixed length [1]-[4]. In this work, we propose NEO Natural language Data Augmentation (NEO-NDA), a more comprehensive tool able to address data generation and rebalancing datasets. It supports data augmentation of minority classes. NEO-NDA is able to work with multiple languages, besides implementing distinct transformations to create new samples. Our results show that NEO-NDA was able to boost the performance of ML models with all datasets evaluated and, in some cases, doubling the performance in comparison with original datasets when no data augmentation method is used.
DOI:10.1109/ICSC52841.2022.00021