We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:


Current browse context:


Change to browse by:

References & Citations

DBLP - CS Bibliography


(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Computer Science > Computation and Language

Title: What Have Been Learned & What Should Be Learned? An Empirical Study of How to Selectively Augment Text for Classification

Abstract: Text augmentation techniques are widely used in text classification problems to improve the performance of classifiers, especially in low-resource scenarios. Whilst lots of creative text augmentation methods have been designed, they augment the text in a non-selective manner, which means the less important or noisy words have the same chances to be augmented as the informative words, and thereby limits the performance of augmentation. In this work, we systematically summarize three kinds of role keywords, which have different functions for text classification, and design effective methods to extract them from the text. Based on these extracted role keywords, we propose STA (Selective Text Augmentation) to selectively augment the text, where the informative, class-indicating words are emphasized but the irrelevant or noisy words are diminished. Extensive experiments on four English and Chinese text classification benchmark datasets demonstrate that STA can substantially outperform the non-selective text augmentation methods.
Subjects: Computation and Language (cs.CL)
Cite as: arXiv:2109.00175 [cs.CL]
  (or arXiv:2109.00175v1 [cs.CL] for this version)

Submission history

From: Biyang Guo [view email]
[v1] Wed, 1 Sep 2021 04:03:11 GMT (789kb,D)

Link back to: arXiv, form interface, contact.