Title:	USING NEURAL MACHINE TRANSLATION FOR DETECTING AND CORRECTING GRAMMATICAL ERRORS
Author(s):	Dongqiang Yang, Xiaodong Sun and Pikun Wang
ISBN:	978-989-8704-34-4
Editors:	Pedro Isaías and Hans Weghorn
Year:	2021
Edition:	Single
Type:	Full
First Page:	11
Last Page:	18
Language:	English
Cover:
Full Contents:	click to dowload
Paper Abstract:	Computer assisted language learning can help ESL/EFL learners facilitate their writings in multiple ways such as spell checking, grammar checking, and style checking. Owning to the complexity of various linguistic errors intertwining in a sentence, it is still a challenging task to detect and correct grammatical errors automatically. Different from previous studies on using pattern matching or statistical language models on this task, we design a Transformer-based neural sequence transduction model to detect and correct grammatical errors. Neural language models are often data-hungry and their performance is also data-dependent. Given the limited size of standard learner corpora and its enormous annotating cost, we employ another Transformer-based encoder-decoder structure to back-translate an error-free sentence into an erroneous one, automating data augmentation for training neural models. We first design some artificial rules to produce a noisy learner dataset to train the back-translation model. The model can then generate more synthesized learner data for training the Transformer-based correction model. In addition to that we also propose an iterative training scheme that unifies the process of error generation and correction. Our state-of-the-art model can reach F0.5, the harmony vale of precision and recall, at 64.3% in the shared task of CoNLL-2014, surpassing all the participating systems.
	Go Back