Dynamic decoding and dual synthetic data for automatic correction of grammar in low-resource scenario

Area of Science:

• Natural Language Processing (NLP)
• Computational Linguistics

Background:

• Grammar error correction (GEC) is vital for NLP and communication.
• Neural machine translation (NMT) shows promise but struggles with data scarcity for low-resource languages like Indonesian.

Purpose of the Study:

• To develop an effective GEC system for Indonesian, addressing data scarcity and complexity.
• To introduce InSpelPoS, a confusion method combining synthetic data generation techniques.

Main Methods:

• Developed InSpelPoS, integrating Inverted Spellchecker and Patterns+POS for synthetic data generation.
• Adapted a seq2seq framework with dynamic decoding and Transformer models for enhanced GEC.
• Leveraged contextual information for accurate error identification and correction.

Main Results:

• Achieved significant improvements in Indonesian GEC accuracy using synthetic data.
• Demonstrated superior performance compared to existing GEC systems.
• Validated the effectiveness of the dynamic decoding method in handling diverse error types.

Conclusions:

• The proposed InSpelPoS framework and adapted seq2seq model offer a robust solution for Indonesian GEC.
• The method effectively overcomes challenges posed by low-resource languages and data scarcity.
• This research advances GEC capabilities, particularly for under-resourced linguistic contexts.