Improving Loanword Identification in Low-Resource Language with Data Augmentation and Multiple Feature Fusion

Area of Science:

• Natural Language Processing (NLP)
• Computational Linguistics
• Low-Resource Language Technologies

Background:

• Loanword identification is crucial for NLP tasks like machine translation, especially for low-resource languages where data scarcity hinders performance.
• Existing research often focuses on high-resource languages, leaving low-resource languages like Uyghur and Mongolian with suboptimal loanword identification models.

Purpose of the Study:

• To address the challenge of poor loanword identification performance in low-resource languages.
• To develop a novel data augmentation technique and a robust identification model tailored for languages with limited annotated data.

Main Methods:

• A lexical constraint-based data augmentation method was proposed to generate synthetic training data for low-resource loanword identification.
• A log-linear recurrent neural network (RNN) model was developed, integrating word and character embeddings, pronunciation similarity, and part-of-speech (POS) features.

Main Results:

• The proposed data augmentation method effectively increased the available training data for Uyghur loanword identification.
• The log-linear RNN model incorporating multiple features achieved superior performance in identifying Arabic, Chinese, Russian, and Turkish loanwords in Uyghur.
• Experimental results demonstrated that the proposed approach outperformed several strong baseline systems.

Conclusions:

• The developed data augmentation and log-linear RNN model offer a significant advancement for loanword identification in low-resource languages.
• This research provides a practical solution to overcome data limitations and improve NLP task performance for under-resourced languages like Uyghur.