Native language identification from text using a fine-tuned GPT-2 model

Area of Science:

• Computational Linguistics
• Artificial Intelligence
• Natural Language Processing

Background:

• Native Language Identification (NLI) is crucial for applications like personalized language learning and machine translation.
• Existing methods often struggle with nuanced language variations.
• Advancements in deep learning offer potential for improved NLI accuracy.

Purpose of the Study:

• To investigate the efficacy of a fine-tuned GPT-2 model for Native Language Identification.
• To enhance the accuracy of classifying the native language of Portuguese learners.
• To compare the performance of GPT-2 against traditional machine learning and other pre-trained models.

Main Methods:

• Utilized the NLI-PT dataset for training and evaluation.
• Preprocessed text data including tokenization and embedding extraction.
• Fine-tuned the GPT-2 model using a multi-layer transformer-based classification approach.

Main Results:

• The fine-tuned GPT-2 model achieved a weighted F1 score of 0.9419 and an accuracy of 94.65%.
• Demonstrated superior performance compared to Support Vector Machines (SVM), Random Forest, BERT, RoBERTa, and BioBERT.
• Significant improvement in NLI accuracy was observed.

Conclusions:

• Large transformer models, like GPT-2, are highly effective for Native Language Identification.
• The fine-tuned GPT-2 model offers a robust solution for NLI tasks.
• Findings can inform the development of AI-driven personalized language learning tools.