Corpus creation and language identification for code-mixed Indonesian-Javanese-English Tweets.

Area of Science:

• Computational Linguistics
• Natural Language Processing (NLP)
• Social Media Analysis

Background:

• Code-mixing, the phenomenon of language mixing, is increasingly prevalent in social media text.
• This linguistic trend poses significant challenges for Natural Language Processing (NLP) tasks, particularly Language Identification (LID).
• Existing LID models often struggle with the complexities of code-mixed data, necessitating specialized approaches.

Purpose of the Study:

• To develop and evaluate a word-level language identification model for code-mixed Indonesian, Javanese, and English tweets.
• To introduce a novel annotated corpus, the Indonesian-Javanese-English Language Identification (IJELID) dataset, for training and evaluation.
• To compare the effectiveness of different modeling strategies, including fine-tuned BERT, BLSTM, and CRF, for this specific task.

Main Methods:

• Creation of the Indonesian-Javanese-English Language Identification (IJELID) corpus, detailing data collection and annotation standards.
• Investigation of various machine learning approaches for word-level language identification.
• Fine-tuning of BERT-based models (IndoBERTweet), Bidirectional Long Short-Term Memory (BLSTM) networks, and Conditional Random Fields (CRF).

Main Results:

• Fine-tuned IndoBERTweet models demonstrated superior performance in identifying languages within code-mixed Indonesian, Javanese, and English tweets compared to BLSTM and CRF.
• The effectiveness of BERT models is attributed to their ability to capture contextual information of each word within a text sequence.
• Sub-word language representation within BERT models proved to be a reliable mechanism for language identification in code-mixed text.

Conclusions:

• Fine-tuned IndoBERTweet models offer a robust solution for language identification in code-mixed social media text.
• The developed IJELID corpus provides a valuable resource for future research in code-mixed language identification.
• BERT's contextual understanding and sub-word representation capabilities are key to successfully addressing the challenges of code-mixing in NLP.