Advancing Language Detection for Ethiopia’s Omotic Languages

TLDR: This research introduces a new method for identifying and distinguishing between two closely related Omotic languages, Wolayta and Gofa, at the word level in mixed texts. The study developed a dataset and found that a combination of a Bert-based pre-trained language model and an LSTM approach achieved the best performance with an F1-score of 0.72. This work is crucial for addressing social media issues and improving communication in multilingual Ethiopian communities.

Language identification (LID) is the process of automatically determining the language of a given text. While traditionally assumed that texts are written in a single language, real-world scenarios, especially in multilingual communities, often involve text containing more than one language. This is where Bilingual Language Identification (BLID) becomes crucial: the task of identifying and distinguishing between two languages within a single text.

A recent research paper, Bilingual Word-Level Language Identification for Omotic Languages, delves into this challenging area, focusing on two closely related languages spoken in the southern part of Ethiopia: Wolayta and Gofa. These languages, part of the Omotic family, present a unique challenge due to their significant similarities and subtle differences, making automated distinction difficult.

The researchers, Mesay Gemeda Yigezu, Girma Yohannis Bade, Atnafu Lambebo Tonja, Olga Kolesnikova, Grigori Sidorov, and Alexander Gelbukh, aimed to overcome these challenges by experimenting with various approaches. Their work is particularly significant because, to their knowledge, no prior research had focused on LID for the Omotic language family of Ethiopian languages. Given that Ethiopian Omotic language groups often use Latin scripts with similar characteristics, identifying one language from another requires careful attention to minor distinctions, often stemming from spoken accents.

Why This Research Matters

The study highlights several key contributions and potential improvements:

• Creation of a word-level open-source dataset for Wolayta and Gofa, which are low-resource languages.
• The proposed model can be applied to other low-resource Ethiopian languages.
• It offers a tool to identify the origin of hate speech, rumors, conspiracy theories, racism, and prejudice on social media.
• Improved communication by reducing language barriers, allowing systems to quickly identify languages without manual specification.
• Enhanced customer service for businesses by supporting a wider range of customers.
• Increased accessibility of technology and applications to a broader audience.

The Languages: Wolayta and Gofa

Ethiopia is a linguistically diverse nation with over 83 distinct languages. The Omotic language family, primarily spoken in the southern rift valley, includes Wolayta and Gofa under its Northern Omotic branch. Wolayta is spoken by approximately 2.48 million people, while Gofa has over 362,000 speakers. Both are used in education and government within their respective zones. The geographical proximity of their speakers and their shared linguistic group contribute to the structural and morphological similarities between the two languages, making BLID a complex task.

Methodology and Experiments

The research involved a comprehensive methodology:

• Dataset Development: A novel dataset was collected from religious, educational, and social media sources.
• Data Pre-processing: The collected data underwent cleaning, including removing URLs, non-alphabetic characters, numbers, and HTML tags, and converting text to lowercase.
• Data Annotation: Native speakers manually annotated the dataset, tagging words as ‘wal’ (Wolayta), ‘gof’ (Gofa), or ‘wal-gof’ (shared words). A majority vote among three annotators determined the final label.

The researchers conducted experiments using four categories of approaches:

1. Machine Learning (ML): Logistic Regression was applied.
2. Deep Learning (DL): Techniques like LSTM (Long Short-Term Memory) with attention, BiLSTM (Bidirectional LSTM) with attention, and CNN (Convolutional Neural Network) were explored.
3. Combination of DL: CNN was used for feature extraction, followed by LSTM and BiLSTM as classifiers.
4. Combination of Pre-trained Language Models (PLM) and DL: Transformer-based models such as BERT-based uncased and RoBERTa-based were used as embedding layers with an LSTM model.

Key Findings

The experimental results showed varying levels of performance across the different approaches. The most significant finding was that the combination of a Bert-based pre-trained language model and an LSTM approach performed the best, achieving an F1-score of 0.72 on the test set. This model effectively learned the language patterns in the bilingual text dataset, allowing it to distinguish between different languages even with short texts and limited information.

While other deep learning combinations like CNN + BiLSTM also showed good performance (F1-score of 0.69), they were slightly outperformed by the Bert-based model. This suggests that leveraging pre-trained language models is a highly effective strategy for identifying languages in bilingual texts, especially for low-resource languages like Wolayta and Gofa.

Also Read:

• Bridging Morphological Gaps: MoVoC for Ge’ez Script NLP
• Unpacking LLM Effectiveness: Fine-Tuning Outperforms In-Context Learning for Misinformation Detection

Conclusion and Future Directions

This study provides a strong foundation for future research in bilingual LID for Omotic languages. The success of the Bert-based uncased model combined with LSTM highlights the potential of advanced deep learning and transformer models in this domain. Future work could focus on developing even more sophisticated algorithms, expanding the size and diversity of training data, and exploring new feature representations. Additionally, the accurate detection of source languages through such models has direct applications in improving machine translation tasks, ensuring more precise and contextually relevant translations.