Enhancing African low-resource languages: Swahili data for language modelling.

Area of Science:

• Computational Linguistics
• African Languages
• Natural Language Processing

Background:

• Neural network-based language modeling requires substantial data for effective word representation in Natural Language Processing (NLP).
• African languages, like Swahili, are often classified as low-resource languages due to insufficient data for NLP tasks.
• Existing NLP resources are scarce for Swahili, hindering its development and application.

Purpose of the Study:

• To address the data scarcity for Swahili in NLP by creating and contributing new datasets.
• To provide essential resources for improving Swahili language models and other NLP applications.
• To support research and development for low-resource languages in the field of NLP.

Main Methods:

• An unannotated Swahili dataset was derived through the pre-processing of raw Swahili text using a Python script.
• A Swahili syllabic alphabet was formulated.
• A Swahili word analogy dataset was developed, drawing inspiration from an existing English dataset.

Main Results:

• The creation and contribution of three novel Swahili datasets: an unannotated dataset, a syllabic alphabet, and a word analogy dataset.
• Demonstrated a method for generating NLP resources for low-resource languages.
• Provided foundational data for advancing Swahili NLP.

Conclusions:

• The newly derived Swahili datasets are expected to significantly benefit Swahili language models.
• These resources will also support various downstream NLP tasks, including part-of-speech tagging, machine translation, and sentiment analysis.
• This work contributes to bridging the resource gap for low-resource African languages in NLP.