Development of the Annotated Swahili Digraph Corpus Using a CNN-Based Digraph Extraction Model

Abstract

This study undertakes the development of the Annotated Swahili Digraph Corpus, utilizing a convolutional neural network-based model specifically designed for the extraction of digraphs. This initiative addresses a significant gap in the availability of dedicated digraph corpora for the Swahili language, which is increasingly needed for various applications in Natural Language Processing (NLP). The CNN-based model was accurately crafted to optimize the extraction and classification of digraphs, taking full advantage of the annotated features within the corpus. Digraphs are pairs of letters that create distinct sounds in a language, and Swahili`s linguistic structure presents unique challenges and requirements in this regard. Therefore, specialized tools and models are essential for ensuring accurate transcription and efficient speech recognition that cater specifically to the nuances of the Swahili language. The resulting Swahili Digraph Corpus comprises a comprehensive collection of 31,197 words, each systematically annotated to highlight their respective digraphs. Notably, this corpus features the nine key Swahili digraphs: "ch," "dh," "gh," "kh," "ng’," "ny," "sh," "th," and "ng." Furthermore, it includes annotations for vowel distribution, showcasing the core vowels "a," "e," "i," "o," and "u." This detailed annotated corpus supports a wide array of NLP applications, enabling researchers and developers to utilize accurate linguistic data for tasks such as text processing, machine translation, and speech synthesis. Through this dedicated effort, we aim to enhance the resources available for processing the Swahili language, ultimately contributing to its greater accessibility in the digital landscape.