This article discusses the contribution of experimental techniques to recording phonetic data in the field. Only a small part of the phonological systems of African languages is described with precision. This is why it is important to collect empirical data in the form of sound, video and physiological recordings. This allows research questions such as patterns of variation to be addressed. Analytical methods show how to interpret data from physical principles and integrate them into appropriate models. The question of linguistic contact between different language families is also addressed. To achieve these general objectives, we present the way we design corpora, and the different ways of recording data with crucial technical considerations during fieldwork. Finally, we focus on 3 languages spoken in the Great African Rift Zone, which includes several linguistic areas belonging to the four major linguistic families of the continent. (1) Hadza is a click language with a very complex consonant system. (2) Iraqw is a Cushitic language with ejective consonants. (3) Maasai is a Nilotic language with implosive consonants and a very elaborate set of interjections, ideophones and animal calls that include sounds not described in the International Phonetic Alphabet.

This work is part of the Kallaama project, whose objective is to produce and disseminate national languages corpora for speech technologies developments, in the field of agriculture. Except for Wolof, which benefits from some language data for natural language processing, national languages of Senegal are largely ignored by language technology providers. However, such technologies are keys to the protection, promotion and teaching of these languages. Kallaama focuses on the 3 main spoken languages by Senegalese people: Wolof, Pulaar and Sereer. These languages are widely spoken by the population, with around 10 million of native Senegalese speakers, not to mention those outside the country. However, they remain under-resourced in terms of machine-readable data that can be used for automatic processing and language technologies, all the more so in the agricultural sector. We release a transcribed speech dataset containing 125 hours of recordings, about agriculture, in each of the above-mentioned languages. These resources are specifically designed for Automatic Speech Recognition purpose, including traditional approaches. To build such technologies, we provide textual corpora in Wolof and Pulaar, and a pronunciation lexicon containing 49,132 entries from the Wolof dataset.

A growing body of research suggests that young children’s early speech and language exposure is associated with later language development (including delays and diagnoses), school readiness, and academic performance. The last decade has seen increasing use of child-worn devices to collect long-form audio recordings by educators, economists, and developmental psychologists. The most commonly used system for analyzing this data is LENA, which was trained on North American English child-centered data and generates estimates of children’s speech-like vocalization counts, adult word counts, and child-adult turn counts. Recently, cheaper and open-source non-LENA alternatives with multilingual training have been proposed. Both kinds of systems have been employed in under-resourced, sometimes multilingual contexts, including Africa where access to printed or digital linguistic resources may be limited. In this paper, we describe each kind of system (LENA, non-LENA), provide information on audio data collected with them that is available for reuse, review evidence of the accuracy of extant automated analyses, and note potential strengths and shortcomings of their use in African communities.

In the current digital age, languages lacking digital presence face an imminent risk of extinction. In addition, the absence of digital resources poses a significant obstacle to the development of Natural Language Processing (NLP) applications for such languages. Therefore, the development of digital language resources contributes to the preservation of these languages and enables application development. This paper contributes to the ongoing efforts of developing language resources for South African languages with a specific focus on Setswana and presents a new English-Setswana bilingual dataset that focuses on the space domain. The dataset was constructed using the expansion method. A subset of space domain English synsets from Princeton WordNet was professionally translated to Setswana. The initial submission of translations demonstrated an accuracy rate of 99% before validation. After validation, continuous revisions and discussions between translators and validators resulted in a unanimous agreement, ultimately achieving a 100% accuracy rate. The final version of the resource was converted into an XML format due to its machine-readable framework, providing a structured hierarchy for the organization of linguistic data.

This paper addresses the pressing need for improved readability assessment in Setswana through the creation of a list of frequently used words in Setswana. The end goal is to integrate this list into the adaptation of traditional readability measures in Setswana, such as the Dale-Chall index, which relies on frequently used words. Our initial list is developed using corpus-based methods utilising frequency lists obtained from five sets of corpora. It is then refined using manual methods. The analysis section delves into the challenges encountered during the development of the final list, encompassing issues like the inclusion of non-Setswana words, proper names, unexpected terms, and spelling variations. The decision-making process is clarified, highlighting crucial choices such as the retention of contemporary terms and the acceptance of diverse spelling variations. These decisions reflect a nuanced balance between linguistic authenticity and readability. This paper contributes to the discourse on text readability in indigenous Southern African languages. Moreover, it establishes a foundation for tailored literacy initiatives and serves as a starting point for adapting traditional frequency-list-based readability measures to Setswana.

The South African Language Identifier (SA-LID) has proven to be a valuable tool for data analysis in the multilingual context of South Africa, particularly in governmental texts. However, its suitability for broader projects has yet to be determined. This paper aims to assess the performance of the SA-LID in identifying isiXhosa in YouTube comments as part of the methodology for research on the expression of cultural identity through linguistic strategies. We curated a selection of 10 videos which focused on the isiXhosa culture in terms of theatre, poetry, language learning, culture, or music. The videos were predominantly in English as were most of the comments, but the latter were interspersed with elements of isiXhosa, identifying the commentators as speakers of isiXhosa. The SA-LID was used to identify all instances of the use of isiXhosa to facilitate the analysis of the relevant items. Following the application of the SA-LID to this data, a manual evaluation was conducted to gauge the effectiveness of this tool in selecting all isiXhosa items. Our findings reveal significant limitations in the use of the SA-LID, encompassing the oversight of unconventional spellings in indigenous languages and misclassification of closely related languages within the Nguni group. Although proficient in identifying the use of Nguni languages, differentiating within this language group proved challenging for the SA-LID. These results underscore the necessity for manual checks to complement the use of the SA-LID when other Nguni languages may be present in the comment texts.

This paper presents the first publicly available UD treebank for Tswana, Tswana-Popapolelo. The data used consists of the 20 Cairo CICLing sentences translated to Tswana. After pre-processing these sentences with detailed POS (XPOS) and converting them to universal POS (UPOS), we proceeded to annotate the data with dependency relations, documenting decisions for the language specific constructions. Linguistic issues encountered are described in detail as this is the first application of the UD framework to produce a dependency treebank for the Bantu language family in general and for Tswana specifically.

This article discusses the adaptation of traditional English readability measures into Sesotho, a Southern African indigenous low-resource language. We employ the use of a translated readability corpus to extract textual features from the Sesotho texts and readability levels from the English translations. We look at the correlation between the different features to ensure that non-competing features are used in the readability metrics. Next, through linear regression analyses, we examine the impact of the text features from the Sesotho texts on the overall readability levels (which are gauged from the English translations). Starting from the structure of the traditional English readability measures, linear regression models identify coefficients and intercepts for the different variables considered in the readability formulas for Sesotho. In the end, we propose ten readability formulas for Sesotho (one more than the initial nine; we provide two formulas based on the structure of the Gunning Fog index). We also introduce intercepts for the Gunning Fog index, the Läsbarhets index and the Readability index (which do not have intercepts in the English variants) in the Sesotho formulas.

IsiZulu and Siswati are mutually intelligible languages that are considered under-resourced despite their status as official languages. Even so, the available digital and computational language resources for isiZulu significantly outstrip those for Siswati, such that it is worth investigating to what degree bootstrapping approaches can be leveraged to develop resources for Siswati. In this paper, we present the development of a computational grammar and parallel treebank, based on parallel linguistic descriptions of the two languages.

Prior to the initiation of the project reported on in this paper, there were no instruments available with which to measure the language skills of young speakers of nine official African languages of South Africa. This limited the kind of research that could be conducted, and the rate at which knowledge creation on child language development could progress. Not only does this result in a dearth of knowledge needed to inform child language interventions but it also hinders the development of child language theories that would have good predictive power across languages. This paper reports on (i) the development of a questionnaire that caregivers complete about their infant’s communicative gestures and vocabulary or about their toddler’s vocabulary and grammar skills, in isiNdebele, isiXhosa, isiZulu, Sesotho, Sesotho sa Leboa, Setswana, Siswati, Tshivenda, and Xitsonga; and (ii) the 24 child language corpora thus far developed with these instruments. The potential research avenues opened by the 18 instruments and 24 corpora are discussed.

Ge’ez is an ancient Semitic language renowned for its unique alphabet. It serves as the script for numerous lan- guages, including Tigrinya and Amharic, and played a pivotal role in Ethiopia’s cultural and religious development during the Aksumite kingdom era. Ge’ez remains significant as a liturgical language in Ethiopia and Eritrea, with much of the national identity documentation recorded in Ge’ez. These written materials are invaluable primary sources for studying Ethiopian and Eritrean philosophy, creativity, knowledge, and civilization. Ge’ez is a complex morphological structure with rich inflectional and derivational morphology, and no usable NLP has been developed and published until now due to the scarcity of annotated linguistic data, corpora, labeled datasets, and lexicons. Therefore, we proposed a rule-based Ge’ez morphological synthesis to generate surface words from root words according to the morphological structures of the language. Consequently, we proposed an automatic morphological synthesizer for Ge’ez using TLM. We used 1,102 sample verbs, representing all verb morphological structures, to test and evaluate the system. Finally, we get a performance of 97.4%. This result outperforms the baseline model, suggesting that other scholars build a comprehensive system considering morphological variations of the language. Keywords: Ge’ez, NLP, morphology, morphological synthesizer, rule-based

Recent research in natural language processing (NLP) has achieved impressive performance in tasks such as machine translation (MT), news classification, and question-answering in high-resource languages. However, the performance of MT leaves much to be desired for low-resource languages. This is due to the smaller size of available parallel corpora in these languages, if such corpora are available at all. NLP in Ethiopian languages suffers from the same issues due to the unavailability of publicly accessible datasets for NLP tasks, including MT. To help the research community and foster research for Ethiopian languages, we introduce EthioMT – a new parallel corpus for 15 languages. We also create a new benchmark by collecting a dataset for better-researched languages in Ethiopia. We evaluate the newly collected corpus and the benchmark dataset for 23 Ethiopian languages using transformer and fine-tuning approaches.

Hate speech on social media has proliferated in Ethiopia. To support studies aimed at investigating the targets and types of hate speech circulating in the Ethiopian context, we developed a new fine-grained annotation scheme that captures three elements of hate speech: the target (i.e., any groups with protected characteristics), type (i.e., the method of abuse) and nature (i.e., the style of the language used). We also developed a new lexicon of hate speech-related keywords in the four most prominent languages found on Ethiopian social media: Amharic, Afaan Oromo, English and Tigrigna. These keywords enabled us to retrieve social media posts (also in the same four languages) from three platforms (i.e., X, Telegram and Facebook), that are likely to contain hate speech. Experts in the Ethiopian context then manually annotated a sample of those retrieved posts, obtaining fair to moderate inter-annotator agreement. The resulting annotations formed the basis of a case study of which groups tend to be targeted by particular types of hate speech or by particular styles of hate speech language.

Question Answering (QA) systems return concise answers or answer lists based on natural language text, which uses a given context document. Many resources go into curating QA datasets to advance the development of robust QA models. There is a surge in QA datasets for languages such as English; this is different for low-resource languages like Amharic. Indeed, there is no published or publicly available Amharic QA dataset. Hence, to foster further research in low-resource QA, we present the first publicly available benchmarking Amharic Question Answering Dataset (Amh-QuAD). We crowdsource 2,628 question-answer pairs from over 378 Amharic Wikipedia articles. Using the training set, we fine-tune an XLM-R-based language model and introduce a new reader model. Leveraging our newly fine-tuned reader run a baseline model to spark open-domain Amharic QA research interest. The best- performing baseline QA achieves an F-score of 80.3 and 81.34 in retriever-reader and reading comprehension settings.

A significant number of research studies have been presented for detecting hate speech in social media during the last few years. However, the majority of these studies are in English. Only a few studies focus on Arabic and its dialects (especially the Algerian dialect) with a smaller number of them targeting sexism detection (or hate speech against women). Even the works that have been proposed on Arabic sexism detection consider two classes only (hateful and non-hateful), and three classes(adding the neutral class) in the best scenario. This paper aims to propose the first fine-grained corpus focusing on 13 classes. However, given the challenges related to hate speech and fine-grained annotation, the Kappa metric is relatively low among the annotators (i.e. 35% ). This work in progress proposes three main contributions: 1) Annotation of different categories related to hate speech such as insults, vulgar words or hate in general. 2) Annotation of 10,000 comments, in Arabic and Algerian dialects, automatically extracted from Youtube. 3) High-lighting the challenges related to manual annotation such as subjectivity, risk of bias, lack of annotation guidelines, etc

This paper introduces a novel approach to spell checking and correction for low-resource and under-represented languages, with a specific focus on an African language, Wolof. By leveraging the capabilities of transformer models and neural networks, we propose an efficient and practical system capable of correcting typos and improving text quality. Our proposed technique involves training a transformer model on a parallel corpus consisting of misspelled sentences and their correctly spelled counterparts, generated using a semi-automatic method. As we fine tune the model to transform misspelled text into accurate sentences, we demonstrate the immense potential of this approach to overcome the challenges faced by resource-scarce and under-represented languages in the realm of spell checking and correction. Our experimental results and evaluations exhibit promising outcomes, offering valuable insights that contribute to the ongoing endeavors aimed at enriching linguistic diversity and inclusion and thus improving digital communication accessibility for languages grappling with scarcity of resources and under-representation in the digital landscape.

The aim of this paper is expose the structural form of the Igala language and the inherent complexity related to the translation of the language to a second language – i.e. the English language, through an inquisition into its the word order, lateral inversions, and unnamed grammatical entities inherent in the language. While this study finds out that there is a preponderance of a linguistic typology with subject-verb-object word order and the total absence of preposition in the speech composition of the Igala language. The implication of these trio of topic sentences (syntactic inversion, word ordering, unnamed entities) have remain within the dark corner of intellectual consideration and worst still the incorporation of this considerations in syntax parsing and annotation in computing. Rising from ongoing abstruseness and incongruity in machine translation of Igala, a comprehension model for automotive identification, application and/or conversion of these structural forms to the English language shall be the focus of this paper.