"Low-resource African languages have not fully benefited from the progress in neural machine translation because of a lack of data. Motivated by this challenge we compare zero-shot learning, transfer learning and multilingual learning on three Bantu languages (Shona, isiXhosa and isiZulu) and English. Our main target is English-to-isiZulu translation for which we have just 30,000 sentence pairs, 28% of the average size of our other corpora. We show the importance of language similarity on the performance of English-to-isiZulu transfer learning based on English-to-isiXhosa and English-to-Shona parent models whose BLEU scores differ by 5.2. We then demonstrate that multilingual learning surpasses both transfer learning and zero-shot learning on our dataset, with BLEU score improvements relative to the baseline English-to-isiZulu model of 9.9, 6.1 and 2.0 respectively. Our best model also improves the previous SOTA BLEU score by more than 10. "

Building effective neural machine translation (NMT) models for very low-resourced and morphologically rich African indigenous languages is an open challenge. Besides the issue of finding available resources for them, a lot of work is put into preprocessing and tokenization. Recent studies have shown that standard tokenization methods do not always adequately deal with the grammatical, diacritical, and tonal properties of some African languages. That, coupled with the extremely low availability of training samples, hinders the production of reliable NMT models. In this paper, using Fon language as a case study, we revisit standard tokenization methods and introduce Word-Expressions-Based (WEB) tokenization, a human-involved super-words tokenization strategy to create a better representative vocabulary for training. Furthermore, we compare our tokenization strategy to others on the Fon-French and French-Fon translation tasks.

From Word2Vec to GloVe, word embedding models have played key roles in the current state-of-the-art results achieved in Natural Language Processing. Designed to give significant and unique vectorized representations of words and entities, those models have proven to efficiently extract similarities and establish relationships reflecting semantic and contextual meaning among words and entities. African Languages, representing more than 31% of the worldwide spoken languages, have recently been subject to lots of research. However, to the best of our knowledge, there are currently very few to none word embedding models for those languages words and entities, and none for the languages under study in this paper. After describing Glove, Word2Vec, and Poincaré embeddings functionalities, we build Word2Vec and Poincaré word embedding models for Fon and Nobiin, which show promising results. We test the applicability of transfer learning between these models as a landmark for African Languages to jointly involve in mitigating the scarcity of their resources, and attempt to provide linguistic and social interpretations of our results. Our main contribution is to arouse more interest in creating word embedding models proper to African Languages, ready for use, and that can significantly improve the performances of Natural Language Processing downstream tasks on them. The official repository and implementation is at https://github.com/bonaventuredossou/afrivec

The Yorùbá group has diverse language speakers across the world, translating the language to other widely spoken languages must be emphasized. We aim to develop an English to Yorùbá machine translation system which can translate English verb phrase text to its Yorùbá equivalent. Words from both languages Source Language and Target Language were collected for the verb phrase group in the home domain. The lexical translation is done by assigning values of the matching word in the dictionary. The syntax of the two languages was realized using Context Free Grammar, we validated the rewrite rules with finite state automata. Human evaluation method was used and expert opinion scored. The evaluation shows the system performed better than that of sampled Google translation with over 70% of the response matching that of the system’s output.

Amharic is the second most Semitic language spoken after Arabic and serves as the official working language of the government of Ethiopia. In Amharic/Ge’ez writing, there are different characters with the same sound but different in shape and meaning, which are called homophones. Even though there are rules and regulations for Amharic writing, the online community tends to use homophone characters randomly. This means, the first character in the word ሃብታም (rich) can be replaced with one of the characters from ሀ(hā), ሐ(ḥā), ሓ(ḥa), ኀ(ḫā), ኃ(ḫā), or ኻ(ẖa). To study the usage of homophone characters, we have collected and analyzed around 5m sentences and build Word2Vec and FastText embedding models. Our analysis shows that 1) the usage of homophone characters in Amharic text is mostly random, and 2) the normalization of homophone characters has a negative impact while computing word similarity using the embedding models.

---

" አማርኛ ከአረብኛ ቀጥሎ በሁለተኛነት በስፋት የሚነገር ሴማዊ ቋንቋ ሲሆን የኢትዮጵያ ፌዴራላዊ ዴሞክራሲያዊ ሪፑብሊክ መንግሥት ሕጋዊ (ይፋዊ) የሥራ ቋንቋ ነው። በአማርኛ ወይም በግእዝ ሥርዓተ ጽሕፈት ውስጥ በቅርጽ እና በትርጕም የተለያዩ ሆነው ተመሳሳይ ድምፅ ያላቸው ብዙ ፊደላት አሉ። እነዚህ ፊደላት “ሞክሼ ሆሄያት” በመባል ይታወቃሉ። የአማርኛ ቋንቋ የሥርዓተ ጽሕፈት ሕግ ያለው ቢሆንም የ”ኦንላይን” ተጠቃሚዎች ግን እነዚህን ሞክሼ ሆሄያት እንደፈለጋቸው የመጠቀም ዝንባሌ ያሳያሉ። ይህም ማለት ለምሳሌ፣ ሃብታም በሚለው ቃል ውስጥ ያለው የመጀመሪያው ሆሄ ወይም ፊደል ከሚከተሉት ሆሄያት በአንዱ ሊተካ ይችላል (ሀ፣ ሐ፣ ሓ፣ ኀ፣ ኃ፣ ኻ)። የሞክሼ ሆሄያትን አጠቃቀም ለማጥናት ወደ 5 ሚሊዮን የሚደርሱ ዐረፍተ ነገሮችን ሰብስበን በመተንተን Word2Vec እና FastText የተሰኙትን የ“ኢምቤዲንግ ሞዴሎችን” ገንብተናል። ይህ የጥናት እና የምርምር ሙከራችን የሚከተሉትን ውጤቶች ያሳያል፡-

1. የሞክሼ ሆሄያት አጠቃቀም በአብዛኛው በዘፈቀደ የሚጻፉ መሆኑ፣

2. ሞክሼ ሆሄያትን በአንድ ፊደል ብቻ እንዲወከሉና እንዲለመድ የሚደረገው ጥረት በኢምቤዲንግ ሞዴሎች ላይ የቃላት ቅመራ ሲደረግ አሉታዊ ተጽዕኖ አለው።"

This paper describes the acquisition, preprocessing, segmentation, and alignment of an Amharic-English parallel corpus. It will be useful for machine translation of a low-resource language, Amharic. The corpus is larger than previously compiled corpora; it is released for research purposes. We trained neural machine translation and phrase-based statistical machine translation models using the corpus. In the automatic evaluation, neural machine translation models outperform phrase-based statistical machine translation models.ition models for other African low-resourced languages, as well as guide future NLP research for Fon and Igbo. The Fon and Igbo models source code have been made publicly available.

In this paper, an approach for hate speech detection against women in Arabic community on social media (e.g. Youtube) is proposed. In the literature, similar works have been presented for other languages such as English. However, to the best of our knowledge, not much work has been conducted in the Arabic language. A new hate speech corpus (Arabic_fr_en) is developed using three different annotators. For corpus validation, three different machine learning algorithms are used, including deep Convolutional Neural Network(CNN), long short-term memory (LSTM) network and Bi-directional LSTM(Bi-LSTM) network. Simulation results demonstrate the best performance of the CNN model, which achieved F1-score up to 86% for the unbalanced corpus as compared to LSTM and Bi-LSTM.

In NLP, text classification is one of the primary problems we try to solve and its uses in language analyses are indisputable. The lack of labeled training data made it harder to do these tasks in low resource languages like Amharic. The task of collecting, labeling, annotating, and making valuable this kind of data will encourage junior researchers, schools, and machine learning practitioners to implement existing classification models in their language. In this short paper, we aim to introduce the Amharic text classification dataset that consists of more than 50k news articles that were categorized into 6 classes. This dataset is made available with easy baseline performances to encourage studies and better performance experiments.

Transformer-based language models have been changing the modern Natural Language Processing (NLP) landscape for high-resource languages such as English, Chinese, Russian, etc. However, this technology does not yet exist for any Ghanaian language. In this paper, we introduce the first of such models for Twi or Akan, the most widely spoken Ghanaian language. The specific contribution of this research work is the development of several pretrained transformer language models for the Akuapem and Asante dialects of Twi, paving the way for advances in application areas such as Named Entity Recognition (NER), Neural Machine Translation (NMT), Sentiment Analysis (SA) and Part-of-Speech (POS) tagging. Specifically, we introduce four different flavours of ABENA -- A BERT model Now in Akan that is fine-tuned on a set of Akan corpora, and BAKO - BERT with Akan Knowledge only, which is trained from scratch. We open-source the model through the Hugging Face model hub and demonstrate its use via a simple sentiment classification example.

We present a parallel machine translation training corpus for English and Akuapem Twi of 25,421 sentence pairs. We used a transformer-based translator to generate initial translations in Akuapem Twi, which were later verified and corrected where necessary by native speakers to eliminate any occurrence of translationese. In addition, 697 higher quality crowd-sourced sentences are provided for use as an evaluation set for downstream Natural Language Processing (NLP) tasks. The typical use case for the larger human-verified dataset is for further training of machine translation models in Akuapem Twi. The higher quality 697 crowd-sourced dataset is recommended as a testing dataset for machine translation of English to Twi and Twi to English models. Furthermore, the Twi part of the crowd-sourced data may also be used for other tasks, such as representation learning, classification, etc. We fine-tune the transformer translation model on the training corpus and report benchmarks on the crowd-sourced test set.

Massively multilingual machine translation (MT) has shown impressive capabilities, including zero and few-shot translation between low-resource language pairs. However, these models are often evaluated on high-resource languages with the assumption that they generalize to low-resource ones. The difficulty of evaluating MT models on low-resource pairs is often due the lack of standardized evaluation datasets. In this paper, we present MENYO-20k, the first multi-domain parallel corpus for the low-resource Yorùbá--English (yo--en) language pair with standardized train-test splits for benchmarking. We provide several neural MT (NMT) benchmarks on this dataset and compare to the performance of popular pre-trained (massively multilingual) MT models, showing that, in almost all cases, our simple benchmarks outperform the pre-trained MT models. A major gain of BLEU +9.9 and +8.6 (en2yo) is achieved in comparison to Facebook's M2M-100 and Google multilingual NMT respectively when we use MENYO-20k to fine-tune generic models.

We take a step towards addressing the under-representation of the African continent in NLP research by creating the first large publicly available high-quality dataset for named entity recognition (NER) in ten African languages, bringing together a variety of stakeholders. We detail characteristics of the languages to help researchers understand the challenges that these languages pose for NER. We analyze our datasets and conduct an extensive empirical evaluation of state-of-the-art methods across both supervised and transfer learning settings. We release the data, code, and models in order to inspire future research on African NLP.

"This paper is an informal collaboration launched during the African Master of Machine Intelligence (AMMI) program in June 2020. After a series of lectures and labs on speech data collection using mobile applications and on self-supervised representation learning from speech, a small group of students and the lecturer continued working on automatic speech recognition (ASR) projects for three languages: Wolof, Ga, and Somali.

This paper describes the data collection process and ASR systems developed with a small amount (1h) of transcribed speech as training data. In these low resource conditions, pretraining a model on a large amount of raw speech is fundamental for developing an efficient ASR system."

"The evolution of the Internet has increased the amount of information that is expressed by people on different platforms.

This information can be product reviews, discussions on forums, or social media platforms.

Accessibility of these opinions and people’s feelings open the door to opinion mining and sentiment analysis.

As language and speech technologies become more advanced, many languages have been used and the best models have been obtained.

However, due to linguistic diversity and lack of datasets, African languages have been left behind.

In this study, by using the current state of the art model, multilingual BERT, we perform sentiment classification on Swahili datasets created by extracting and annotating 8.2k reviews and comments on different social media platforms.

The data were classified as either positive or negative.

The model was fine-tuned and achieve the best accuracy of 87.59\%."

This work empirically demonstrates the ability of Text Graph Convolutional Network (Text GCN) to outperform traditional natural language processing benchmarks for the task of semi-supervised Swahili news classification. In particular, we focus our experimentation on the sparsely-labelled semi-supervised context which is representative of the practical constraints facing low-resourced African languages. We follow up on this result by introducing a variant of the Text GCN model which utilises a bag of words embedding rather than a naive one-hot encoding to reduce the memory footprint of Text GCN whilst demonstrating similar predictive performance.

We, in this work, perform an empirical study of natural language generation (NLG) of dialogues for a low-resource language. We do so by fine-tuning DialoGPT-medium on Yorùbá, a foreign language, as the target language. DialoGPT is a state-of-the-art (SotA) model pre-trained in English. Two variants of the Yorùbá language are evaluated (diacritized and undiacritized). The results, in terms of low perplexity and human evaluation, show that good performance can be achieved for the Yorùbá language when fine-tuned on the DialoGPT model, though this was pre-trained in English. As further contribution to the research community, we make available the dialogue dataset, given that data, particularly for low-resource languages, can be hard to come by.

"We present a preprocessed, ready-to-use automatic speech recognition corpus, BembaSpeech, consisting over 24 hours of read speech in the Bemba language, a written but low-resourced language spoken by over 30% of the population in Zambia. To assess its usefulness for training and testing ASR systems for Bemba, we train an end-to-end Bemba

ASR system by fine-tuning a pre-trained DeepSpeech English model on the training portion

of the BembaSpeech corpus. Our best model achieves a word error rate (WER) of 54.78%.

The results show that the corpus can be used for building ASR systems for Bemba."

---

"Ili ipepala lilelanda pamashiwi mu Cibemba na ifyebo fyalembwa ifyabikwa pamo nga mashiwi yakopwa elyo na yalembwa ukupanga iileitwa BembaSpeech. iikwete amashiwi ayengabelengwa ukufika kuma awala amakumi yabili na yane mu lulimi lwa Cibemba, ululandwa na impendwa ya bantu ba mu Zambia ukufika cipendo ca 30%. Pakufwaisha ukumona ubukankala bwakubomfiwa mu mukupanga ifya mibombele ya ASR mu Cibemba, tupanga imibombele ya ASR iya mu Cibemba ukufuma pantendekelo ukufika na pampela, kubomfya elyo na ukuwaminisha icilangililo ca mibomfeshe yamashiwi na ifyebo ifyabikwa pamo mu Cisungu icitwa DeepSpeech na ukupangako iciputulwa ca mashiwi na ifyebo fyalembwa mu Cibemba (BembaSpeech corpus). Imibobembele yesu iyakunuma ilangisha icipimo ca kupusa nelyo ukulufyanya kwa mashiwi ukwa 54.78%. Ifyakufumamo filangisha ukuti ifyalembwa kuti fyabomfiwa ukupanga imibombele ya ASR mu Cibemba.

"

This paper presents a manually annotated spelling error corpus for Amharic, lingua franca in Ethiopia. The corpus is designed to be used for the evaluation of spelling error detection and correction. The misspellings are tagged as non-word and real-word errors. In addition, the contextual information available in the corpus makes it useful in dealing with both types of spelling errors.

Most of the common applications of Named Entity Recognition (NER) is on English and other highly available languages. In this work, we present our findings on Named Entity Recognition for 5 Nigerian Languages (Nigerian English, Nigerian Pidgin English, Igbo, Yoruba and Hausa). These languages are considered low-resourced, and very little openly available Natural Language Processing work has been done in most of them. In this work, individual NER models were trained and metrics recorded for each of the languages. We also worked on a combined model that can handle Named Entity Recognition (NER) for any of the five languages. The combined model works well for Named Entity Recognition(NER) on each of the languages and with better performance compared to individual NER models trained specifically on annotated data for the specific language. The aim of this work is to share our learning on how information extraction using Named Entity Recognition can be optimized for the listed Nigerian Languages for inclusion, ease of deployment in production and reusability of models. Models developed during this project are available on Github(https://git.io/JY0kk) and an interactive web app(https://nigner.herokuapp.com/)

In this paper we describe our efforts to make a bidirectional Congolese Swahili (SWC) to French (FRA) neural machine translation system with the motivation of improving humanitarian translation workflows. For training, we created a 25,302-sentence general domain parallel corpus and combined it with publicly available data. Experimenting with low-resource methodologies like cross-dialect transfer and semi-supervised learning, we recorded improvements of up to 2.4 and 3.5 BLEU points in the SWC–FRA and FRA–SWC directions, respectively. We performed human evaluations to assess the usability of our models in a COVID-domain chatbot that operates in the Democratic Republic of Congo (DRC). Direct assessment in the SWC–FRA direction demonstrated an average quality ranking of 6.3 out of 10 with 75% of the target strings conveying the main message of the source text. For the FRA–SWC direction, our preliminary tests on post-editing assessment showed its potential usefulness for machine-assisted translation. We make our models, datasets containing up to 1 million sentences, our development pipeline, and a translator web-app available for public use. eds.

This paper proposes a generative language model called AfriKI. Our approach is based on an LSTM architecture trained on a small corpus of contemporary fiction. With the aim of promoting human creativity, we use the model as an authoring tool to explore machine-in-the-loop Afrikaans poetry generation. To our knowledge, this is the first study to attempt creative text generation in Afrikaans.

NLP Ghana is an open-source non-profit organisation aiming to advance the development and adoption of state-of-the-art NLP techniques and digital language tools to Ghanaian languages and problems. In this paper, we first present the motivation and necessity for the efforts of the organisation; by introducing some popular Ghanaian languages while presenting the state of NLP in Ghana. We then present the NLP Ghana organisation and outline its aims, scope of work, some of the methods employed and contributions made thus far in the NLP community in Ghana.

Amharic is the second most Semitic language spoken after Arabic and serves as the official working language of the government of Ethiopia. In Amharic/Ge’ez writing, there are different characters with the same sound but different in shape and meaning, which are called homophones. Even though there are rules and regulations for Amharic writing, the online community tends to use homophone characters randomly. This means, the first character in the word ሃብታም (rich) can be replaced with one of the characters from ሀ(hā), ሐ(ḥā), ሓ(ḥa), ኀ(ḫā), ኃ(ḫā), or ኻ(ẖa). To study the usage of homophone characters, we have collected and analyzed around 5m sentences and build Word2Vec and FastText embedding models. Our analysis shows that 1) the usage of homophone characters in Amharic text is mostly random, and 2) the normalization of homophone characters has a negative impact while computing word similarity using the embedding models.

---

" አማርኛ ከአረብኛ ቀጥሎ በሁለተኛነት በስፋት የሚነገር ሴማዊ ቋንቋ ሲሆን የኢትዮጵያ ፌዴራላዊ ዴሞክራሲያዊ ሪፑብሊክ መንግሥት ሕጋዊ (ይፋዊ) የሥራ ቋንቋ ነው። በአማርኛ ወይም በግእዝ ሥርዓተ ጽሕፈት ውስጥ በቅርጽ እና በትርጕም የተለያዩ ሆነው ተመሳሳይ ድምፅ ያላቸው ብዙ ፊደላት አሉ። እነዚህ ፊደላት “ሞክሼ ሆሄያት” በመባል ይታወቃሉ። የአማርኛ ቋንቋ የሥርዓተ ጽሕፈት ሕግ ያለው ቢሆንም የ”ኦንላይን” ተጠቃሚዎች ግን እነዚህን ሞክሼ ሆሄያት እንደፈለጋቸው የመጠቀም ዝንባሌ ያሳያሉ። ይህም ማለት ለምሳሌ፣ ሃብታም በሚለው ቃል ውስጥ ያለው የመጀመሪያው ሆሄ ወይም ፊደል ከሚከተሉት ሆሄያት በአንዱ ሊተካ ይችላል (ሀ፣ ሐ፣ ሓ፣ ኀ፣ ኃ፣ ኻ)። የሞክሼ ሆሄያትን አጠቃቀም ለማጥናት ወደ 5 ሚሊዮን የሚደርሱ ዐረፍተ ነገሮችን ሰብስበን በመተንተን Word2Vec እና FastText የተሰኙትን የ“ኢምቤዲንግ ሞዴሎችን” ገንብተናል። ይህ የጥናት እና የምርምር ሙከራችን የሚከተሉትን ውጤቶች ያሳያል፡-

1. የሞክሼ ሆሄያት አጠቃቀም በአብዛኛው በዘፈቀደ የሚጻፉ መሆኑ፣

2. ሞክሼ ሆሄያትን በአንድ ፊደል ብቻ እንዲወከሉና እንዲለመድ የሚደረገው ጥረት በኢምቤዲንግ ሞዴሎች ላይ የቃላት ቅመራ ሲደረግ አሉታዊ ተጽዕኖ አለው።"

"Translating between languages where certain features are marked morphologically in one but absent or marked contextually in the other is an important test case for machine translation. When translating into English which marks (in)definiteness morphologically, from Yorùbá which uses bare nouns but mark these features contextually, ambiguities arise. In this work, we perform fine-grained analysis on how an SMT system compares with two NMT systems (BiLSTM and Transformer) when translating bare nouns in Yorùbá into English. We investigate how the systems what extent they identify BNs, correctly translate them, and compare with human translation patterns. We also analyze the type of errors each model makes and provide a linguistic description of these errors. We glean insights for evaluating model performance in low-resource settings. In translating bare nouns, our results show the transformer model outperforms the SMT and BiLSTM models for 4 categories, the BiLSTM outperforms the SMT model for 3 categories while the SMT outperforms the NMT models for 1 category."

This work supports further development of language technology for the languages of Africa by providing a Wikidata-derived resource of name lists corresponding to common entity types (person, location, and organization). While we are not the first to mine Wikidata for name lists, our approach emphasizes scalability and replicability and addresses data quality issues for languages that do not use Latin scripts. We produce lists containing approximately 1.9 million names across 28 African languages. We describe the data, the process used to produce it, and its limitations, and provide the software and data for public use. Finally, we discuss the ethical considerations of producing this resource and others of its kind.

With the success of large-scale pre-training and multilingual modeling in Natural Language Processing (NLP), recent years have seen a proliferation of large, web-mined text datasets covering hundreds of languages. However, to date there has been no systematic analysis of the quality of these publicly available datasets, or whether the datasets actually contain content in the languages they claim to represent. In this work, we manually audit the quality of 205 language-specific corpora released with five major public datasets (CCAligned, ParaCrawl, WikiMatrix, OSCAR, mC4), and audit the correctness of language codes in a sixth (JW300). We find that lower-resource corpora have systematic issues: at least 15 corpora are completely erroneous, and a significant fraction contains less than 50% sentences of acceptable quality. Similarly, we find 82 corpora that are mislabeled or use nonstandard/ambiguous language codes. We demonstrate that these issues are easy to detect even for non-speakers of the languages in question, and supplement the human judgements with automatic analyses. Inspired by our analysis, we recommend techniques to evaluate and improve multilingual corpora and discuss the risks that come with low-quality data releases.

It has been shown that the performance of neural machine translation (NMT) drops starkly in low-resource conditions, often requiring large amounts of auxiliary data to achieve competitive results. An effective method of generating auxiliary data is back-translation of target language sentences. In this work, we present a case study of Tigrinya where we investigate several back-translation methods to generate synthetic source sentences. We find that in low-resource conditions, back-translation by pivoting through a higher-resource language related to the target language proves most effective resulting in substantial improvements over baselines.

Major advancement in the performance of machine translation models has been made possible in part thanks to the availability of large-scale parallel corpora. But for most languages in the world, the existence of such corpora is rare. Emakhuwa, a language spoken in Mozambique, is like most African languages low-resource in NLP terms. It lacks both computational and linguistic resources and, to the best of our knowledge, few parallel corpora including Emakhuwa already exist. In this paper we describe the creation of the Emakhuwa-Portuguese parallel corpus, which is a collection of texts from the Jehovah's Witness website and a variety of other sources including the African Story Book website, the Universal Declaration of Human Rights and Mozambican legal documents. The dataset contains 47,415 sentence pairs, amounting to 699,976 word tokens of Emakhuwa and 877,595 word tokens in Portuguese. After normalization processes which remain to be completed, the corpus will be made freely available for research use.

In this digital era, almost in every discipline people are using automated systems that generate information represented in document format in different natural languages. As a result, there is a growing interest towards better solutions for finding, organizing and analyzing these documents. In this paper, we propose a system that clusters Amharic text documents using Encyclopedic Knowledge (EK) with neural word embedding. EK enables the representation of related concepts and neural word embedding allows us to handle the contexts of the relatedness. During the clustering process, all the text documents pass through preprocessing stages. Enriched text document features are extracted from each document by mapping with EK and word embedding model. TF-IDF weighted vector of enriched feature was generated. Finally, text documents are clustered using popular spherical K-means algorithm. The proposed system is tested with Amharic text corpus and Amharic Wikipedia data. Test results show that the use of EK with word embedding for document clustering improves the average accuracy over the use of only EK. Furthermore, changing the size of the class has a significant effect on accuracy.

Faced with a considerable lack of resources in African languages to carry out work in Natural Language Processing (NLP), Natural Language Understanding (NLU) and artificial intelligence, the research teams of NTeALan association has set itself the objective of building open-source platforms for the collaborative construction of lexicographic data in African languages. In this article, we present our first reports after 2 years of collaborative construction of lexicographic resources useful for African NLP tools.

Training data for machine learning models can come from many different sources, which can be of dubious quality. For resource-rich languages like English, there is a lot of data available, so we can afford to throw out the dubious data. For low-resource languages where there is much less data available, we can’t necessarily afford to throw out the dubious data, in case we end up with a training set which is too small to train a model. In this study, we examine the effects of text normalization and data set quality for a set of low-resource languages of Africa—Afrikaans, Amharic, Hausa, Igbo, Malagasy, Somali, Swahili, and Zulu. We describe our text normalizer which we built in the Pynini framework, a Python library for finite state transducers, and our experiments in training language models for African languages using the Natural Language Toolkit (NLTK), an open-source Python library for NLP.

Social desirability bias is a type of response bias which arises when the answers to a survey or interview are influenced by what the interviewee believes to be more socially acceptable. The effect of social desirability bias may be more pronounced in traditional communities in Low-Income Countries, where social pressures can be ripe. While this type of bias has been well understood within social sciences, it has not yet been systematically studied within the Natural Language Processing (NLP) field. In this paper, we make a first attempt to quantify the impact of social desirability on people’s utterances in different contexts. To achieve this, we build on top of an existing corpus of annotated interviews from rural Ugandan communities. We then implement and test a range of BERT-based architectures which leverage information on the specific setting in which a given interview was collected. We find that models which are exposed to those variables tend to perform better. This suggests that the interview setting plays a relevant role in influencing what is being said, and opens new interesting research directions.

---

Social desirability bias kitegeeza okudibwamu okuvudde mukunonnyereza okukoledwa abakugu naye nga abawa endowooza bagobelela ebyo abantu abasinga byebakiririzame. Endowooza eno ey'okukiririza mwebyo abikiririzibwa abantu abasinga obungyi eri nyo mu mawanga aga kyakula olwesonga nti twesigama nnyo kwe'bbyo abalala byebakirirrizaamu. Wadde nga abasoma ebikyata nku bbela za'bantu bafubye okutegera nti waliwo ebbela ya abantu okwagala endowooza zabye obutawukana nazabalala, etabbi lya Natural Language Processing (NLP) telinakitekako nnyo sila. Okunonyereza kuno tukwesigamiza nnyo kumbeera z'omubantu ababera mu byalo mu Uganda, nga kino tukizimbidde kundowooza azatuweebwa okuva eri abantu abo. Endowooza za'bantu ze'kebegyezebwa mu mitedela egyengyawulo nga twesigama nku BERT-architectures ezituyamba okunnyonnyoka okwenjawulo nku bitundu abyo okunonnyereza mwekwakolebwa. Okwekebeja kwetwakola kukakasa nti ebyuma ebitendekebwa n'okumanya okwenjawulo ku bitundu abyo okunonnyereza mwekukolebwa bituyamba okola okusalawo okusinga kyebyo ebitandekwedwa n'okumanya kunno. Kino kitegeeza nti ebitundu okunonyereza mwekukolebwa biyina kinene kye bizanya kugeli abantu gyebawa endowooza zabwe.

"The increasing occurrence, forms, and negative effects of misinformation on social media platforms has necessitated more misinformation detection tools. Currently, work is being done addressing COVID-19 misinformation however, there are no misinformation detection tools for any of the 40 distinct indigenous Ugandan languages. This paper addresses this gap by presenting basic language resources and a misinformation detection data set based on code-mixed Luganda-English messages sourced from the Facebook and Twitter social media platforms. Several machine learning methods are applied on the misinformation detection data set to develop classification models for detecting whether a code-mixed Luganda-English message contains misinformation or not. A 10-fold cross validation evaluation of the classification methods in an experimental misinformation detection task shows that a Discriminative Multinomial Naive Bayes (DMNB) method achieves the highest accuracy and F-measure of 78.19% and 77.90% respectively. Also, Support Vector Machine and Bagging ensemble classification models achieve comparable results. These results are promising since the machine learning models are based on n-gram features from only the misinformation detection data set."

The application of deep neural networks to the task of acoustic modeling for automatic speech recognition (ASR) has resulted in dramatic decreases in ASR word error rates, enabling the use of this technology for interacting with smart phones and personal home assistants in high-resource languages like English or Mandarin. Developing ASR models of this caliber, however, requires thousands of hours of transcribed speech recordings, which presents challenges to the vast majority of the world's languages. In this paper, we apply a non-neural and a simple neural ASR architecture, as well as a fully convolutional ASR architecture originally developed for an endangered Native American language, to two under-resourced but widely spoken African languages, Wolof and Amharic. We explore the impact of typological features, audio recording quality, and speaker diversity on the accuracy of ASR output using these two architectures in order to learn ways in which the development of a low-resource ASR system can be customized to fit the characteristics of a particular corpus.

Translating to and from low-resource languages is a challenge for machine translation (MT) systems due to a lack of parallel data. In this paper we address the issue of domain-specific MT for Bambara, an under-resourced Mande language spoken in Mali. We present the first domain-specific parallel dataset for MT of Bambara into and from French. We discuss challenges in working with small quantities of domain-specific data for a low-resource language and we present the results of machine learning experiments on this data.

"The recent raft of high-profile gaffes involving neural machine translation technology has brought to light the unreliability and brittleness of this fledgling technology. These revelations have worryingly coincided with two other developments: The rise of back-translated text being increasingly used to augment training data in so termed low-resource natural language processing scenarios (such as those in the African context) and the emergence of 'AI-enhanced legal-tech' as a panacea that promises 'disruptive democratization' of access to legal services. In the backdrop of these quandaries, we present this cautionary tale where we shed light on the specifics of the risks surrounding cavalier deployment of this technology by exploring two specific failings: Androcentrism and Enantiosemy.

In this regard, we empirically investigate the fate of pronouns and a list of contronyms when subjected to back-translation using the state-of-the-art Google translate API. Through this, we seek to highlight the prevalence of the defaulting-to-the-masculine phenomenon in the context of gendered profession-related translations and also empirically demonstrate the scale and nature of threats pertaining to contronymous phrases covering both current-affairs and legal issues."

Advances in speech and language technologies enable tools such as voice-search, text-to-speech, speech recognition and machine trans-lation. These are however only available forhigh resource languages like English, Frenchor Chinese. Without foundational digital re-sources for African languages, which are con-sidered low-resource in the digital context,these advanced tools remain out of reach. Thiswork details the AI4D - African LanguageProgram, a 3-part project that 1) incentivisedthe crowd-sourcing, collection and curation oflanguage datasets through an online quantita-tive and qualitative challenge, 2) supported re-search fellows for a period of 3-4 months tocreate datasets annotated for NLP tasks, and3) hosted competitive Machine Learning chal-lenges on the basis of these datasets. Keyoutcomes of the work so far include 1) thecreation of 9+ open source, African languagedatasets annotated for a variety of ML tasks,and 2) the creation of baseline models for thesedatasets through hosting of competitive ML challenges.

"Models pre-trained on multiple languages have shown significant promise for improving speech recognition, particularly for low-resource languages.

In this work, we focus on phoneme recognition using Allosaurus, a method for multilingual recognition based on phonetic annotation, which incorporates phonological knowledge through a language-dependent allophone layer that associates a universal narrow phone-set with the phonemes that appear in each language.

To evaluate in a challenging real-world scenario, we curate phone recognition datasets for Bukusu and Saamia, two varieties of the Luhya language cluster of western Kenya and eastern Uganda. To our knowledge, these datasets are the first of their kind. We carry out similar experiments on the dataset of an endangered Tangkhulic language, East Tusom, a Tibeto-Burman language variety spoken mostly in India.

We explore both zero-shot and few-shot recognition by fine-tuning using datasets of varying sizes (10 to 1000 utterances). We find that fine-tuning of Allosaurus, even with just 100 utterances, leads to significant improvements in phone error rates."

Morphological Segmentation involves decomposing words into morphemes, the smallest meaning-bearing units of language. This is an important NLP task for morphologically-rich agglutinative languages such as the Southern African Nguni language group. In this paper, we investigate supervised and unsupervised models for two variants of morphological segmentation: canonical and surface segmentation. We train sequence-to-sequence models for canonical segmentation, where the underlying morphemes may not be equal to the surface form of the word, and Conditional Random Fields (CRF) for surface segmentation. Transformers outperform LSTMs with attention on canonical segmentation, obtaining an average F1 score of 72.5% across 4 languages. Feature-based CRFs outperform bidirectional LSTM-CRFs to obtain an average of 97.1% F1 on surface segmentation. In the unsupervised setting, an entropy-based approach using a character-level LSTM language model fails to outperforms a Morfessor baseline, while on some of the languages neither approach perform much better than a random baseline. We hope that the high performance of the supervised segmentation models will help to facilitate the development of better NLP tools for Nguni languages.

---

Morfologiese segmentasie is die taak om woorde op te breek in morfeme, die kleinste betekenisdraende eenhede van 'n taal. Dit it 'n belangrike taak vir Natuurlike Taalverwerking (NTV) vir agglutinatiewe tale met 'n ryk morfologie, soos die Suider-Afrikaanse Nguni tale. In hierdie artikel ondersoek ons afgerigte en onafgerigte modelle vir twee formulerings van morfologiese segmentasie: kanonieke en oppervlaksegmentasie. On rig ry-na-ry modelle af vir kanonieke segmentasie, waar die onderliggende morfeme nie noodwendig gelyk is aan die oppervlaksform van die woord nie, asook Voorwaardelike Kansfelde (VFKs) vir oppervlaksegmentasie. Transformators doen beter as Lang Korttermyn Geheue netwerke (LKTGs) met aandag op kanonieke segmentasie, met 'n gemiddelde F1 van 72.5% oor 4 tale. Kenmerk-gebasseerde VFKs doen beter as twee-rigting LKTG-VFKs, met 'n gemiddelde F1 van 97.1% vir oppervlaksegmentasie. In die onafgerigte opstelling slaag 'n entropie-gebasseerde benadering met 'n karakter LKTG taalmodel nie daarin om beter te doen as 'n Morfessor basismodel nie, terwyl vir sommige tale nie een van hierdie benaderings beter doen as 'n kansbenadering nie. Ons hoop dat die hoë akkuraatheid van die afgerigte segmentasiemodelle sal help om die ontwikkeling van beter NTV toepassings vir Nguni tale te bevorder.

Language models are the foundation of current neural network-based models for natural language understanding and generation. However, research on the intrinsic performance of language models on African languages has been extremely limited, which is made more challenging by the lack of large or standardised training and evaluation sets that exist for English and other high-resource languages. In this paper, we evaluate the performance of open-vocabulary language models on low-resource South African languages, using byte-pair encoding to handle the rich morphology of these languages. We evaluate different variants of n-gram models, feedforward neural networks, recurrent neural networks (RNNs), and Transformers on small-scale datasets. Overall, well-regularized RNNs give the best performance across two isiZulu and one Sepedi datasets. Multilingual training further improve performance on these datasets. We hope that this research will open new avenues for research into multilingual and low-resource language modelling for African languages.