Examinando por Autor "Osei, Salomey"

Mostrando 1 - 3 de 3
  • Miniatura
    Ítem
    AFRIHATE: a multilingual collection of hate speech and abusive language datasets for African languages
    (Association for Computational Linguistics (ACL), 2025) Muhammad,S.H.; Abdulmumin,I.; Ayele,A.A.; Adelani,D.I.; Ahmad,I.S.; Aliyu,S.M.; Onyango,N.O.; Wanzare,L.D.A.; Rutunda,S.; Aliyu,L.J.; Alemneh,E.; Hourrane,O.; Gebremichael,H.T.; Ismail,E.A.; Beloucif,M.; Jibril,E.C.; Bukula,A.; Mabuya,R.; Osei, Salomey; Oppong,A.; Belay,T.D.; Guge,T.K.; Asfaw,T.T.; Chukwuneke,C.I.; Röttger,P.; Yimam,S.M.; Ousidhoum,N.
    Hate speech and abusive language are global phenomena that need socio-cultural background knowledge to be understood, identified, and moderated. However, in many regions of the Global South, there have been several documented occurrences of (1) absence of moderation and (2) censorship due to the reliance on keyword spotting out of context. Further, high-profile individuals have frequently been at the center of the moderation process, while large and targeted hate speech campaigns against minorities have been overlooked. These limitations are mainly due to the lack of high-quality data in the local languages and the failure to include local communities in the collection, annotation, and moderation processes. To address this issue, we present AFRIHATE: a multilingual collection of hate speech and abusive language datasets in 15 African languages, annotated by native speakers. We report the challenges related to the construction of the datasets and present various classification baseline results with and without using LLMs. We find that model performance highly depends on the language and that multilingual models can help boost the performance in low-resource settings.
  • Miniatura
    Ítem
    IrokoBench: a new benchmark for African languages in the age of Large Language Models
    (Association for Computational Linguistics (ACL), 2025) Adelani,D.I.; Ojo,J.; Azime,I.A.; Zhuang,J.Y.; Alabi,J.O.; He,X.; Ochieng,M.; Hooker,S.; Bukula,A.; Lee,E.-S.A.; Chukwuneke,C.; Buzaaba,H.; Sibanda,B.; Kalipe,G.; Mukiibi,J.; Kabongo,S.; Yuehgoh,F.; Setaka,M.; Ndolela,L.; Odu,N.; Mabuya,R.; Muhammad,S.H.; Osei, Salomey; Samb,S.; Guge,T.K.; Sherman,T.V.; Stenetorp,P.
    Despite the widespread adoption of Large language models (LLMs), their remarkable capabilities remain limited to a few high-resource languages. Additionally, many low-resource languages (e.g., African languages) are often evaluated only on basic text classification tasks due to the lack of appropriate or comprehensive benchmarks outside of high-resource languages. In this paper, we introduce IrokoBench-a human-translated benchmark dataset for 17 typologically-diverse low-resource African languages covering three tasks: natural language inference (AfriXNLI), mathematical reasoning (AfriMGSM), and multi-choice knowledge-based question answering (AfriMMLU). We use IrokoBench to evaluate zero-shot, few-shot, and translate-test settings (where test sets are translated into English) across 10 open and six proprietary LLMs. Our evaluation reveals a significant performance gap between high-resource languages (such as English and French) and low-resource African languages. We observe a significant performance gap between open and proprietary models, with the highest performing open model, Gemma 2 27B only at 63% of the best-performing proprietary model GPT-4o performance. In addition, machine translating the test set to English before evaluation helped to close the gap for larger models that are English-centric, such as Gemma 2 27B and LLaMa 3.1 70B. These findings suggest that more efforts are needed to develop and adapt LLMs for African languages.
  • Miniatura
    Ítem
    Understanding the role of diversity in ensemble-based AutoML methods for classification tasks
    (Institute of Electrical and Electronics Engineers Inc., 2025-04-17) Osei, Salomey; Masegosa, Andrés R.; Masegosa Arredondo, Antonio David
    Ensemble-based Automated Machine Learning (AutoML) methods have gained prominence for their ability to combine diverse machine learning models, achieving superior generalization performance. Despite their empirical success, the underlying mechanisms driving this performance, particularly the role of model diversity, are not yet adequately understood. This study uses novel theoretical frameworks related to the role of diversity in ensembles, which were recently proposed, to shed light on this issue. In this work, we focus on AutoML methods for classification tasks. We use AUTO-SKLEARN (a widely used AutoML ensemble-based method) as a basis. More specifically, we examine how individual model diversity and performance evolves across the four key phases of AUTO-SKLEARN (base-learners, meta-learning, Bayesian Optimization (BO), and Caruana Ensemble). We also examine how they contribute to the diversity and performance of the final ensemble produced by the AutoML method. Using datasets from the AutoML benchmark, we empirically validate these insights by analyzing error rates and diversity measures across the mentioned phases. Our findings highlight the trade-off between individual model accuracy and ensemble diversity, showing that phases like BO improve the mean error rate of classifiers by nearly 50% percent but reduce their mean diversity by 20%. However, the Caruana phase increases the diversity by a 50% compared to the BO phase, allowing better generalization despite the higher mean error rate of the selected individual models (48% higher than BO). This work provides theoretical and empirical evidence that diversity is critical to the success of ensemble-based AutoML methods and a deeper understanding of diversity’s impact on generalization performance and the role of the different AutoML phases. These findings can contribute to advance the development of more robust and theoretically grounded AutoML frameworks