Publications Scientifiques
Explorez l'excellence académique à travers nos travaux de recherche, allant de la santé publique à l'intelligence artificielle responsable.
A Low Resource Language Translation: French To Moore
ABSTRACT
Natural Language Processing (NLP) is an exciting field of artificial intelligence with the goal of enabling machines to understand human language in a natural way. Neural Machine Translation (NMT) stands out as one of the most promising applications of NLP, offering the ability to ef fectively translate text from a source language to a target language. In recent years, NMT has experienced significant advances, marking a major milestone in the development of automatic translation systems. Through the use of neural networks, NMT has demonstrated an ability to capture the nuances of language, thereby improving the quality of translations and making the experience of multilingual communication more seamless and precise. This evolution has opened new perspectives in areas such as international collaboration, intercultural understanding, and the global dissemination of information. However, most African languages, especially those in Burk ina Faso, have received very little research attention in this context. In this article, we propose automated translation models French to Mooré based on Transformers. We achieved a BLEU score of 71.18 for the automated for the second model, French to Mooré translation.
AUTHEURS & AFFILIATIONS
AI BASED APPROACH FOR EARLY DIAGNOSIS SUPPORT IN HEMORRHAGIC STROKE
ABSTRACT
A hemorrhagic stroke is a life-threatening medical condition that happens when a blood vessel in your brain ruptures and bleeds. It constitutes a burden on health services and the victim's family. The current definitive diagnosis of stroke is based on brain scanning. However, the clinical diagnosis of hemorrhagic stroke is complex and depends on the skills and experience of the practitioner. Human diagnostic errors lead to delays in treatment and thus compromise clinical outcomes. Our vision is to propose an artificial intelligence approach for medical assistance in the early clinical diagnosis of hemorrhagic strokes. We studied and compared three machine learning models, namely logistic regression, Random Forest and artificial neural networks, to choose the best one after setting up a stroke dataset and identifying the most important characteristics. We can conclude that our system designed with artificial intelligence is important with satisfactory results to help health workers make the rapid diagnosis of hemorrhagic stroke and promote rapid treatment of suspected patients.
AUTHEURS & AFFILIATIONS
AI driven Generation of News Summaries
ABSTRACT
The escalating volume of online information presents a formidable challenge in swiftly accessing pertinent news. The profusion of articles across diverse sources necessitates however automated methods for efficient extraction and summarization of information. In re sponse to this challenge, our research delves into the design and implementation of an online press synthesis tool, harnessing advanced artificial intelligence models. This paper explores the viability of an AI-based online press synthesis tool incorporating two pre-trained models, GPT-3.5 Turbo 16k and Pegasus Summarizer. These models generate high quality automatic summaries from scraped articles on various news sites. Our methodology encompasses a robust web scraping process utilizing Beautiful Soup, integration of pre-trained models, calculation of article similarity, and audio matching of produced summaries. Evaluation metrics such as ROUGE or BLEU gauge model performance. Experimentation results reveal the GPT-3.5 Turbo 16k model’s superior accuracy. In a test involv ing a renowned article on the declaration of human rights, GPT-3.5 Turbo 16k achieved a BLEU score of 16.39% and a ROUGE score of 0.66%. Notably, the turner007/pegasus-summarizer model demonstrated significant performance, registering a BLEU score of 15.45% and a ROUGE score of 0.45%. These scores were obtained through tests on reference summaries authored by humans. Beyond the results, we identify potential areas for enhancement, including enriching the database with expert-authored human summaries and developing a dynamic approach to adapt to evolving news developments. Furthermore, we explore the unified models approach. In conclusion, our work lays the groundwork for an advanced AI-powered online press synthesis tool, paving the way for continuous improvements in an ever-evolving media landscape.
AUTHEURS & AFFILIATIONS
Artificial Intelligence for the Analysis of the Security Situation in Burkina Faso
ABSTRACT
Face à la situation d’insécurité due au terrorisme que vit le Burkina Faso depuis 2015, la population ne manque de manifester ses sentiments. Les diverses réactions de la population se font par les commentaires sur les différentes plateformes sociales créant ainsi une grande quantité de données. L’analyse de ces opinions pourrait être une aide à la prise de décision appliquée à la sécurité. Cette analyse pourrait se faire grâce aux techniques et méthodes offertes par l’intelligence artificielle (IA). Pour nos travaux dans cet article, nous proposons un outil de web scraping pour collecter des données. Puis, nous utilisons des modèles d’analyse des sentiments et de détection de propos haineux basés sur les transformers [2]. A travers ces recherches, nos contributions sont les suivantes :la mise en place d’une base de données textuelles ou un corpus sur la situation sécuritaire, la proposition d’une approche méthodologique basée sur l’IA pour l’analyse de ces données et la proposition d’un tableau de bord pour la visualisation des résultats d’analyse. Nos travaux sont disponibles sur le dépôt suivant https://github.com/abdoulfataoh/security-situation-analysis Index Terms—situation sécuritaire, intelligence artificielle, transformers, web scraping .
AUTHEURS & AFFILIATIONS
Fakipedia: Building and Exploiting Commonsense Knowledge
ABSTRACT
Misinformation poses a significant challenge, especially in developing countries with low literacy rates. The rapid spread on social media, coupled with their lack of robust verification mecha nisms, makes distinguishing between credible and false information increasingly difficult. This document outlines our efforts to address this challenge in Burkina Faso using the pre-trained lin guistic model CAMEMBERT. Our preliminary results show high performance (up to a 99.34% F1 score) on Fakipedia, suggesting that certain blatant instances of false information can be effec tively identified and stopped by automated tools.
AUTHEURS & AFFILIATIONS
Resource Constrained Environments Video Analytics
ABSTRACT
Developing countries face a growing demand for video analytics, yet often lack sufficient computational resources. This paper addresses this challenge by proposing and evaluating optimization techniques for efficient video stream processing on resource-constrained devices, including edge systems. We introduce and evaluate several techniques, including image resizing, frame skipping, parallel processing, threading, queue management, memory optimization, and buffering. Experimental results demonstrate substantial improvements in frames per second (FPS) and memory usage, enabling real-time video analytics without compromising accuracy. By effectively balancing performance and resource consumption, our methods facilitate the deployment of advanced AI-driven video analysis in resource-limited environments, paving the way for practical real-time monitoring and alert systems.
AUTHEURS & AFFILIATIONS
AI-driven Generation of News Summaries Leveraging GPT and Pegasus
ABSTRACT
The escalating volume of online information presents a formidable challenge in swiftly accessing pertinent news. The profusion of articles across diverse sources necessitates however automated methods for efficient extraction and summarization of information. In re sponse to this challenge, our research delves into the design and implementation of an online press synthesis tool, harnessing advanced artificial intelligence models. This paper explores the viability of an AI-based online press synthesis tool incorporating two pre-trained models, GPT-3.5 Turbo 16k and Pegasus Summarizer. These models generate high quality automatic summaries from scraped articles on various news sites. Our methodology encompasses a robust web scraping process utilizing Beautiful Soup, integration of pre-trained models, calculation of article similarity, and audio matching of produced summaries. Evaluation metrics such as ROUGE or BLEU gauge model performance. Experimentation results reveal the GPT-3.5 Turbo 16k model’s superior accuracy. In a test involv ing a renowned article on the declaration of human rights, GPT-3.5 Turbo 16k achieved a BLEU score of 16.39% and a ROUGE score of 0.66%. Notably, the turner007/pegasus-summarizer model demonstrated significant performance, registering a BLEU score of 15.45% and a ROUGE score of 0.45%. These scores were obtained through tests on reference summaries authored by humans. Beyond the results, we identify potential areas for enhancement, including enriching the database with expert-authored human summaries and developing a dynamic approach to adapt to evolving news developments. Furthermore, we explore the unified models approach. In conclusion, our work lays the groundwork for an advanced AI-powered online press synthesis tool, paving the way for continuous improvements in an ever-evolving media landscape.
AUTHEURS & AFFILIATIONS
Evaluating Acoustic Parameters for DeepFake Audio Identification
ABSTRACT
The progress made in the field of machine learning applied to signal processing offers interesting perspectives in terms of technological evolution but also causes some troubles in terms of ethics and security. For example, we are witnessing the emergence of audio deepFakes used to orchestrate scams. However, although the tools used in the generation of these deepFake audios show good results which can sometimes produce audios that seem to be confused with real audio, it is not impossible to dissect them. In order to detect them, many methods exist, in particular the analysis of the acoustic param eters which can attest to the authenticity of an audio extract. These parameters include energy, power, pitch, signal spectrum, cepstral coefficients, etc. However, these acoustic parameters are numerous and not all of them are suitable for detecting deepFake audio. This paper presents a comparative review of acoustic parameters useful in detecting DeepFake audio. Among them, we highlight the relevance of the study of cepstral parameters such as MFCC compared to other acoustic parameters such as mel-spectograms. The objective is to provide reliable leads in the detection of deepFake audio. Index Terms—deepFake audio, detection, mel-spectogram, MFCC.
AUTHEURS & AFFILIATIONS
Detecting Illicit Data Leaks on Android Smartphones
ABSTRACT
In today’s digital landscape, hackers and espionage agents are increasingly targeting Android, the world’s most prevalent mobile operating system. We introduce DeepDetector- a system based on artifi cial intelligence to recognize data thefts in Android. This model is based upon a large dataset comprising of clean and tainted network traffic trained using a Random Forest Classifier. DeepDetector scores high in two main areas as it achieves 82.9% accuracy for connection anomaly detection and 89.9% recall in connection anomaly detection whereas it gets 78.9% accuracy and 81.6 recall in terms of detection of under the sys tem mounted with Raspberry Pi, automatic data collection, preparing of a dataset, training and testing of the model, as well as leak detection are ensured. In this regard, DeepDetector offers a viable way of enhancing Android user security.
AUTHEURS & AFFILIATIONS
L'intelligence artificielle pour l'analyse de la situation sécuritaire du Burkina Faso
ABSTRACT
Face à la situation d’insécurité due au terrorisme que vit le Burkina Faso depuis 2015, la population ne manque de manifester ses sentiments. Les diverses réactions de la population se font par les commentaires sur les différentes plateformes sociales créant ainsi une grande quantité de données. L’analyse de ces opinions pourrait être une aide à la prise de décision appliquée à la sécurité. Cette analyse pourrait se faire grâce aux techniques et méthodes offertes par l’intelligence artificielle (IA). Pour nos travaux dans cet article, nous proposons un outil de web scraping pour collecter des données. Puis, nous utilisons des modèles d’analyse des sentiments et de détection de propos haineux basés sur les transformers [2]. A travers ces recherches, nos contributions sont les suivantes : la mise en place d’une base de données textuelles ou un corpus sur la situation sécuritaire, la proposition d’une approche méthodologique basée sur l’IA pour l’analyse de ces données et la proposition d’un tableau de bord pour la visualisation des résultats d’analyse. Nos travaux sont disponibles sur le dépôt suivant https://github.com/abdoulfataoh/security-situation-analysis Index Terms—situation sécuritaire, intelligence artificielle, transformers, web scraping .
AUTHEURS & AFFILIATIONS
MetaTPTrans: A Meta-Learning Approach for Multilingual Code Representation Learning
ABSTRACT
Representation learning of source code is essential for apply ing machine learning to software engineering tasks. Learning code representation from a multilingual source code dataset has been shown to be more effective than learning from single language datasets separately, since more training data from multilingual dataset improves the model’s ability to extract language-agnostic information from source code. However, existing multilingual training overlooks the language-specific information which is crucial for modeling source code across different programming languages, while only focusing on learning a unified model with shared parameters among differ ent languages for language-agnostic information modeling. To address this problem, we propose MetaTPTrans, a meta learn ing approach for multilingual code representation learning. MetaTPTrans generates different parameters for the feature extractor according to the specific programming language type of the input code snippet, enabling the model to learn both language-agnostic and language-specific information with dy namic parameters in the feature extractor. We conduct ex periments on the code summarization and code completion tasks to verify the effectiveness of our approach. The results demonstrate the superiority of our approach with significant improvements on state-of-the-art baselines.
AUTHEURS & AFFILIATIONS
SNEToolkit: Spatial Named Entities Disambiguation Toolkit
ABSTRACT
‘‘Can you tell me where San Jose is located?’’ ‘‘Uh! Do you know that there are more than 1700 locations named San Jose in the world?’’ The official name of a location is often not the name with which we are familiar. Spatial named entity (SNE) disambiguation is the process of identifying and assigning precise coordinates to a place name that can be identified in a text. This task is not always straightforward, especially when the place name in question is ambiguous for various reasons. In this context, we are interested in the disambiguation of spatial named entities that can be identified in a textual document on a country level. The solution that we propose is based on a set of techniques that allow us to disambiguate the spatial entity considering the context in which it is mentioned from a certain number of characteristics that are specific to it. The solution uses as input a textual document and extricates the named entities identified therein while associating them with the correct coordinates. SNE disambiguation is designed to support the process of fast exploration of spatiotemporal data analysis, most often for event tracking. The proposed approach was tested on 1360 SNEs extracted from the GeoVirus dataset. The results show that SNEToolkit outperformed the baseline, the standard Geonames geocoder, with a recall value of 0.911 against a recall value of 0.871 for the baseline. A flexible Python package is provided for end users.
AUTHEURS & AFFILIATIONS
Text-to-OWL: Automated Ontology Construction for Tuberculosis Treatment Recommendation
ABSTRACT
This paper presents an automated approach for building ontologies to improve treat ment recommendations for tuberculosis (TB), in particular multidrug-resistant tuberculosis (MDR TB) cases in Burkina Faso, using generative language models such as GPT-3. The aim is to facilitate the personalization of treatments according to the patient profile and drug resistance. Two approaches were explored: a automated approach based on the DaVinci GPT-3 model to generate OWL axioms from natural language sentences and a semi-automated approach using text extraction and natural language processing (NLP) techniques. The automated approach was fine-tuned with a dataset consisting of technical guidelines on TB management. The automated approach created an ontology composed of 158 classes, 55 object properties and 57 data prop erties, outperforming the semi-automated approach in terms of efficiency and accuracy. The axioms generated were validated using Protégé and integrated into a formal knowledge base. The study demonstrates that the use of language models such as GPT-3 can efficiently automate ontology generation, reducing human intervention.This approach is particularly well-suited to the management of complex MDR-TB cases and paves the way for standardization of treatment recommendations, while remaining adaptable to local specificities.
AUTHEURS & AFFILIATIONS
Voice Interaction in Moore Language: Study on Isolated Word Recognition
ABSTRACT
This paper explores the optimization of telephone functionalities through voice interaction in the Moore language, prevalent in Burkina Faso. Data gathered from 492 individuals in Ouagadougou, representing diverse dialects and vocal intensities across age groups, informs the study. Employing K-Nearest Neighbor (KNN), Random Forest (RF), and Recurrent Neural Networks (RNNs), the analysis focuses on 29 Moore language commands, prioritizing practical ity and user interaction. The findings suggest promising prospects for RNNs, achieving a 63% accuracy in recognizing isolated words. This success hints at potential advancements in RNNs, incorporating attention mechanisms and end-to-end technology, catering to the voice-controlled mobile device needs of Moore speakers.
AUTHEURS & AFFILIATIONS
Neural Machine Translation for French–Mooré: Adapting Large Language Models to Low-Resource Languages
ABSTRACT
This work focuses on neural machine transla tion between French and Mooré, leveraging the capabilities of Large Language Models (LLMs) in a low-resource language context. Mooré is a local language widely spoken in Burkina Faso but remains underrepresented in digital resources. Alongside Mooré, French, now a working language, remains widely used in ad ministration, education, justice, etc. The coexis tence of these two languages creates a growing demand for effective translation tools. How ever, Mooré, like many low-resource languages, poses significant challenges for machine trans lation due to the scarcity of parallel corpora and its complex morphology. The main objective of this work is to adapt LLMs for French–Mooré translation. Three pre-trained models were selected: No Lan guage Left Behind (NLLB-200), mBART50, and AfroLM. A corpus of approximately 83,000 validated sentence pairs was compiled from an initial collection of 97,060 pairs through pre-processing, semantic filtering, and human evaluation. Specific adaptations to tok enizers and model architectures were applied to improve translation quality. The results show that the fine-tuned NLLB model outperforms the others, highlighting the importance of native language support. mBART50 achieves comparable performance after fine-tuning, while AfroLM remains less ef fective. Despite existing limitations, this study demonstrates the potential of fine-tuned LLMs for African low-resource languages.
AUTHEURS & AFFILIATIONS
Contributing to Speech-to-Speech Translation for African Low-Resource Languages: Study of French-Mooré Pair
ABSTRACT
Most of African low-resource languages are primarily spoken rather than written and lack large, standardized textual resources. In many communities, low literacy rates and limited access to formal education mean that text based translation technologies alone are insuffi cient for effective communication. As a result, speech-to-speech translation systems play a cru cial role by enabling direct and natural interac tion across languages without requiring reading or writing skills. Such systems are essential for improving access to information, public ser vices, healthcare, and education. The goal of our work is to build powerful transcription and speech synthesis models for Mooré language. Then, these models have been used to build a cascaded voice translation system between French and Mooré, since we already got a French-Mooré machine translation model. We collected Mooré audio-text pairs, reaching a to tal audio duration of 150 hours. Then, We fine tuned Orpheus-3B and XTTS-v2 for speech synthesis and Wav2Vec-Bert-2.0 for transcrip tion task. After fine-tuning and evaluation by 36 Mooré native speakers, XTTS-v2 achieved a MOS of 4.36 out of 5 compared to 3.47 out of 5 for Orpheus-3B. The UTMOS evaluation resulted in 3.47 out of 5 for XTTS-v2 and 2.80 out of 5 for Orpheus-3B. The A/Btests revealed that the evaluators preferred XTTS-v2 Mooré audios in 77.8% of cases compared to 22.2% for Orpheus-3B. After fine-tuning on Mooré, Wav2Vec-Bert-2.0 achieved a WER of 4.24% and a CER of 1.11%. Using these models, we successfully implemented a French-Mooré Speech-to-Speech Translation system.
AUTHEURS & AFFILIATIONS