Adaptive Translation Routing
When should multilingual text be classified directly, and when should it first be translated into a higher-resource language?
Research profile
Multilingual NLP under real-world constraints
My research focuses on multilingual natural language processing and short-text classification, with particular attention to performance differences across languages. I study when translation through a pivot language improves classification and how routing decisions can account for predictive performance, inference cost, latency, privacy constraints, and statistical uncertainty.
My current work combines compact language models with statistical decision methods to develop multilingual systems that are efficient, locally deployable, and supported by reliable evidence.
View ORCID profileCurrent Research Agenda
Statistical Evidence for Routing
How can e-values and sequential statistical methods identify translation-worthy languages while controlling erroneous routing decisions?
Efficient and Private NLP
How can compact, self-hosted models deliver useful multilingual performance under cost, latency, security, and confidentiality constraints?
Selected Research Projects
2026
Statistical foundations for adaptive translation routing
Question Which languages benefit reliably from translation before classification?
Contribution Formulation of routing as a decision problem under uncertainty, with a paired evaluation protocol and language-selection procedure based on e-values and e-processes.
Result Statistically controlled selection of translation-worthy languages and reduced translation use on SIB-200.
Research output Wajdi Ben Saad and Safa Madiouni, Discovering Translation-Worthy Languages with E-Values. Accepted at the NeurIPS 2026 Workshop “E-Values: From Statistics to ML.”
2025–26
Multilingual NLP under cost and latency constraints
Question Where should the boundary between direct multilingual classification and translation-based classification be placed?
Contribution Design and evaluation of self-hosted routing strategies using compact pretrained models, without task-specific fine-tuning.
Result Selective translation improves weaker-language performance, while the best quality-latency policy varies by task.
Research outputs Wajdi Ben Saad and Safa Madiouni, A Cost-Efficient Routing Pipeline for Multilingual Short-Text Classification Using Small Language Models, published in the proceedings of the 16th International Conference on Advanced Computer Information Technologies (ACIT 2026). The probabilistic extension, When Should We Translate? A Translation-Aware Probabilistic Routing Framework for Cost-Efficient Multilingual Text Classification, was presented orally at the 4th International Conference on Applied Mathematics in Engineering (ICAME 2026); its abstract appears on page 73 of the conference proceedings, eISBN 978-625-5540-34-8.
2022–25
Specialized financial text classification
Question How can short, noisy, domain-specific transaction descriptions be classified reliably?
Contribution Method design, scientific leadership, experimental supervision, and statistical validation.
Result A published classification method followed by industrial deployment within a European banking group.
Research outputs D. T. Ta, Wajdi Ben Saad, and J. Y. Oh, Specialized Text Classification: An Approach to Classifying Open Banking Transactions, published in the proceedings of the IEEE 18th International Conference on Computer Science and Information Technologies (CSIT 2023). This research was extended in Wajdi Ben Saad, Income and Salary Detection from Open Banking Transaction and Payment Data: A Comparative Methodological Perspective, presented orally at the International Fintech Research Conference, University of Perugia, on 31 January 2025.
Publications
Accepted Workshop Paper
Discovering Translation-Worthy Languages with E-Values
Wajdi Ben Saad and Safa Madiouni. Accepted at the NeurIPS 2026 Workshop “E-Values: From Statistics to ML.”
Peer-Reviewed Publications
A Cost-Efficient Routing Pipeline for Multilingual Short-Text Classification Using Small Language Models
Wajdi Ben Saad and Safa Madiouni. 16th International Conference on Advanced Computer Information Technologies (ACIT), 2026.
Specialized Text Classification: An Approach to Classifying Open Banking Transactions
D. T. Ta, Wajdi Ben Saad, and J. Y. Oh. IEEE 18th International Conference on Computer Science and Information Technologies (CSIT), 2023.
Conference Communications
When Should We Translate? A Translation-Aware Probabilistic Routing Framework for Cost-Efficient Multilingual Text Classification
Wajdi Ben Saad and Safa Madiouni. 4th International Conference on Applied Mathematics in Engineering (ICAME), 2026.
Income and Salary Detection from Open Banking Transaction and Payment Data: A Comparative Methodological Perspective
Wajdi Ben Saad. International Fintech Research Conference, University of Perugia, 2025.
Methods and Research Resources
Methods
- E-values and e-processes
- Paired statistical comparisons
- Uncertainty-aware decision rules
- Zero-shot text classification
- Multilingual benchmarking
- Translation-routing policies
Models and Data
- Compact sentence encoders
- SIB-200 multilingual benchmark
- MASSIVE multilingual intent data
- Specialized financial short text
- Self-hosted inference pipelines
- Privacy-constrained environments
Academic Service and Supervision
Scientific Service
- Reviewer, EMNLP 2026 Industry Track
- Reviewer, FLTA 2026
- Reviewer, IEEE AGCS 2026
- Reviewer, CSIT 2023
- Member, IEEE Computer Society
- Member, French Statistical Society
Teaching and Supervision
- Lecturer in data analysis and visualization, Université Paris Cité
- University supervisor for data-science apprentices
- Member of B.U.T. Data Science thesis-defense juries
- Former statistics software lecturer at ESSAI, University of Carthage
- Accredited European Statistician (AEUStat)