FactAlign: Fact-Level Hallucination Detection and Classification Through Knowledge Graph Alignment
Recasts hallucination detection as knowledge-graph alignment, which also separates intrinsic from extrinsic hallucinations.
Research
Papers and degrees
Three peer-reviewed papers: two on Arabic speech — dialect identification, then pronunciation scoring, which became my M.Sc. thesis — and one on catching hallucinations by aligning what a model says against a knowledge graph.
Fine-Tuning Self-Supervised Learning Models for End-to-End Pronunciation Scoring
E2E-R, an end-to-end pronunciation scoring architecture that reaches state of the art on less data and less compute.
A Character Level Convolutional BiLSTM for Arabic Dialect Identification
Dialect identification straight from characters, avoiding the tokenisation problem that Arabic dialects create.
Education
M.Sc. in Computer Science
Cairo University, Faculty of Computers and Information
- Thesis: “Enhancement of Mispronunciation Detection Using Deep Learning Techniques”.
- Proposed E2E-R, an end-to-end pronunciation scoring architecture using fine-tuned SSL speech models, achieving state-of-the-art performance while reducing data and compute needs.
B.Sc. in Computer Engineering
Cairo University, Faculty of Engineering
- Graduation project: Animtractor, a marker-less motion capture system that needs no depth camera. Won first place in the Innovation track of the Microsoft Imagine Cup 2016 national finals.
- Prepared cloud computing coursework and summer training for CMP303B — Distributed Operating Systems.
- Member of the Academic Committee in IEEE’s Cairo University Student Branch.