Hassan Soliman

Updates

2024

[09-2024] Presented my publication with the title "Scalable Mentoring Support with a Large Language Model Chatbot" at the ECTEL'24 conference, and won the 2nd Best Demo Paper.

[09-2024] Presented my publication with the title "Generative KI zur Lernenbegleitung in den Bildungswissenschaften: Implementierung eines LLM-basierten Chatbots im Lehramtsstudium" at the DELFI'24 conference.

[07-2024] Presented my publication with the title "Using Large Language Models for Adaptive Dialogue Management in Digital Telephone Assistants" at the UMAP'24 conference (HAAPIE).

[07-2024] Attended the UMAP'24 (User Modelling Adaptation and Personalization) conference in Cagliari, Italy, and attended several keynotes and workshops.

[05-2024] Our demo paper in DFKI with the title "Scalable Mentoring Support with a Large Language Model Chatbot" got accepted to be published at the ECTEL'24 conference in September.

[05-2024] My work with chatbot design in DFKI with the title "Generative KI zur Lernenbegleitung in den Bildungswissenschaften: Implementierung eines LLM-basierten Chatbots im Lehramtsstudium" got accepted to be published at the DELFI'24 conference in September.

[04-2024] My work with dialogue utterance adaptation based on user's context in DFKI with the title "Using Large Language Models for Adaptive Dialogue Management in Digital Telephone Assistants" got accepted to be published at the UMAP'24 conference (HAAPIE).

[03-2024] Presented recent advancements in Chatbot Design for university-level courses during my work in DFKI in collaboration with our partners from the University of Leipzig.

Interests

Artificial Intelligence (AI)

Natural Language Processing (LLM)

Information Retrieval (IR)

Question Answering (QA)

Retrieval Augmented Generation (RAG)

Education

Awards

Master of Science in Computer Science

Saarland Informatics Campus, Saarland University

2019 — 2022

Thesis Title: Cross-Domain Neural Entity Linking
Supervision: Prof. Dietrich Klakow

Saarbrücken, Germany

Bachelor of Science in Computer and Communication Engineering

Faculty of Engineering, Alexandria University

2013 — 2018

Thesis Title: Egyptian Car License Plate Information Detection
Supervision: Prof. Marwan Torki

Alexandria, Egypt

Second Best Demo Paper

ECTEL

2024

Krems, Austria

First Class Honor Degree

Alexandria University

2018

Alexandria, Egypt

Philosophy Doctor in Computer Science

Uniquoutious Knowledge Processing Lab,
Technische Universität Darmstadt

2025 — Present

Thesis Title: Safeguarding Multi-Modal LLMs against Misleading Evidence Attacks
Supervision: Prof. Iryna Gurevych

Darmstadt, Germany

Selected Publications

[09-2024] Hassan Soliman, Miloš Kravčík, Alexander Tobias Neumann, Yue Yin, Norbert Pengel and Maike Haag. 2024. Scalable Mentoring Support with a Large Language Model Chatbot. Technology Enhanced Learning for Inclusive and Equitable Quality Education (ECTEL), September 16–20, 2024, Krems, Austria, 6 pages.

ECTEL'24

ACL'24
RepL4NLP

[05-2022] Hassan Soliman, Heike Adel, Mohamed H. Gad-Elrab, Dragan Milchevski, and Jannik Strötgen. 2022. A Study on Entity Linking Across Domains: Which Data is Best for Fine-Tuning?. In Proceedings of the 7th Workshop on Representation Learning for NLP, ACL, 184–190, Dublin, Ireland.

Experience

AI Researcher

German Research Center for AI (DFKI)

Jan 2023 — Dec 2024

• Led two projects in the Educational Technology lab, managing technical implementation and supervising two students.

• Developed a chatbot for a graduate-level course that answered student queries with 87% accuracy. One of the two papers published on the project was nominated for the Best Demo Award at ECTEL 2024.

• Applied advanced Retrieval-Augmented Generation (RAG) techniques, including Hybrid Ensemble Search and Reranking Mechanism, to enhance chatbot interactions and improve the retrieval of course materials.

• Supported mentoring-style conversations by leveraging flexible agentic workflows with LangGraph, utilizing multiple small open-source models hosted on Azure, and using databases for user usage tracking and monitoring.

• Implemented a sub-module for adaptive dialogue systems, customizing responses based on user emotional state and demographics, and benchmarking performance using OpenAI LLMs and open-source models.

Berlin, Germany

Applied Scientist Intern

Bosch Center for AI (BCAI)

May 2022 — Aug 2022

• Contributed to the NLP & Semantic Reasoning group, applying findings from my master’s thesis on Neural Entity Linking to a high-impact industrial project using real data at Bosch.

• Refactored, tested, and documented production-level code for machine learning models, ensuring scalability and efficiency for real-world deployment, leveraging the in-house GPU cluster for model fine-tuning.

• Trained and evaluated machine learning models on a large-scale domain-specific dataset, achieving 77% end-to-end recall for top-3 entity predictions, outperforming existing models.

Renningen, Germany

Research Assistant

Max Planck Institute for Informatics (MPII)

Nov 2020 — May 2021

• Developed a model prototype within the Database & Information Systems group to identify diverse peer groups for entities, contributing to advanced set expansion techniques.

• Implemented a baseline model for entity set expansion, leveraging Wikipedia lists as a knowledge source to enhance the model’s accuracy and comprehensiveness in the expanded sets.

• Optimized the algorithm’s performance by achieving a 3x faster runtime using efficient sparse matrix multiplication techniques, significantly improving computational efficiency.

Saarbrücken, Germany

Master's Thesis Student

Bosch Center for AI (BCAI)

Jun 2021 — Jan 2022

• Joined the NLP & Semantic Reasoning group and worked on a unified system for linking named entities to general-domain (Wikipedia) and domain-specific knowledge bases (KBs), using context-aware embeddings (BERT) to learn a joint vector space. A pre-print of the thesis is available on arXiv. https://arxiv.org/abs/2210.15616.

• Optimized a state-of-the-art model for cross-domain applications, supporting domain extension and identifying optimal data sources for fine-tuning, and improved GPU memory utilization for efficient embedding calculations.

• Achieved a 9% increase in Average Precision for the top-1 entity and a 20% gain in Mean Average Precision (MAP) for top-10 entity linking across four domain-specific KBs, resulting in a workshop publication at ACL 2022.

Renningen, Germany

Software Development Engineer Intern

Amazon

Aug 2019 — Feb 2020

• Maintained a web-based simulation tool for the Fulfillment Acceleration team using the AWS cloud platform, working as a full-stack software engineer.

• Enhanced delivery speed simulations for prime customers, contributing to a successful report on fulfillment operations and improving delivery efficiency.

• Collaborated as a system administrator in an Agile environment, managing server infrastructure and providing technical support for team tools.

Luxembourg, Luxembourg

Projects

SmolLM: Implementing, Fine-Tuning, and Aligning a LLM for Grammatical Error Correction

• Implemented the SmolLM-135M (by HuggingFace) language model architecture, including components like Rotary Positional Embeddings, KV Cache, and Grouped-Query Attention, RMS Normalization, and SwiGLU Activation.

• Fine-tuned the model on the Grammatical Error Correction (GEC) task using the Grammarly CoEdIT dataset.

• Applied RLAIF through Direct Preference Optimization (DPO) to align model outputs with desired corrections.

• Created a Colab notebook to guide users through implementation, fine-tuning, and evaluation processes.

• Achieved significant improvements in grammatical error correction accuracy, scoring an expected BLUE score of ∼ 0.48.

• Leveraged Python libraries such as PyTorch, Transformers, Datasets, and TRL to build and train the model effectively.

• The code is open-source and published on Github.

LinguaLexMatch: Enhanced Document Language Detection

• Developed and evaluated three language detection models, including an Embedding-Based approach, a TF-IDF-based Multinomial Naive Bayes model, and a fine-tuned Transformer-Based methodology.

• Implemented an embedding-based approach using the intfloat/multilingual-e5-large-instruct model by generating a representative embedding for each language and classifying documents based on cosine similarity.

• Benchmarked models on the papluca/language-identification dataset, achieving 99.81% accuracy with the embedding-based model.

• Analyzed performance metrics such as Accuracy, F1 Scores, and Confusion Matrices across 20 different languages.

• Developed a Colab notebook for replicable implementation and evaluation of different language detection models.

• Utilized Python libraries including Datasets, Transformers, and Scikit-learn for model development and evaluation.

• The code is open-source and published on Github.

Scalable Mentoring Support with a LLM Chatbot

• Designed and implemented a chatbot based on an LLM-based Agent to provide scalable educational support and timely feedback to students of education sciences, demonstrating significant potential of generative AI in education.

• Utilized Advanced techniques in Retrieval Augmented Generation (RAG) to enhance chatbot interactions, e.g., Hybrid Ensemble Search and Reranking Mechanism, enabling it to retrieve and analyze course materials effectively.

• The code is subject to a Non-Disclosure Agreement.

Using LLMs for Adaptive Dialogue Management

• Adapted user-directed utterances using LLMs based on the user's parameters like gender, age, and sentiment, aiming to optimize user satisfaction in conversational AI systems, focusing on healthcare patient-practice interactions.

• Evaluated different LLMs and open-source tools for effectiveness in utterance adaptation, in terms of speed, cost-effectiveness, and quality of the generated text based on the adaptation relevancy and adaptation adequacy.

• The code is subject to a Non-Disclosure Agreement.