How can an idea be captured?
through WORDS or maybe with an IMAGE? with a GRAPH? a LOGIC?
I am an associate professor of Artificial Intelligence, Department of Computer Science, Technical University of Cluj-Napoca. My current work involves research on meaning representation through different modalities: text, vision, knowledge graphs and, of course, their combination. We have dense vectors, but how can learning be done such that the result is robust, general, adaptable?
Relevant Projects
Director:
- Vocabulary mapping and duplicates identification - Research project in collaboration with msg systems ag (Munich), Germany : 2023-2024
- Improve the existing cognitive search approach by exploring semistructured job descriptions - Research project in collaboration with msg systems ag (Munich), Germany : 2022-2023
- Improve the existing cognitive search approach by extending the information extracted from semi-structured job descriptions and quickly learning from user feedback - Research project in collaboration with minnosphere, Germany : 2021-2022
Member:
- New Optical Coherence Tomography Biomarkers Identified with Deep Learning for Risk
Stratification of
Patients with Age-related Macular Degeneration DeLArMaD, PED616 – UEFISCDI: 2022-2024
See details: Application for resident training
GO TO LLM answers for AMD questions - Extensive capitalization of experience in activities of Space and Security, PNIII-P1-1.2-1.2.1 PCCDI 2018 – member – UEFISCDI - 2018-2020
- Cooperative Advanced Driving Assistance System Based on Smart Mobile Platforms and Road Side Units (SmartCoDrive), PCCA – member - UEFISCDI:2014-2016
Courses:
Artificial Intelligence
Heuristics for search space. Adversarial Search. Propositional Logic. First Order Logic. Probabilities. Hidden Markov Models
Intelligent Systems
Supervised machine learning. Decision Trees. Linear regression. Neural networks. Convolutional neural Networks. Transformer. Natural Language Processing
Formal Languages and Translators
Grammars. Regular grammars. Context free grammars. LL(k). LR(k).
Activity:
HybridOM tool is an unsupervised approach for ontology matching that blend knowledge graphs with LLM for ontology matching. HybridOM was evaluated within Bio-ML 2024 track for the task of concept matching and it achieved the highest values for F1-score and Recall for most of the ontology pairs while maintaining a balance between precision and recall. The proposed method has been adapted for industrial usage in a human capital management product called msg.ProfileMap.
Our finetunned smolLM2 obtained 7th place, at 0.025 distance to the 1st team, in Subtask3: given a news article and a dominant narrative of the text of this article, generate a free-text explanation supporting the choice of this dominant narrative. The same smolLM2 was also finetuned to assign a role from a predefined role-taxonomy to list of mentions of named entities in the article.
Romanian strategy for Open Science and High Performance Computing
Strategic framework for the adoption and use of innovative technologies in public
administration 2021 – 2027 –
solutions for streamlining the activity, SIPOCA 704,
Activity A7.1. National
strategic framework and
financing instruments for Romania's participation in European initiatives and networks -
support for evidence-based policies at the central level
- member of the HPC team working on Romanian strategy for Open Science and High
Performance Computing (HPC)
Transilvania Digital Innovation Hub
2024 - Consulting activities on machine learning, Large Language Models, Knowledge Graphs, Retrieval Augmented Generation
2021 - member of the 2nd place team among Romanian teams, 37th place place worldwide
2023-2024 - advisor for TUCN teams participating in WiDS Datathon++ 2024 University
Challenge
Global Winners: Andreea Onaci, Petrut Paul!
CLEF competition
1st place in the Task 2 - Biomedical question answering over interlinked data - QALD-4 (Question Answering over Linked Data 2014) challenge.
The winning system is presented in Q1 journal paper “Question answering over biomedical linked data with Grammatical Framework” and it consists of a controlled natural language for querying biomedical linked data from Diseasome, Sider, and DrugBank datasources
Visiting scholar
2016 - Department of Mathematics, West Virginia University Eberly College of Arts and Science, Morgantown, WV, USA, for maintenance activities of CoNtRol. CoNtRol is an open source framework for the analysis of chemical reaction networks. It is described in the paper published in Bioinformatics 2014: "CoNtRol: an open source framework for the analysis of chemical reaction networks"
Best paper
best paper award for B. Varga, A.Trambitas-Miron, A. Roth, A. Marginean, R. Slavescu, A. Groza - A natural language processing system for Romanian tourism, ASIR@ Fedcsis., Warsaw, Polonia, 2014