A5069950074- Topic Modeling
- Natural Language Processing Techniques
- Domain Adaptation and Few-Shot Learning
- Semantic Web and Ontologies
- Biomedical Text Mining and Ontologies
- Multimodal Machine Learning Applications
- Advanced Text Analysis Techniques
- Data Quality and Management
- Neural Networks and Applications
- Web Data Mining and Analysis
- Systemic Sclerosis and Related Diseases
- Advanced Image and Video Retrieval Techniques
- Skin Diseases and Diabetes
- Radiology practices and education
- Radiomics and Machine Learning in Medical Imaging
- Advanced Graph Neural Networks
- Privacy-Preserving Technologies in Data
- Dermatologic Treatments and Research
- Text and Document Classification Technologies
- Data Management and Algorithms
- Complex Network Analysis Techniques
- Explainable Artificial Intelligence (XAI)
- Machine Learning in Healthcare
- Speech and dialogue systems
- Sentiment Analysis and Opinion Mining
ZHAW Zurich University of Applied Sciences
2025
University of Zurich
2018-2024
University Hospital of Zurich
2021-2024
ETH Zurich
2023
Kyoto University
2023
Dalle Molle Institute for Artificial Intelligence Research
2023
University of Oslo
2018-2020
Institut national de recherche en informatique et en automatique
2016-2018
Research Centre Inria Sophia Antipolis - Méditerranée
2016
Universitatea Națională de Știință și Tehnologie Politehnica București
2013
Learning what to share between tasks has become a topic of great importance, as strategic sharing knowledge been shown improve downstream task performance. This is particularly important for multilingual applications, most languages in the world are under-resourced. Here, we consider setting training models on multiple different at same time, when little or no data available other than English. We show that this challenging setup can be approached using meta-learning: addition source...
Inspired by Curriculum Learning, we propose a consecutive (i.e., image-to-text-to-text) generation framework where divide the problem of radiology report into two steps. Contrary to generating full from image at once, model generates global concepts in first step and then reforms them finer coherent texts using transformer-based architecture. We follow sequence-to-sequence paradigm each step. improve upon state-of-the-art on benchmark datasets.
The widespread use of chest X-rays (CXRs), coupled with a shortage radiologists, has driven growing interest in automated CXR analysis and AI-assisted reporting. While existing vision-language models (VLMs) show promise specific tasks such as report generation or abnormality detection, they often lack support for interactive diagnostic capabilities. In this work we present RadVLM, compact, multitask conversational foundation model designed interpretation. To end, curate large-scale...
Many different methods for prompting large language models have been developed since the emergence of OpenAI's ChatGPT in November 2022. In this work, we evaluate six few-shot methods. The first set experiments evaluates three frameworks that focus on quantity or type shots a prompt: baseline method with simple prompt and small number shots, random 10, 20, 30 similarity-based prompting. second target optimizing enhancing through Large Language Model (LLM)-generated explanations, using...
The first objective of this study was to implement and assess the performance reliability a vision transformer (ViT)-based deep-learning model, an 'off-the-shelf' artificial intelligence solution, for identifying distinct signs microangiopathy in nailfold capilloroscopy (NFC) images patients with SSc. second compare ViT's analysis that practising rheumatologists.NFC prospectively enrolled our European Scleroderma Trials Research group (EUSTAR) Very Early Diagnosis Systemic Sclerosis (VEDOSS)...
The main objective of this paper is to compare the sentiments that prevailed before and after presidential elections, held in both US France year 2012. To achieve we extracted content information from a social medium such as Twitter used tweets electoral candidates public users (voters), collected by means crawling during course election. In order gain useful insights about scored for each tweet using different metrics performed time series analysis topics (identified specific keywords)....
Existing named entity recognition (NER) systems rely on large amounts of human-labeled data for supervision. However, obtaining large-scale annotated is challenging particularly in specific domains like health-care, e-commerce and so on. Given the availability domain knowledge resources, (e.g., ontologies, dictionaries), distant supervision a solution to generate automatically labeled training reduce human effort. The outcome NER, however, often noisy. False positive false negative instances...
This article presents the SIRIUS-LTG-UiO system for SemEval 2018 Task 7 on Semantic Relation Extraction and Classification in Scientific Papers. First we extract shortest dependency path (sdp) between two entities, then introduce a convolutional neural network (CNN) which takes embeddings as input performs relation classification with differing objectives each subtask of shared task. approach achieved overall F1 scores 76.7 83.2 clean noisy data, respectively. Furthermore, combined...
While several benefits were realized for multilingual vision-language pretrained models, recent benchmarks across various tasks and languages showed poor cross-lingual generalisation when multilingually pre-trained models are applied to non-English data, with a large gap between (supervised) English performance (zero-shot) transfer. In this work, we explore the of these on zero-shot visual question answering (VQA) task, where fine-tuned visual-question data evaluated 7 typologically diverse...
Sanghwan Kim, Farhad Nooralahzadeh, Morteza Rohanian, Koji Fujimoto, Mizuho Nishio, Ryo Sakamoto, Fabio Rinaldi, Michael Krauthammer. The 22nd Workshop on Biomedical Natural Language Processing and BioNLP Shared Tasks. 2023.
This article presents the SIRIUS-LTG system for Fact Extraction and VERification (FEVER) Shared Task. It consists of three components: 1) Wikipedia Page Retrieval: First we extract entities in claim, then find potential URI candidates each using a SPARQL query over DBpedia 2) Sentence selection: We investigate various techniques i.e. Smooth Inverse Frequency (SIF), Word Mover’s Distance (WMD), Soft-Cosine Similarity, Cosine similarity with unigram Term Document (TF-IDF) to rank sentences by...
The potential for improvements brought by Large Language Models (LLMs) in Text-to-SQL systems is mostly assessed on monolingual English datasets. However, LLMs' performance other languages remains vastly unexplored. In this work, we release the StatBot.Swiss dataset, first bilingual benchmark evaluating based real-world applications. dataset contains 455 natural language/SQL-pairs over 35 big databases with varying level of complexity both and German. We evaluate state-of-the-art LLMs such...
International enterprises, organizations, or hospitals collect large amounts of multi-modal data stored in databases, text documents, images, and videos. While there has been recent progress the separate fields exploration as well database systems that automatically translate natural language questions to query languages, research challenge querying combined with other unstructured modalities such images is widely unexplored. In this paper, we propose XMODE - a system enables explainable,...
We investigate the use of different syntactic dependency representations in a neural relation classification task and compare CoNLL, Stanford Basic Universal Dependencies schemes. further with syntax-agnostic approach perform an error analysis order to gain better understanding results.
Background: Federated learning methods offer the possibility of training machine models on privacy-sensitive data sets, which cannot be easily shared. Multiple regulations pose strict requirements storage and usage healthcare data, leading to being in silos (i.e. locked-in at facilities). The application federated algorithms these datasets could accelerate disease diagnostic, drug development, as well improve patient care. Methods: We present an extensive evaluation impact different...
While several benefits were realized for multilingual vision-language pretrained models, recent benchmarks across various tasks and languages showed poor cross-lingual generalisation when multilingually pre-trained models are applied to non-English data, with a large gap between (supervised) English performance (zero-shot) transfer. In this work, we explore the of these on zero-shot visual question answering (VQA) task, where fine-tuned visual-question data evaluated 7 typologically diverse...
Inspired by Curriculum Learning, we propose a consecutive (i.e., image-to-text-to-text) generation framework where divide the problem of radiology report into two steps. Contrary to generating full from image at once, model generates global concepts in first step and then reforms them finer coherent texts using transformer architecture. We follow transformer-based sequence-to-sequence paradigm each step. improve upon state-of-the-art on benchmark datasets.