PhD Candidate in Computer Science
Bar-Ilan University · Advised by Yoav Goldberg
My research focuses on natural language processing — specifically on how we can build systems that understand and retrieve information from text in more flexible, human-like ways. I’m interested in named entity recognition beyond fixed categories, question answering that requires complex reasoning, and tools that help researchers explore scientific literature more effectively.
Researchers often struggle to explore unfamiliar scientific topics — keyword search only works when you already know what to look for. Knowledge Navigator uses an LLM to iteratively browse a paper's citation graph, reading abstracts and deciding which references to follow, much like a curious researcher would. It generates structured, multi-faceted summaries of a research area from a single seed paper.
Traditional NER is limited to a fixed set of entity types (person, location, org). But what if you want to find 'renewable energy companies' or 'neurodegenerative diseases' in text? NERetrieve reframes NER as an open-ended retrieval problem — given any free-text entity type description, retrieve matching entities from a document. We release a large-scale dataset and benchmarks for this task.
When a transformer is trained on multiple tasks at once, what happens inside its attention heads? We find that heads naturally specialize — some become task-specific experts, others learn shared representations, and some go dormant. This work provides a systematic analysis of how multi-task learning shapes the internal structure of transformers, revealing emergent division of labor that no one explicitly designed.
Bar-Ilan University
Advisor: Prof. Yoav Goldberg
Tel Aviv University
Advisor: Prof. Jonathan Berant
Weizmann Institute of Science
Advisor: Prof. Rony Paz