🔭 Computer Science MSc candidate researching NLP @HUJI (publications), former Data Scientist @PayPal, Research Assistant @AI2, NLP Research Intern @IBM Research

💬 Feel free to ask me about any of my repos, I love getting messages about my work!
☕ If my code or my notes helped you, you can buy me a coffee if you'd like

I have collected all of the detailed notes I wrote during my studies at HebrewU as well as courses I studied independently, and published them as a part of my goal to make Data Science & NLP topics more accessible to Hebrew speakers.

This colelction contains detailed notes in Hebrew on subjects such as Math (Calculus, Linear Algebra, Probability, Discrete Math), foundations of Computer Science (Data Structures, Algorithms, Complexity), as well as advanced Data Science (Machine Learning, NLP).

This includes my recent detailed notes (90 pages) for Stanford's CS224N (NLP with DL) course, that gained more than 1K likes across Israeli DS & ML communities.

Recently I decided to share my private Notion hub where I organize all of my NLP knowledge (mostly in Hebrew). This hub is meant for my personal use, but since many people found it useful I decided to share it. It conatains some notes by topic (such as NLP tasks, architectures or uses) that has significant ovarlaps with my CS224N notes, as well as noted I wrote for a few dozens NLP papers I have read in the past year for my studies and my reasearch.

I shared my simple-but-useful system for queueing and reviewing papers I read (or plan to read). This Notion template is free to use, and also contains tips on how to personalize it to work for your needs.

Wikipedia's revision history records how scientific knowledge evolves over time, but scientifically meaningful edits are buried under a much larger volume of routine ones. In this project we built a system that takes each section-level edit, compares the before and after with an LLM, and labels it with an expert-designed taxonomy of 10 scientifically meaningful edit types (such as addition or removal of scientific information, change in scientific narrative, and addition or removal of academic references), along with an explanation for each label. Interactive views then trace an article's history at any granularity, from aggregate trends across sections and years down to individual edits.

In a user study, experts from three domains used the system to surface patterns and research questions they could not have reached before. This work was published as a paper (arXiv:2607.12441), and we release the system, the code, and a human-annotated benchmark.

Methods used in this NLP project: LLM-based multi-label classification, prompt design, semantic similarity matching, interactive visualization.

In this project, we introduce the novel NLP task of corpy textual style-transfer, which involves the transformation of casual English text into a style suited for a professional workplace setting. We constructed an original parallel corpus comprising 634 sentences in casual English and their corporate-style paraphrases.

This project includes the dataset itself, the code for fine-tuning the style transfer models, 2 of the best performing fine-tuned models, and code for fine-tuning a style detection model for detecting corpy style in text.

Methods used in this NLP projects: Textual style transfer, text classification.

In this project, I started with a raw dataset of quotes (in Hebrew) gathered from protocoles of the Knesset (the Israeli parliment). In the first stage of the project, I used unsupervised topic modeling methods in order to cluster quotes by topics. The topic assignment that was created during the first stage were used to prioritize qoutes for manual tagging process - quotes with the highest confidence score were sent to mannual tagging. This process created ~2,700 quotes that were manually tagged into 8 topics (in addition to a "no topic" tag). Then, in the second phase of this project, I trained a supervised classifier to predict quotes topics.

Methods used in this NLP projects: Topic modeling (unsupervised), Topic classification (supervised).

In this project, we used restaurants menus gathered via Wolt's API and created a smart system that helps groups of friends or co-workers find a single restaurant that matches everyone's needs and preferences (such as vegeterianism, price limits, prefered cuisines etc). We examined several different algorithms (neither are ML-based), all of them provided solutions who were incredibly close to the optimal solution (that could be found by iterating over the entire 30M combinations dataset) in a fraction of the time (up to 11K times faster)!

Methods used in this AI projects: local search, genetic algorithms.