Welcome to Urchade Zaratiana's Webpage
I am a PhD student at Laboratoire Informatique de Paris Nord (LIPN) 🏫, under the supervision of Nadi Tomeh and Thierry Charnois. My research focuses on structured prediction for Natural Language Processing 🧠📊.
I am also a Member of Technical Staff at Fastino 🦊🇺🇸, where I am leading the Data structuring team 🧠. I am currently based in Île de la Réunion 🇫🇷🇷🇪.
I am passionate about the science of deep learning, with particular interest in topics such as domain generalization 🌐, zero/few-shot/instruction learning 🛠️📚, learning under noisy data/labels 📉🔍, scaling laws 📈, and simplicity bias 🎛️.
During my PhD, I have worked on the following research problems:
Structured decoding (2021-)
Named Entity Recognition as Structured Span Prediction (Zaratiana et al., EMNLP 2022 UM-IoS)
EnriCO: Constrained decoding of information extraction using logical rules (Zaratiana et al., ArXiv 2024)
Graph structure learning (2021-)
GraphER: End-to-end graph structure learning of joint entity and relation extraction (Zaratiana et al., ArXiv 2024)
GNNer: Using graph (and Graph Neural Network) to implicitly constrain the output of neural network (Zaratiana et al., ACL 2022 SRW) (Link)
Structured loss functions (2022-)
Filtered Semi-Markov CRF (Zaratiana et al., EMNLP 2023)
Global Span Selection (Zaratiana et al., EMNLP 2022 UM-IoS)
Constrained decoding of language model (2022-)
Autoregressive text-to-graph model for Information extraction (Zaratiana et al., AAAI 2024)
Zero-shot Learning for Information Extraction (2023-)
GLiNER (Zaratiana et al., NAACL 2024): Zero-shot Named Entity Recognition ()
GraphER (Zaratiana et al., ArXiv 2024): An end-to-end model for zero-shot joint entity and relation extraction
Arabic Information Extraction (2023-)
Cross-Dialectal Named Entity Recognition in Arabic (El Khbir et al., ArabNLP 2023)
A Span-Based Approach for Flat and Nested Arabic Named Entity Recognition: Top 1 system at Wojood NER shared task (El Khbir et al., ArabNLP 2023)
Language model pretraining (2024-)
Training a BERT-like model without positional encoding (Zaratiana et al., ICLR 2024 Tiny Paper)
Detection of machine-generated text (2024-)
Our system ranked top 2 at SemEval "Multidomain, Multimodal and Multilingual Machine-Generated Text Detection" shared task using a 300M parameters model (Ben Fares et al, 2024)
Selected OSS Projects
A lightweight model for Named Entity Recognition (NER) using a BERT-like transformer.
End-to-end zero-shot entity and relation extraction.
An autoregressive text-to-graph framework for joint entity and relation extraction.
Structured Information Extraction with Large Language Models.