Towards Reliable Code Understanding and Synthesis with LLMs
Talk at University of Reading (online), University of Reading, Reading, UK (online)
CV
Education
PhD in Computer Science & Research Fellow
Computer Science Department,, Sapienza University of Rome, 2014 - 2018
Advisee of Prof. Roberto Navigli. Build state-the-art WSD system IMS-Embed.
Created the state-of-the-art system for word similarity based on sense embeddings, SensEmbedM.Sc. in Computer Science
Faculty of Exact and Naturals Sciences, University of Buenos Aires 2003 - July 2012
Master’s thesis: Neural Networks for Optimization: The Linear Ordering Problem
Thesis valuation: 10/10
Advisor: PhD. Enrique C. Segura
Work experience
ELM Europe Nov 2024 - Ongoing
- London, United Kingdom
- Senior NLP Researcher
Huawei Noah’s Ark Lab Sep 2018 - Oct 2024
- London, United Kingdom
- Expert NLP / Team Leader
- Led a team of 7 to 10 members made by BSc./MSc. interns, contractors and Research Scientist.
- Co-authored, developed and advised 20+ research papers on top level conferences such as ACL, EMNLP, NAACL, EACL, AAAI, etc.
- Led the PanGu-Coder . project which was state-of-the-art model on code generation (HumanEval) and was used as backbone for CodeArts Snap, Code AI assistant from Huawei.
- Coordination and development of InfoXLM-XFT, runner up system on XTREME benchmark for crosslingual NLU (Oct 2021) via data augmentation and post training.
- Coordination and development Nezha Plus, runner up system on SuperGLUE benchmark for General NLU (2020).
- Interviewed 60+ candidates on 6 years. Helped to create the team from scratch.
Google Zurich July 2016 - Sep 2016
- Zurich, Switzerland
- Software Engineer Intern
- C++ development on the YouTube Content ID team. Improving detecting copyrighted material by incorporating metadata (textual) information.
Flowics, April 2012 - April 2014
- Buenos Aires, Argentina
- Machine Learning Engineer
- Build a Word Cloud generator with Collocations for Streaming (Apache Storm) and offline (Hadoop).
- User demographic identification, i.e. geolocation, gender, age range, preferred language, etc.
- Build several domain specific sentiment analysis models.
Keepcon July 2010 - March 2012
- Buenos Aires, Argentina
- R+D Developer
- Build a Machine Learning based Sentiment Analizer for social network sanitization.
- Category prediction of ads for international online retail client.
Talks
Program Synthesis with Pretrained Language Models
Talk at Cambridge Language Technology Lab Seminars, University of Cambridge, Cambridge, UK
Pretrained Language Models for Code Understanding and Generation
Talk at 1st Cardiff NLP Summer Workshop, University of Cardiff, Cardiff, Wales, UK
Dialogue Systems and Natural Language Understanding
Talk at Huawei Moscow Speech and Language Seminar, Moscow, Russia (online)
An Overview on Conversational AI
Talk at Data and knowledge engineering seminars, School of Computer Science and Informatics, Cardiff University, Cardiff, Wales
Semantic Representations of Word Senses and Concepts
Tutorial at 54rd Annual Meeting of the Association for Computational Linguistics (ACL 2016), Berlin, Germany
Supervision
Department of Computer Science, University College London 2019 - 2024
Co-supervision MSc. thesis under the UCL IXN programme.
- 2024 Abhineet Kumar - Code Synthesis Via Planning (UCL supervisor: Prof. Jun Wang)
- 2023 Qi Wang - Improving Text-to-Code Generation with Curriculum Learning (cosupervised with Dr. Fenia Christopoulou and UCL Prof. Mark Herbster)
- 2022 Jingwei Ni - Novel Auxiliary Objectives for Code Generation/Understanding Tasks (Prof. Carlo Ciliberto)
- 2021 Han Zhou Zero-shot Transfer Learning for Multi-Domain and Multi-Lingual Dialogue State Tracking (UCL supervisors: Dr. Pasquale Minervini and Prof. Pontus Stenetorp)
- 2021 Nils Zerrer - Task-oriented Conversational Modelling with Unstructured Knowledge Access (UCL supervisor: Prof. Pontus Stenetorp) • 2020 Matteo Donati - On the Robustness of Neural Reading Comprehension Models (UCL supervisor: Dr. Pasquale Minervini)