SemiAmor : Amortized Lagrangian Prediction for NLP
ANR CE23-2023-0005
Table of Contents
SemiAmor is an ANR-funded project spanning 4 years (2024-2027). It aims to bridge semi-amortized methods developed in probabilistic inference and Lagrangian relaxation for Natural Language Processing. More generally, we develop new models for structured prediction in NLP, with applications in other domain such as Machine Learning and Combinatorial Optimization.
Publications
- Nested Named Entity Recognition as Single-Pass Sequence Labeling, Findings EMNLP 2025
- Bregman Conditional Random Fields: Sequence Labeling with Parallelizable Inference Algorithms, ACL 2025
- Scaling Graph-Based Dependency Parsing with Arc Vectorization and Attention-Based Refinement, NAACL 2025
- Predicting Lagrangian Multipliers for Mixed Integer Linear Programs, ICML 2024
Meetings
Meeting 07/06/24
Finally the finances are sorted out, the project can begin! For our first meeting we discussed:
- Joseph Le Roux (PI) summarized the structure of the project.
- Francesca Demelas presented some work on the Prediction of Lagrangian Multipliers for MILPs, accepted at ICML24.
- Joseph Le Roux presented some preliminary work on Amortization and LR for supervised POS-tagging.
- Caio Corro presented some preliminary results on using Bregman divergence for weakly supervised POS tagging.