|
Graph-based semantic parsing, compositional generalization and loss functions| title | Graph-based semantic parsing, compositional generalization and loss functions |
|---|
| start_date | 2023/04/28 |
|---|
| schedule | 11h |
|---|
| online | no |
|---|
| location_info | salle C434 & en ligne |
|---|
| details | Ce séminaire aura lieu en français. |
|---|
| summary | Semantic parsing aims to transform a natural language utterance into a structured representation that can be easily manipulated by a software (for example to query a database). As such, it is a central task in human-computer interfaces. It has recently been observed that sequence-to-sequence models struggle on settings that require compositional generalization. On the contrary, previous work has shown that span-based parsers are more robust.
In this talk, I will present our work on reentrancy-free semantic parsing. We proposed a novel graph-based formulation of this problem which addresses the search space issue of span-based models. We proved that both MAP inference and latent tag anchoring (required for weakly-supervised learning) are NP-hard problems. Therefore, we developed approximation algorithms based on combinatorial optimization techniques. Our approach delivers novel state-of-art results on standard benchmarks that test for compositional generalization. Finally, I will discuss the theoretical limitations of token-separable losses that are commonly used in the literature (and in this work!) to bypass expensive computation of the log-partition function, followed by ongoing work on weakly-supervised learning.
The research presented in this talk is a joint work with Alban Petit, Emile Chapuis and myself. |
|---|
| responsibles | Bawden |
|---|
Workflow history| from state (1) | to state | comment | date |
|---|
| submitted | published | | 2023/04/25 10:32 UTC |
| |
|