|
Temporality and Modality in Entailment Graph Learning| title | Temporality and Modality in Entailment Graph Learning |
|---|
| start_date | 2023/01/13 |
|---|
| schedule | 11h |
|---|
| online | no |
|---|
| visio | https://docs.google.com/document/d/1bnuIxo9F2WLPsdI9VIn233fMVwybFSUequrUh4vlYkk/edit?usp=sharing |
|---|
| location_info | salle C434 & visio-conférence |
|---|
| summary | The ability to recognise textual entailment and paraphrase is crucial in many downstream tasks, including Open-domain Question Answering from text. Entailment Graphs, constructed via unsupervised learning techniques over large text corpora, provide a solution to learning and encoding this information. Entailment Graphs comprise nodes representing linguistic predicates and edges representing the entailment relations between them.
Despite recent progress, methods for learning Entailment Graphs produce many spurious entailment relations. In this talk I propose the incorporation of temporal information and the linguistic modality of predicates as signals for refining the entailment graph learning process. I will show that these phenomena are useful in disentangling highly correlated but contradictory predicates, such as "winning » and "losing”. |
|---|
| responsibles | Bawden |
|---|
Workflow history| from state (1) | to state | comment | date |
|---|
| submitted | published | | 2023/01/12 14:29 UTC |
| |
|