|
How to choose the test set size? Some observations on the evaluation of PoS taggers on the Universal Dependencies project, by Guillaume Wisniewski (LLF & Univ. Paris VII)| old_uid | 17998 |
|---|
| title | How to choose the test set size? Some observations on the evaluation of PoS taggers on the Universal Dependencies project, by Guillaume Wisniewski (LLF & Univ. Paris VII) |
|---|
| start_date | 2019/10/18 |
|---|
| schedule | 11h30 |
|---|
| online | no |
|---|
| summary | This presentation questions the usual framework of statistical learning in which test set and train sets are fixed arbitrarily and independently of the model considered. Taking the evaluation of PoS taggers on the UD project as an example, we show that, in many cases, it is possible to consider smaller test sets than those generally available without hurting evaluation quality and that the examples that have been `saved' can be added to the train set to improve system performance, especially in the context of domain adaptation. |
|---|
| responsibles | Seddah |
|---|
| |
|