Data wydarzenia:

Copy-mechanisms in sequence-to-sequence models

Data: wtorek, 19.07.2022, godz. 11:00-12:00

Prelegent: Tomasz Dwojak (Applica)

Abstrakt: Sequence-to-sequence models are a default architecture in many NLP tasks such as machine translation or summarization. It was discovered that current backbones (RNN, transformers) cannot learn to copy a fragment from the source text to the target. A copy-mechanism is a simple way to solve this problem.
I will present a short story and the latest approaches to the copy-mechanism. In addition, I want to discuss whether the copy-networks are proper solutions to copy-paste problems.

Miejsce: B1-7/8 oraz online