Explainability in NLP II
April 14th, 2021. 12:00 - 1:30 pm
Speaker: Adam KovacsThe attention mechanism has recently been boosting performance on a variety of NLP tasks. These mechanisms are also one of the main building blocks of Transformer-based models that are achieveing state-of-the-art perfomance on most of the NLP datasets. The attention layers of a deep learning architecture explicitly weight the representations of the input components. Because of this it is often assumed that attention can be used as an explanation of the model. Multiple works have appeared to try to test this hypothesis, including:
The seminar tries to give a brief introduction to the attention mechanism and an overview of the listed papers.Location: Zoom