31 - Tying Word Vectors and Word Classifiers: A Loss Framework for Language Modeling
from NLP Highlights Hosted by Allen Institute for Artificial Intelligence
ICLR 2017 paper by Hakan Inan, Khashayar Khosravi, Richard Socher, presented by Waleed. The paper presents some tricks for training better language models. It introduces a modified loss function for language modeling, where producing a word that is similar to the target word is not penalized as much as producing a word that is very different to the target (I've seen this in other places, e.g., image classification, but not in language modeling). They also give theoretical and empirical justification for tying input and output embeddings. https://www.semanticscholar.org/paper/Tying-Word-Vectors-and-Word-Classifiers-A-Loss-Fra-Inan-Khosravi/424aef7340ee618132cc3314669400e23ad910ba
Embed this episode
No transcript yet.