76 - Increasing In-Class Similarity by Retrofitting Embeddings with Demographics, with Dirk Hovy
NLP Highlights
00:00
The Convergence of Embeddings in Neural Networks
In one case for using retrofitting with a very low alpha on age prediction, it does look as if the convergence would happen a lot faster. The other paper is also accepted at EMNOP where we're actually trying to capture dialects. What we see is that there are very, very clear distinctions between different regions. And what we sees is that if we modify the document embeddings for the different cities that we found by using geographic information, all of these cities are closed together. So it's very likely that they're dialects or the words used there and the language there is more similar to each other. We actually get very, very good results and match sociolinguistic
Transcript
Play full episode
