Shuoyang Ding

After a day of meetings, I just realized that using the surface headphone I got from a Microsoft internship to work for Amazon is an idea that’s doomed to backfire.

This drove us to non-autoregressive models, and a specific architecture called *Levenshtein Transformer* works really similarly to a word-level quality estimation model. Look at the following figure and notice how it resolves both of the weaknesses we mentioned above. (5/N)

Ever staring at a saliency heatmap like this and wonder how reliable those interpretations are? We had the same question and took a deep dive into that in our #NAACL2021 paper (aclweb.org/anthology/2021.na…). Come to our presentation in session 14C starting at 5pm PDT today!

I intentionally chose to read the Chinese version of this book cuz the author was also an expert in Mandarin and was involved in the translation process!

tl;dr for our new paper "Parallelizable Stack Long Short-Term Memory" appearing at NAACL Workshop on Structured Prediction for NLP preprint: arxiv.org/abs/1904.03409 code: github.com/shuoyangd/hoolock #NAACL2019 #SPNLP