Making Better Use of (Large) Language and Translation Models with Simple Inference Improvements.

Add to your list(s) Download to your calendar using vCal

• Rico Sennrich, University of Zurich
• Thursday 22 February 2024, 11:00-12:00
• https://cam-ac-uk.zoom.us/j/97599459216?pwd=QTRsOWZCOXRTREVnbTJBdXVpOXFvdz09.

If you have a question about this talk, please contact Panagiotis Fytas.

In a field where the state of the art is often advanced by scale – building larger models on more data – I will make the argument that a surprising amount of progress can be achieved with simple modifications to inference algorithms. In this talk, I will focus on machine translation, where massively multilingual models and large language models have been shown to handle many translation directions, but which still suffer from problems such as hallucinations or translations in the wrong language. I will show how these issues can be reduced massively with contrastive decoding methods that pair each input with appropriate contrastive inputs. I will also discuss Minimum Bayes Risk (MBR) Decoding, a decoding method that has received renewed interest because it avoids common pitfalls in machine translation, but which suffers from a major increase in computational cost. However, I will show how the computational complexity of MBR decoding can be reduced from quadratic to linear to the number of samples by using reference aggregation.

This talk is part of the Language Technology Lab Seminars series.

This talk is included in these lists:

• Cambridge Centre for Data-Driven Discovery (C2D3)
• Cambridge Forum of Science and Humanities
• Cambridge Language Sciences
• Cambridge talks
• Chris Davis' list
• Guy Emerson's list
• Interested Talks
• Language Sciences for Graduate Students
• Language Technology Lab Seminars
• Simon Baker's List
• Trust & Technology Initiative - interesting events
• bld31
• https://cam-ac-uk.zoom.us/j/97599459216?pwd=QTRsOWZCOXRTREVnbTJBdXVpOXFvdz09
• ndk22's list
• ob366-ai4er
• rp587

Note that ex-directory lists are not shown.