Search

1 Sep 2023: The Eagle 2023. THE EAGLE 2023 1. The Eagle 2023VOLUME 105. WELCOME. 2. Published in the United Kingdom in 2023 by St John’s College, CambridgeFirst published in the United Kingdom in 1858 by St John’s College, Cambridge. Cover photo by: Rich MarshamPrevious page photo: Ascension Day carol, St John’s College Choir directed by Christopher Gray, May 2023. Credit: Joe Giddens, Press Association. Designed by Out of the Bleu (07759 919440; www.outofthebleu.co.uk)Printed by CDP (01517 247000; www.cdp.co.uk). The Eagle is published annually by St John’s College, Cambridge, and is provided free of chargeto members of the College and other interested parties. THE EAGLE 2023 3. Photo: Spring at St John’s, March 2022 Credit: Paul Everest WELCOME. ContentsWelcome. Contributors. 6. Editorial. 7. Message from the Master. 8. Features. St John’s in the 1970s. 14. Cleaning up: converting waste streams into green energy. 17. Writer in Residence. 21. Making way for tigers: decoding the relocation of people from India’s Protected Areas. 24. Ax:son Johnson Centre for the Study of Classical Architecture. 28. College life. Senior Tutor’s report on student welfare support. 34. Education report – Mathematics. 36. The SBR. 39. A day in

24 Nov 2023: Multilingual Models in Neural MachineTranslation. Guangyu Yang. Department of EngineeringUniversity of Cambridge. This dissertation is submitted for the degree ofMaster of Philosophy in Machine Learning and Machine Intelligence. Clare Hall August 2023. Dedicated to my loving parents. Declaration. I, Guangyu Yang of Clare Hall, being a candidate for the MPhil in Machine Learning andMachine Intelligence, hereby declare that this report and the work described in it are my ownwork, unaided except as may be specified below, and that the report does not contain materialthat has already been used to any substantial extent for a comparable purpose. All software used in this thesis was written in Python and PyTorch. The open-sourceHugging Face transformers library1 [47] was used for downloading pre-trained models andinference. The sacreBLEU library2 [26] was used to download datasets and evaluate theperformance of the models using the BLEU metric. The BLEURT library3 [34] was used toevaluate the performance of the models using the BLEURT metric. The OpenICL frameworkwas used to experiment with in-context learning. And the Direct Preference Optimizationcodebase4 [28] was used for unsupervised preference optimization with MBR. The