学生论坛-大语言模型的跨语言可解释性 科研选题和经验分享-澳门大学.pdf

1、语模型的跨语可解释性：科研选题和经验分享CCMT 2024 学论坛 詹润哲 导师：輝副教授 Prof.Derek F.Wong 澳学 科技学院 然语与中葡机器翻译实验室2024/11/102Outline Research Workflow,Dilemma Regarding MT/LLMs(8 mins)Identifying Research Directions Addressing Research Dilemma in the era of LLM Notable examples of MT-related Research practices My Research Method

2、ology&Practices(5 mins)My routines in cross-lingual/multilingual research Brief introduction to two research projects Possible Research Directions(2 mins)3Common Research WorkflowQuestions/ObservationResearch Topic IdeaForm HypothesisTest with ExperimentsAnalyze DataReport ConclusionsConcepts Adapte

3、d From Wikipedia.Application-driven(Top-down)Be the first to 1+1,Make models/systems better Release models/benchmarksIdentifying Research Directions 3Common Research WorkflowQuestions/ObservationResearch Topic IdeaForm HypothesisTest with ExperimentsAnalyze DataReport ConclusionsConcepts Adapted Fro

4、m Wikipedia.Application-driven(Top-down)Be the first to 1+1,Make models/systems better Release models/benchmarks Curiosity-driven(Bottom-up)Why does work?How can we achieve?Identifying Research Directions 4Common Research WorkflowXu,H.,Kim,Y.J.,Sharaf,A.,&Awadalla,H.H.(2023).A paradigm shift in mach

5、ine translation:Boosting translation performance of large language models.ICLR 2023.Alves,D.M.,Pombal,J.,Guerreiro,N.M.,Martins,P.H.,Alves,J.,Farajian,A.,.&Martins,A.F.(2024).Tower:An open multilingual large language model for translation-related tasks.COLM 2024.Identifying Research Directions ALMA

6、continual pre-training and then fine-tune on MT instruction data.Application-driven research(Top-down)Representative machine translation papers:ALMA(ICLR23),Tower(COLM24).4Common Research WorkflowXu,H.,Kim,Y.J.,Sharaf,A.,&Awadalla,H.H.(2023).A paradigm shift in machine translation:Boosting translati