Thanks a lot for reading!

Later, I'll summerize my thoughts on Keynote presentations and second day. Open for further discussions 😊
8/8
#EMNLP24 #NLP #MT #AI #science

Word Alignment as Preference for Machine Translation!

aclanthology.org/2024.emnlp-m...

🎯 Goal: Combating hallucinations and omissions in process
➡️LLMs as MT
➡️Word Alignment correctly impacts hallucinations and omissions
➡️Mixed results depending on translation directions

7/8

What do large language models need for machine translation evaluation?

🎯 Goal: Approaches for Automatic Evaluation of MT using LLMs.
➡️0-shot, few-shot, and Chain-Of-Thought prompting
➡️Comet>LLMs
➡️CoT require larger models
➡️Larger models ≠ better evaluators

aclanthology.org/2024.emnlp-m...
6/8

Towards Cross-Cultural MT with RAG from Multilingual Knowledge Graphs

🎯 Goal: Transcreation of Named Entities
➡️Use RAG for injecting knowledge to MT system
➡️Aproaches: Implicit RAG and Explicit RAG
➡️No quality degradation in general domain

arxiv.org/abs/2410.14057
5/8

Simul-MuST-C: Simultaneous Multilingual Speech Translation Corpus Using LLM

🎯 Goal: Simultaneous Multilingual Speech MT
➡️Reducing latancy
➡️Linguistic "Salami technique" for data preparation
➡️LLMs for chunking the input
➡️Translate chunks separately and rejoin

aclanthology.org/2024.emnlp-m...
4/8

PsFuture: A Pseudo-Future-based Zero-Shot Adaptive Policy for Simultaneous Machine Translation

🎯 Goal: Simultaneous MT - translating incomplete or streaming sentences
➡️ Adaptively assess if there is enough data to translate
➡️ Use suffixes to guide MT

aclanthology.org/2024.emnlp-m...

3/8

Neuron Specialization: Leveraging intrinsic task modularity for multilingual machine translation
arxiv.org/abs/2404.11201

goal 🎯: reducing Language specific interference
➡️ applied neural masking on MLP layers, 0 added parameters
➡️ increased directional 0-shot performance
2/8

Can't wait for this years #EMNLP24
Going to highlight interesting papers here. See y'all soon!
#throwback #EMNLP23