Day 8 - More Reading LRL MT Related Papers

August 31, 2025

• Thoughts reading Teaching Large Language Models to Translate on Low-resource Languages with Textbook Prompting - ACL Anthology (opens in a new tab):
  • Initial thoughts: i like connection to human approaches of learning, but so far not seeing how all 5 senses play a huge role in how we learn. Like, imagine trying to learn to write deaf, blind, can’t smell, can’t taste, can’t feel, and can’t speak(speaking is included because it reinforces thoughts among other things). Seems like only some multimodal approaches consider a few of these aspects, i hope i see some mentioned in this paper.
  • Im not so sure how much “language-specific nuances” can be derived from textbooks to be honest. Also, giving language examples(as the paper does) so the LLM can parse syntax patterns rather than directly providing syntax patterns seems contradictory to the prior statement of “These instructional materials serve as valuable guides, assisting individuals on their journey to fluency, eliminating the need to awkwardly deduce language-specific nuances from extensive language examples”.
  • Just briefly looking at babelnet for Oromo, its not that great and unreliable. For example “happening” translated to “happening” and “before” didn’t have a translation.
  • So far a tad sad not to see comparison of results against other translation models like NLLB or Google Translate or MADLAD-400
  • Retrieval Utility seems like a good way to evaluate how useful some given extra knowledge really is (in this case the extra context of vocab, syntax, and example sentences). Retrieval Quality metric though seems sort of brute forcey as not all vocab words need to be in the translation for it to have a high quality translation, thats because synonyms may be used for example.