Multilingual LLMs also seem to adopt different strategies to handle the complexity of interlingual homographs, particularly in cross-lingual sentence contexts involving English and non-English words.

6/

Furthermore, we found no correlation between the semantic judgment capacity of LLMs and their word pair disambiguation ability. Indicating the lack of utility of semantic cues in the disambiguation task and the lack of utility of orthographic signals in semantic retrieval.

5/

However, this does not extend to the semantic retrieval ability of models, where no single category is easier for the models to retrieve the meaning of, which is contrary to the expectation of cognate facilitation.

4/

LLMs demonstrate a facilitation effect when disambiguating cognates, mirroring human bilingual processing patterns, where shared orthographic and lexical representations simplify word recognition and understanding.

3/

How do multilingual LLMs process bilingual words?
Do they approach it like humans?

Our paper explores LLM's processing of cognates, non-cognates, and interlingual homographs, uncovering LLMs' strengths & limits in bilingual processing! w/
Tanmoy Chakraborty
and Gayatri Oke.

1/🧵