Researchers identify shared meanings in Bangla idioms across languages

Researchers studied how meanings from Bangla, a language spoken in South Asia, transfer between different language models.

The team analyzed 10,359 senses of Bangla idioms using six frozen encoders to see how much information shared between languages survives when models are trained on English. The study found that class-mean directions provided the most consistent evidence of shared coarse meanings. These shared meanings outperformed random-subspace and permuted-label nulls in all 12 model-label tests.

The results suggest that while some meaning is shared, the ability to transfer specific classifiers depends heavily on input maps and regularization. The researchers also observed an asymmetry in how information is erased. When they used English-fitted LEACE, it removed less than one percentage point of Bangla figurative-class accuracy. In contrast, using Bangla-fitted LEACE removed between 4 to 11 points of accuracy. These findings suggest that researchers should evaluate reading, classifier transfer, erasure, and overlap as separate operations.