Study tests cross-lingual concept transfer in Bangla idiom subspaces
Original title: What Travels Across Languages? Reading, Transfer and Erasure of Cross-Lingual Concept Subspaces in Bangla Idioms
English class-mean readouts beat zero-shot nulls in 8 of 12 tests, but English-fitted erasure barely changes Bangla accuracy.
A paper studies reading, classifier transfer and erasure of concept subspaces across six frozen encoders using 10,359 Bangla idiom senses, eight input conditions, and figurative-class and sentiment labels. Class-mean directions give the most consistent evidence of shared coarse meaning, outperforming nulls in all 12 model-label tests. English-fitted LEACE removes less than one percentage point of Bangla figurative-class accuracy, versus 4–11 points for Bangla-fitted LEACE. The authors argue reading, transfer, erasure and overlap should be evaluated separately.
- The study covers 10,359 Bangla idiom senses and eight input conditions.
- Confirmatory logistic regression rejects both planned nulls in 4 of 12 tests.
- English directions gain only 3.0–7.3 accuracy points above the majority-class floor.
- Explicit glosses show greater coarse-class sharing than idiom counterparts.
- The multiclass ordering appears in 11 of 12 cells.
From the original article
1 min read at arxiv.orgCross-lingual concept analysis requires distinguishing the information that a subspace retains from the classifiers and erasers that transfer through it. We study these operations in six frozen encoders using 10,359 Bangla (Bengali) idiom senses, eight input conditions, and figurative-class and sentiment labels.
By Syed Rifat Raiyan · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.