Determine whether constituent-family asymmetry explains CAM’s weaker generalization for three-character compounds

Determine whether asymmetry between the left and right constituent-family sizes causes the Compound Analogy Model to generalize less effectively to held-out three-character Mandarin compounds than to held-out compounds of other lengths.

Background

CAM performs better than CAOSS on most held-out data but performs worse for three-character compounds. The authors note that three-character compounds have especially small and imbalanced constituent families, whereas four-character compounds show a different pattern of family-size asymmetry.

Because the observed differences could also reflect the particular compounds included in the dataset, the authors do not establish a causal explanation. They explicitly raise constituent-family asymmetry as a possible reason why analogical generalization is less successful for three-character compounds.

References

The better performance of CAOSS for 3-character words and the better performance of CAM for 4-character words may be just a coincidence that is contingent on the longer words that happened to be included in the CLD. It is also possible that an asymmetry in the constituent family sizes is at issue.

— Compound interpretation is based on analogy  (2610.01688 - Shen et al., 1 Oct 2026) in Section 5, subsection “Discussion,” paragraph beginning “In the light of these considerations...”