Evaluate alternative generation and denoising methods for MBR overfitting mitigation

Investigate whether alternative candidate-generation methods or denoising algorithms, including Non-Negative Matrix Factorization (NMF), provide complementary benefits to Singular Value Decomposition Minimum Bayes Risk (SVD-MBR) for mitigating metric overfitting.

Background

The paper evaluates SVD-MBR using ε-sampling for candidate generation and Singular Value Decomposition for denoising the pairwise utility matrix. Although SVD-MBR is reported to mitigate metric overfitting, its experimental scope does not compare alternative generation procedures or matrix-denoising techniques.

The unresolved question is whether other candidate-generation methods or denoising algorithms—specifically including Non-Negative Matrix Factorization (NMF)—could provide complementary benefits within the proposed framework. This question is relevant to determining whether the observed improvements depend on the selected sampling and factorization methods or extend to broader combinations of MBR components.