Case-Based Decision-Theoretic Decoding with Quality Memories

2025-09-16

AI Summary

Traditional Minimum Bayes Risk (MBR) decoding relies on internal model sampling, limiting its ability to incorporate out-of-domain knowledge and thus suffering from poor cross-domain robustness. To address this, we propose Case-Based Decision-Theoretic (CBDT) decoding—a novel framework that integrates high-quality external domain examples into the MBR paradigm for semantic-aware decoding decisions via case retrieval and expected utility estimation. Unlike conventional MBR, CBDT bypasses internal sampling and instead dynamically evaluates candidate outputs’ utility using semantically similar cases retrieved from an external memory bank, substantially improving cross-domain generalization. Empirical results across seven De↔En/Ja↔En machine translation tasks and MSCOCO/nocaps image captioning demonstrate consistent superiority of CBDT over both MAP and standard MBR decoding—particularly under low-resource and out-of-distribution conditions. Our core contribution lies in systematically embedding case-based reasoning into Bayesian decoding, establishing a new paradigm for controllable text generation grounded in external memory.

Technology Category

• Machine Learning: Bayesian Learning
• Reasoning under Uncertainty: Relational Probabilistic Models
• Cognitive Modeling & Cognitive Systems: Conceptual Inference and Reasoning

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking

Abstract

Minimum Bayes risk (MBR) decoding is a decision rule of text generation, which selects the hypothesis that maximizes the expected utility and robustly generates higher-quality texts than maximum a posteriori (MAP) decoding. However, it depends on sample texts drawn from the text generation model; thus, it is difficult to find a hypothesis that correctly captures the knowledge or information of out-of-domain. To tackle this issue, we propose case-based decision-theoretic (CBDT) decoding, another method to estimate the expected utility using examples of domain data. CBDT decoding not only generates higher-quality texts than MAP decoding, but also the combination of MBR and CBDT decoding outperformed MBR decoding in seven domain De--En and Ja$leftrightarrow$En translation tasks and image captioning tasks on MSCOCO and nocaps datasets.

Problem

• Improve text generation quality using decision rules
• Address out-of-domain knowledge capture limitations
• Combine case-based and risk-based decoding methods

Innovation

• Case-based decision-theoretic decoding method
• Uses domain examples for utility estimation
• Combines with MBR for improved performance