rescoring

Rescoring is the act of re-evaluating previously computed scores or rankings, often to improve accuracy or adapt to new criteria. The term is used across several domains, including information retrieval, natural language processing, and audio-visual production. In computational contexts, rescoring usually means applying a secondary scoring process to a subset of results rather than recomputing everything from scratch.

In information retrieval and related AI tasks, rescoring typically operates on a set of top candidates produced by an initial retriever. A second, more expensive model (such as a cross-encoder or a domain-specific transformer) re-scores these candidates to produce a refined ranking. This approach balances speed and accuracy, leveraging the efficiency of the initial stage with the precision of the re-ranking model. Rescoring may use additional features, context, or multi-task signals and can be trained on task-specific data to adapt to user intent or domain requirements.

In speech recognition and machine translation, rescoring is the process of re-evaluating the N-best hypothesis list with a higher-capacity language model or domain-tuned scorer. The goal is to select the most probable output by applying more nuanced linguistic information or external knowledge, thereby improving error rates.

In music and film production, rescoring refers to creating or commissioning a new musical score for a work, such as a silent film, game, or re-release. This can involve re-orchestrating, rewriting motifs, or updating instrumentation and style, often to suit contemporary audiences or different cultural contexts. Rescoring may also be used in licensing or restoration projects.

Common considerations include computational cost, latency, risk of overfitting to a training corpus, and the need to calibrate rescored outputs with evaluation metrics appropriate to the domain, such as precision and recall, NDCG, or perceptual quality in audio.

See also: re-ranking, cross-encoder, language model rescoring, music scoring.