Brevity-penalty normalization for sequence scoring
Mathematical Methods AIMech:BrevityPenaltyNormalization DRAFT
Aliases: BP-Norm
A sequence rescoring rule that adds a predicted-length brevity penalty in log space to average token log probability: S(x,y)/|y| + min(1−Lpred(x)/|y|,0).
Identity Notes & Concept Scope
BP-Norm is Yang et al.’s predicted-length decoding score. It resembles the algebraic structure of BLEU’s brevity penalty but does not compute n-gram precision, use an actual reference translation at inference, or equal the BLEU evaluation metric. It differs from plain average log probability and additive word rewards. No normalized identity match was found across the 442 AIO and 146 prior native records; both checkout text searches included ignored and hidden files and returned no matches for the selected exact IDs, full labels, aliases and source titles.
Claims & Grounded Evidence
2 Claims
Mechanism
#method
Predict a target length from the input using the paper’s encoder-state length predictor. For a nonempty hypothesis of length n, combine average log probability S/n with log brevity penalty min(1−Lpred/n,0). For decoding with a genuine finite maximum output length R, a current raw log score S_t≤0 yields an upper bound S_t/R on descendant BP-Norm scores. Compare the maximum active-prefix bound with the best completed BP-Norm score before stopping.
Limitation
#predicted-versus-maximum
Equation (13) defines R as a maximum generation length, whereas the experiments identify it with predicted length. The bound is curated only when that value truly bounds the permitted completion lengths; prediction alone is not a mathematical upper bound.
⚠️ Curated Limitations & Negative Results
- Lpred is a prediction, not a reference length or a proven upper bound on output length. The S_t/R stopping bound requires that admissible completions actually have length at most R.
- The experiments set R to the predicted length. This supports the stopping guarantee only under an enforced length cap or another justification that all allowed completions obey that bound; a forecast alone is insufficient.
- The formula requires nonempty hypotheses and a valid length prediction. A BOS-only prefix needs separate handling; length and EOS conventions must be specified.
- The score is neither a normalized probability nor BLEU itself. Improved search score does not imply better human translation quality.
- The source’s hyperparameter-free comparison concerns removal of a separately tuned reward constant, not absence of trained models, a length predictor or decoding choices. Finite beam pruning remains approximate.
Artificial Intelligence Ontology (AIO) Grounding
Mapped AIO Classes
| AIO Target Class | Mapping Relation | Mapping Rationale |
|---|---|---|
AIO:NaturalLanguageProcessing
Natural Language Processing
|
related_match | The primary paper defines this sequence scoring, stopping or pruning method for natural-language generation, with neural machine translation as its central setting. The exact AIO CLASS describes the language-processing task. This is an method-to-application-domain relationship, not equivalence or an is-a assertion; the method is not a trained language-model instance and may also apply outside NLP. |
Evidence Sources & Primary Literature
2 Sources| Source ID | Publication / Document | Type / Access | Reference & Checksum |
|---|---|---|---|
method-source
|
Yilin Yang, Liang Huang, Mingbo Ma
|
primary paper
full_text
|
ID: doi:10.18653/v1/D18-1342
sha256:019ba14d74ad…
|
predicted-versus-maximum
|
Yilin Yang, Liang Huang, Mingbo Ma
|
primary paper
full_text
|
ID: doi:10.18653/v1/D18-1342
sha256:019ba14d74ad…
|
Curation Provenance
Curator: Codex
Method: ai_assisted
Model: GPT-6
Created: 2026-10-11T06:36:44Z
Updated: 2026-10-11T06:37:11Z
Source File:
data/curated/BrevityPenaltyNormalization.yaml