meaningweight

Meaningweight is a proposed quantitative metric intended to express the degree to which the semantic content of a unit of information contributes to a task or interpretation. It treats meaning as a separable component from form, syntax, or noise, and aims to capture how much of a signal’s value derives from its meaning rather than its surface features.

In natural language processing and information retrieval, meaningweight can be used to adjust models, prioritize semantically rich signals, and diagnose where models rely on incidental cues. It can be computed in several ways, including information-theoretic measures such as normalized mutual information between a feature and the target, model-based attributions (for example, SHAP or gradient-based scores) aggregated across a unit, or attention-based weights interpreted as semantic importance.

Applications include feature weighting in text classification, improving semantic search, abstractive summarization, and disambiguation. It can help in corpus analysis by highlighting semantically dense terms or phrases and by guiding data curation toward semantically informative content.

Limitations include subjectivity, dependence on task, and context sensitivity; meaningweight values may vary with data distribution, annotation schemes, and model architecture, making cross-task comparability challenging.

See also: semantic weighting, term weighting, TF-IDF, feature attribution, attention mechanisms.