Abstract: In sentiment analysis (SA), a vague assignment of a text to a set of n-ary discrete classes is insufficient. A great deal of research is concentrated on the automated assignment of strength to both terms and the finer-grained term senses, but these strength values rely purely on statistical means, and there is no semantic mechanism involved, leading to potentially biased results. As a solution, this works proposes a model that utilizes only the semantic information manually encoded within the human-defined glosses of term senses, a semantic network, and a set of predefined degree adverbs, in order to quantify their ‘natural’ sentiment strength (NSS) values. The ‘natural’ sentiment strength of a term sense here refers to the strength value derived in a ‘semantically natural’ manner, i.e. the NSS is assigned based on the agreed-upon meanings that humans have naturally assigned to words; and not ‘artificially statistical’, i.e. based a simple metric of probabilistic computation. Intrinsic evaluation against a manually-annotated gold standard benchmark demonstrates that the model outperforms related sense-level lexicon generation models against this same benchmark, and that it is in agreement with human intuition.