Semantic Window Slicing

Semantic slicing is framed as a deployment optimization that preserves accuracy while reducing sequence length and self-attention cost. Full text and semantic slicing performed similarly, while random windows substantially degraded performance. Hints guided model attention but did not constrain outputs with rules. Inputs used plus-or-minus five-line windows around prioritized clinical headers such as FINAL DIAGNOSIS. Each model input followed a four-block schema containing Field, Classes, Report, and Hints.