Constraint Model Features. In addition to the string-to-tree features, we add two features related to constraint evaluation: Q• exp(f ), where f is the derivation’s constraint set failure count. This serves as a penalty fea- ture in a soft constraint variant of the model: for each constraint set in which a unification failure occurs, this count is increased and an empty feature structure is produced, permitting decoding to continue. n pcase(cn), the product of the derivation’s case model probabilities. Where the case value is ambiguous we take the highest possible prob- ability.
Appears in 2 contracts
Sources: Research Paper, Research Paper