Constraint Model Setup Sample Clauses

Constraint Model Setup. A feature structure lexicon was generated by run- ning the Morphisto3 morphological analyser over the training vocabulary and then extracting feature values from the output. The constraint rules were extracted using the agreement relation identification and filtering meth- ods described in section 3.3. We tested two constraint model systems, one us- ing the rules as hard constraints and the other as soft constraints. The former discarded all hypotheses that failed constraints and used the modified cube pruning search algorithm. The latter allowed con- straint failure but used the failure count feature as a penalty. Both systems used the NP case probabil- ity feature. The weights for these two features were optimised using MERT (with all baseline weights fixed). The systems were otherwise identical to the baseline.