Training Data Generation
Step 1 – antecedent filtering -, in which different kinds of restrictions for eliminating impossible antecedents (in particular, agreement in person/number/gender and syntactic disjoint reference) are applied, is immediately taken over from the original ROSANA algorithm. In step 2, however, no salience ranking of the remaining antecedent candidates is performed. Rather, each remaining anaphor-candidate pair (A,C) is mapped to a feature vector fv(A,C), the attributes f1,…,fk of which comprise individual and relational features derived from the descriptions of the occurrences A and C. The signature of the feature vectors, i.e. the inventory of features to be taken into account has to be chosen carefully in order to fulfill the conditions of robust processing: instead of requiring complete and unambiguous descriptions, they should be computable from potentially partial representations such as fragmentary syntactic representations.
Anaphor Resolution through Learned Classifiers
Again, step 1 is identical with the antecedent filtering phase of the manually designed ROSANA algorithm. Step 2, however, is modified. For a specific instance (A,C) of anaphor and antecedent candidate, after the computation of the feature vector fv(A,C), the decision tree lookup takes place; basically, its result consists in a prediction {COSPEC,NON_COSPEC}. In the subsequent step, these predictions are employed for computing a ranking over the candidate sets of each anaphor. In its base version, candidates which are classified to COSPECify with the anaphor rank higher than candidates that are predicted to NON_COSPECify; surface nearness (i.e. word distance) serves as the secondary criterion. Among the possible refinements are: further ranking the candidates according to the classification error probability yielded by the decision tree lookup, and eliminating candidates which are (fuzzily) classified as NON_COSPECifying. There is a final step 3 in which the actual antecedent selection takes place. The remaining candidates are considered in the order determined by the ranking procedure; additional means are taken to avoid combinations of antecedent decisions that are mutually incompatible.