The system starts by finding the target verb (which is marked in the corpus as such). Then, it finds the complete form of the verb (that in the corpus is tagged as verb group, infinitive, gerund, etc.) and the clause boundaries in order to look for the siblings of the verb that are under the same clause. Our as­sumption is that all siblings of the verb are potential candidates for semantic roles. The focus selection process produces two groups of focus tokens: on the one hand, the verbs and, on the other, the siblings of the verbs. These tokens will be the instances in each training set. Table 1 shows the number of training and test instances for each subtask.

Table 1: Number of instances per corpus for each task ('Ca' stands for Catalan, 'Sp' stands for Spanish).

In both systems we approach the classification task in one step, predicting directly the SR and the SC class. This means that in the SR task we do not perform a previous classification to select the tokens that might be assigned a role. We assume that all verbs belong to a class. As for the SR, we assume that most siblings of the verb will have a class, ex­cept for those that have syntactic functions AO, ET, MOD, NEG, IMPERS, PASS, and VOC. The sib­lings that do not have a semantic role are assigned the NONE tag. Because the corpus is small and be­cause the amount of instances with a NONE class is proportionally low, we do not consider it necessary to filter these cases.

Regarding the learning algorithm, we use the IB1 classifier as implemented in TiMBL (version 5.1) (Daelemans et al., 2004), a supervised induc­tive algorithm for learning classification tasks based on the k nearest neighbor (k-nn) algorithm. In IB1, similarity is defined by a feature-level distance met­ric between a test instance and a memorized training instance. The metric combines a per-feature value-based distance metric with global feature weights that account for relative differences in importance of the features.

The TiMBL parameters used in the systems are the IB 1 algorithm, the Jeffrey Divergence as feature metric, MVDM threshold at level 1, weighting us­ing GainRatio, k=11, and weighting neighbors as function of their Inverse Linear Distance (for details we refer the reader to the TiMBL reference guide (Daelemans et al., 2004)).

As for the features, we started by using the same feature set for both classifiers and then, after some experimentation, we decided to use slightly different feature sets for the two sub-tasks. Most of the features we designed are features that have become standard for the SRL task (Gildea and Jurafsky, 2002; Xue and Palmer, 2004; Carreras and Marquez, 2004; Carreras and Marquez, 2005). In our system, the features relate to the verb, the verb siblings, what we take to be the content word of the siblings, the clause, and the relation verb-arguments. Addition­ally, we added lexical features extracted from the verb lexicon provided for the task, and from Word­Net.

After experimenting with 323 features, we se­lected 98 for the SR task and 77 for the SC subclass. In order to select the features, we started with a basic system, the results of which were used as a baseline. Every new feature that was added to the basic system was evaluated in terms of average accuracy in 10­fold cross-validation experiments; if it improved the performance on held-out data, it was added to the se­lection. One problem with this hill-climbing method is that the selection of features is determined by the order in which the features have been introduced. We also performed experiments applying the feature selection process reported in (Tjong Kim Sang et al., 2005), a bi-directional hill climbing process. How­ever, experiments with this advanced method did not produce a better selection of features.

The features for the SR prediction subtask are the following:

• Features on the verb (6). They are shared by all the instances that represent phrases belonging to the same clause:

VForm; VLemma; VCau: binary feature that indicate if the verb is in a causative construction with hacer, fer or if the main verb is causar; VPron, VImp, VPass: binary features that indi­cate if the verb is pronominal, impersonal, and in passive form respectively.

• Features on the sibling in focus (12):

SibSynCat: syntactic category; SibSynFunc: syntactic function; SibPrep: preposition; SibLemW1, SibPOSW1, SibLemW2, SibPOSW2, SibLemW3, SibPOSW3: lemma and POS tag of the first, second and third words of the sibling; SibRelPos: position of the sibling in relation to the verb (PRE or POST); Sib+1RelPos: position of the sibling next to the cur­rent phrase in relation to the verb (PRE or POST); SibAbsPos: absolute position of the sibling in the clause.

• Features that describe the properties of the content word (CW) of the focus sibling (13): in the case of prepositional phrases the CW is the head of the first noun phrase; in cases of coordination, we only take the first element of the coordination.

As for the postprocessing phase, it consists of six simple rules to correct some basic errors in predict­ing some types of ArgM arguments. It only applies to the SR task. The rules are the following ones:

1. If prediction = ArgM-LOC, ArgM-MNR or ArgM-ADV, and either {SibPrep = 'durante' or 'durant'}, or {SibSynCat = sn and one of the WN semantic classes = 28}, then prediction =

ArgM-TMP.