The extraction ofpredicative structures is carried out at the sentence level. Given a predicate within a natural language sentence, its arguments have to be properly labeled. This problem is usually divided in two subtasks: (a) the detection of the boundaries, i.e. the word spans of the arguments, and (b) the classification of their type, e.g. Arg0 and ArgM in

^jjia/project NNP

/nations /United ïL^Vgrace-period JLitt^/final Jfor NN NP

Predicate ARG1

ê-tt/allowing NN ^Ll/before NNP

/the-chance /Cyprus _J

ARGM-PRP

PropBank or Agent and Goal in FrameNet.

The standard approach to learn both the detection and the classification of predicate arguments is sum­marized by the following steps:

1. Given a sentence from the training-set, gener­ate a full syntactic parse-tree;

2. let P and A be the set of predicates and the set of parse-tree nodes (i.e. the potential argu­ments), respectively;• extract the feature representation set, Fpa;

• if the subtree rooted in a covers exactly the words of one argument of p, put Fpain T + (positive examples), otherwise put it in T- (negative examples).

For instance, in Figure 1, for each combination of the predicate instated with the nodes NP, S, VP, VPB, NNP, NN, PP, JJ or IN the instances Finstated,a are generated. In case the node a ex­actly covers "project nations United", "grace-period final" or "for allowing the chance before Cyprus", Fpa will be a positive instance otherwise it will be a negative one, e.g. Finstated,iN.

The T + and T- sets are used to train the bound­ary classifier. To train the multi-class classifier, T+ can be reorganized as positive T+fl. and negative T~rg. examples for each argument i. In this way, an individual ONE-vs-ALL classifier for each argu­ment i can be trained. We adopted this solution, ac­cording to (Pradhan et al., 2003), since it is simple and effective. In the classification phase, given an unseen sentence, all its Fp a are generated and clas­sified by each individual classifier Ci. The argument associated with the maximum among the scores pro­vided by the individual classifiers is eventually se­lected.

The above approach assigns labels independently for the different arguments in the predicate argument structure. As a consequence the classifier output may generate overlapping arguments. Thus, to make the annotations globally consistent, we apply a dis-ambiguating heuristic that selects only one argument among multiple overlapping arguments. The heuris­tic is based on the following steps:

• if more than two nodes are involved, i.e. a node d and two or more of its descendants niare classified as arguments, then assume that d is not an argument. This choice is justified by previous studies (Moschitti et al., 2005) show­ing that for lower nodes, the role classification is generally more accurate than for upper ones;

• if only two nodes are involved, i.e. they dom­inate each other, then keep the one with the higher SVM classification score.

The discovery of relevant features is, as usual, a complex task. However, there is a common con­sensus on the set of basic features that should be adopted. Among them, we select the following sub­set: (a) Phrase Type, Predicate Word, Head Word, Position and Voice as defined in (Gildea and Ju-rafsky, 2002); (b) Partial Path, No Direction Path, Head Word POS, First and Last Word/POS in Con­stituent and SubCategorization as proposed in (Prad-han et al., 2003); and (c) Syntactic Frame as de­signed in (Xue and Palmer, 2004).

For example, Phrase Type indicates the syntactic type of the phrase labeled as a predicate argument, NP for Arg1 in Figure 1 whereas the Parse Tree Path contains the path in the parse tree between the pred­icate and the argument phrase, expressed as a se­quence of nonterminal labels linked by direction (up or down) symbols, VPB | VP | S j NP for Arg1 in Figure 1.

We use the dataset released in the SEMEVAL 2 0 07 Task 18 on Arabic Semantic Labeling, which is sampled from the Pilot Arabic PropBank. Such data covers the 95 most frequent verbs in the Arabic Treebank III ver. 2 (ATB) newswire data from Annhar newspaper from the months of July through November 2002.

An important characteristic of the dataset is the use of unvowelized Arabic in the Buckwalter transliteration scheme. We used the gold standard parses in the ATB as a source for syntactic parses for the data. The data comprises a development set of 886 sentences, a test set of 902 sentences, and a training set of 8,402 sentences. The development setcomprises 1,725 argument instances, the testdata comprises 1,661 argument instances, and training data comprises 21,194 argument instances. These

Table 1: Boundary detection F1 results on the development and test sets.

instances are distributed over 26 different role types.

The training instances for the boundary detection task relate to parse-tree nodes that do not correspond to correct boundaries. For efficiency reasons, we use only the first 350K training instances for the bound­ary classifier out of more than 700K available.

The experiments are carried out with the SVM-light-TK software available at which encodes tree kernels in the SVM-light soft­ware. This allows us to design a system which can exploit tree kernels in future research. To implement the boundary classifier and the individual argument classifiers, we use a polynomial kernel with the default regularization parameter (of SVM-light), and a cost-factor equal to 1.

Our system is evaluated using the official CoNLL evaluator (Carreras and Marquez, 2005), avail­able at http://www.lsi.upc.es/~srlconll/ soft.html.

Table 1 shows the F1 scores obtained on the de­velopment and test data. We note that the F1 on the development set, i.e. 93.68, is slightly lower than the result on the test set, i.e. 94.06. This suggests that the test data is easier than the development set.

Similarbehaviorcanbe observed forthe role clas­sification task in tables 2 and 3.

Again, the overall F1 on the development set (77.85) is lower than the result on the test set (81.43). This confirms that the test data is, indeed, easier than the development set.

Regarding the F1 of individual arguments, we note that, as for English SRL, ARG0 shows high values, 95.42 and 96.69 on the development and test sets, respectively. Interestingly, ARG1 seems

set.

Table 3 : Argument classification results on the test set.

more difficult classify in Arabic than it is in En­glish. In our current experiments, the F1 for ARG1 is only 89.83 (compared to 95.42 for ARG0). This may be attributed to two main factors. Ara­bic allows for different types of syntactic config­urations, subject-verb-object, object-verb-subject, verb-subject-object, hence the logical object of a predicate is highly confusable with the logical sub­ject. Moreover, around 30% of the ATB data is pro-dropped, where the subject is morphologically marked on the verb and its absence is marked in the gold trees with an empty trace. In the current version of the data, the traces are annotated with the ARG0 semantic role consistently allowing for the high rel­ative performance yielded.

The F1 of the other arguments seems to follow the

English SRL behavior as their lower value depends on the lower number of available training examples.