In Semantic Role Labeling (SRL) the goal is to iden­tify word sequences or arguments accompanying the predicate and assign them labels depending on their semantic relation. In this task we disambiguate ar­gument structures in two ways: predicting VerbNet (Kipper et al., 2000) thematic roles and PropBank (Palmer et al., 2005) numbered arguments, as well as adjunct arguments.

In this paper we describe our system for the SRL subtask of the Semeval2007 task 17. It is based on the architecture and features of the system named 'model 2' of (Surdeanu et al., forthcoming), but it introduces two changes: we use Maximum Entropy for learning instead of AdaBoost and we enlarge the feature set with combined features and other seman­tic features.

Traditionally, most of the features used in SRL are extracted from automatically generated syntac­tic and lexical annotations. In this task, we also ex­periment with provided hand labeled semantic information for each verb occurrence such as the Prop-Bank predicate sense and the Levin class. In addi­tion, we use automatically learnt Selectional Prefer­ences based on WordNet to generate a new kind of semantic based features.

We participated in both the "close" and the "open" tracks of Semeval2007 with the same system, mak­ing use, in the second case, of the larger CoNLL-2005 training set.

Selectional Preferences (SP) try to capture the fact that linguistic elements prefer arguments of a cer­tain semantic class, e.g. a verb like 'eat' prefers as subject edible things, and as subject animate entities, as in "She was eating an apple" They can be learned from corpora, generalizing from the observed argu­ment heads (e.g. 'apple', 'biscuit', etc.) into ab­stract classes (e.g. edible things). In our case we follow (Agirre and Martinez, 2001) and use Word­Net (Fellbaum, 1998) as the generalization classes (the concept <food,nutrient>).

The aim of using Selectional Preferences (SP) in SRL is to generalize from the argument heads in the training instances into general word classes. In theory, using word classes might overcome the data sparseness problem for the head-based features, but at the cost of introducing some noise.

More specifically, given a verb, we study the oc­currences ofthe target verb in a training corpus (e.g. the PropBank corpus), and learn a set of SPs for each argument and adjunct of that verb. For in­stance, given the verb 'kill' we would have 2 SPs for each argument type, and 4 SPs for some of the observed adjuncts: kiii_A0, kiii_Ai, kiii_AM-

LOC, kill_AM-MNR, ki 11_AM-PNC and kill_AM-TMP.

Rather than coding the SPs directly as features, we code the predictions instead, i.e. for each propo­sition in the training and testing set, we check the SPs for all the argument (and adjunct) headwords, and the SP which best fits the headword (see below) is the one that is selected. We codify the predicted argument (or adjunct) label as features, and we insert them among the corresponding argument features.

For instance, let's assume that the word 'railway' appears as the headword of a candidate argument of 'kill'. WordNet 1.6 yields the following hypernyms for 'railway' (from most general to most specific, we include the WordNet 1.6 concept numbers preceded by their specifity level);

Note that we do not care about the sense ambigu­ity and the explosion of concepts that it carries. Our algorithm will check each of the hypernyms of rail­way and match them with the concepts in the SPs of 'kill', giving preference to the most specific concept. In case that equally specific concepts match different SPs, we will choose the SP that has the concept that ranks highest in the SP, and code the SP feature with the label of the SP where the match succeeds. In the example, these are the most specific matches:

There is a tie in the level, so we choose the one with the highest rank. All in all, this means that ac­cording to the learntSPs we wouldpredictthat'rail-way' is a location feature for 'kill', and we would therefore insert the 'SP:AM-LOC' feature among the argument features.

If 'railway' appears as the headword of other verbs, the predicted argument might be different.

See for instance, the following verbs:

Note that our training examples did not contain 'railway' as an argument of any of these verbs, but due to the SPs we are able to code into a feature that 'railway' belongs to a concrete semantic class which contains conceptually similar headwords.

We decided to code the prediction of the SPs, rather than the SPs themselves, in order to be more robust to noise.

There is a further subtlety with our SP system. In order to label training and testing sets in similar con­ditions and avoid overfitting problems as much as possible, we split the training set into five folds and tagged each one with SPs learnt from the other four. For extracting SP features from test set examples, we use SPs learnt in the whole training set.

We participated in the "close" and the "open" tracks with the same classification model, but using dif­ferent training sets in each one. In the close track we only use the provided training set, and in the open, the CoNLL-2005 training set (without Verb­Net classes or thematic roles).

Before our participation, we tested the system in the CoNLL-2005 close track setting and it achieved competitive performance in comparison to the state-of-the-art results published in that challenge.

We have presented a sequential semantic role la­beling system for the Semeval-2007 task 17 (SRL). Based on Maximum Entropy Markov Models, it ob­tains competitive and promising results. We also have introduced semantic features extracted from Selectional Restrictions but we only have prelimi­nary evidence of their usefulness.

We thank David Martinez for kindly providing the software that learnt the selectional preferences. This work has been partially funded by the Spanish ed­ucation ministry (KNOW). Befiat is supported by a PhD grant from the University of the Basque Coun­try.