The Syntax of Semantic Relations in Italian Fabio Celli CLIC, University of Trento October 24, 2009 Abstract The experiment reported in this paper is a feature engeneering op- eration that explores the syntactic structures of three different semantic relation types with tree kernels. This is done in order to find whether syntax can be a good feature to improve semantic relation separability in classification tasks. The average accuracy obtained is 63.09% . 1 Introduction And related work In a previous experiment [1] it was found that three semantic relation classes (”role”, ”location”, ”social”) yielded better results in a classification task with respect to the seven semantic relation classes in ACE 2004 (see [3]), since those three classes are easily separable. The three semantic relations are: Role (entity 1 is involved/included in entity 2), Location (entity1 has a position with respect to entity 2), Social (entity 1 has some interaction with entity 2). In another experiment ([2]) those relation classes were annotated on I-CAB, an Italian corpus, in order to test their performance on Italian. In this paper I want to test whether or not syntax could be a good feature for a further improvement in the separability of those three semantic relation classes in Italian. In this paper I will use tree kernels: a computational technique able to extract all the subtrees (tree fragments) from target tree structures and compare tree structures using subtrees, this technique already showed good performances in semantic role labeling tasks ([5]). The paper is structured as follows: in the next section there is the experiment, then a discussion follows in the third section. 2 Experiment The data used for the experiment are taken from I-CAB, an Italian corpus previously annotated with named entities (person, organizations, locations and geo-political entities) and the three semantic relations, defined as information undelying between two named entities, described above (role, location, social). The Shortest Path-enclosed Trees (SPTs) between the named entities involved in a semantic relation were given to the system as inputs, since Zhang et Al ([7]) found that those context-free trees work better than the context-dependent ones. Thirty SPTs (18 training, 12 testing) were randomly sampled from I-CAB and manually parsed, other features, inputted as vectors, were named entity 1 type (it takes the four values described above) and named entity 2 type (that takes 1