The starting point of our research is an approach for identifying relations between named entities ex­ploiting only shallow linguistic information, such as tokenization, sentence splitting, part-of-speech tag­ging and lemmatization (Giuliano et al., 2006). A combination of kernel functions is used to represent two distinct information sources: (i) the global con­text where entities appear and (ii) their local con­texts. The whole sentence where the entities appear (global context) is used to discover the presence of a relation between two entities. Windows of limited size around the entities (local contexts) provide use­ful clues to identify the roles played by the entities within a relation (e.g., agent and target of a gene in­teraction). In the task of detecting protein-protein interactions, we obtained state-of-the-art results on two biomedical data sets. In addition, promising re­sults have been recently obtained for relations such as work for and org based in in the news domain.

In this paper, we investigate the use of the above approach to discover semantic relations between nominals. In addition to the original feature rep­resentation, we have integrated deep syntactic pro­cessing of the global context and semantic informa­tion for each candidate nominals using WordNet as external knowledge source. Each source of informa­tion is represented by kernel functions. A tree kernel (Moschitti, 2004) is used to exploit the deep syn­tactic processing obtained using the Charniak parser (Charniak, 2000). On the other hand, bag of syn­onyms and hypernyms is used to enhance the repre­sentation of the candidate nominals. The final sys­tem is based on five basic kernel functions (bag-of-words kernel, global context kernel, tree kernel, su­persense kernel, bag of synonyms and hypernyms kernel) linearly combined and weighted under dif­ferent conditions. The experiments were carried out using support vector machines (Vapnik, 1998) as classifier.

We present results on the Classification of Seman­tic Relations between Nominals task at SemEval-2007, in which sentences containing ordered pairs of marked nominals, possibly semantically related, have to be classified. On this task, we achieve an overall Fi of 71.8% (B category evaluation), largely outperforming all the baselines.

In order to implement the approach based on syntac­tic and semantic information, we employed a linear weighted combination of kernels, using support vec­tor machines as classifier. We designed two families of basic kernels: syntactic kernels and semantic ker­nels. These basic kernels are combined by exploit­ing the closure properties of kernels. We define our composite kernel KC (x1 ,x2) as follows where each basic kernel Ki is normalized and wi G (0,1} is the kernel weight. The normalization factor plays an important role in allowing us to integrate in­formation from heterogeneous knowledge sources.

All basic kernels, but the tree kernel (see Section 2.1.3), are explicitly calculated as follows where </>(•) is the embedding vector. Even though the resulting feature space has high dimensionality, an efficient computation of Equation 2 can be carried out explicitly since the input representations defined below are extremely sparse.

Sentences have been tokenized, lemmatized, and POS tagged with TextPro. We considered each re­lation as a different binary classification task, and each sentence in the data set is a positive or negative example for the relation. The direction of the rela­tion is considered labelling the first argument of the relation as agent and the second as target.

All the experiments were performed using the SVM package SVMLight-TK, customized to em­bed our own kernels. We optimized the linear com­bination weights wi and regularization parameter c using 10-fold cross-validation on the training set. We set the cost-factor j to be the ratio between the number of negative and positive examples.

Table 1 shows the performance on the test set. We achieve an overall F1 of 71.8% (B category evalua­tion), largely outperforming all the baselines, rang­ing from 48.5% to 57.0%. The average training plus test running time for a relation is about 10 seconds on a Intel Pentium M755 2.0 GHz. Figure 2 shows the learning curves on the test set. For all relations but theme-tool, accurate classifiers can be learned using a small fraction of training.

Experimental results show that our kernel-based ap­proach is appropriate also to detect semantic rela­tions between nominals. However, differently from relation extraction between named entities, there is not a common kernel setup for all relations. E.g.,

Learning Curve

for content-container we obtain the best perfor­mance combining the tree kernel and the bag of syn­onyms and hypernyms kernel; on the other hand, for instrument-agency the best performance is obtained by combining the global kernel and the supersense kernel. Surprisingly, the supersense kernel alone works quite well and obtains results comparable to the bag of synonyms and hypernyms kernel. This result is particularly interesting as a supersense tag­ger can easily provide a satisfactory accuracy (Cia­ramita and Altun, 2006). On the other hand, ob­taining an acceptable accuracy in word sense disam­biguation (required for a realistic application of the bag of synonyms and hypernyms kernel) is imprac­tical as a sufficient amount of training for at least all nouns is currently not available. Hence, the super­sense could play a crucial role to improve the perfor­mance when approaching this task without the nomi-nals disambiguated. To model the global context us­ing the Fore-Between, Between and Between-After contexts did not produce a significant improvement with respect to the bag-of-words model. This is mainly due to the fact that examples have been collected from the Web using heuristic patterns/queries, most of which implying Between patterns/contexts (e.g., for the cause-effect relation "* comes from *", "* out of *" etc.).

Claudio Giuliano, Alberto Lavelli and Lorenza Ro­mano are supported by the X-Media project (http : //www.x-media-project .org), sponsored by the European Commission as part of the Infor­mation Society Technologies (IST) programme un­der EC grant number IST-FP6-026978.