The survey of the state-of-art reveals an increas­ing interest in automatically discovering the un­derlying semantics in natural language. In this interdisciplinary field, the growing interest is justified by the number of applications which can directly benefit from introducing semantic information. Question Answering, Information Retrieval and Text Summarization are examples of these applications (Turney and Littman, 2005; Girju et al., 2005).

In the present work and for the purpose of the SEMEVAL task 4, our scope is limited to the semantic relationships between nominals. By this definition, we understand it is the process of discovering the underlying relations between two concepts expressed by two nominals.

Within the framework of SEMEVAL, nomi-nals can occur either on the phrase, clause or the sentence level. This fact constitutes the major challenge in this task since most of the previous research limited their approaches to certain types of nominals mainly the "compound nomi-nals"(Girju et al. 2005).

The paper is divided as follows; section 2 is a brief introduction to SMO used as the classifier for the task. Section 3 is dedicated to the de­scription of the set of features applied in our ex­periments. In section 4, we discuss the experi­ment's results compared to the baselines of the SEMEVAL task and the top scores. Finally, we summarize our approach, pointing out conclu­sions and future directions of our work.

We decided to use Support Vector Machine (SVM), as one of the most successful Machine Learning techniques, achieving the best per­formances for many classification tasks. Algo­rithm performance and time efficiency are key issues in our task, considering that our final goal is to apply this classification in a Question An­swering System.

Sequential Minimal Optimization (SMO) is a fast method to train SVM. SMO breaks the large quadratic programming (QP) optimization prob­lem needed to be resolved in SVM into a series of smallest possible QP problems. These small QP problems are analytically solved, avoiding, in this way, a time-consuming numerical QP optimization as an inner loop. We used Weka (Witten and Frank, 2005) an implementation of the SMO (Platt, 1998).

Prior to the classification of semantic rela­tions, characteristics of each sentence are auto­matically extracted using GATE (Cunningham et al., 2002). GATE is an infrastructure for de­veloping and deploying software components for Language Engineering. We used the following GATE components: English Tokenizer, Part-Of-Speech (POS) tagger and Morphological ana­lyser.

The set of features used for the classification of semantic relations includes information from different levels: word tokens, POS tags, verb lemmas, semantic information from WordNet, etc. Semantic features are only applied in cate­gories B and D.

On the lexical level, the set of word features include the two nominals, their heads in case one of the nominals in question or both are com­pound nominals ( e.g. the relation between <e1>tumor shrinkage</e1> and <e2>radiation therapy </e2> is actually between the head of the first "shrinkage' and "radiation therapy"). More features include: the two words before the first nominal, the two words after the second nominal, and the word list in-between (Wang et al., 2006).

On the POS level, we opted for using a set of POS features since word features are often too sparse. This set includes POS tags of the two words occurring before the first nominal and the two words occurring after the second nominal together with the tag list of the words in-between (Wang et al., 2006). POS tags of nomi-nals are considered redundant information.

Information regarding verbs and prepositions, occurring in-between the two nominals, is highly considered. In case of the verb, the system takes into account the verb token and the information concerning the voice and lemma. In the same way, the system keeps track of the prepositions occurring between both nominals. In addition, a feature, called numinter, indicating the number of words between nominals is considered.

Other important feature is the path from the first nominal to the second nominal. This feature is built by the concatenation of the POS tags be­tween both nominals.

The feature related to the query provided for each sentence is only considered in the categories C and D according to the SEMEVAL re­strictions.

On the semantic level, we used features ob­tained from WordNet. In addition to the Word­Net sense keys, provided for each nominal, we extracted its synset number and its lexical file number.

Based on the work of Rosario, Hearst and Fillmore (2002), we suppose that these lexical file numbers can help to determine if the nomi-nals satisfy the restrictions for each relation. For example, in the relation Theme-Tool, the theme should be an object, an event, a state of being, an agent, or a substance. Else, it is possible to af­firm that the relation is false.

For the Part-Whole relation and due to its relevance in this classification task, a feature indicating metonymy relation in WordNet was taken into account.

Furthermore, we designed a new feature, called WordNet vector. For constructing this vector, we selected the synsets of the third level of depth in WordNet and we detected if each is ancestor or not of the nominal. It is a binary vec­tor, i.e. if the synset is ancestor of the nominal it is assigned the value 1, else it is assigned the value 0. In this way, we worked with two vec­tors, one for each nominal. Each vector has a dimension of 13 coordinates. Each coordinate represents one of the 13 nodes in the third level of depth in WordNet. Our initial hypothesis con­siders that this representation for the nominals could perform well on unseen data.

Cross validation is a way to test the ability of the model to classify unseen examples. We trained the system using 10-fold cross-validation; the fold number recommended for small training datasets. For each relation and for each category (A, B, C, D) we selected the set of features that obtained the best results using the indicated cross validation.

140).

In our first approach to automatic classifica­tion of semantic relations between nominals and as expected from the training phase, our system achieved its best performance using WordNet information. In general, we obtained better scores in category 4 (size of training: 1 to 140), i.e., when all the training examples are used.

On the other hand, overfitting the training data (most probably due to the small size of training dataset) is the main reason behind the low scores obtained by our system.

These facts lead us to the conclusion that se­mantic features from WordNet, in general, play a key role in the classification task. However, the relevance of WordNet-related features var­ies. For example, lexical file numbers proved to be highly effective, while the use of the Word-Net Vector did not improve significantly the re­sults. Thus, we consider that a level 3 WordNet Vector is rather abstract to represent each nomi­nal. Developing a WordNet Vector with a deeper level (> 3) could be more effective as the repre­sentation of nouns is more descriptive. Query features, on the other hand, did not im­prove the performance of the system. This is due to the fact that the same query could represent both positive and negative examples of the rela­tion. However, to improve results in categories A and C, more features need to introduced, es­pecially context and syntactic information such as chunks or dependency relations.

To improve results across the whole dataset, wider use of semantic information is necessary. For example, the immediate hypernym for each synset obtained from WordNet could help in improving the system performance (Nastase et al., 2006). Besides, information regarding the entity features could help in the classification of some relations like Origin-Entity or Product-Producer. Other semantic resources such as VerbNet, FrameNet, PropBank, etc. could also be used.

Furthermore, we consider introducing a Word Sense Disambiguation module to obtain the cor­responding synsets of the nominals. Also, in­formation concerning the synsets of the list of the context words could be of great value for the classification task (Wang et al., 2006).