The SemEval 2007 task on Semantic Relations be­tween Nominals is to identify the underlying se­mantic relation between two nouns in the context of a sentence. The dataset provided consists of a definition file and 140 training and about 70 test sentences for each of the seven relations consid­ered: Cause-Effect, Instrument-Agency, Product-Producer, Origin-Entity, Theme-Tool, Part-Whole, and Content-Container. The task is defined as a binary classification problem. Thus, given a pair of nouns and their sentential context, the classifier decides whether the nouns are linked by the target semantic relation. In each training and test exam­ple sentence, the nouns are identified and manu­ally labeled with their corresponding WordNet 3.0 senses. Moreover, each example is accompanied by the heuristic pattern (query) the annotators used to extract the sentence from the web and the position of the arguments in the relation.

Based on the information employed, systems can be classified in four types of classes: (A) systems that use neither the given WordNet synsets nor the queries, (B) systems that use only WordNet senses, (C) systems that use only the queries, and (D) sys­tems that use both.

In this paper we present a type-B system that re­lies on various sets of new and previously used lin­guistic features employed in a supervised learning model.

Semantic relations between nominals can be en­coded by different syntactic constructions. We extend here over previous work that has focused mainly on noun compounds and other noun phrases, and noun-verb-noun constructions.

We selected a list of 18 lexico-syntactic and se­mantic features split here into three sets: feature set #1 (core features), feature set #2 (context features), and the feature set #3 (special features). Table 1 shows all three sets of features along with their defi­nitions; a detailed description is presented next. For some features, we list previous works where they proved useful. While features F1 - F4 were selected from our previous experiments, all the other features are entirely the contribution of this research.

Feature set #1 : Core features

This set contains six features that were employed in all seven relation classifiers. The features take into consideration only lexico-semantic information

Table 1 : The three sets of features used for the automatic semantic relation classification.

about the two target nouns.

Argument position (F1) indicates the position of the semantic arguments in the relation. This infor­mation is very valuable, since some relations have a particular argument arrangement depending on the lexico-syntactic construction in which they occur. For example, most of the noun compounds encod­ing Stuff-Object / Part-Whole relations have e\ as the part and e2 as the whole (e.g., silk dress).

Semantic specialization (F2) is a binary feature representing the prediction of a semantic specializa­tion learning model. The method consists of a set of iterative procedures of specialization of the train­ing examples on the WordNet IS-A hierarchy. Thus, after all the initial noun-noun pairs are mapped through generalization to entity - entity pairs in WordNet, a set of necessary specialization iterations is applied until it finds a boundary that separates pos­itive and negative examples. This boundary is tested on new examples for relation prediction.

The nominalization features (F3, F4) indicate if the target noun is a nominalization and, if yes, of what type. We distinguish here between agential nouns, other nominalizations, and neither. The features were identified based on WordNet and NomLex-Plus and were introduced to filter some of negative examples, such as car owner/theme.

Spatio-Temporal features (F5, F6) were also in­troduced to recognize some near miss examples, such as Temporal and Location relations. For in­stance, activation by summer (near-miss for Cause-Effect) and mouse in the field (near-miss for Content­Container). Similarly, for Theme-Tool, a word act­ing as a Theme should not indicate a period of time, as in <e\>the appointment</e\> was for more than one <e2>year</e2>. For this we used the in­formation provided by WordNet and special classes generated from the works of (Herskovits, 1987), (Linstromberg, 1997), and (Tyler and Evans, 2003).

For our experiments we chose libSVM, an open source SVM package. Since some of our features are nominal, we followed the standard practice of representing a nominal feature with n discrete val­ues as n binary features. We used the RBF kernel.

We built a binary classifier for each of the seven relations. Since the size of the task training data per relation is small, we expanded it with new examples from various sources. We added a new corpus of 3,000 sentences of news articles from the TREC-9 text collection (Girju, 2003) encoding Cause-Effect (1,320) and Product-Producer (721). Another col-

Table 2: Performance obtained per relation. Precision, Recall, F-measure, Accuracy, and Total (number ofexamples) are macro-averaged for system's performance on all 7 relations. Base-F shows the baseline F measure (all true), while Base-Acc shows the baseline accuracy score (majority).

lection of 3,129 sentences from Wall Street Journal (Moldovan et al., 2004; Girju et al., 2004) was con­sidered for Part-Whole (1,003), Origin-Entity (167), Product-Producer (112), and Theme-Tool (91). We also extracted 552 Product-Producer instances from eXtended WordNet (noun entries and their gloss definition). Moreover, for Theme-Tool and Content­Container we used special lists of constraints. Be­sides the selectional restrictions imposed on the nouns by special features such as F15 and F16 (psy­chological feature), we created lists of containers from various thesauri and identified selectional re­strictions that differentiate between containers and locations relying on taxonomies of spatial entities discussed in detail in (Herskovits, 1987) and (Tyler and Evans, 2003).

Each instance in this text collection had the tar­get nouns identified and annotated with WordNet senses. Since the annotations used different Word­Net versions, senses were mapped to sense keys.

Table 2 shows the performance of our system for each semantic relation. Base-F indicates the base­line F-measure (all true), while Base-Acc shows the baseline accuracy score (majority). The Average score of precision, recall, F-measure, and accuracy is macroaveraged over all seven relations. Overall, all features contributed to the performance, with a different contribution per relation (cf. Table 2).

This paper describes a method for the automatic identification of a set of seven semantic relations based on support vector machines (SVMs). The ap­proach benefits from an extended dataset on which binary classifiers were trained for each relation. The feature sets fed into the SVMs produced very good results.

We would like to thank Brian Drexler for his valu­able suggestions on the set of semantic relations.