Classification of semantic relations is important to NLP as it would benefit many NLP applications, such as machine translation and information re­trieval.

Researchers have already proposed various schemes. For example, Hearst (1992) manually de­signed lexico-syntactic patterns for extracting is-a relations. Berland and Charniak (1999) proposed a similar method for part-whole relations. Brin (1998) employed a bootstrapping algorithm for more spe­cific relations (author-book relations). Kim and Baldwin (2006) and Moldovan et al.(2004) focused on nominal relations in compound nouns. Turney (2005) measured relation similarity between two words. While these methods differ, they all utilize lexical patterns between two entities.

Within this context, our goal was to utilize infor­mation specific to an entity. Although entities con­tain many types of information, we focused on the physical size of an entity. Here, physical size refers to the typical width/height of an entity. For example, we consider book to have a physical size of 20 x 25 cm, and book to have a size of 10 x 10 m, etc.

We chose to use physical size for the following reasons:

1. Most entities (except abstract entities) have a physical size.

2. Several semantic relations are sensitive to phys­ical size. For example, a content-container rela­tion (el content-container e2) naturally means that el has a smaller size than e2.A book is also smaller than its container, library. A part-whole relation has a similar constraint.

Our next problem was how to determine physi­cal sizes. First, we used Google to conduct Web searches using queries such as "book (*cm x*cmj" and "library (*m x*mf \ Next, we extracted numeric expressions from the search results and used the av­erage value as the physical size.

Experimental results revealed that our proposed approach is feasible and prevents the problems in­herent in other methods.

We used a corpus provided by SemEval2007 Task #4 training set. This corpus consisted of 980 anno­tated sentences (140 sentencesx 7 relations). Table 1 presents an example.

Although the corpus contained a large quantity of information such as WordNet sense keys, comments, etc., we used only the most pertinent information: entity1 (el ), entity2 (e2), and its relation (true/false)

We applied support vector machine (SVM)-based learning (Vapnik, 1999) using three types of fea­tures: (1) basic pattern features (Section 3.1), (2) se­lected pattern features (Section 3.2), and (3) physical size features (Section 3.3). Figure 1 presents some examples of these features.

pattern features.

For SVM learning, we used TinySVM with a lin­ear kernel .

2. In contrast, entities such as "toy car"or"mini car" frequently appear with a size.

Figure 2 presents the size distribution of "car." Few instances appeared of real cars sized approxi­mately 500 x 400 cm, while very small cars smaller than 100 x 100 cm appeared frequently. Our current method of calculating average size is ineffective un­der this type of situation.

In the future, using physical size as a clue for de­termining a semantic relation will require resolving this problem.

We briefly presented a method for obtaining the size of an entity and proposed a method for classifying semantic relations using entity size. Experimental results revealed that the proposed approach yielded slightly higher performance than a baseline, demon­strating its feasibility. If we are able to estimate en-

"#" indicates the number of obtained size expressions.

"-" indicates a "NO-SIZE" entity.

tity sizes more precisely in the future, the system will become much more accurate.