Using the Structure of a Conceptual Network in Computing Semantic Relatedness

Using the Structure of a Conceptual Network in Computing Semantic Relatedness Iryna Gurevych EML Research gGmbH, Schloss-Wolfsbrunnenweg 33, 69118, Heidelberg, Germany http://www.eml-research.de/∼gurevych Abstract. We present a new method for computing semantic relatedness of con- cepts. The method relies solely on the structure of a conceptual network and eliminates the need for performing additional corpus analysis. The network struc- ture is employed to generate artificial conceptual glosses. They replace textual definitions proper written by humans and are processed by a dictionary based metric of semantic relatedness [1]. We implemented the metric on the basis of GermaNet, the German counterpart of WordNet, and evaluated the results on a German dataset of 57 word pairs rated by human subjects for their semantic re- latedness. Our approach can be easily applied to compute semantic relatedness based on alternative conceptual networks, e.g. in the domain of life sciences. 1 Introduction Semantic relatedness of words represents important information for many applications dealing with processing of natural language. A more narrowly defined phenomenon of semantic similarity has been extensively studied in psychology, cognitive science, artificial intelligence, and computational linguistics. In the context of linguistics, it is typically defined via the lexical relation of synonymy. While synonymy is indeed an example of extreme similarity (suggesting that two words are interchangeable in a cer- tain context), many natural language processing applications require knowledge about semantic relatedness rather than just similarity [2]. Departing from that, we define se- mantic relatedness as any kind of lexical or functional association that may exist be- tween two words. For example, the words “car” and “journey” apparently display a close semantic relationship, while they are not synonymous. Many natural language processing applications, e.g. word sense disambiguation or information retrieval, do not need to determine the exact type of a semantic relation, but rather to judge if two words are closely semantically related or not. For example, for an application in the domain of career consultancy it might be important to conclude that the words “baker” and “bagel” are closely related, while the exact type of a semantic relation does not need to be assigned. Metrics of semantic relatedness are increasingly embedded into natural language processing applications for English due to the availability of free software, e.g. [3] and pre-computed information content values from English corpora. The evaluation of all approaches to compute semantic relatedness has so far been done for the task of semantic similarity. The underlying data was based on the English language [4,5]. We propose the following classification of the metrics of semantic relatedness: R. Dale et al. (Eds.): IJCNLP 2005, LNAI 3651, pp. 767–778, 2005. c⃝ Springer-Verlag Berlin Heidelberg 2005 768 I. Gurevych – intrinsic or extrinsic. Intrinsic metrics employ no external evidence, i.e. no know- ledge sources except for the conceptual network itself [1,6,7]. Extrinsic metrics re- quire additional knowledge, e.g. information content values of concepts computed from corpora [8,9,10]. – the type of knowledge source employed, e.g. a dictionary or a conceptual net- work. Metrics can either employ a …

Content truncated to 500 words