Construction of an Objective Hierarchy of Abstract Concepts via Directional Similarity

Construction of an Objective Hierarchy of Abstract Concepts via Directional Similarity Kyoko Kanzaki Eiko Yamamoto Hitoshi Isahara Computational Linguistics Group, National Institute of Information and Communications Technology 3-5 Hikari-dai, Seika-cho, Souraku-gun, Kyoto, Japan, {kanzaki, eiko, isahara}@nict.go.jp Qing Ma Faculty of Science and Technology Ryukoku University Seta, Otsu,520-2194, Japan qma@math.ryukoku.ac.jp Abstract The method of organization of word mean- ings is a crucial issue with lexical databases. Our purpose in this research is to extract word hierarchies from corpora automatically. Our initial task to this end is to determine adjec- tive hyperonyms. In order to find adjective hyperonyms, we utilize abstract nouns. We constructed linguistic data by extracting se- mantic relations between abstract nouns and adjectives from corpus data and classifying abstract nouns based on adjective similarity using a self-organizing semantic map, which is a neural network model (Kohonen 1995). In this paper we describe how to hierarchi- cally organize abstract nouns (adjective hy- peronyms) in a semantic map mainly using CSM. We compare three hierarchical organi- zations of abstract nouns, according to CSM, frequency (Tf.CSM) and an alternative simi- larity measure based on coefficient overlap, to estimate hyperonym relations between words. 1. Introduction A lexical database is necessary for computers, and even humans, to fully understand a word's meaning because the lexicon is the origin of lan- guage understanding and generation. Progress is being made in lexical database research, notably with hierarchical semantic lexical databases such as WordNet, which is used for NLP research worldwide. When compiling lexical databases, it is impor- tant to consider what rules or phenomena should be described as lexical meanings and how these lexical meanings should be formalized and stored electronically. This is a common topic of discus- sion in computational linguistics, especially in the domain of computational lexical semantics. The method of organization of word meanings is also a crucial issue with lexical databases. In current lexical databases and/or thesauri, abstract nouns indicating concepts are identified manually and words are classified in a top-down manner based on human intuition. This is a good way to make a lexical database for users with a specific purpose. However, word hierarchies based on human intuition tend to vary greatly depending on the lexicographer, and there is often dis- agreement as to the make-up of the hierarchy. If we could find an objective method to organize word meanings based on real data, we would avoid this variability. Our purpose in this research is to extract word hierarchies from corpora automatically. Our ini- tial task to this end is to determine adjective hy- peronyms. In order to find adjective hyperonyms, we utilize abstract nouns. Past linguistic research has focused on classifying the semantic relation- ship between abstract nouns and adjectives (Nemoto 1969, Takahashi 1975). We constructed linguistic data by extracting semantic relations between abstract nouns and adjectives from corpus data and classifying ab- stract nouns based on adjective similarity using a self-organizing semantic map (SOM), which is a neural network model (Kohonen 1995). The rela- tive proximity of words in …

Content truncated to 500 words