Dictionaries as Networks: Identifying the graph structure of Ogden’s Basic English

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 3565–3576, Osaka, Japan, December 11-17 2016. Dictionaries as Networks: Identifying the graph structure of Ogden’s Basic English Camilo Garrido and Claudio Gutierrez Center for Semantic Web Research & Department of Computer Science, Universidad de Chile {cgarrido, cgutierr}@dcc.uchile.cl Abstract We study the network structure underlying dictionaries. We systematize the properties of such networks and show their relevance for linguistics. As case of study, we apply this technique to identify the graph structure of Ogden’s Basic English. We show that it constitutes a strong core of the English language network and that classic centrality measures fail to capture this set of words. 1 Introduction Dictionaries are rich sources to investigate the semantic structure of natural language. The purpose of dictionaries, writes Wilks et al. (1993), “is to provide definitions of senses of words and, in so doing, supply knowledge about not just language, but the world.” The definition of a word involves recursively new words, and thus, new senses and meanings. In this way, a dictionary can naturally be viewed as a network where each word w is related to the set of words w1, . . . , wn that define it: for each word w so defined, consider the relationship w → wi for each i = 1, . . . , n, and proceed recursively with all the entries of the dictionary. The idea is not new and was already proposed by K. C. Litkowski (1978). Clearly this basic idea must be refined. There are words that are in inflected form (e.g. verbs); that are the same but have different meanings (e.g. singer: the machine, the musician, etc.); that are in plural or singular; that are the same adjective with different gender, etc. In order to make the network con- ceptually coherent, one should define classes of words; for example, all the inflected forms of the verb “play” define one class whose representative is the word “play”. There are several other simple process- ing decisions to be made. This naive version can be further refined by incorporating more elaborated linguistic features, like labeling the edges according to parts of speech to which they point, e.g. nouns, verbs, adjectives, adverbs, etc., or giving different nodes and weights to different meanings of a word, and so on. The surprising fact is that even using a naive approach, the network obtained gives highly relevant and interesting information about the language. See Figure 1 for a small example. Although the idea of using mathematical and computational tools to capture the semantics information in dictionaries has been broadly explored (Amsler, 1980; Calzolari, 1984; Wilks et al., 1988), the idea fuel fire burn state combustion consume use produce heat matter FIRE: Fuel in a state of combustion. FUEL: Any matter used to produce heat by burning. BURN: To consume with fire. Figure 1: Entries in the dictionary for the words burn, fire, and fuel, and their corresponding subgraph built from them. …

Content truncated to 500 words