Growing Finely-Discriminating Taxonomies from Seeds of Varying Quality and Size

Proceedings of the 12th Conference of the European Chapter of the ACL, pages 835–842, Athens, Greece, 30 March – 3 April 2009. c⃝2009 Association for Computational Linguistics Growing Finely-Discriminating Taxonomies from Seeds of Varying Quality and Size Tony Veale School of Computer Science University College Dublin Ireland tony.veale@ucd.ie Guofu Li School of Computer Science University College Dublin Ireland guofu.li@ucd.ie Yanfen Hao School of Computer Science University College Dublin Ireland yanfen.hao@ucd.ie Abstract Concept taxonomies offer a powerful means for organizing knowledge, but this organiza- tion must allow for many overlapping and fine-grained perspectives if a general-purpose taxonomy is to reflect concepts as they are actually employed and reasoned about in ev- eryday usage. We present here a means of bootstrapping finely-discriminating tax- onomies from a variety of different starting points, or seeds, that are acquired from three different sources: WordNet, ConceptNet and the web at large. 1 Introduction Taxonomies provide a natural and intuitive means of organizing information, from the bio- logical taxonomies of the Linnaean system to the layout of supermarkets and bookstores to the or- ganizational structure of companies. Taxonomies also provide the structural backbone for ontolo- gies in computer science, from common-sense ontologies like Cyc (Lenat and Guha, 1990) and SUMO (Niles and Pease, 2001) to lexical ontolo- gies like WordNet (Miller et al., 1990). Each of these uses is based on the same root-branch-leaf metaphor: the broadest terms with the widest scope occupy the highest positions of a taxono- my, near the root, while specific terms with the most local concerns are located lower in the hier- archy, nearest the leaves. The more interior nodes that a taxonomy possesses, the finer the conceptual distinctions and the more gradated the similarity judgments it can make (e.g., Budanit- sky and Hirst, 2006). General-purpose computational taxonomies are called upon to perform both coarse-grained and fine-grained judgments. In NLP, for in- stance, the semantics of “eat” requires just enough knowledge to discriminate foods like tofu and cheese from non-foods like wool and steel, while specific applications in the domain of cooking and recipes (e.g., Hammond’s (1986) CHEF) require enough discrimination to know that tofu can be replaced with clotted cheese in many recipes because each is a soft, white and bland food. So while much depends on the domain of us- age, it remains an open question as to how many nodes a good taxonomy should possess. Prince- ton WordNet, for instance, strives for as many nodes as there are word senses in English, yet it also contains a substantial number of composite nodes that are lexicalized not as single words, but as complex phrases. Print dictionaries intend- ed for human consumption aim for some econo- my of structure, and typically do not include the meaning of phrases that can be understood as straightforward compositions of the meaning of their parts (Hanks, 2004). But WordNet also serves another purpose, as a lexical knowledge- base for computers, not humans, a context in which concerns about space seem quaint. When space is not …

Content truncated to 500 words