📄 A Taxonomy for English Nouns and Verbs

Summary

The author tests whether a coherent classification of English vocabulary can be derived from the definitions of an ordinary dictionary. The material is the machine-readable Merriam-Webster New Pocket Dictionary (MPD). In each definition of a noun or verb, the head (genus) word was identified manually, along with the sense in which it is used. From the resulting "defined sense — defining sense" pairs, two "tangled hierarchies" were assembled automatically. These are semilattices in which a node can have several parents. The noun hierarchy has about 24,000 nodes, and the verb hierarchy about 11,000. The author analyzes the most general top nodes, as well as the size and depth of the subtrees. The conclusion is that the dictionary is, on the whole, consistent in describing word senses. In addition, the ISA relation alone proved insufficient. Words denoting a part or a collection point to a second transitive relation, "part-whole" (ISPART). The paper is one of the first attempts to build a taxonomy of the entire lexicon from an existing dictionary. It shows that semantic resources for language understanding systems can be extracted from dictionaries rather than built from scratch.

Key points

• The data source is the MPD. Its machine-readable version, like that of the Seventh Collegiate Dictionary, was prepared in the late 1960s by John Olney and colleagues at System Development Corporation. The project was funded by the NSF (grant MCS77-01315).
• Scale: about 24,000 noun senses and 11,000 verb senses. The largest noun hierarchy has the top node ONE-2.1A (3379 nodes), and the largest verb hierarchy has REMAIN-.4A (4175 nodes). The most general top nodes include senses of the words "cause", "thing", "class", "being".
• The size of a hierarchy is measured in two ways: its maximum depth and the number of reachable nodes. To determine how central a node is, the number of its immediate descendants is also taken into account.
• The content categories largely coincide with Nida's semantic classification. The verb taxonomy supports the semantic primitives proposed by the LNR group (San Diego).
• Partitives (words for a part) and collectives (words for a collection) were classified by Gretchen Hazard and John White. Their groups include group names, piece units, space shapes, and others. These words motivate the ISPART relation alongside ISA. Example: LEAF:PLANT :: FROND:FERN.