Discussion of [ Grefenstette , 1993 ] for CS 388 HW 4

1 Auto-generated Thesaurus Thesaurus is not only a useful resource to help understand related words and phrases in a document , it can also catalog correlation between words across multiple domains that were previously not obvious. For example, the word " mouse " in a document about animals is related to " cat " and " rabbit " , while in an article from a different domain it may relate with the words " printer " and " keyboard " [Dorow and Widdows, 2003]. Both concepts should appear in a thesaurus. Creating a thesaurus therefore requires a high degree of knowledge across many domains. Doing so manually requires domain experts and therefore can be expensive, yet incomplete. [Grefenstette, 1993] explores the possibility of automatically generating thesauri with domain-independent algorithm. Domain independence is what is being referred to as " knowledge-poor ". Given a large source of documents that contain sentences about some topic, the idea is to comb through this data and find terms that may relate to each other. In [Grefenstette, 1993], four specific strategies to mine related concepts were investigated. Before proceeding, the words were syntactically analyzed. They were put through a part of speech tagger and labeled as the appropriate part of speech. Their morphological deviations were standardized to reduce the number of different words. Each word now becomes a feature on its own. Words that are modifiers for other words (such as adjectives and verbs for nouns) were considered to be additional attributes for that word feature. • Related nouns Using the attributes as additional information about the noun words, each word pair can be assigned a similarity value. Words with attributes that are highly similar implies they were often used in similar ways, and therefore there is a high probability the two words are related. For additional robustness, two words are only considered related if they are among the top-N number of similar words for each other. • Related verbs Similarly for verbs, the attributes are in turn the nouns and phrases that are subjects or objects of the verbs. Verbs with similar attributes are related. • Noun phrases Noun phrases are similar to nouns, except that a larger number of adjacent words are used as attributes. This was found to produce more relevant attributes than a more stringent, smaller scope. • Morphology discovery Creative morphological development varied in form depending on the domain. A …