A Wordnet from the Ground Up

3.4. Measures of Semantic Relatedness 71features was also discussed in (Lapata, 2001, Boleda et al., 2004, 2005), but applied inthe semantic classification of adjectives. We have identified three types of constraintsas the potential semantic descriptors of adjectives:ANmod – an occurrence of a particular noun as modified by the given adjective,AAdv – an adverb in close proximity to the given adjective,AA – the co-occurrence with an adjective that agrees on case, number and gender asa potential co-constituent of the same noun phrase.ANmod is symmetrical to the AdjC constraint used for nominal LUs, but this timelexical elements are nouns instead of adjectives. AAdv is very similar to VAdv: lexicalelements are adverbs and we test the presence of an adverb in a distance not greaterthan 2. The implementation of AA, where lexical elements are adjectival LUs, hasbeen based on the scheme of ANmod, but we are looking for an occurrences of anotheradjectival LU which agrees on case, number and gender and which can be a co-modifierof the same nominal LU.The latter feature was advocated by Hatzivassiloglou and McKeown (1993) asexpressing negative semantic information: only unrelated adjectives can sit in the samenoun phrase. Our corpus data (collected from IPIC), however, suggest that it is toostrong a bias. In addition, our AA constraint also accepts coordination of adjectives,and then related adjectives can co-occur in a noun phrase. In the end, we used theAA feature in a positive way, just like the other features. Features of all three types,weighted and filtered by the RWF weight function discussed in Section 3.4.4, wereused in the discovery of contexts of occurrences of particular adjectives.The AA constraint was applied in two different ways:• as part of a joint large matrix together with the two other constraints: differentparts (columns) of row vectors generated by different constraints, but the matrixprocessed as a whole – this usage is encoded ANmod+AAdv+AA in Table 3.13,• two separate matrices were created: one joint for ANmod+AAdv and anotherfor AA only.In the second situation, the semantic relatedness values were calculated separatelyon the basis of both matrices separately processed and next linearly combined (Brodaet al., 2008):MSR Adj (l 1 , l 2 ) =α MSR ANmod+AAdv (l 1 , l 2 ) + β MSR AA (l 1 , l 2 )(3.5)The values of the coefficients were selected experimentally; α = β = 0.5 gave thebest results.