06.08.2015 Views

A Wordnet from the Ground Up

A Wordnet from the Ground Up - School of Information Technology ...

A Wordnet from the Ground Up - School of Information Technology ...

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

84 Chapter 3. Discovering Semantic RelatednessFeatures Lin CRMI PMI RW F zscore GRW F LinNArg(acc) 60.47 58.13 55.53 56.89 63.35NArg(dat) 38.59 23.98 36.40 26.29 39.06NArg(inst) 55.60 45.59 50.75 42.54 57.57NArg(loc) 51.42 46.72 48.54 41.47 54.50Nsb 53.15 54.54 47.68 57.79 54.78VPart 46.70 44.69 44.89 47.28 48.32VAdv 65.32 53.77 58.04 64.19 66.50NArg(acc+dat+inst+loc) 65.13 65.04 60.94 64.17 68.05NSb+NArg+VPart+VAdv 67.10 67.80 62.42 62.41 71.85AAdv 57.60 20.86 53.27 58.96 58.71AA 74.24 71.86 71.75 72.32 76.87ANmod 76.12 74.77 73.99 79.18 77.75ANmod+AAdv 76.97 75.41 74.80 81.13 78.93ANmod+AA 78.18 78.32 78.43 83.05 79.89ANmod+AAdv+AA 79.71 78.32 78.39 83.26 82.48Table 3.14: Experiments with MSRs for all lemmasThe result of our best adjective MSR is very close to <strong>the</strong> result achieved by humans(Section 3.3.1). For verbs, <strong>the</strong> difference is comparable to that observed fornouns (Piasecki et al., 2007b) (but <strong>the</strong> result of verb MSR still approaches humanperformance).The constructed MSRs are intended to assist linguists in selecting LUs semanticallyrelated to <strong>the</strong> LU being edited. Lexicographers can find missing synonyms or instancesof lexico-semantic relations while browsing <strong>the</strong> MSRlist (x,k) lists (according to <strong>the</strong>MSRs).Long suggestion lists may preclude careful analysis. We chose k = 20 for a smallexperiment to test a possible future use of both MSRs by linguists. We randomlyselected two subsets of lemmas, verbs and adjectives. We determined sample sizes insuch a way that <strong>the</strong> results of <strong>the</strong> manual evaluation performed on <strong>the</strong> samples couldbe ascribed to <strong>the</strong> whole sets with <strong>the</strong> 95% confidence level, according to <strong>the</strong> methoddiscussed in (Israel, 1992). For every LU in each subset, we generated <strong>the</strong> list of <strong>the</strong>k = 20 LUs most related to <strong>the</strong> given one. One of <strong>the</strong> co-authors manually assessedall elements on all lists, distinguishing any elements that are in some wordnet relationto <strong>the</strong> head LU.The evaluated LU lists were classified into:• very useful – a half, or almost a half, of <strong>the</strong> LUs on <strong>the</strong> list are in some semanticrelation to <strong>the</strong> given one,• useful – a sizable part of <strong>the</strong> list is somehow related,

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!