24.07.2013 Views

Onto.PT: Towards the Automatic Construction of a Lexical Ontology ...

Onto.PT: Towards the Automatic Construction of a Lexical Ontology ...

Onto.PT: Towards the Automatic Construction of a Lexical Ontology ...

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

7.2. <strong>Onto</strong>logising performance 123<br />

Table 7.2 presents <strong>the</strong> scores obtained for each measure, according to algorithm<br />

and relation type. In RP and RP+AC, <strong>the</strong> threshold θ was empirically set to 0 and<br />

0.55, respectively. In NT+AC, θ was set to 3.<br />

Algorithm<br />

Hypernym-<strong>of</strong> (210 tb-triples)<br />

Precision (%) Recall (%) F1(%) F0.5(%) RF1(%)<br />

RP 53.8 12.4 20.2 32.3 50.3<br />

AC 60.1 15.7 24.9 38.4 59.8<br />

RP+AC 58.5 15.6 24.6 37.7 58.5<br />

NT 57.7 17.3 26.6 39.4 57.7<br />

NT+AC 58.7 15.3 24.3 37.4 58.6<br />

PR 46.2 11.5 18.5 28.9 45.7<br />

MD 58.6 15.8 24.9 38.0 58.6<br />

Part-<strong>of</strong> (175 tb-triples)<br />

Precision (%) Recall (%) F1(%) F0.5(%) RF1(%)<br />

RP 56.9 10.6 17.9 30.4 47.0<br />

AC 58.7 14.9 23.8 37.0 58.7<br />

RP+AC 64.1 16.6 26.3 40.7 64.1<br />

NT 50.7 15.8 24.1 35.2 50.7<br />

NT+AC 59.2 15.2 24.2 37.5 59.2<br />

PR 50.6 12.6 20.2 31.6 49.9<br />

MD 59.1 15.3 24.4 37.6 59.1<br />

Purpose-<strong>of</strong> (67 tb-triples)<br />

Precision (%) Recall (%) F1(%) F0.5(%) RF1(%)<br />

RP 51.5 5.1 9.3 18.3 32.6<br />

AC 63.2 13.0 21.5 35.6 63.2<br />

RP+AC 63.4 13.6 22.3 36.5 63.4<br />

NT 48.1 15.4 23.3 33.7 48.1<br />

NT+AC 62.2 13.9 22.7 36.6 62.2<br />

PR 56.3 10.8 18.2 30.6 56.3<br />

MD 60.9 12.7 20.9 34.5 60.9<br />

Table 7.2: <strong>Onto</strong>logising algorithms performance results.<br />

The comparison shows that <strong>the</strong> best performing algorithms for hypernymy are<br />

AC and NT, which have close F1 and RF1. NT is more likely to originate ties<br />

for <strong>the</strong> best attachments than AC, and thus to have higher recall. However, its<br />

precision is lower than AC’s. For part-<strong>of</strong>, RP+AC is clearly <strong>the</strong> best performing<br />

algorithm. For purpose-<strong>of</strong>, RP+AC has also <strong>the</strong> best precision and RF1, but its<br />

scores are very close to AC. Moreover, it is outperformed by NT and NT+AC in<br />

<strong>the</strong> o<strong>the</strong>r measures. Once again, NT has higher recall and thus higher F1. NT+AC<br />

combines good precision and recall in an interesting way and has <strong>the</strong>refore <strong>the</strong> best<br />

F0.5. However, as that <strong>the</strong> set <strong>of</strong> purpose-<strong>of</strong> tb-triples contains only 67 instances, <strong>the</strong><br />

results for this relation might not be significant enough to take strong conclusions.<br />

These results show as well that PR has <strong>the</strong> worst performance for hypernymy<br />

and part-<strong>of</strong> tb-triples, which suggests that PageRank is not adequate for this task.<br />

For purpose-<strong>of</strong>, RP is <strong>the</strong> worst algorithm, especially due to <strong>the</strong> low recall.<br />

7.2.3 Performance against an existing gold standard<br />

In <strong>the</strong> second performance evaluation, we used <strong>the</strong> proposed algorithms to ontologise<br />

antonymy relations. For this purpose, <strong>the</strong> antonymy sb-triples <strong>of</strong> TeP were converted<br />

to tb-triples. This resulted in 46,339 unique antonymy pairs – 7,633 between nouns,

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

Saved successfully!

Ooh no, something went wrong!