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.

120 Chapter 4. Extracting Relation InstancesThe former introduces a bias – plWordNet still is relatively small – enables testing<strong>the</strong> whole set of instances, while manual evaluation is always laborious and can beperformed only on a sample. Yet, <strong>the</strong> samples have been chosen as for <strong>the</strong> manualpatterns (Israel, 1992), so <strong>the</strong> results can apply to <strong>the</strong> whole sets with a 95% confidence.In both types of comparison we applied <strong>the</strong> standard measures of precision andrecall (Manning and Schütze, 2001) 9 .Precision and recall are defined in <strong>the</strong> standard way: tp is <strong>the</strong> number of truepositives (extracted pairs of LUs which are instances of <strong>the</strong> target relation), fp is <strong>the</strong>number of false positives (incorrect instances marked by algorithms as correct), fn –false negatives (correct instances in text but not extracted by <strong>the</strong> algorithm).P =R =tptp + fptptp + fn(4.7)(4.8)Note that <strong>the</strong> denominator in R accounts for correct patterns or instances thatwere ei<strong>the</strong>r marked as incorrect or not extracted at all. We cannot treat <strong>the</strong> limitedcore plWordNet as <strong>the</strong> exhaustive description of relations. That is why recall in ourapproach only measures <strong>the</strong> ratio of rediscovery of <strong>the</strong> plWordNet structure. It is not arecall in terms of all correct instances in <strong>the</strong> corpus or patterns that <strong>the</strong> corpus supports.Thus, following Pantel and Pennacchiotti (2006), we also use <strong>the</strong> relative recallmeasured in relation to <strong>the</strong> results of some o<strong>the</strong>r algorithm (Kurc, 2008, pp. 72):R A|B = R AR B=tp ACtp BC= tp Atp B= P A × (tp A + fp A )P B × (tp B + fp B )(4.9)where R A and R B denote <strong>the</strong> recall of <strong>the</strong> algorithms A and B, and C is <strong>the</strong> unknownnumber of instances occurring in <strong>the</strong> corpus.We extracted a ranked list of possible instances which can be sorted in descendingorder by reliability. The values are real numbers and <strong>the</strong>re is no characteristic pointbelow which we can cut off <strong>the</strong> rest of pairs according to some analytical properties.Thus, instead of pure precision and recall, we prefer to use cut-off precision and cut-offrecall calculated only in relation to some n first positions on <strong>the</strong> sorted list of results(instances or patterns).In <strong>the</strong> end, <strong>the</strong>n, we used three evaluation measures.1. Cut-off precision based on plWordNet marks as correct only those instances andpatterns that were found both in plWordNet and on an additional list provided a9 The F-measure could not be applied because of <strong>the</strong> limitations of recall based on plWordNet, to bediscussed later.

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

Saved successfully!

Ooh no, something went wrong!