Skip to content

smartinsightsfromdata/SemEval-2015

 
 

Repository files navigation

SemEval-2015

Making an attempt at SemEval 2015 task 1

Hoanh and I decided to take a crack at the SemEval paraphrase evaluation task 1. We split up the work as follows, Hoanh wrote the neural network which integrated each of the word feature extrators. The final configuration of this network contained 2 hidden layers with back propogation. Our results from this are listed in the tables below. I on the other hand worked on the data analsys writing both a parser and preprocessor to feed the neural network in addition to handling all documents and references.

Final Results (run 1)

Training:

tp = 2467.0, tn = 6791.0, fp = 743.0, fn = 1529.0

9258 of 11530 correctly tagged... 0.802948829141370

measure score
precision 0.7685358255451713
recall 0.6173673673673674
f1 0.6847071884540661

Dev:

tp = 640.0, tn = 2468.0, fp = 204.0, fn = 830.0

3108 of 4142 correctly tagged... 0.7503621438918396

measure score
precision 0.7582938388625592
recall 0.43537414965986393
f1 0.5531547104580813

Test:

tp = 95.0, tn = 613.0, fp = 50.0, fn = 80.0

708 of 838 correctly tagged... 0.8448687350835322

measure score
precision 0.6551724137931034
recall 0.5428571428571428
f1 0.5937499999999999

Final Results (run 2)

Training:

tp = 2470.0, tn = 6769.0, fp = 765.0, fn = 1526.0

9239 of 11530 correctly tagged... 0.8013009540329575

measure score
precision 0.7635239567233385
recall 0.6181181181181181
f1 0.6831696860738488

Dev:

tp = 651.0, tn = 2466.0, fp = 206.0, fn = 819.0

3117 of 4142 correctly tagged... 0.7525350072428778

measure score
precision 0.7596266044340724
recall 0.44285714285714284
f1 0.559518693596906

Test:

tp = 94.0, tn = 621.0, fp = 42.0, fn = 81.0

715 of 838 correctly tagged... 0.8532219570405728

measure score
precision 0.6911764705882353
recall 0.5371428571428571
f1 0.6045016077170418

About

Making an attempt at SemEval 2015 task 1

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Perl 54.4%
  • Prolog 43.8%
  • Python 0.6%
  • PostScript 0.4%
  • Java 0.3%
  • HTML 0.2%
  • Other 0.3%