Stupid ideas - STS-NTNU/STS13 GitHub Wiki

Interpolation

Id like to find a way to use additional training such as paraphrases. But how to assign a score to them? Perhaps give a max score (5) to paraphrase pairs and min score (0) to unrelated random pairs? What happens if you train SVR on only maximum and minimum scores? Does it somehow manage to interpolate between the scores? I tried this with the TakeLab system on MSRpar data. I mapped the scores to just two extremes:

if score < 3:
     score = 1.0
else:
     score = 5.0

Does the trained SVR produce any scores in between 1.0 and 5.0. Hardly:

cut -c-3  msrpar-output-ext.txt |sort |uniq -c
   2 0.6
   7 0.7
  16 0.8
   8 0.9
  10 1.0
   7 1.1
  15 1.2
   9 1.3
  10 1.4
  12 1.5
  11 1.6
   4 1.7
   5 1.8
   2 1.9
   1 4.0
   2 4.1
   3 4.2
   6 4.3
   9 4.4
  43 4.5
  69 4.6
  97 4.7
 125 4.8
 152 4.9
 125 5.0