README

Mark Johnson, 16th August 2003

This program evaluates how the oracle score varies with beam size
and conditional probability threshold.

lugha [162] % main ../train.dat.bz2 
# Read 35540 sentences from ../train.dat.bz2
# Beam-size     Cum-freq        f-score
1       1       0.889067
5       0.892853        0.931632
10      0.759961        0.943433
15      0.663506        0.948784
20      0.587113        0.951959
25      0.518852        0.953942
30      0.448734        0.955338
35      0.370118        0.956211
40      0.274845        0.956836
45      0.156922        0.957161
50      0.0571469       0.957275
55      0.0385763       0.957307
60      0.0302757       0.957342
65      0.0243669       0.957354
70      0.0195554       0.957359
75      0.0156162       0.957363
80      0.0119584       0.957376
85      0.0087507       0.957384
90      0.0056556       0.957389
95      0.0025605       0.957389
100     0.00073157      0.957389

# percentile    min-cond-prob
0       2.01864e-06
0.05    1.53321e-05
0.1     3.32996e-05
0.15    8.2883e-05
0.2     0.00021316
0.25    0.000382411
0.3     0.00052441
0.35    0.000636331
0.4     0.000754905
0.45    0.000881554
0.5     0.00102076
0.55    0.00118796
0.6     0.00139984
0.65    0.00166871
0.7     0.00201246
0.75    0.00250743
0.8     0.00318711
0.85    0.0043278
0.9     0.0063322
0.95    0.0134335
1       1

# cond-prob-cutoff      f-score
0       0.957389
0.001   0.956844
0.002   0.956023
0.003   0.955131
0.004   0.954212
0.005   0.953319
0.006   0.952436
0.007   0.951474
0.008   0.950558
0.009   0.949765
0.01    0.948971
# utime 289.34s, vsize 4.93359 Mb.
381.270u 2.230s 6:56.34 92.1%   0+0k 0+0io 570pf+0w

lugha [161] % main ../dev.dat.bz2
# Read 3676 sentences from ../dev.dat.bz2
# Beam-size     Cum-freq        f-score
1       1       0.885598
5       0.893362        0.92914
10      0.765234        0.941086
15      0.662133        0.94626
20      0.593308        0.949634
25      0.51877 0.951843
30      0.447225        0.952923
35      0.366431        0.953796
40      0.274755        0.954436
45      0.164309        0.954828
50      0.0652884       0.954932
55      0.0448857       0.955008
60      0.0359086       0.955015
65      0.0288357       0.955022
70      0.0247552       0.955022
75      0.0193145       0.955066
80      0.0125136       0.955066
85      0.00979325      0.955083
90      0.0054407       0.955083
95      0.00217628      0.955083
100     0.000816104     0.955083

# percentile    min-cond-prob
0       2.18112e-06
0.05    1.45718e-05
0.1     2.85824e-05
0.15    6.68804e-05
0.2     0.0001857
0.25    0.000341583
0.3     0.000483883
0.35    0.000608685
0.4     0.000719489
0.45    0.000845115
0.5     0.000999269
0.55    0.00114795
0.6     0.00137011
0.65    0.00164907
0.7     0.00201006
0.75    0.00249685
0.8     0.00315291
0.85    0.00423701
0.9     0.00599749
0.95    0.0130043
1       1

# cond-prob-cutoff      f-score
0       0.955083
0.001   0.954434
0.002   0.95368
0.003   0.952956
0.004   0.951998
0.005   0.951232
0.006   0.950353
0.007   0.949259
0.008   0.948447
0.009   0.947641
0.01    0.946598
# utime 30.06s, vsize 3.22266 Mb.
