blob: 8dcac1360948c3b7172c272a4293fd0b75b302ee (
plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
|
- require 'zipf'
!!!
%html
%head
%title debug view (for session ##{session_key})
%link(rel="stylesheet" type="text/css" href="debug.css")
%body
%h1 debug view (for session ##{session_key})
%table
%tr
%td.noborder
%strong source:
%td.left #{data["source"]}
%tr
%td.noborder
%strong post-edit:
%td.left #{data["target"]}
%tr
%td.noborder
%strong original mt:
%td.left #{data["1best"]}
%tr
%td.noborder
%strong best match (bleu=#{data["best_match_score"]}):
%td.left #{data["best_match"]}
%h2 meta
%p <strong>k:</strong> #{data["samples_size"]}
%p <strong>number of updates:</strong> #{data["num_up"]}
%p <strong>updated features:</strong> #{data["updated_features"]}
%p <strong>learning rate:</strong> #{data["learning_rate"]}
%h2 k-best
%p bleu | model score | original rank | \|e\| | translation \n features
%p.red update needed
%ol
- kbest = []
- data["kbest"].each { |i| x=splitpipe(i); kbest << [ x[0].to_f, x[1].to_f, x[2].to_i, x[3], x[4] ] }
- kbest.sort! { |i,j| j[0] <=> i[0] }
- kbest.each_with_index do |k,j|
- b = kbest[0,j].map { |l| l[0]>k[0] && l[1]<k[1] }.include? true
-if b
%li.red
%strong #{"%.2f"%(k[0].to_f*100)} | #{k[1]} | #{k[2]} | #{k[4].split.size} |#{k[4]} <br/>
%pre #{k[3]}
- else
%li
%strong #{"%.2f"%(k[0].to_f*100)} | #{k[1]} | #{k[2]} | #{k[4].split.size} | #{k[4]} <br/>
%pre #{k[3]}
- if [9,89].include? j
%hr
%h2 weight updates
%table
%tr
%th feature
%th before
%th after
%th diff
%th raw diff
- data["weights_after"].keys.each.sort { |a,b| a[0] <=> b[0] }.each do |k|
%tr
%td.noborder <strong> #{k} </strong>
%td #{"%+.3f"%data["weights_before"][k].round(4)}
%td #{"%+.3f"%data["weights_after"][k].round(4)}
- diff = data["weights_before"][k].abs-data["weights_after"][k].abs
- if diff < 0
%td.red #{"%+.3f"%(diff).round(4)}
- elsif diff > 0
%td.green #{"%+.3f"%(diff).round(4)}
- else
%td #{"%+.3f"%(diff).round(4)}
%td #{"%+.1f"%((data["weights_before"][k].abs-data["weights_after"][k].abs)/data["learning_rate"]).round(2)}
%h3 dense features explained
%table
%tr
%td.noborder EgivenFCoherent (TM)
%td.left -log10[ c(e, f) / sample c(f) ]
%tr
%td.noborder Glue (per edge/rule) (dyn)
%td.left absolute count of used rules from glue grammar
%tr
%td.noborder IsSingletonF/E (per edge/rule) (TM)
%td.left true|false (1|0) (sum)
%tr
%td.noborder IsSingletonFE (per edge/rule) (TM)
%td.left true|false (1|0) (sum)
%tr
%td.noborder LanguageModel
%td.left -log?[ score ]
%tr
%td.noborder MaxLexFgivenE
%td.left Sum_f -log10(maxScore) (maxScore = max_e(ttable(f))
%tr
%td.noborder MaxLexEgivenF
%td.left Sum_e -log10(maxScore) (maxScore = max_f(ttable(e))
%tr
%td.noborder PassThrough
%td.left absolute count of used PassThrough rules (1 per word)
%tr
%td.noborder SampleCountF
%td.left log10 [ sample c(f) ]
%tr
%td.noborder WordPenalty (per edge/rule) (dyn)
%td.left log_10(e)*|e| = 1/log(10) * |e| (*-1) = -0.43429448190325176*|e|
%tr
%td.noborder SourceWordPenalty (per edge/rule) (dyn)
%td.left ^^^ (|e| <=> |f|)
|
