line |
stmt |
bran |
cond |
sub |
pod |
time |
code |
1
|
|
|
|
|
|
|
package Treex::Tool::Parser::MSTperl::ModelLabelling; |
2
|
|
|
|
|
|
|
{ |
3
|
|
|
|
|
|
|
$Treex::Tool::Parser::MSTperl::ModelLabelling::VERSION = '0.11949'; |
4
|
|
|
|
|
|
|
} |
5
|
|
|
|
|
|
|
|
6
|
1
|
|
|
1
|
|
4198
|
use Moose; |
|
0
|
|
|
|
|
|
|
|
0
|
|
|
|
|
|
|
7
|
|
|
|
|
|
|
use Carp; |
8
|
|
|
|
|
|
|
|
9
|
|
|
|
|
|
|
extends 'Treex::Tool::Parser::MSTperl::ModelBase'; |
10
|
|
|
|
|
|
|
|
11
|
|
|
|
|
|
|
# basic MLE from data |
12
|
|
|
|
|
|
|
# unigrams->{label} = prob |
13
|
|
|
|
|
|
|
# to be used for smoothing and/or backoff |
14
|
|
|
|
|
|
|
# (can be used both for emissions and transitions) |
15
|
|
|
|
|
|
|
# It also contains the SEQUENCE_BOUNDARY_LABEL prob |
16
|
|
|
|
|
|
|
# (the SEQUENCE_BOUNDARY_LABEL is counted once for each sequence) |
17
|
|
|
|
|
|
|
# which might be unappropriate in some cases (eg. for emission probs) |
18
|
|
|
|
|
|
|
has 'unigrams' => ( |
19
|
|
|
|
|
|
|
is => 'rw', |
20
|
|
|
|
|
|
|
isa => 'HashRef', |
21
|
|
|
|
|
|
|
default => sub { {} }, |
22
|
|
|
|
|
|
|
); |
23
|
|
|
|
|
|
|
|
24
|
|
|
|
|
|
|
# transition scores for Viterbi with the structure (if MIRA-computed): |
25
|
|
|
|
|
|
|
# transitions->{feature}->{label_prev}->{label_this} = score |
26
|
|
|
|
|
|
|
# or probabilties (if obtained by MLE): |
27
|
|
|
|
|
|
|
# transitions->{label_prev}->{label_this} = prob |
28
|
|
|
|
|
|
|
# (if MLE is used for transitions, during the precomputing phase |
29
|
|
|
|
|
|
|
# counts are temporarily stored instead of probs |
30
|
|
|
|
|
|
|
# and they are converted to probs on calling prepare_for_mira() ); |
31
|
|
|
|
|
|
|
has 'transitions' => ( |
32
|
|
|
|
|
|
|
is => 'rw', |
33
|
|
|
|
|
|
|
isa => 'HashRef', |
34
|
|
|
|
|
|
|
default => sub { {} }, |
35
|
|
|
|
|
|
|
); |
36
|
|
|
|
|
|
|
|
37
|
|
|
|
|
|
|
# smoothing parameters of transition probabilities |
38
|
|
|
|
|
|
|
# (to be computed by EM algorithm) |
39
|
|
|
|
|
|
|
# PROB(label|prev_label) = |
40
|
|
|
|
|
|
|
# smooth_bigrams * transitions->{prev_label}->{label} + |
41
|
|
|
|
|
|
|
# smooth_unigrams * unigrams->{label} + |
42
|
|
|
|
|
|
|
# smooth_uniform |
43
|
|
|
|
|
|
|
|
44
|
|
|
|
|
|
|
has 'smooth_bigrams' => ( |
45
|
|
|
|
|
|
|
is => 'rw', |
46
|
|
|
|
|
|
|
isa => 'Num', |
47
|
|
|
|
|
|
|
default => 0.6, |
48
|
|
|
|
|
|
|
); |
49
|
|
|
|
|
|
|
|
50
|
|
|
|
|
|
|
has 'smooth_unigrams' => ( |
51
|
|
|
|
|
|
|
is => 'rw', |
52
|
|
|
|
|
|
|
isa => 'Num', |
53
|
|
|
|
|
|
|
default => 0.3, |
54
|
|
|
|
|
|
|
); |
55
|
|
|
|
|
|
|
|
56
|
|
|
|
|
|
|
has 'smooth_uniform' => ( |
57
|
|
|
|
|
|
|
is => 'rw', |
58
|
|
|
|
|
|
|
isa => 'Num', |
59
|
|
|
|
|
|
|
default => 0.1, |
60
|
|
|
|
|
|
|
); |
61
|
|
|
|
|
|
|
|
62
|
|
|
|
|
|
|
# = 1 / ( keys %{ $self->unigrams } ) |
63
|
|
|
|
|
|
|
# set in compute_smoothing_params |
64
|
|
|
|
|
|
|
has 'uniform_prob' => ( |
65
|
|
|
|
|
|
|
is => 'rw', |
66
|
|
|
|
|
|
|
isa => 'Num', |
67
|
|
|
|
|
|
|
default => 0.02, |
68
|
|
|
|
|
|
|
); |
69
|
|
|
|
|
|
|
|
70
|
|
|
|
|
|
|
# emission scores for Viterbi with the structure |
71
|
|
|
|
|
|
|
# emissions->{feature}->{label} = score |
72
|
|
|
|
|
|
|
has 'emissions' => ( |
73
|
|
|
|
|
|
|
is => 'rw', |
74
|
|
|
|
|
|
|
isa => 'HashRef', |
75
|
|
|
|
|
|
|
default => sub { {} }, |
76
|
|
|
|
|
|
|
); |
77
|
|
|
|
|
|
|
|
78
|
|
|
|
|
|
|
# just an array ref with the sentences that represent the heldout data |
79
|
|
|
|
|
|
|
# to be able to run the EM algorithm in prepare_for_mira() |
80
|
|
|
|
|
|
|
has 'EM_heldout_data' => ( |
81
|
|
|
|
|
|
|
is => 'rw', |
82
|
|
|
|
|
|
|
isa => 'ArrayRef[Treex::Tool::Parser::MSTperl::Sentence]', |
83
|
|
|
|
|
|
|
default => sub { [] }, |
84
|
|
|
|
|
|
|
); |
85
|
|
|
|
|
|
|
|
86
|
|
|
|
|
|
|
sub BUILD { |
87
|
|
|
|
|
|
|
my ($self) = @_; |
88
|
|
|
|
|
|
|
|
89
|
|
|
|
|
|
|
$self->featuresControl( $self->config->labelledFeaturesControl ); |
90
|
|
|
|
|
|
|
|
91
|
|
|
|
|
|
|
return; |
92
|
|
|
|
|
|
|
} |
93
|
|
|
|
|
|
|
|
94
|
|
|
|
|
|
|
# STORING AND LOADING |
95
|
|
|
|
|
|
|
|
96
|
|
|
|
|
|
|
sub get_data_to_store { |
97
|
|
|
|
|
|
|
my ($self) = @_; |
98
|
|
|
|
|
|
|
|
99
|
|
|
|
|
|
|
return { |
100
|
|
|
|
|
|
|
'unigrams' => $self->unigrams, |
101
|
|
|
|
|
|
|
'transitions' => $self->transitions, |
102
|
|
|
|
|
|
|
'emissions' => $self->emissions, |
103
|
|
|
|
|
|
|
'smooth_uniform' => $self->smooth_uniform, |
104
|
|
|
|
|
|
|
'smooth_unigrams' => $self->smooth_unigrams, |
105
|
|
|
|
|
|
|
'smooth_bigrams' => $self->smooth_bigrams, |
106
|
|
|
|
|
|
|
'uniform_prob' => $self->uniform_prob, |
107
|
|
|
|
|
|
|
}; |
108
|
|
|
|
|
|
|
} |
109
|
|
|
|
|
|
|
|
110
|
|
|
|
|
|
|
sub load_data { |
111
|
|
|
|
|
|
|
|
112
|
|
|
|
|
|
|
my ( $self, $data ) = @_; |
113
|
|
|
|
|
|
|
|
114
|
|
|
|
|
|
|
$self->unigrams( $data->{'unigrams'} ); |
115
|
|
|
|
|
|
|
$self->transitions( $data->{'transitions'} ); |
116
|
|
|
|
|
|
|
$self->emissions( $data->{'emissions'} ); |
117
|
|
|
|
|
|
|
|
118
|
|
|
|
|
|
|
$self->smooth_uniform( $data->{'smooth_uniform'} ); |
119
|
|
|
|
|
|
|
$self->smooth_unigrams( $data->{'smooth_unigrams'} ); |
120
|
|
|
|
|
|
|
$self->smooth_bigrams( $data->{'smooth_bigrams'} ); |
121
|
|
|
|
|
|
|
$self->uniform_prob( $data->{'uniform_prob'} ); |
122
|
|
|
|
|
|
|
|
123
|
|
|
|
|
|
|
my $unigrams_ok = scalar( keys %{ $self->unigrams } ); |
124
|
|
|
|
|
|
|
my $transitions_ok = scalar( keys %{ $self->transitions } ); |
125
|
|
|
|
|
|
|
my $emissions_ok = scalar( keys %{ $self->emissions } ); |
126
|
|
|
|
|
|
|
|
127
|
|
|
|
|
|
|
my $smooth_sum = $self->smooth_uniform + $self->smooth_unigrams |
128
|
|
|
|
|
|
|
+ $self->smooth_bigrams; |
129
|
|
|
|
|
|
|
|
130
|
|
|
|
|
|
|
my $smooth_ok = ( |
131
|
|
|
|
|
|
|
|
132
|
|
|
|
|
|
|
# should be 1 but might be a little shifted |
133
|
|
|
|
|
|
|
$smooth_sum > 0.999 |
134
|
|
|
|
|
|
|
&& $smooth_sum < 1.001 |
135
|
|
|
|
|
|
|
|
136
|
|
|
|
|
|
|
# must be between 0 and 1 |
137
|
|
|
|
|
|
|
&& $self->uniform_prob > 0 |
138
|
|
|
|
|
|
|
&& $self->uniform_prob < 1 |
139
|
|
|
|
|
|
|
); |
140
|
|
|
|
|
|
|
|
141
|
|
|
|
|
|
|
my $ALGORITHM = $self->config->labeller_algorithm; |
142
|
|
|
|
|
|
|
|
143
|
|
|
|
|
|
|
if ($ALGORITHM == 0 |
144
|
|
|
|
|
|
|
|| $ALGORITHM == 1 |
145
|
|
|
|
|
|
|
|| $ALGORITHM == 2 |
146
|
|
|
|
|
|
|
|| $ALGORITHM == 3 |
147
|
|
|
|
|
|
|
|| $ALGORITHM == 4 |
148
|
|
|
|
|
|
|
|| $ALGORITHM == 8 |
149
|
|
|
|
|
|
|
|| $ALGORITHM == 9 |
150
|
|
|
|
|
|
|
|| $ALGORITHM == 10 |
151
|
|
|
|
|
|
|
|| $ALGORITHM == 11 |
152
|
|
|
|
|
|
|
|| $ALGORITHM == 14 |
153
|
|
|
|
|
|
|
) |
154
|
|
|
|
|
|
|
{ |
155
|
|
|
|
|
|
|
|
156
|
|
|
|
|
|
|
# these algorithms do not use lambda smoothing |
157
|
|
|
|
|
|
|
# (smoothing is kind of part of the learning) |
158
|
|
|
|
|
|
|
$smooth_ok = 1; |
159
|
|
|
|
|
|
|
} |
160
|
|
|
|
|
|
|
|
161
|
|
|
|
|
|
|
if ( $ALGORITHM >= 20 ) { |
162
|
|
|
|
|
|
|
|
163
|
|
|
|
|
|
|
# these algorithms do not use separate transitions |
164
|
|
|
|
|
|
|
# (transitions are included in emissions) |
165
|
|
|
|
|
|
|
$transitions_ok = 1; |
166
|
|
|
|
|
|
|
} |
167
|
|
|
|
|
|
|
|
168
|
|
|
|
|
|
|
if ( $unigrams_ok && $transitions_ok && $emissions_ok && $smooth_ok ) { |
169
|
|
|
|
|
|
|
return 1; |
170
|
|
|
|
|
|
|
} else { |
171
|
|
|
|
|
|
|
return 0; |
172
|
|
|
|
|
|
|
} |
173
|
|
|
|
|
|
|
} |
174
|
|
|
|
|
|
|
|
175
|
|
|
|
|
|
|
# TRANSITION AND EMISSION COUNTS AND PROBABILITIES |
176
|
|
|
|
|
|
|
# (more or less standard MLE) |
177
|
|
|
|
|
|
|
|
178
|
|
|
|
|
|
|
sub add_unigram { |
179
|
|
|
|
|
|
|
my ( $self, $label ) = @_; |
180
|
|
|
|
|
|
|
|
181
|
|
|
|
|
|
|
if ( $self->config->DEBUG >= 2 ) { |
182
|
|
|
|
|
|
|
print "add_unigram($label)\n"; |
183
|
|
|
|
|
|
|
} |
184
|
|
|
|
|
|
|
|
185
|
|
|
|
|
|
|
# increment number of unigrams |
186
|
|
|
|
|
|
|
$self->unigrams->{$label} += 1; |
187
|
|
|
|
|
|
|
|
188
|
|
|
|
|
|
|
return; |
189
|
|
|
|
|
|
|
} |
190
|
|
|
|
|
|
|
|
191
|
|
|
|
|
|
|
sub add_transition { |
192
|
|
|
|
|
|
|
|
193
|
|
|
|
|
|
|
# Str, Str, Maybe[Str] |
194
|
|
|
|
|
|
|
my ( $self, $label_this, $label_prev, $feature ) = @_; |
195
|
|
|
|
|
|
|
|
196
|
|
|
|
|
|
|
if ( defined $feature ) { |
197
|
|
|
|
|
|
|
if ( $self->config->DEBUG >= 2 ) { |
198
|
|
|
|
|
|
|
print "add_transition($label_this, $label_prev, $feature)\n"; |
199
|
|
|
|
|
|
|
} |
200
|
|
|
|
|
|
|
|
201
|
|
|
|
|
|
|
# increment number of bigrams |
202
|
|
|
|
|
|
|
$self->transitions->{$feature}->{$label_prev}->{$label_this} += 1; |
203
|
|
|
|
|
|
|
} else { |
204
|
|
|
|
|
|
|
if ( $self->config->DEBUG >= 2 ) { |
205
|
|
|
|
|
|
|
print "add_transition($label_this, $label_prev)\n"; |
206
|
|
|
|
|
|
|
} |
207
|
|
|
|
|
|
|
|
208
|
|
|
|
|
|
|
# increment number of bigrams |
209
|
|
|
|
|
|
|
$self->transitions->{$label_prev}->{$label_this} += 1; |
210
|
|
|
|
|
|
|
} |
211
|
|
|
|
|
|
|
|
212
|
|
|
|
|
|
|
return; |
213
|
|
|
|
|
|
|
} |
214
|
|
|
|
|
|
|
|
215
|
|
|
|
|
|
|
sub add_emission { |
216
|
|
|
|
|
|
|
my ( $self, $feature, $label ) = @_; |
217
|
|
|
|
|
|
|
|
218
|
|
|
|
|
|
|
if ( $self->config->DEBUG >= 3 ) { |
219
|
|
|
|
|
|
|
print "add_emission($feature, $label)\n"; |
220
|
|
|
|
|
|
|
} |
221
|
|
|
|
|
|
|
|
222
|
|
|
|
|
|
|
$self->emissions->{$feature}->{$label} += 1; |
223
|
|
|
|
|
|
|
|
224
|
|
|
|
|
|
|
return; |
225
|
|
|
|
|
|
|
} |
226
|
|
|
|
|
|
|
|
227
|
|
|
|
|
|
|
# called after preprocessing training data, before entering the MIRA phase |
228
|
|
|
|
|
|
|
sub prepare_for_mira { |
229
|
|
|
|
|
|
|
|
230
|
|
|
|
|
|
|
my ( $self, $trainer ) = @_; |
231
|
|
|
|
|
|
|
|
232
|
|
|
|
|
|
|
# $trainer used only in algoprithm no. 9 for emissions initialization |
233
|
|
|
|
|
|
|
|
234
|
|
|
|
|
|
|
my $ALGORITHM = $self->config->labeller_algorithm; |
235
|
|
|
|
|
|
|
|
236
|
|
|
|
|
|
|
if ( $ALGORITHM == 9 ) { |
237
|
|
|
|
|
|
|
|
238
|
|
|
|
|
|
|
# no need to recompute to probabilities (counts are OK) |
239
|
|
|
|
|
|
|
# but have to update emissions_summed |
240
|
|
|
|
|
|
|
# and transitions_summed appropriately |
241
|
|
|
|
|
|
|
|
242
|
|
|
|
|
|
|
my $sumUpdateWeight = $trainer->number_of_inner_iterations; |
243
|
|
|
|
|
|
|
|
244
|
|
|
|
|
|
|
# emissions->{feature}->{label} |
245
|
|
|
|
|
|
|
foreach my $feature ( keys %{ $self->emissions } ) { |
246
|
|
|
|
|
|
|
foreach my $label ( keys %{ $self->emissions->{$feature} } ) { |
247
|
|
|
|
|
|
|
$trainer->emissions_summed->{$feature}->{$label} |
248
|
|
|
|
|
|
|
= $sumUpdateWeight * $self->emissions->{$feature}->{$label}; |
249
|
|
|
|
|
|
|
} |
250
|
|
|
|
|
|
|
} |
251
|
|
|
|
|
|
|
|
252
|
|
|
|
|
|
|
# transitions->{feature}->{label_prev}->{label_this} |
253
|
|
|
|
|
|
|
foreach my $feature ( keys %{ $self->transitions } ) { |
254
|
|
|
|
|
|
|
foreach my $label_prev ( |
255
|
|
|
|
|
|
|
keys %{ $self->transitions->{$feature} } |
256
|
|
|
|
|
|
|
) |
257
|
|
|
|
|
|
|
{ |
258
|
|
|
|
|
|
|
foreach my $label_this ( |
259
|
|
|
|
|
|
|
keys %{ $self->transitions->{$feature}->{$label_prev} } |
260
|
|
|
|
|
|
|
) |
261
|
|
|
|
|
|
|
{ |
262
|
|
|
|
|
|
|
$trainer->transitions_summed |
263
|
|
|
|
|
|
|
->{$feature}->{$label_prev}->{$label_this} |
264
|
|
|
|
|
|
|
= $sumUpdateWeight * $self->transitions |
265
|
|
|
|
|
|
|
->{$feature}->{$label_prev}->{$label_this}; |
266
|
|
|
|
|
|
|
} |
267
|
|
|
|
|
|
|
} |
268
|
|
|
|
|
|
|
} |
269
|
|
|
|
|
|
|
|
270
|
|
|
|
|
|
|
} elsif ( $ALGORITHM == 1 || $ALGORITHM == 8 || $ALGORITHM >= 20 ) { |
271
|
|
|
|
|
|
|
|
272
|
|
|
|
|
|
|
# no recomputing taking place |
273
|
|
|
|
|
|
|
|
274
|
|
|
|
|
|
|
} elsif ( |
275
|
|
|
|
|
|
|
$ALGORITHM == 0 |
276
|
|
|
|
|
|
|
|| $ALGORITHM == 2 |
277
|
|
|
|
|
|
|
|| $ALGORITHM == 3 |
278
|
|
|
|
|
|
|
|| $ALGORITHM == 4 |
279
|
|
|
|
|
|
|
|| $ALGORITHM == 5 |
280
|
|
|
|
|
|
|
|| $ALGORITHM == 10 |
281
|
|
|
|
|
|
|
|| $ALGORITHM == 11 |
282
|
|
|
|
|
|
|
|| $ALGORITHM == 12 |
283
|
|
|
|
|
|
|
|| $ALGORITHM == 13 |
284
|
|
|
|
|
|
|
|| $ALGORITHM == 14 |
285
|
|
|
|
|
|
|
|| $ALGORITHM == 15 |
286
|
|
|
|
|
|
|
|| $ALGORITHM == 16 |
287
|
|
|
|
|
|
|
|| $ALGORITHM == 17 |
288
|
|
|
|
|
|
|
|| $ALGORITHM == 18 |
289
|
|
|
|
|
|
|
|| $ALGORITHM == 19 |
290
|
|
|
|
|
|
|
) |
291
|
|
|
|
|
|
|
{ |
292
|
|
|
|
|
|
|
|
293
|
|
|
|
|
|
|
# compute unigram probs |
294
|
|
|
|
|
|
|
$self->compute_probs_from_counts( $self->unigrams ); |
295
|
|
|
|
|
|
|
|
296
|
|
|
|
|
|
|
# compute transition probs |
297
|
|
|
|
|
|
|
foreach my $label ( keys %{ $self->transitions } ) { |
298
|
|
|
|
|
|
|
$self->compute_probs_from_counts( $self->transitions->{$label} ); |
299
|
|
|
|
|
|
|
} |
300
|
|
|
|
|
|
|
|
301
|
|
|
|
|
|
|
if ($ALGORITHM == 4 |
302
|
|
|
|
|
|
|
|| $ALGORITHM == 5 |
303
|
|
|
|
|
|
|
) |
304
|
|
|
|
|
|
|
{ |
305
|
|
|
|
|
|
|
|
306
|
|
|
|
|
|
|
# compute emission probs (MLE) |
307
|
|
|
|
|
|
|
foreach my $feature ( keys %{ $self->emissions } ) { |
308
|
|
|
|
|
|
|
$self->compute_probs_from_counts( |
309
|
|
|
|
|
|
|
$self->emissions->{$feature} |
310
|
|
|
|
|
|
|
); |
311
|
|
|
|
|
|
|
} |
312
|
|
|
|
|
|
|
} # end if $ALGORITHM == 4|5 |
313
|
|
|
|
|
|
|
|
314
|
|
|
|
|
|
|
if ($ALGORITHM == 5 |
315
|
|
|
|
|
|
|
|| $ALGORITHM == 12 |
316
|
|
|
|
|
|
|
|| $ALGORITHM == 13 |
317
|
|
|
|
|
|
|
|| $ALGORITHM == 15 |
318
|
|
|
|
|
|
|
|| $ALGORITHM == 16 |
319
|
|
|
|
|
|
|
|| $ALGORITHM == 17 |
320
|
|
|
|
|
|
|
|| $ALGORITHM == 18 |
321
|
|
|
|
|
|
|
|| $ALGORITHM == 19 |
322
|
|
|
|
|
|
|
) |
323
|
|
|
|
|
|
|
{ |
324
|
|
|
|
|
|
|
|
325
|
|
|
|
|
|
|
# run the EM algorithm to compute |
326
|
|
|
|
|
|
|
# transtition probs smoothing params |
327
|
|
|
|
|
|
|
$self->compute_smoothing_params(); |
328
|
|
|
|
|
|
|
} # end if $ALGORITHM == 5|12|12|>=16 |
329
|
|
|
|
|
|
|
|
330
|
|
|
|
|
|
|
} else { # $ALGORITHM not in 0~9 |
331
|
|
|
|
|
|
|
croak "ModelLabelling->prepare_for_mira not implemented" |
332
|
|
|
|
|
|
|
. " for algorithm no. $ALGORITHM!"; |
333
|
|
|
|
|
|
|
} |
334
|
|
|
|
|
|
|
|
335
|
|
|
|
|
|
|
return; |
336
|
|
|
|
|
|
|
} # end prepare_for_mira |
337
|
|
|
|
|
|
|
|
338
|
|
|
|
|
|
|
# basic MLE |
339
|
|
|
|
|
|
|
sub compute_probs_from_counts { |
340
|
|
|
|
|
|
|
my ( $self, $hashref ) = @_; |
341
|
|
|
|
|
|
|
|
342
|
|
|
|
|
|
|
my $sum = 0; |
343
|
|
|
|
|
|
|
foreach my $key ( keys %{$hashref} ) { |
344
|
|
|
|
|
|
|
$sum += $hashref->{$key}; |
345
|
|
|
|
|
|
|
} |
346
|
|
|
|
|
|
|
foreach my $key ( keys %{$hashref} ) { |
347
|
|
|
|
|
|
|
$hashref->{$key} = $hashref->{$key} / $sum; |
348
|
|
|
|
|
|
|
} |
349
|
|
|
|
|
|
|
|
350
|
|
|
|
|
|
|
return; |
351
|
|
|
|
|
|
|
} |
352
|
|
|
|
|
|
|
|
353
|
|
|
|
|
|
|
# EM algorithm to estimate linear interpolation smoothing parameters |
354
|
|
|
|
|
|
|
# for smoothing of transition probabilities |
355
|
|
|
|
|
|
|
sub compute_smoothing_params { |
356
|
|
|
|
|
|
|
my ($self) = @_; |
357
|
|
|
|
|
|
|
|
358
|
|
|
|
|
|
|
# only progress and/or debug info |
359
|
|
|
|
|
|
|
if ( $self->config->DEBUG >= 1 ) { |
360
|
|
|
|
|
|
|
print "Running EM algorithm to estimate lambdas...\n"; |
361
|
|
|
|
|
|
|
} |
362
|
|
|
|
|
|
|
|
363
|
|
|
|
|
|
|
# uniform probability is 1 / number of different labels |
364
|
|
|
|
|
|
|
$self->uniform_prob( 1 / ( keys %{ $self->unigrams } ) ); |
365
|
|
|
|
|
|
|
|
366
|
|
|
|
|
|
|
my $change = 1; |
367
|
|
|
|
|
|
|
while ( $change > $self->config->EM_EPSILON ) { |
368
|
|
|
|
|
|
|
|
369
|
|
|
|
|
|
|
#compute "expected counts" |
370
|
|
|
|
|
|
|
my $expectedCounts = $self->count_expected_counts_all(); |
371
|
|
|
|
|
|
|
my $expectedCountsSum = $expectedCounts->[0] + $expectedCounts->[1] |
372
|
|
|
|
|
|
|
+ $expectedCounts->[2]; |
373
|
|
|
|
|
|
|
|
374
|
|
|
|
|
|
|
#compute new lambdas |
375
|
|
|
|
|
|
|
my @new_lambdas = map { $_ / $expectedCountsSum } @$expectedCounts; |
376
|
|
|
|
|
|
|
|
377
|
|
|
|
|
|
|
#compute the change (sum of changes of lambdas) |
378
|
|
|
|
|
|
|
$change = abs( $self->smooth_uniform - $new_lambdas[0] ) |
379
|
|
|
|
|
|
|
+ abs( $self->smooth_unigrams - $new_lambdas[1] ) |
380
|
|
|
|
|
|
|
+ abs( $self->smooth_bigrams - $new_lambdas[2] ); |
381
|
|
|
|
|
|
|
|
382
|
|
|
|
|
|
|
# set new lambdas |
383
|
|
|
|
|
|
|
$self->smooth_uniform( $new_lambdas[0] ); |
384
|
|
|
|
|
|
|
$self->smooth_unigrams( $new_lambdas[1] ); |
385
|
|
|
|
|
|
|
$self->smooth_bigrams( $new_lambdas[2] ); |
386
|
|
|
|
|
|
|
|
387
|
|
|
|
|
|
|
# only progress and/or debug info |
388
|
|
|
|
|
|
|
if ( $self->config->DEBUG >= 2 ) { |
389
|
|
|
|
|
|
|
print "Last change: $change\n"; |
390
|
|
|
|
|
|
|
} |
391
|
|
|
|
|
|
|
} |
392
|
|
|
|
|
|
|
|
393
|
|
|
|
|
|
|
# only progress and/or debug info |
394
|
|
|
|
|
|
|
if ( $self->config->DEBUG >= 2 ) { |
395
|
|
|
|
|
|
|
print "Final lambdas:\n" |
396
|
|
|
|
|
|
|
. "uniform: " . $self->smooth_uniform |
397
|
|
|
|
|
|
|
. "unigram: " . $self->smooth_unigrams |
398
|
|
|
|
|
|
|
. "bigram: " . $self->smooth_bigrams; |
399
|
|
|
|
|
|
|
} |
400
|
|
|
|
|
|
|
if ( $self->config->DEBUG >= 1 ) { |
401
|
|
|
|
|
|
|
print "Done.\n"; |
402
|
|
|
|
|
|
|
} |
403
|
|
|
|
|
|
|
|
404
|
|
|
|
|
|
|
return; |
405
|
|
|
|
|
|
|
} |
406
|
|
|
|
|
|
|
|
407
|
|
|
|
|
|
|
#count "expected counts" of lambdas |
408
|
|
|
|
|
|
|
sub count_expected_counts_all { |
409
|
|
|
|
|
|
|
my ($self) = @_; |
410
|
|
|
|
|
|
|
|
411
|
|
|
|
|
|
|
my $expectedCounts = [ 0, 0, 0 ]; |
412
|
|
|
|
|
|
|
my $sentence_counts; |
413
|
|
|
|
|
|
|
|
414
|
|
|
|
|
|
|
foreach my $sentence ( @{ $self->EM_heldout_data } ) { |
415
|
|
|
|
|
|
|
$sentence_counts = $self->count_expected_counts_tree( |
416
|
|
|
|
|
|
|
$sentence->nodes_with_root->[0] |
417
|
|
|
|
|
|
|
); |
418
|
|
|
|
|
|
|
$expectedCounts->[0] += $sentence_counts->[0]; |
419
|
|
|
|
|
|
|
$expectedCounts->[1] += $sentence_counts->[1]; |
420
|
|
|
|
|
|
|
$expectedCounts->[2] += $sentence_counts->[2]; |
421
|
|
|
|
|
|
|
} |
422
|
|
|
|
|
|
|
|
423
|
|
|
|
|
|
|
return $expectedCounts; |
424
|
|
|
|
|
|
|
} |
425
|
|
|
|
|
|
|
|
426
|
|
|
|
|
|
|
#count "expected counts" of lambdas for a parse (sub)tree, recursively |
427
|
|
|
|
|
|
|
sub count_expected_counts_tree { |
428
|
|
|
|
|
|
|
my ( $self, $root_node ) = @_; |
429
|
|
|
|
|
|
|
|
430
|
|
|
|
|
|
|
my @edges = @{ $root_node->children }; |
431
|
|
|
|
|
|
|
|
432
|
|
|
|
|
|
|
# get sequence of labels |
433
|
|
|
|
|
|
|
my @labels = map { $_->child->label } @edges; |
434
|
|
|
|
|
|
|
|
435
|
|
|
|
|
|
|
# counts for this sequence |
436
|
|
|
|
|
|
|
my $expectedCounts = $self->count_expected_counts_sequence( \@labels ); |
437
|
|
|
|
|
|
|
|
438
|
|
|
|
|
|
|
# recursion |
439
|
|
|
|
|
|
|
my $subtree_counts; |
440
|
|
|
|
|
|
|
foreach my $edge (@edges) { |
441
|
|
|
|
|
|
|
$subtree_counts = $self->count_expected_counts_tree( $edge->child ); |
442
|
|
|
|
|
|
|
$expectedCounts->[0] += $subtree_counts->[0]; |
443
|
|
|
|
|
|
|
$expectedCounts->[1] += $subtree_counts->[1]; |
444
|
|
|
|
|
|
|
$expectedCounts->[2] += $subtree_counts->[2]; |
445
|
|
|
|
|
|
|
} |
446
|
|
|
|
|
|
|
|
447
|
|
|
|
|
|
|
return $expectedCounts; |
448
|
|
|
|
|
|
|
} |
449
|
|
|
|
|
|
|
|
450
|
|
|
|
|
|
|
# count "expected counts" of lambdas for a sequence of labels |
451
|
|
|
|
|
|
|
# (including the boundaries) |
452
|
|
|
|
|
|
|
sub count_expected_counts_sequence { |
453
|
|
|
|
|
|
|
|
454
|
|
|
|
|
|
|
my ( $self, $labels_sequence ) = @_; |
455
|
|
|
|
|
|
|
|
456
|
|
|
|
|
|
|
# to be computed here |
457
|
|
|
|
|
|
|
my $expectedCounts = [ 0, 0, 0 ]; |
458
|
|
|
|
|
|
|
|
459
|
|
|
|
|
|
|
# boundary at the beginning |
460
|
|
|
|
|
|
|
my $label_prev = $self->config->SEQUENCE_BOUNDARY_LABEL; |
461
|
|
|
|
|
|
|
|
462
|
|
|
|
|
|
|
# boundary at the end |
463
|
|
|
|
|
|
|
push @$labels_sequence, $self->config->SEQUENCE_BOUNDARY_LABEL; |
464
|
|
|
|
|
|
|
|
465
|
|
|
|
|
|
|
foreach my $label_this (@$labels_sequence) { |
466
|
|
|
|
|
|
|
|
467
|
|
|
|
|
|
|
# get probs |
468
|
|
|
|
|
|
|
my $ngramProbs = |
469
|
|
|
|
|
|
|
$self->get_transition_probs_array( $label_this, $label_prev ); |
470
|
|
|
|
|
|
|
my $finalProb = $ngramProbs->[0] * $self->smooth_uniform |
471
|
|
|
|
|
|
|
+ $ngramProbs->[1] * $self->smooth_unigrams |
472
|
|
|
|
|
|
|
+ $ngramProbs->[2] * $self->smooth_bigrams; |
473
|
|
|
|
|
|
|
|
474
|
|
|
|
|
|
|
# update expected counts |
475
|
|
|
|
|
|
|
$expectedCounts->[0] += |
476
|
|
|
|
|
|
|
$self->smooth_uniform * $ngramProbs->[0] / $finalProb; |
477
|
|
|
|
|
|
|
$expectedCounts->[1] += |
478
|
|
|
|
|
|
|
$self->smooth_unigrams * $ngramProbs->[1] / $finalProb; |
479
|
|
|
|
|
|
|
$expectedCounts->[2] += |
480
|
|
|
|
|
|
|
$self->smooth_bigrams * $ngramProbs->[2] / $finalProb; |
481
|
|
|
|
|
|
|
|
482
|
|
|
|
|
|
|
$label_prev = $label_this; |
483
|
|
|
|
|
|
|
} |
484
|
|
|
|
|
|
|
|
485
|
|
|
|
|
|
|
return $expectedCounts; |
486
|
|
|
|
|
|
|
} |
487
|
|
|
|
|
|
|
|
488
|
|
|
|
|
|
|
sub get_all_labels { |
489
|
|
|
|
|
|
|
my ($self) = @_; |
490
|
|
|
|
|
|
|
|
491
|
|
|
|
|
|
|
my @labels = keys %{ $self->unigrams }; |
492
|
|
|
|
|
|
|
|
493
|
|
|
|
|
|
|
return \@labels; |
494
|
|
|
|
|
|
|
} |
495
|
|
|
|
|
|
|
|
496
|
|
|
|
|
|
|
# ACCESS TO SCORES |
497
|
|
|
|
|
|
|
|
498
|
|
|
|
|
|
|
sub get_label_score { |
499
|
|
|
|
|
|
|
|
500
|
|
|
|
|
|
|
# (Str $label, Str $label_prev, ArrayRef[Str] $features) |
501
|
|
|
|
|
|
|
my ( $self, $label, $label_prev, $features ) = @_; |
502
|
|
|
|
|
|
|
|
503
|
|
|
|
|
|
|
my $ALGORITHM = $self->config->labeller_algorithm; |
504
|
|
|
|
|
|
|
|
505
|
|
|
|
|
|
|
if ( $ALGORITHM == 8 || $ALGORITHM == 9 ) { |
506
|
|
|
|
|
|
|
|
507
|
|
|
|
|
|
|
my $result = 0; |
508
|
|
|
|
|
|
|
|
509
|
|
|
|
|
|
|
# foreach present feature |
510
|
|
|
|
|
|
|
foreach my $feature (@$features) { |
511
|
|
|
|
|
|
|
|
512
|
|
|
|
|
|
|
# add "emission score" and "transition score" |
513
|
|
|
|
|
|
|
$result += |
514
|
|
|
|
|
|
|
$self->get_emission_score( $label, $feature ) |
515
|
|
|
|
|
|
|
+ |
516
|
|
|
|
|
|
|
$self->get_transition_score( |
517
|
|
|
|
|
|
|
$label, $label_prev, $feature |
518
|
|
|
|
|
|
|
) |
519
|
|
|
|
|
|
|
; |
520
|
|
|
|
|
|
|
} # end foreach $feature |
521
|
|
|
|
|
|
|
|
522
|
|
|
|
|
|
|
return $result; |
523
|
|
|
|
|
|
|
|
524
|
|
|
|
|
|
|
} elsif ( $ALGORITHM == 14 || $ALGORITHM == 15 ) { |
525
|
|
|
|
|
|
|
|
526
|
|
|
|
|
|
|
my $label_scores = $self->get_emission_scores($features); |
527
|
|
|
|
|
|
|
|
528
|
|
|
|
|
|
|
my $result = $label_scores->{$label}; |
529
|
|
|
|
|
|
|
if ( !defined $result ) { |
530
|
|
|
|
|
|
|
$result = 0; |
531
|
|
|
|
|
|
|
} |
532
|
|
|
|
|
|
|
|
533
|
|
|
|
|
|
|
# multiply by transitions score |
534
|
|
|
|
|
|
|
$result *= $self->get_transition_score( $label, $label_prev ); |
535
|
|
|
|
|
|
|
|
536
|
|
|
|
|
|
|
return $result; |
537
|
|
|
|
|
|
|
|
538
|
|
|
|
|
|
|
} elsif ( $ALGORITHM == 16 || $ALGORITHM == 18 ) { |
539
|
|
|
|
|
|
|
|
540
|
|
|
|
|
|
|
my $result = 0; |
541
|
|
|
|
|
|
|
|
542
|
|
|
|
|
|
|
# sum of emission scores |
543
|
|
|
|
|
|
|
foreach my $feature (@$features) { |
544
|
|
|
|
|
|
|
$result += $self->get_emission_score( $label, $feature ); |
545
|
|
|
|
|
|
|
} |
546
|
|
|
|
|
|
|
|
547
|
|
|
|
|
|
|
# multiply by transitions score |
548
|
|
|
|
|
|
|
$result *= $self->get_transition_score( $label, $label_prev ); |
549
|
|
|
|
|
|
|
|
550
|
|
|
|
|
|
|
return $result; |
551
|
|
|
|
|
|
|
|
552
|
|
|
|
|
|
|
} elsif ( $ALGORITHM == 19 ) { |
553
|
|
|
|
|
|
|
|
554
|
|
|
|
|
|
|
my $result = 0; |
555
|
|
|
|
|
|
|
|
556
|
|
|
|
|
|
|
# sum of emission scores |
557
|
|
|
|
|
|
|
foreach my $feature (@$features) { |
558
|
|
|
|
|
|
|
$result += $self->get_emission_score( $label, $feature ); |
559
|
|
|
|
|
|
|
} |
560
|
|
|
|
|
|
|
|
561
|
|
|
|
|
|
|
# sigmoid transformation |
562
|
|
|
|
|
|
|
$result = 1 / ( 1 + exp( -$result * $self->config->SIGM_LAMBDA ) ); |
563
|
|
|
|
|
|
|
|
564
|
|
|
|
|
|
|
# multiply by transitions score |
565
|
|
|
|
|
|
|
$result *= $self->get_transition_score( $label, $label_prev ); |
566
|
|
|
|
|
|
|
|
567
|
|
|
|
|
|
|
return $result; |
568
|
|
|
|
|
|
|
|
569
|
|
|
|
|
|
|
} elsif ( $ALGORITHM == 17 ) { |
570
|
|
|
|
|
|
|
|
571
|
|
|
|
|
|
|
my $result = 0; |
572
|
|
|
|
|
|
|
|
573
|
|
|
|
|
|
|
# sum of emission scores |
574
|
|
|
|
|
|
|
foreach my $feature (@$features) { |
575
|
|
|
|
|
|
|
$result += $self->get_emission_score( $label, $feature ); |
576
|
|
|
|
|
|
|
} |
577
|
|
|
|
|
|
|
|
578
|
|
|
|
|
|
|
# multiply by transitions score |
579
|
|
|
|
|
|
|
if ( $result > 0 ) { |
580
|
|
|
|
|
|
|
$result *= $self->get_transition_score( $label, $label_prev ); |
581
|
|
|
|
|
|
|
} else { |
582
|
|
|
|
|
|
|
|
583
|
|
|
|
|
|
|
# For negative scores this works the other way round, |
584
|
|
|
|
|
|
|
# eg. if I had two labels, both with emission score -5 |
585
|
|
|
|
|
|
|
# and their transition probs were 0.2 and 0.9, |
586
|
|
|
|
|
|
|
# then the latter should get a higher score; |
587
|
|
|
|
|
|
|
# simple mltiplication won't help as that would yield scores |
588
|
|
|
|
|
|
|
# of -1.0 and -4.5, thus inverting the order. |
589
|
|
|
|
|
|
|
# What I do is that for transition prob p I use (1-p) |
590
|
|
|
|
|
|
|
# which yields 0.8 and 0.1 transition probabilities here, |
591
|
|
|
|
|
|
|
# giving scores of -4.0 and -0.5, which is much better. |
592
|
|
|
|
|
|
|
# Still, a label with negative emission score, even if very close |
593
|
|
|
|
|
|
|
# to 0 and with a high transition prob, cannot outscore any label |
594
|
|
|
|
|
|
|
# with a positive emission score, even if low with a low transition |
595
|
|
|
|
|
|
|
# prob - normalizing scores to be non-negative would be necessary |
596
|
|
|
|
|
|
|
# for this, as is alg 0 and similar. |
597
|
|
|
|
|
|
|
# $result *= |
598
|
|
|
|
|
|
|
# ( 1 - $self->get_transition_score( $label, $label_prev ) ); |
599
|
|
|
|
|
|
|
|
600
|
|
|
|
|
|
|
# TODO trying new variant - setting negative scores to 0 |
601
|
|
|
|
|
|
|
$result = 0; |
602
|
|
|
|
|
|
|
} |
603
|
|
|
|
|
|
|
|
604
|
|
|
|
|
|
|
return $result; |
605
|
|
|
|
|
|
|
|
606
|
|
|
|
|
|
|
} elsif ( $ALGORITHM >= 20 ) { |
607
|
|
|
|
|
|
|
|
608
|
|
|
|
|
|
|
my $result = 0; |
609
|
|
|
|
|
|
|
|
610
|
|
|
|
|
|
|
# sum of emission scores |
611
|
|
|
|
|
|
|
foreach my $feature (@$features) { |
612
|
|
|
|
|
|
|
$result += $self->get_emission_score( $label, $feature ); |
613
|
|
|
|
|
|
|
} |
614
|
|
|
|
|
|
|
|
615
|
|
|
|
|
|
|
# TODO: could also compute using $label_prev, |
616
|
|
|
|
|
|
|
# using transitions to store these; |
617
|
|
|
|
|
|
|
# would allow to use full Viterbi |
618
|
|
|
|
|
|
|
|
619
|
|
|
|
|
|
|
return $result; |
620
|
|
|
|
|
|
|
|
621
|
|
|
|
|
|
|
} else { |
622
|
|
|
|
|
|
|
croak "ModelLabelling->get_label_score not implemented" |
623
|
|
|
|
|
|
|
. " for algorithm no. $ALGORITHM!"; |
624
|
|
|
|
|
|
|
|
625
|
|
|
|
|
|
|
# usually because it needs to know scores of all possible labels |
626
|
|
|
|
|
|
|
# to normalize them properly |
627
|
|
|
|
|
|
|
} |
628
|
|
|
|
|
|
|
} |
629
|
|
|
|
|
|
|
|
630
|
|
|
|
|
|
|
sub get_emission_score { |
631
|
|
|
|
|
|
|
|
632
|
|
|
|
|
|
|
# (Str $label, Str $feature) |
633
|
|
|
|
|
|
|
my ( $self, $label, $feature ) = @_; |
634
|
|
|
|
|
|
|
|
635
|
|
|
|
|
|
|
my $ALGORITHM = $self->config->labeller_algorithm; |
636
|
|
|
|
|
|
|
|
637
|
|
|
|
|
|
|
if ($ALGORITHM == 8 |
638
|
|
|
|
|
|
|
|| $ALGORITHM == 9 |
639
|
|
|
|
|
|
|
|| $ALGORITHM == 16 |
640
|
|
|
|
|
|
|
|| $ALGORITHM == 17 |
641
|
|
|
|
|
|
|
|| $ALGORITHM == 18 |
642
|
|
|
|
|
|
|
|| $ALGORITHM == 19 |
643
|
|
|
|
|
|
|
|| $ALGORITHM >= 20 |
644
|
|
|
|
|
|
|
) |
645
|
|
|
|
|
|
|
{ |
646
|
|
|
|
|
|
|
|
647
|
|
|
|
|
|
|
if ($self->emissions->{$feature} |
648
|
|
|
|
|
|
|
&& $self->emissions->{$feature}->{$label} |
649
|
|
|
|
|
|
|
) |
650
|
|
|
|
|
|
|
{ |
651
|
|
|
|
|
|
|
return $self->emissions->{$feature}->{$label}; |
652
|
|
|
|
|
|
|
} else { |
653
|
|
|
|
|
|
|
return 0; |
654
|
|
|
|
|
|
|
} |
655
|
|
|
|
|
|
|
|
656
|
|
|
|
|
|
|
} else { |
657
|
|
|
|
|
|
|
croak "ModelLabelling->get_emission_score not implemented" |
658
|
|
|
|
|
|
|
. " for algorithm no. $ALGORITHM!"; |
659
|
|
|
|
|
|
|
} |
660
|
|
|
|
|
|
|
} |
661
|
|
|
|
|
|
|
|
662
|
|
|
|
|
|
|
sub get_transition_score { |
663
|
|
|
|
|
|
|
|
664
|
|
|
|
|
|
|
# (Str $label_this, Str $label_prev, Maybe[Str] $feature) |
665
|
|
|
|
|
|
|
my ( $self, $label_this, $label_prev, $feature ) = @_; |
666
|
|
|
|
|
|
|
|
667
|
|
|
|
|
|
|
my $ALGORITHM = $self->config->labeller_algorithm; |
668
|
|
|
|
|
|
|
|
669
|
|
|
|
|
|
|
if ( $ALGORITHM == 8 || $ALGORITHM == 9 ) { |
670
|
|
|
|
|
|
|
if ($self->transitions->{$feature} |
671
|
|
|
|
|
|
|
&& $self->transitions->{$feature}->{$label_prev} |
672
|
|
|
|
|
|
|
&& $self->transitions->{$feature}->{$label_prev}->{$label_this} |
673
|
|
|
|
|
|
|
) |
674
|
|
|
|
|
|
|
{ |
675
|
|
|
|
|
|
|
return $self->transitions->{$feature}->{$label_prev}->{$label_this}; |
676
|
|
|
|
|
|
|
} else { |
677
|
|
|
|
|
|
|
|
678
|
|
|
|
|
|
|
# no smoothing as it is used in addition, not in multiplication |
679
|
|
|
|
|
|
|
return 0; |
680
|
|
|
|
|
|
|
} |
681
|
|
|
|
|
|
|
} elsif ( |
682
|
|
|
|
|
|
|
$ALGORITHM == 5 |
683
|
|
|
|
|
|
|
|| $ALGORITHM == 12 || $ALGORITHM == 13 |
684
|
|
|
|
|
|
|
|| $ALGORITHM == 15 |
685
|
|
|
|
|
|
|
|| $ALGORITHM == 16 || $ALGORITHM == 17 |
686
|
|
|
|
|
|
|
|| $ALGORITHM == 18 || $ALGORITHM == 19 |
687
|
|
|
|
|
|
|
) |
688
|
|
|
|
|
|
|
{ |
689
|
|
|
|
|
|
|
|
690
|
|
|
|
|
|
|
# smoothing by linear combination |
691
|
|
|
|
|
|
|
# PROB(label|prev_label) = |
692
|
|
|
|
|
|
|
# smooth_bigrams * transitions->{prev_label}->{label} + |
693
|
|
|
|
|
|
|
# smooth_unigrams * unigrams->{label} + |
694
|
|
|
|
|
|
|
# smooth_uniform |
695
|
|
|
|
|
|
|
|
696
|
|
|
|
|
|
|
my $probs = |
697
|
|
|
|
|
|
|
$self->get_transition_probs_array( $label_this, $label_prev ); |
698
|
|
|
|
|
|
|
|
699
|
|
|
|
|
|
|
my $result = $probs->[0] * $self->smooth_uniform |
700
|
|
|
|
|
|
|
+ $probs->[1] * $self->smooth_unigrams |
701
|
|
|
|
|
|
|
+ $probs->[2] * $self->smooth_bigrams; |
702
|
|
|
|
|
|
|
|
703
|
|
|
|
|
|
|
return $result; |
704
|
|
|
|
|
|
|
|
705
|
|
|
|
|
|
|
} elsif ( |
706
|
|
|
|
|
|
|
$ALGORITHM == 0 |
707
|
|
|
|
|
|
|
|| $ALGORITHM == 1 |
708
|
|
|
|
|
|
|
|| $ALGORITHM == 2 |
709
|
|
|
|
|
|
|
|| $ALGORITHM == 3 |
710
|
|
|
|
|
|
|
|| $ALGORITHM == 4 |
711
|
|
|
|
|
|
|
|| $ALGORITHM == 10 |
712
|
|
|
|
|
|
|
|| $ALGORITHM == 11 |
713
|
|
|
|
|
|
|
|| $ALGORITHM == 14 |
714
|
|
|
|
|
|
|
) |
715
|
|
|
|
|
|
|
{ |
716
|
|
|
|
|
|
|
|
717
|
|
|
|
|
|
|
# no real smoothing |
718
|
|
|
|
|
|
|
if ($self->transitions->{$label_prev} |
719
|
|
|
|
|
|
|
&& $self->transitions->{$label_prev}->{$label_this} |
720
|
|
|
|
|
|
|
) |
721
|
|
|
|
|
|
|
{ |
722
|
|
|
|
|
|
|
return $self->transitions->{$label_prev}->{$label_this}; |
723
|
|
|
|
|
|
|
} else { |
724
|
|
|
|
|
|
|
return 0.00001; |
725
|
|
|
|
|
|
|
} |
726
|
|
|
|
|
|
|
} else { |
727
|
|
|
|
|
|
|
croak "ModelLabelling->get_transition_score not implemented" |
728
|
|
|
|
|
|
|
. " for algorithm no. $ALGORITHM!"; |
729
|
|
|
|
|
|
|
} |
730
|
|
|
|
|
|
|
} # end get_transition_score |
731
|
|
|
|
|
|
|
|
732
|
|
|
|
|
|
|
# $result->[0] = uniform prob |
733
|
|
|
|
|
|
|
# $result->[1] = unigram prob |
734
|
|
|
|
|
|
|
# $result->[2] = bigram prob |
735
|
|
|
|
|
|
|
sub get_transition_probs_array { |
736
|
|
|
|
|
|
|
|
737
|
|
|
|
|
|
|
# (Str $label_this, Str $label_prev) |
738
|
|
|
|
|
|
|
my ( $self, $label_this, $label_prev ) = @_; |
739
|
|
|
|
|
|
|
|
740
|
|
|
|
|
|
|
my $result = [ 0, 0, 0 ]; |
741
|
|
|
|
|
|
|
|
742
|
|
|
|
|
|
|
# uniform |
743
|
|
|
|
|
|
|
$result->[0] = $self->uniform_prob; |
744
|
|
|
|
|
|
|
|
745
|
|
|
|
|
|
|
if ( $self->unigrams->{$label_this} ) { |
746
|
|
|
|
|
|
|
|
747
|
|
|
|
|
|
|
# unigram |
748
|
|
|
|
|
|
|
$result->[1] = $self->unigrams->{$label_this}; |
749
|
|
|
|
|
|
|
|
750
|
|
|
|
|
|
|
if ( $self->transitions->{$label_prev}->{$label_this} ) { |
751
|
|
|
|
|
|
|
|
752
|
|
|
|
|
|
|
# bigram |
753
|
|
|
|
|
|
|
$result->[2] = $self->transitions->{$label_prev}->{$label_this}; |
754
|
|
|
|
|
|
|
} |
755
|
|
|
|
|
|
|
} |
756
|
|
|
|
|
|
|
|
757
|
|
|
|
|
|
|
return $result; |
758
|
|
|
|
|
|
|
} |
759
|
|
|
|
|
|
|
|
760
|
|
|
|
|
|
|
# get scores of all possible labels based on all the features |
761
|
|
|
|
|
|
|
# (gives different numbers for different algorithms, |
762
|
|
|
|
|
|
|
# often they are not real probabilities but general scores) |
763
|
|
|
|
|
|
|
sub get_emission_scores { |
764
|
|
|
|
|
|
|
|
765
|
|
|
|
|
|
|
# (ArrayRef[Str] $features) |
766
|
|
|
|
|
|
|
my ( $self, $features ) = @_; |
767
|
|
|
|
|
|
|
|
768
|
|
|
|
|
|
|
# a hashref of the structure $result->{label} = prob |
769
|
|
|
|
|
|
|
# where prob might or might not be a real probability |
770
|
|
|
|
|
|
|
# (i.e. may or may not fulfill 0 <= prob <= 1 & sum(probs) == 1), |
771
|
|
|
|
|
|
|
# depending on the algorithm used |
772
|
|
|
|
|
|
|
# (but always a higher prob means a better scoring (more probable) label |
773
|
|
|
|
|
|
|
# and all of the probs are non-negative) TODO does it hold? |
774
|
|
|
|
|
|
|
my $result = {}; |
775
|
|
|
|
|
|
|
|
776
|
|
|
|
|
|
|
my $ALGORITHM = $self->config->labeller_algorithm; |
777
|
|
|
|
|
|
|
|
778
|
|
|
|
|
|
|
if ($ALGORITHM == 0 |
779
|
|
|
|
|
|
|
|| $ALGORITHM == 1 |
780
|
|
|
|
|
|
|
|| $ALGORITHM == 2 |
781
|
|
|
|
|
|
|
|| $ALGORITHM == 3 |
782
|
|
|
|
|
|
|
|| $ALGORITHM == 10 |
783
|
|
|
|
|
|
|
|| $ALGORITHM == 11 |
784
|
|
|
|
|
|
|
|| $ALGORITHM == 12 |
785
|
|
|
|
|
|
|
|| $ALGORITHM == 13 |
786
|
|
|
|
|
|
|
|| $ALGORITHM == 14 |
787
|
|
|
|
|
|
|
|| $ALGORITHM == 15 |
788
|
|
|
|
|
|
|
) |
789
|
|
|
|
|
|
|
{ |
790
|
|
|
|
|
|
|
$result = $self->get_emission_scores_basic_MIRA($features); |
791
|
|
|
|
|
|
|
} elsif ( $ALGORITHM == 4 || $ALGORITHM == 5 ) { |
792
|
|
|
|
|
|
|
$result = $self->get_emission_scores_no_MIRA($features); |
793
|
|
|
|
|
|
|
} else { |
794
|
|
|
|
|
|
|
croak "ModelLabelling->get_emission_scores not implemented" |
795
|
|
|
|
|
|
|
. " for algorithm no. $ALGORITHM!"; |
796
|
|
|
|
|
|
|
} |
797
|
|
|
|
|
|
|
|
798
|
|
|
|
|
|
|
# the boundary label is NOT a valid label |
799
|
|
|
|
|
|
|
delete $result->{ $self->config->SEQUENCE_BOUNDARY_LABEL }; |
800
|
|
|
|
|
|
|
|
801
|
|
|
|
|
|
|
return $result; |
802
|
|
|
|
|
|
|
} |
803
|
|
|
|
|
|
|
|
804
|
|
|
|
|
|
|
sub get_emission_scores_basic_MIRA { |
805
|
|
|
|
|
|
|
|
806
|
|
|
|
|
|
|
my ( $self, $features ) = @_; |
807
|
|
|
|
|
|
|
|
808
|
|
|
|
|
|
|
my $ALGORITHM = $self->config->labeller_algorithm; |
809
|
|
|
|
|
|
|
|
810
|
|
|
|
|
|
|
my $result = {}; |
811
|
|
|
|
|
|
|
|
812
|
|
|
|
|
|
|
my $warnNoEmissionProbs = "!!! WARNING !!! " |
813
|
|
|
|
|
|
|
. "Based on the training data, no possible label was found" |
814
|
|
|
|
|
|
|
. " for an edge. This usually means that either" |
815
|
|
|
|
|
|
|
. " your training data are not big enough or that" |
816
|
|
|
|
|
|
|
. " the set of features you are using" |
817
|
|
|
|
|
|
|
. " is not well constructed - either it is too small" |
818
|
|
|
|
|
|
|
. " or it lacks features that would be general enough" |
819
|
|
|
|
|
|
|
. " to cover all possible sentences." |
820
|
|
|
|
|
|
|
. " Using blind emission probabilities instead.\n"; |
821
|
|
|
|
|
|
|
|
822
|
|
|
|
|
|
|
# "pure MIRA", i.e. no MLE |
823
|
|
|
|
|
|
|
|
824
|
|
|
|
|
|
|
if ( $ALGORITHM == 11 || $ALGORITHM == 13 ) { |
825
|
|
|
|
|
|
|
|
826
|
|
|
|
|
|
|
# initialize all label scores with 0 (so that all labels get some score) |
827
|
|
|
|
|
|
|
my $all_labels = $self->get_all_labels(); |
828
|
|
|
|
|
|
|
foreach my $label (@$all_labels) { |
829
|
|
|
|
|
|
|
$result->{$label} = 0; |
830
|
|
|
|
|
|
|
} |
831
|
|
|
|
|
|
|
} |
832
|
|
|
|
|
|
|
|
833
|
|
|
|
|
|
|
# get scores |
834
|
|
|
|
|
|
|
foreach my $feature (@$features) { |
835
|
|
|
|
|
|
|
if ( $self->emissions->{$feature} ) { |
836
|
|
|
|
|
|
|
foreach my $label ( keys %{ $self->emissions->{$feature} } ) { |
837
|
|
|
|
|
|
|
$result->{$label} += $self->emissions->{$feature}->{$label}; |
838
|
|
|
|
|
|
|
} |
839
|
|
|
|
|
|
|
} |
840
|
|
|
|
|
|
|
} |
841
|
|
|
|
|
|
|
|
842
|
|
|
|
|
|
|
# subtracting the minimum from the score |
843
|
|
|
|
|
|
|
if ($ALGORITHM == 0 |
844
|
|
|
|
|
|
|
|| $ALGORITHM == 1 |
845
|
|
|
|
|
|
|
|| $ALGORITHM == 2 |
846
|
|
|
|
|
|
|
|| $ALGORITHM == 10 |
847
|
|
|
|
|
|
|
|| $ALGORITHM == 11 |
848
|
|
|
|
|
|
|
|| $ALGORITHM == 12 |
849
|
|
|
|
|
|
|
|| $ALGORITHM == 13 |
850
|
|
|
|
|
|
|
|| $ALGORITHM == 14 |
851
|
|
|
|
|
|
|
|| $ALGORITHM == 15 |
852
|
|
|
|
|
|
|
) |
853
|
|
|
|
|
|
|
{ |
854
|
|
|
|
|
|
|
|
855
|
|
|
|
|
|
|
# find min and max score |
856
|
|
|
|
|
|
|
my $min = 1e300; |
857
|
|
|
|
|
|
|
my $max = -1e300; |
858
|
|
|
|
|
|
|
foreach my $label ( keys %$result ) { |
859
|
|
|
|
|
|
|
if ( $result->{$label} < $min ) { |
860
|
|
|
|
|
|
|
$min = $result->{$label}; |
861
|
|
|
|
|
|
|
} |
862
|
|
|
|
|
|
|
if ( $result->{$label} > $max ) { |
863
|
|
|
|
|
|
|
$max = $result->{$label}; |
864
|
|
|
|
|
|
|
} |
865
|
|
|
|
|
|
|
|
866
|
|
|
|
|
|
|
# else is between $min and $max -> keep the values as they are |
867
|
|
|
|
|
|
|
} |
868
|
|
|
|
|
|
|
|
869
|
|
|
|
|
|
|
if ( $min > $max ) { |
870
|
|
|
|
|
|
|
|
871
|
|
|
|
|
|
|
# $min > $max, i.e. nothing has been generated -> backoff |
872
|
|
|
|
|
|
|
if ( $self->config->DEBUG >= 2 ) { |
873
|
|
|
|
|
|
|
print $warnNoEmissionProbs; |
874
|
|
|
|
|
|
|
} |
875
|
|
|
|
|
|
|
|
876
|
|
|
|
|
|
|
# backoff by using unigram probabilities |
877
|
|
|
|
|
|
|
# (or unigram counts in some algorithms) |
878
|
|
|
|
|
|
|
$result = $self->unigrams; |
879
|
|
|
|
|
|
|
} else { |
880
|
|
|
|
|
|
|
|
881
|
|
|
|
|
|
|
# something has been generated, now 0 and 1 start to differ |
882
|
|
|
|
|
|
|
if ($ALGORITHM == 0 |
883
|
|
|
|
|
|
|
|| $ALGORITHM == 10 |
884
|
|
|
|
|
|
|
|| $ALGORITHM == 11 |
885
|
|
|
|
|
|
|
|| $ALGORITHM == 12 |
886
|
|
|
|
|
|
|
|| $ALGORITHM == 13 |
887
|
|
|
|
|
|
|
|| $ALGORITHM == 14 |
888
|
|
|
|
|
|
|
|| $ALGORITHM == 15 |
889
|
|
|
|
|
|
|
) |
890
|
|
|
|
|
|
|
{ |
891
|
|
|
|
|
|
|
|
892
|
|
|
|
|
|
|
# 0 MIRA-trained scores recomputed by +abs(min) |
893
|
|
|
|
|
|
|
# and converted to probs |
894
|
|
|
|
|
|
|
if ( $min < $max ) { |
895
|
|
|
|
|
|
|
|
896
|
|
|
|
|
|
|
# the typical case |
897
|
|
|
|
|
|
|
# my $subtractant = $min; |
898
|
|
|
|
|
|
|
my $divisor = 0; |
899
|
|
|
|
|
|
|
|
900
|
|
|
|
|
|
|
foreach my $label ( keys %$result ) { |
901
|
|
|
|
|
|
|
$result->{$label} = ( $result->{$label} - $min ); |
902
|
|
|
|
|
|
|
$divisor += $result->{$label}; |
903
|
|
|
|
|
|
|
} |
904
|
|
|
|
|
|
|
foreach my $label ( keys %$result ) { |
905
|
|
|
|
|
|
|
$result->{$label} = $result->{$label} / $divisor; |
906
|
|
|
|
|
|
|
} |
907
|
|
|
|
|
|
|
} else { |
908
|
|
|
|
|
|
|
|
909
|
|
|
|
|
|
|
# $min == $max |
910
|
|
|
|
|
|
|
|
911
|
|
|
|
|
|
|
# uniform prob distribution |
912
|
|
|
|
|
|
|
my $prob = 1 / scalar( keys %$result ); |
913
|
|
|
|
|
|
|
foreach my $label ( keys %$result ) { |
914
|
|
|
|
|
|
|
$result->{$label} = $prob; |
915
|
|
|
|
|
|
|
} |
916
|
|
|
|
|
|
|
} |
917
|
|
|
|
|
|
|
|
918
|
|
|
|
|
|
|
# end $ALGORITHM == 0|10|11|12|13|14|15 |
919
|
|
|
|
|
|
|
} else { |
920
|
|
|
|
|
|
|
|
921
|
|
|
|
|
|
|
# $ALGORITHM == 1|2 |
922
|
|
|
|
|
|
|
# 1 dtto, NOT converted to probs |
923
|
|
|
|
|
|
|
# (but should behave the same as 0) |
924
|
|
|
|
|
|
|
# 2 dtto, sum in Viterbi instead of product |
925
|
|
|
|
|
|
|
# (new_prob = old_prob + emiss*trans) |
926
|
|
|
|
|
|
|
# (for 1 and 2 the emission probs are completely the same, |
927
|
|
|
|
|
|
|
# they are just handled differently by the Labeller) |
928
|
|
|
|
|
|
|
|
929
|
|
|
|
|
|
|
if ( $min < $max ) { |
930
|
|
|
|
|
|
|
|
931
|
|
|
|
|
|
|
# the typical case |
932
|
|
|
|
|
|
|
# my $subtractant = $min; |
933
|
|
|
|
|
|
|
|
934
|
|
|
|
|
|
|
foreach my $label ( keys %$result ) { |
935
|
|
|
|
|
|
|
$result->{$label} = ( $result->{$label} - $min ); |
936
|
|
|
|
|
|
|
} |
937
|
|
|
|
|
|
|
} else { |
938
|
|
|
|
|
|
|
|
939
|
|
|
|
|
|
|
# $min == $max |
940
|
|
|
|
|
|
|
# uniform prob distribution |
941
|
|
|
|
|
|
|
|
942
|
|
|
|
|
|
|
if ( $min <= 0 ) { |
943
|
|
|
|
|
|
|
|
944
|
|
|
|
|
|
|
# we would like to keep the values |
945
|
|
|
|
|
|
|
# but this is not possible in this case |
946
|
|
|
|
|
|
|
foreach my $label ( keys %$result ) { |
947
|
|
|
|
|
|
|
|
948
|
|
|
|
|
|
|
# so lets just assign ones |
949
|
|
|
|
|
|
|
$result->{$label} = 1; |
950
|
|
|
|
|
|
|
} |
951
|
|
|
|
|
|
|
} |
952
|
|
|
|
|
|
|
|
953
|
|
|
|
|
|
|
# else there is already a uniform distribution |
954
|
|
|
|
|
|
|
# so let's keep it as it is |
955
|
|
|
|
|
|
|
} |
956
|
|
|
|
|
|
|
|
957
|
|
|
|
|
|
|
# end $ALGORITHM == 1|2 |
958
|
|
|
|
|
|
|
} |
959
|
|
|
|
|
|
|
} |
960
|
|
|
|
|
|
|
|
961
|
|
|
|
|
|
|
# end $ALGORITHM == 0|1|2|10|11|12|13|14|15 |
962
|
|
|
|
|
|
|
} else { |
963
|
|
|
|
|
|
|
|
964
|
|
|
|
|
|
|
# $ALGORITHM == 3 |
965
|
|
|
|
|
|
|
# no subtraction of minimum, just throw away <= 0 |
966
|
|
|
|
|
|
|
|
967
|
|
|
|
|
|
|
foreach my $label ( keys %$result ) { |
968
|
|
|
|
|
|
|
if ( $result->{$label} <= 0 ) { |
969
|
|
|
|
|
|
|
delete $result->{$label}; |
970
|
|
|
|
|
|
|
} |
971
|
|
|
|
|
|
|
|
972
|
|
|
|
|
|
|
# else > 0 -> just keep it there and that's it |
973
|
|
|
|
|
|
|
} |
974
|
|
|
|
|
|
|
} # end $ALGORITHM == 3 |
975
|
|
|
|
|
|
|
|
976
|
|
|
|
|
|
|
return $result; |
977
|
|
|
|
|
|
|
} # end get_emission_scores_basic_MIRA |
978
|
|
|
|
|
|
|
|
979
|
|
|
|
|
|
|
sub get_emission_scores_no_MIRA { |
980
|
|
|
|
|
|
|
|
981
|
|
|
|
|
|
|
my ( $self, $features ) = @_; |
982
|
|
|
|
|
|
|
|
983
|
|
|
|
|
|
|
my $result = {}; |
984
|
|
|
|
|
|
|
|
985
|
|
|
|
|
|
|
my $warnNoEmissionProbs = "!!! WARNING !!! " |
986
|
|
|
|
|
|
|
. "Based on the training data, no possible label was found" |
987
|
|
|
|
|
|
|
. " for an edge. This usually means that either" |
988
|
|
|
|
|
|
|
. " your training data are not big enough or that" |
989
|
|
|
|
|
|
|
. " the set of features you are using" |
990
|
|
|
|
|
|
|
. " is not well constructed - either it is too small" |
991
|
|
|
|
|
|
|
. " or it lacks features that would be general enough" |
992
|
|
|
|
|
|
|
. " to cover all possible sentences." |
993
|
|
|
|
|
|
|
. " Using blind emission probabilities instead.\n"; |
994
|
|
|
|
|
|
|
|
995
|
|
|
|
|
|
|
# basic or full MLE, no MIRA |
996
|
|
|
|
|
|
|
|
997
|
|
|
|
|
|
|
my %counts = (); |
998
|
|
|
|
|
|
|
my %prob_sums = (); |
999
|
|
|
|
|
|
|
|
1000
|
|
|
|
|
|
|
# get scores |
1001
|
|
|
|
|
|
|
foreach my $feature (@$features) { |
1002
|
|
|
|
|
|
|
if ( $self->emissions->{$feature} ) { |
1003
|
|
|
|
|
|
|
|
1004
|
|
|
|
|
|
|
# !!! TODO tady by mÄl bejt souÄin !!! |
1005
|
|
|
|
|
|
|
foreach my $label ( keys %{ $self->emissions->{$feature} } ) { |
1006
|
|
|
|
|
|
|
$prob_sums{$label} += |
1007
|
|
|
|
|
|
|
$self->emissions->{$feature}->{$label}; |
1008
|
|
|
|
|
|
|
$counts{$label}++; |
1009
|
|
|
|
|
|
|
} |
1010
|
|
|
|
|
|
|
} |
1011
|
|
|
|
|
|
|
} |
1012
|
|
|
|
|
|
|
|
1013
|
|
|
|
|
|
|
if ( keys %prob_sums ) { |
1014
|
|
|
|
|
|
|
foreach my $label ( keys %prob_sums ) { |
1015
|
|
|
|
|
|
|
|
1016
|
|
|
|
|
|
|
# something like average pobability |
1017
|
|
|
|
|
|
|
# = all features have the score of 1 |
1018
|
|
|
|
|
|
|
# (or more precisely 1/number_of_features) |
1019
|
|
|
|
|
|
|
$result->{$label} = $prob_sums{$label} / $counts{$label}; |
1020
|
|
|
|
|
|
|
} |
1021
|
|
|
|
|
|
|
} else { |
1022
|
|
|
|
|
|
|
|
1023
|
|
|
|
|
|
|
# backoff |
1024
|
|
|
|
|
|
|
if ( $self->config->DEBUG >= 2 ) { |
1025
|
|
|
|
|
|
|
print $warnNoEmissionProbs; |
1026
|
|
|
|
|
|
|
} |
1027
|
|
|
|
|
|
|
|
1028
|
|
|
|
|
|
|
# backoff by using unigram probabilities |
1029
|
|
|
|
|
|
|
# (or unigram counts in some algorithms) |
1030
|
|
|
|
|
|
|
$result = $self->unigrams; |
1031
|
|
|
|
|
|
|
} |
1032
|
|
|
|
|
|
|
|
1033
|
|
|
|
|
|
|
return $result; |
1034
|
|
|
|
|
|
|
} # end get_emission_scores_no_MIRA |
1035
|
|
|
|
|
|
|
|
1036
|
|
|
|
|
|
|
# sets emission score (if $label_prev is not set) |
1037
|
|
|
|
|
|
|
# or transition score (if it is) |
1038
|
|
|
|
|
|
|
# of the $feature to $score |
1039
|
|
|
|
|
|
|
sub set_feature_score { |
1040
|
|
|
|
|
|
|
|
1041
|
|
|
|
|
|
|
# (Str $feature, Num $score, Str $label, Maybe[Str] $label_prev) |
1042
|
|
|
|
|
|
|
my ( $self, $feature, $score, $label, $label_prev ) = @_; |
1043
|
|
|
|
|
|
|
|
1044
|
|
|
|
|
|
|
if ( defined $label_prev ) { |
1045
|
|
|
|
|
|
|
$self->transitions->{$feature}->{$label_prev}->{$label} = $score; |
1046
|
|
|
|
|
|
|
} else { |
1047
|
|
|
|
|
|
|
$self->emissions->{$feature}->{$label} = $score; |
1048
|
|
|
|
|
|
|
} |
1049
|
|
|
|
|
|
|
|
1050
|
|
|
|
|
|
|
return; |
1051
|
|
|
|
|
|
|
} |
1052
|
|
|
|
|
|
|
|
1053
|
|
|
|
|
|
|
# updates emission score (if $label_prev is not set) |
1054
|
|
|
|
|
|
|
# or transition score (if it is) |
1055
|
|
|
|
|
|
|
# of the $feature by adding $update |
1056
|
|
|
|
|
|
|
sub update_feature_score { |
1057
|
|
|
|
|
|
|
|
1058
|
|
|
|
|
|
|
# (Str $feature, Num $update, Str $label, Maybe[Str] $label_prev) |
1059
|
|
|
|
|
|
|
my ( $self, $feature, $update, $label, $label_prev ) = @_; |
1060
|
|
|
|
|
|
|
|
1061
|
|
|
|
|
|
|
if ( defined $label_prev ) { |
1062
|
|
|
|
|
|
|
$self->transitions->{$feature}->{$label_prev}->{$label} += $update; |
1063
|
|
|
|
|
|
|
} else { |
1064
|
|
|
|
|
|
|
$self->emissions->{$feature}->{$label} += $update; |
1065
|
|
|
|
|
|
|
} |
1066
|
|
|
|
|
|
|
|
1067
|
|
|
|
|
|
|
return; |
1068
|
|
|
|
|
|
|
} |
1069
|
|
|
|
|
|
|
|
1070
|
|
|
|
|
|
|
# returns number of features in the model (where a "feature" can stand for |
1071
|
|
|
|
|
|
|
# various things depending on the algorithm used) |
1072
|
|
|
|
|
|
|
sub get_feature_count { |
1073
|
|
|
|
|
|
|
|
1074
|
|
|
|
|
|
|
my ($self) = @_; |
1075
|
|
|
|
|
|
|
|
1076
|
|
|
|
|
|
|
my $ALGORITHM = $self->config->labeller_algorithm; |
1077
|
|
|
|
|
|
|
|
1078
|
|
|
|
|
|
|
# result = $emissions_count + $transitions_count |
1079
|
|
|
|
|
|
|
my $emissions_count = 0; |
1080
|
|
|
|
|
|
|
my $transitions_count = 0; |
1081
|
|
|
|
|
|
|
|
1082
|
|
|
|
|
|
|
# structure: emissions->{feature}->{label} |
1083
|
|
|
|
|
|
|
my @emission_features = keys %{ $self->emissions }; |
1084
|
|
|
|
|
|
|
foreach my $feature (@emission_features) { |
1085
|
|
|
|
|
|
|
$emissions_count += scalar( keys %{ $self->emissions->{$feature} } ); |
1086
|
|
|
|
|
|
|
} |
1087
|
|
|
|
|
|
|
|
1088
|
|
|
|
|
|
|
if ( $ALGORITHM == 8 || $ALGORITHM == 9 ) { |
1089
|
|
|
|
|
|
|
|
1090
|
|
|
|
|
|
|
# structure: transitions->{feature}->{label_prev}->{label} |
1091
|
|
|
|
|
|
|
|
1092
|
|
|
|
|
|
|
my @transition_features = keys %{ $self->transitions }; |
1093
|
|
|
|
|
|
|
foreach my $feature (@transition_features) { |
1094
|
|
|
|
|
|
|
|
1095
|
|
|
|
|
|
|
my @labels = keys %{ $self->transitions->{$feature} }; |
1096
|
|
|
|
|
|
|
foreach my $label_prev (@labels) { |
1097
|
|
|
|
|
|
|
|
1098
|
|
|
|
|
|
|
$transitions_count += scalar( |
1099
|
|
|
|
|
|
|
keys %{ $self->transitions->{$feature}->{$label_prev} } |
1100
|
|
|
|
|
|
|
); |
1101
|
|
|
|
|
|
|
} |
1102
|
|
|
|
|
|
|
} |
1103
|
|
|
|
|
|
|
|
1104
|
|
|
|
|
|
|
} else { |
1105
|
|
|
|
|
|
|
|
1106
|
|
|
|
|
|
|
# structure: transitions->{label_prev}->{label} |
1107
|
|
|
|
|
|
|
|
1108
|
|
|
|
|
|
|
my @labels = keys %{ $self->transitions }; |
1109
|
|
|
|
|
|
|
foreach my $label_prev (@labels) { |
1110
|
|
|
|
|
|
|
|
1111
|
|
|
|
|
|
|
$transitions_count += |
1112
|
|
|
|
|
|
|
scalar( keys %{ $self->transitions->{$label_prev} } ); |
1113
|
|
|
|
|
|
|
} |
1114
|
|
|
|
|
|
|
} |
1115
|
|
|
|
|
|
|
|
1116
|
|
|
|
|
|
|
return $emissions_count + $transitions_count; |
1117
|
|
|
|
|
|
|
|
1118
|
|
|
|
|
|
|
} # end get_feature_count |
1119
|
|
|
|
|
|
|
|
1120
|
|
|
|
|
|
|
1; |
1121
|
|
|
|
|
|
|
|
1122
|
|
|
|
|
|
|
__END__ |
1123
|
|
|
|
|
|
|
|
1124
|
|
|
|
|
|
|
=pod |
1125
|
|
|
|
|
|
|
|
1126
|
|
|
|
|
|
|
=for Pod::Coverage BUILD |
1127
|
|
|
|
|
|
|
|
1128
|
|
|
|
|
|
|
=encoding utf-8 |
1129
|
|
|
|
|
|
|
|
1130
|
|
|
|
|
|
|
=head1 NAME |
1131
|
|
|
|
|
|
|
|
1132
|
|
|
|
|
|
|
Treex::Tool::Parser::MSTperl::ModelLabelling |
1133
|
|
|
|
|
|
|
|
1134
|
|
|
|
|
|
|
=head1 VERSION |
1135
|
|
|
|
|
|
|
|
1136
|
|
|
|
|
|
|
version 0.11949 |
1137
|
|
|
|
|
|
|
|
1138
|
|
|
|
|
|
|
=head1 DESCRIPTION |
1139
|
|
|
|
|
|
|
|
1140
|
|
|
|
|
|
|
This is an in-memory represenation of a labelling model, |
1141
|
|
|
|
|
|
|
extended from L<Treex::Tool::Parser::MSTperl::ModelBase>. |
1142
|
|
|
|
|
|
|
|
1143
|
|
|
|
|
|
|
=head1 FIELDS |
1144
|
|
|
|
|
|
|
|
1145
|
|
|
|
|
|
|
=head2 Inherited from base package |
1146
|
|
|
|
|
|
|
|
1147
|
|
|
|
|
|
|
Fields inherited from L<Treex::Tool::Parser::MSTperl::ModelBase>. |
1148
|
|
|
|
|
|
|
|
1149
|
|
|
|
|
|
|
=over 4 |
1150
|
|
|
|
|
|
|
|
1151
|
|
|
|
|
|
|
=item config |
1152
|
|
|
|
|
|
|
|
1153
|
|
|
|
|
|
|
Instance of L<Treex::Tool::Parser::MSTperl::Config> containing settings to be |
1154
|
|
|
|
|
|
|
used for the model. |
1155
|
|
|
|
|
|
|
|
1156
|
|
|
|
|
|
|
Currently the settings most relevant to the model are the following: |
1157
|
|
|
|
|
|
|
|
1158
|
|
|
|
|
|
|
=over 8 |
1159
|
|
|
|
|
|
|
|
1160
|
|
|
|
|
|
|
=item EM_EPSILON |
1161
|
|
|
|
|
|
|
|
1162
|
|
|
|
|
|
|
See L<Treex::Tool::Parser::MSTperl::Config/EM_EPSILON>. |
1163
|
|
|
|
|
|
|
|
1164
|
|
|
|
|
|
|
=item labeller_algorithm |
1165
|
|
|
|
|
|
|
|
1166
|
|
|
|
|
|
|
See L<Treex::Tool::Parser::MSTperl::Config/labeller_algorithm>. |
1167
|
|
|
|
|
|
|
|
1168
|
|
|
|
|
|
|
=item labelledFeaturesControl |
1169
|
|
|
|
|
|
|
|
1170
|
|
|
|
|
|
|
See L<Treex::Tool::Parser::MSTperl::Config/labelledFeaturesControl>. |
1171
|
|
|
|
|
|
|
|
1172
|
|
|
|
|
|
|
=item SEQUENCE_BOUNDARY_LABEL |
1173
|
|
|
|
|
|
|
|
1174
|
|
|
|
|
|
|
See L<Treex::Tool::Parser::MSTperl::Config/SEQUENCE_BOUNDARY_LABEL>. |
1175
|
|
|
|
|
|
|
|
1176
|
|
|
|
|
|
|
=back |
1177
|
|
|
|
|
|
|
|
1178
|
|
|
|
|
|
|
=item featuresControl |
1179
|
|
|
|
|
|
|
|
1180
|
|
|
|
|
|
|
Provides access to labeller features, especially enabling their computation. |
1181
|
|
|
|
|
|
|
Intance of L<Treex::Tool::Parser::MSTperl::FeaturesControl>. |
1182
|
|
|
|
|
|
|
|
1183
|
|
|
|
|
|
|
=back |
1184
|
|
|
|
|
|
|
|
1185
|
|
|
|
|
|
|
=head2 Label scoring |
1186
|
|
|
|
|
|
|
|
1187
|
|
|
|
|
|
|
=over 4 |
1188
|
|
|
|
|
|
|
|
1189
|
|
|
|
|
|
|
=item emissions |
1190
|
|
|
|
|
|
|
|
1191
|
|
|
|
|
|
|
Emission scores for Viterbi. They follow the edge-based factorization |
1192
|
|
|
|
|
|
|
and provide scores for various labels for an edge based on its features. |
1193
|
|
|
|
|
|
|
|
1194
|
|
|
|
|
|
|
The structure is: |
1195
|
|
|
|
|
|
|
|
1196
|
|
|
|
|
|
|
emissions->{feature}->{label} = score |
1197
|
|
|
|
|
|
|
|
1198
|
|
|
|
|
|
|
Scores may or may not be probabilities, based on the algorithm used. |
1199
|
|
|
|
|
|
|
Also based on the algorithm they may be MIRA-computed |
1200
|
|
|
|
|
|
|
or they might be obtained by standard MLE. |
1201
|
|
|
|
|
|
|
|
1202
|
|
|
|
|
|
|
=item transitions |
1203
|
|
|
|
|
|
|
|
1204
|
|
|
|
|
|
|
Transition scores for Viterbi. They follow the |
1205
|
|
|
|
|
|
|
first order Markov chain edge-based factorization |
1206
|
|
|
|
|
|
|
and provide scores for various labels for an edge |
1207
|
|
|
|
|
|
|
probably based on its features |
1208
|
|
|
|
|
|
|
and always based on previous edge label. |
1209
|
|
|
|
|
|
|
|
1210
|
|
|
|
|
|
|
Scores may or may not be probabilities, based on the algorithm used. |
1211
|
|
|
|
|
|
|
Also based on the algorithm they may be obtained by standard MLE |
1212
|
|
|
|
|
|
|
or they might be MIRA-computed. |
1213
|
|
|
|
|
|
|
|
1214
|
|
|
|
|
|
|
The structure is: |
1215
|
|
|
|
|
|
|
|
1216
|
|
|
|
|
|
|
transitions->{label_prev}->{label_this} = prob |
1217
|
|
|
|
|
|
|
|
1218
|
|
|
|
|
|
|
or |
1219
|
|
|
|
|
|
|
|
1220
|
|
|
|
|
|
|
transitions->{feature}->{label_prev}->{label_this} = score |
1221
|
|
|
|
|
|
|
|
1222
|
|
|
|
|
|
|
=back |
1223
|
|
|
|
|
|
|
|
1224
|
|
|
|
|
|
|
=head2 Transitions smoothing |
1225
|
|
|
|
|
|
|
|
1226
|
|
|
|
|
|
|
In some algorithms linear combination smoothing is used |
1227
|
|
|
|
|
|
|
for transition probabilities. |
1228
|
|
|
|
|
|
|
The resulting transition probability is then obtained as: |
1229
|
|
|
|
|
|
|
|
1230
|
|
|
|
|
|
|
PROB(label|prev_label) = |
1231
|
|
|
|
|
|
|
smooth_bigrams * transitions->{prev_label}->{label} + |
1232
|
|
|
|
|
|
|
smooth_unigrams * unigrams->{label} + |
1233
|
|
|
|
|
|
|
smooth_uniform |
1234
|
|
|
|
|
|
|
|
1235
|
|
|
|
|
|
|
=over 4 |
1236
|
|
|
|
|
|
|
|
1237
|
|
|
|
|
|
|
=item smooth_bigrams |
1238
|
|
|
|
|
|
|
|
1239
|
|
|
|
|
|
|
=item smooth_unigrams |
1240
|
|
|
|
|
|
|
|
1241
|
|
|
|
|
|
|
=item smooth_uniform |
1242
|
|
|
|
|
|
|
|
1243
|
|
|
|
|
|
|
The actual smoothing parameters computed by EM algorithm. |
1244
|
|
|
|
|
|
|
Each of them is between 0 and 1 and together they sum up to 1. |
1245
|
|
|
|
|
|
|
|
1246
|
|
|
|
|
|
|
=item uniform_prob |
1247
|
|
|
|
|
|
|
|
1248
|
|
|
|
|
|
|
Unifrom probability of a label, computed as |
1249
|
|
|
|
|
|
|
C<1 / ( keys %{ $self->unigrams } )>. |
1250
|
|
|
|
|
|
|
|
1251
|
|
|
|
|
|
|
Set in C<compute_smoothing_params>. |
1252
|
|
|
|
|
|
|
|
1253
|
|
|
|
|
|
|
=item unigrams |
1254
|
|
|
|
|
|
|
|
1255
|
|
|
|
|
|
|
Basic MLE from data, the structure is |
1256
|
|
|
|
|
|
|
|
1257
|
|
|
|
|
|
|
unigrams->{label} = prob |
1258
|
|
|
|
|
|
|
|
1259
|
|
|
|
|
|
|
To be used for transitions smoothing and/or backoff |
1260
|
|
|
|
|
|
|
(can be used both for emissions and transitions) |
1261
|
|
|
|
|
|
|
It also contains the C<SEQUENCE_BOUNDARY_LABEL> prob |
1262
|
|
|
|
|
|
|
(the SEQUENCE_BOUNDARY_LABEL is counted once for each sequence) |
1263
|
|
|
|
|
|
|
which might be unappropriate in some cases (eg. for emission probs). |
1264
|
|
|
|
|
|
|
|
1265
|
|
|
|
|
|
|
=item EM_heldout_data |
1266
|
|
|
|
|
|
|
|
1267
|
|
|
|
|
|
|
Just an array ref with the sentences that represent the heldout data |
1268
|
|
|
|
|
|
|
to be able to run the EM algorithm in C<prepare_for_mira()>. |
1269
|
|
|
|
|
|
|
Used only in training. |
1270
|
|
|
|
|
|
|
|
1271
|
|
|
|
|
|
|
=back |
1272
|
|
|
|
|
|
|
|
1273
|
|
|
|
|
|
|
=head1 METHODS |
1274
|
|
|
|
|
|
|
|
1275
|
|
|
|
|
|
|
=head2 Inherited |
1276
|
|
|
|
|
|
|
|
1277
|
|
|
|
|
|
|
Subroutines inherited from L<Treex::Tool::Parser::MSTperl::ModelBase>. |
1278
|
|
|
|
|
|
|
|
1279
|
|
|
|
|
|
|
=head3 Load and store |
1280
|
|
|
|
|
|
|
|
1281
|
|
|
|
|
|
|
=over 4 |
1282
|
|
|
|
|
|
|
|
1283
|
|
|
|
|
|
|
=item store |
1284
|
|
|
|
|
|
|
|
1285
|
|
|
|
|
|
|
See L<Treex::Tool::Parser::MSTperl::ModelBase/store>. |
1286
|
|
|
|
|
|
|
|
1287
|
|
|
|
|
|
|
=item store_tsv |
1288
|
|
|
|
|
|
|
|
1289
|
|
|
|
|
|
|
See L<Treex::Tool::Parser::MSTperl::ModelBase/store_tsv>. |
1290
|
|
|
|
|
|
|
|
1291
|
|
|
|
|
|
|
=item load |
1292
|
|
|
|
|
|
|
|
1293
|
|
|
|
|
|
|
See L<Treex::Tool::Parser::MSTperl::ModelBase/load>. |
1294
|
|
|
|
|
|
|
|
1295
|
|
|
|
|
|
|
=item load_tsv |
1296
|
|
|
|
|
|
|
|
1297
|
|
|
|
|
|
|
See L<Treex::Tool::Parser::MSTperl::ModelBase/load_tsv>. |
1298
|
|
|
|
|
|
|
|
1299
|
|
|
|
|
|
|
=back |
1300
|
|
|
|
|
|
|
|
1301
|
|
|
|
|
|
|
=head2 Overriden |
1302
|
|
|
|
|
|
|
|
1303
|
|
|
|
|
|
|
Subroutines overriding stubs in L<Treex::Tool::Parser::MSTperl::ModelBase>. |
1304
|
|
|
|
|
|
|
|
1305
|
|
|
|
|
|
|
=head3 Load and store |
1306
|
|
|
|
|
|
|
|
1307
|
|
|
|
|
|
|
=over 4 |
1308
|
|
|
|
|
|
|
|
1309
|
|
|
|
|
|
|
=item $data = get_data_to_store(), $data = get_data_to_store_tsv() |
1310
|
|
|
|
|
|
|
|
1311
|
|
|
|
|
|
|
Returns the model data, containing the following fields: |
1312
|
|
|
|
|
|
|
C<unigrams>, |
1313
|
|
|
|
|
|
|
C<transitions>, |
1314
|
|
|
|
|
|
|
C<emissions>, |
1315
|
|
|
|
|
|
|
C<smooth_uniform>, |
1316
|
|
|
|
|
|
|
C<smooth_unigrams>, |
1317
|
|
|
|
|
|
|
C<smooth_bigrams>, |
1318
|
|
|
|
|
|
|
C<uniform_prob> |
1319
|
|
|
|
|
|
|
|
1320
|
|
|
|
|
|
|
=item load_data($data), load_data_tsv($data) |
1321
|
|
|
|
|
|
|
|
1322
|
|
|
|
|
|
|
Tries to get all necessary data from C<$data> |
1323
|
|
|
|
|
|
|
(see C<get_data_to_store> to see what data are stored). |
1324
|
|
|
|
|
|
|
Also does basic checks on the data, eg. for non-emptiness, but nothing |
1325
|
|
|
|
|
|
|
sophisticated. Is algorithm-sensitive, i.e. if some data are not needed |
1326
|
|
|
|
|
|
|
for the algorithm used, they do not have to be contained in the hash. |
1327
|
|
|
|
|
|
|
|
1328
|
|
|
|
|
|
|
=back |
1329
|
|
|
|
|
|
|
|
1330
|
|
|
|
|
|
|
=head3 Training support |
1331
|
|
|
|
|
|
|
|
1332
|
|
|
|
|
|
|
=over 4 |
1333
|
|
|
|
|
|
|
|
1334
|
|
|
|
|
|
|
=item prepare_for_mira |
1335
|
|
|
|
|
|
|
|
1336
|
|
|
|
|
|
|
Called after preprocessing training data, before entering the MIRA phase. |
1337
|
|
|
|
|
|
|
|
1338
|
|
|
|
|
|
|
Function varies depending on algorithm used. |
1339
|
|
|
|
|
|
|
Usually recomputes counts stored in C<emissions>, C<transitions> and C<unigrams> |
1340
|
|
|
|
|
|
|
to probabilities that have been computed by C<add_emission>, |
1341
|
|
|
|
|
|
|
C<add_transition> and C<add_unigram>. |
1342
|
|
|
|
|
|
|
Also calls C<compute_smoothing_params> to estimate smoothing parameters |
1343
|
|
|
|
|
|
|
for smoothing of transition probabilities. |
1344
|
|
|
|
|
|
|
|
1345
|
|
|
|
|
|
|
=item get_feature_count |
1346
|
|
|
|
|
|
|
|
1347
|
|
|
|
|
|
|
Only to provide information about the model. |
1348
|
|
|
|
|
|
|
Returns number of features in the model (where a "feature" can stand for |
1349
|
|
|
|
|
|
|
various things depending on the algorithm used). |
1350
|
|
|
|
|
|
|
|
1351
|
|
|
|
|
|
|
=back |
1352
|
|
|
|
|
|
|
|
1353
|
|
|
|
|
|
|
=head2 Technical methods |
1354
|
|
|
|
|
|
|
|
1355
|
|
|
|
|
|
|
=over 4 |
1356
|
|
|
|
|
|
|
|
1357
|
|
|
|
|
|
|
=item BUILD |
1358
|
|
|
|
|
|
|
|
1359
|
|
|
|
|
|
|
my $model = Treex::Tool::Parser::MSTperl::ModelLabelling->new( |
1360
|
|
|
|
|
|
|
config => $config, |
1361
|
|
|
|
|
|
|
); |
1362
|
|
|
|
|
|
|
|
1363
|
|
|
|
|
|
|
Creates an empty model. If you are training the model, this is probably what you |
1364
|
|
|
|
|
|
|
want, otherwise you can use C<load> or C<load_tsv> |
1365
|
|
|
|
|
|
|
to load an existing labelling model from a file. |
1366
|
|
|
|
|
|
|
|
1367
|
|
|
|
|
|
|
However, most often you would probably use a model for a labeller |
1368
|
|
|
|
|
|
|
(L<Treex::Tool::Parser::MSTperl::Labeller>) |
1369
|
|
|
|
|
|
|
or a labelling trainer |
1370
|
|
|
|
|
|
|
(L<Treex::Tool::Parser::MSTperl::TrainerLabelling>) |
1371
|
|
|
|
|
|
|
which both automatically create the model on build. |
1372
|
|
|
|
|
|
|
The labeller also provides wrapping methods |
1373
|
|
|
|
|
|
|
L<Treex::Tool::Parser::MSTperl::Labeller/load_model> |
1374
|
|
|
|
|
|
|
and |
1375
|
|
|
|
|
|
|
L<Treex::Tool::Parser::MSTperl::Labeller/load_model_tsv> |
1376
|
|
|
|
|
|
|
which you can call to load the model from a file. |
1377
|
|
|
|
|
|
|
(Btw. as you might expect, the trainer provides methods |
1378
|
|
|
|
|
|
|
L<Treex::Tool::Parser::MSTperl::TrainerLabelling/store_model> |
1379
|
|
|
|
|
|
|
and |
1380
|
|
|
|
|
|
|
L<Treex::Tool::Parser::MSTperl::TrainerLabelling/store_model_tsv>.) |
1381
|
|
|
|
|
|
|
|
1382
|
|
|
|
|
|
|
=back |
1383
|
|
|
|
|
|
|
|
1384
|
|
|
|
|
|
|
=head2 MLE on training data |
1385
|
|
|
|
|
|
|
|
1386
|
|
|
|
|
|
|
C<emissions> and C<transitions> can be either MIRA-trained |
1387
|
|
|
|
|
|
|
or estimated directly from training data using MLE |
1388
|
|
|
|
|
|
|
(Maximum Likelihood Estimate). |
1389
|
|
|
|
|
|
|
C<unigrams> are always estimated by MLE. |
1390
|
|
|
|
|
|
|
|
1391
|
|
|
|
|
|
|
=over 4 |
1392
|
|
|
|
|
|
|
|
1393
|
|
|
|
|
|
|
=item add_unigram ($label) |
1394
|
|
|
|
|
|
|
|
1395
|
|
|
|
|
|
|
Increment count for the label in C<unigrams>. |
1396
|
|
|
|
|
|
|
|
1397
|
|
|
|
|
|
|
=item add_transition ($label_this, $label_prev) |
1398
|
|
|
|
|
|
|
|
1399
|
|
|
|
|
|
|
=item add_transition ($label_this, $label_prev, $feature) |
1400
|
|
|
|
|
|
|
|
1401
|
|
|
|
|
|
|
Increment count for the transition in C<transitions>, possible including a |
1402
|
|
|
|
|
|
|
feature on "this" edge if the algorithm uses features with transitions. |
1403
|
|
|
|
|
|
|
|
1404
|
|
|
|
|
|
|
=item add_emission ($feature, $label) |
1405
|
|
|
|
|
|
|
|
1406
|
|
|
|
|
|
|
Increment count for this label on an edge with this feature in C<emissions>. |
1407
|
|
|
|
|
|
|
|
1408
|
|
|
|
|
|
|
=item compute_probs_from_counts ($self->emissions) |
1409
|
|
|
|
|
|
|
|
1410
|
|
|
|
|
|
|
Takes a hash reference with label counts and chnages the counts |
1411
|
|
|
|
|
|
|
to probabilities (this is the actual MLE). |
1412
|
|
|
|
|
|
|
May be called in C<prepare_for_mira> on |
1413
|
|
|
|
|
|
|
C<emissions>, C<transitions> and C<unigrams>. |
1414
|
|
|
|
|
|
|
|
1415
|
|
|
|
|
|
|
=back |
1416
|
|
|
|
|
|
|
|
1417
|
|
|
|
|
|
|
=head2 EM algorithm |
1418
|
|
|
|
|
|
|
|
1419
|
|
|
|
|
|
|
=over 4 |
1420
|
|
|
|
|
|
|
|
1421
|
|
|
|
|
|
|
=item compute_smoothing_params() |
1422
|
|
|
|
|
|
|
|
1423
|
|
|
|
|
|
|
The main method containing an implementation of the Expectation Maximization |
1424
|
|
|
|
|
|
|
Algorithm to compute smoothing parameters (C<smooth_bigrams>, |
1425
|
|
|
|
|
|
|
C<smooth_unigrams>, C<smooth_uniform>) for transition probabilities |
1426
|
|
|
|
|
|
|
smoothing by linear combination of bigram, unigram and uniform probability. |
1427
|
|
|
|
|
|
|
Iteratively tries to find |
1428
|
|
|
|
|
|
|
such parameters that the probabilities from training data |
1429
|
|
|
|
|
|
|
(C<transitions>, C<unigrams> and C<uniform_prob>) combined together by |
1430
|
|
|
|
|
|
|
the smoothing parameters model well enough the heldout data |
1431
|
|
|
|
|
|
|
(C<EM_heldout_data>), i.e. tries to maximize the probability of the heldout |
1432
|
|
|
|
|
|
|
data given the training data probabilities by adjusting the smoothing |
1433
|
|
|
|
|
|
|
parameters values. |
1434
|
|
|
|
|
|
|
|
1435
|
|
|
|
|
|
|
Uses C<EM_EPSILON> as a stopping criterion, i.e. stops when the sum of |
1436
|
|
|
|
|
|
|
absolute values of changes to all smoothing parameters are lower |
1437
|
|
|
|
|
|
|
than the value of C<EM_EPSILON>. |
1438
|
|
|
|
|
|
|
|
1439
|
|
|
|
|
|
|
=item count_expected_counts_all() |
1440
|
|
|
|
|
|
|
|
1441
|
|
|
|
|
|
|
=item count_expected_counts_tree($root_node) |
1442
|
|
|
|
|
|
|
|
1443
|
|
|
|
|
|
|
=item count_expected_counts_sequence($labels_sequence) |
1444
|
|
|
|
|
|
|
|
1445
|
|
|
|
|
|
|
Support methods to C<compute_smoothing_params>, in the order in which they |
1446
|
|
|
|
|
|
|
call each other. |
1447
|
|
|
|
|
|
|
|
1448
|
|
|
|
|
|
|
=back |
1449
|
|
|
|
|
|
|
|
1450
|
|
|
|
|
|
|
=head2 Scoring |
1451
|
|
|
|
|
|
|
|
1452
|
|
|
|
|
|
|
A bunch of methods to score the likelihood of a label being assigned to an |
1453
|
|
|
|
|
|
|
edge based on the edge's features and the label assigned to the previous |
1454
|
|
|
|
|
|
|
edge. |
1455
|
|
|
|
|
|
|
|
1456
|
|
|
|
|
|
|
=over 4 |
1457
|
|
|
|
|
|
|
|
1458
|
|
|
|
|
|
|
=item get_all_labels() |
1459
|
|
|
|
|
|
|
|
1460
|
|
|
|
|
|
|
Returns (a reference to) an array of all labels found in the training data |
1461
|
|
|
|
|
|
|
(eg. C<['Subj', 'Obj', 'Atr']>). |
1462
|
|
|
|
|
|
|
|
1463
|
|
|
|
|
|
|
=item get_label_score($label, $label_prev, $features) |
1464
|
|
|
|
|
|
|
|
1465
|
|
|
|
|
|
|
Computes a score of assigning the given label to an edge, |
1466
|
|
|
|
|
|
|
given the features of the edge and the label assigned to the previous edge. |
1467
|
|
|
|
|
|
|
|
1468
|
|
|
|
|
|
|
Always a higher score means a more likely label for the edge. |
1469
|
|
|
|
|
|
|
Some algorithms may give a negative score. |
1470
|
|
|
|
|
|
|
|
1471
|
|
|
|
|
|
|
Is semantically equivalent to calling C<get_emission_score> |
1472
|
|
|
|
|
|
|
and C<get_transition_score> and then combining it together somehow. |
1473
|
|
|
|
|
|
|
|
1474
|
|
|
|
|
|
|
=item get_emission_score($label, $feature) |
1475
|
|
|
|
|
|
|
|
1476
|
|
|
|
|
|
|
Computes the "emission score" of assigning the given label to an edge, |
1477
|
|
|
|
|
|
|
given one of the feature of the edge and disregarding |
1478
|
|
|
|
|
|
|
the label assigned to the previous edge. |
1479
|
|
|
|
|
|
|
|
1480
|
|
|
|
|
|
|
=item get_transition_score($label_this, $label_prev, $feature) |
1481
|
|
|
|
|
|
|
|
1482
|
|
|
|
|
|
|
Computes the "transition score" of assigning the given label to an edge, |
1483
|
|
|
|
|
|
|
given the label assigned to the previous edge |
1484
|
|
|
|
|
|
|
and possibly also one of the features of the edge |
1485
|
|
|
|
|
|
|
but NOT including the emission score returned by C<get_emission_score>. |
1486
|
|
|
|
|
|
|
|
1487
|
|
|
|
|
|
|
=item $result = get_transition_probs_array ($label_this, $label_prev) |
1488
|
|
|
|
|
|
|
|
1489
|
|
|
|
|
|
|
Returns (a reference to) an array of the probabilities of the transition |
1490
|
|
|
|
|
|
|
from label_prev to label_this (to be smoothed together), |
1491
|
|
|
|
|
|
|
having the following structure: |
1492
|
|
|
|
|
|
|
|
1493
|
|
|
|
|
|
|
$result->[0] = uniform prob |
1494
|
|
|
|
|
|
|
$result->[1] = unigram prob |
1495
|
|
|
|
|
|
|
$result->[2] = bigram prob |
1496
|
|
|
|
|
|
|
|
1497
|
|
|
|
|
|
|
=item $result = get_emission_scores($features) |
1498
|
|
|
|
|
|
|
|
1499
|
|
|
|
|
|
|
Get scores of assigning each of the possible labels to an edge |
1500
|
|
|
|
|
|
|
based on all the features of the edge. Is semantically equivalent |
1501
|
|
|
|
|
|
|
to doing: |
1502
|
|
|
|
|
|
|
|
1503
|
|
|
|
|
|
|
foreach label |
1504
|
|
|
|
|
|
|
foreach feature |
1505
|
|
|
|
|
|
|
get_emission_score(label, feature) |
1506
|
|
|
|
|
|
|
|
1507
|
|
|
|
|
|
|
The structure is: |
1508
|
|
|
|
|
|
|
|
1509
|
|
|
|
|
|
|
$result->{label} = score |
1510
|
|
|
|
|
|
|
|
1511
|
|
|
|
|
|
|
Actually only serves as a switch for several implementations of the method |
1512
|
|
|
|
|
|
|
(C<get_emission_scores_basic_MIRA> and C<get_emission_scores_no_MIRA>); |
1513
|
|
|
|
|
|
|
the method to be used is selected based on the algorithm being used. |
1514
|
|
|
|
|
|
|
|
1515
|
|
|
|
|
|
|
=item get_emission_scores_basic_MIRA($features) |
1516
|
|
|
|
|
|
|
|
1517
|
|
|
|
|
|
|
A C<get_emission_scores> implementation used with algorithms |
1518
|
|
|
|
|
|
|
where the emission scores are computed by MIRA (this is currently |
1519
|
|
|
|
|
|
|
the most successful implementation). |
1520
|
|
|
|
|
|
|
|
1521
|
|
|
|
|
|
|
=item get_emission_scores_no_MIRA($features) |
1522
|
|
|
|
|
|
|
|
1523
|
|
|
|
|
|
|
A C<get_emission_scores> implementation using only MLE. Probably obsolete now. |
1524
|
|
|
|
|
|
|
|
1525
|
|
|
|
|
|
|
=back |
1526
|
|
|
|
|
|
|
|
1527
|
|
|
|
|
|
|
=head2 Changing the scores |
1528
|
|
|
|
|
|
|
|
1529
|
|
|
|
|
|
|
Methods used by the trainer |
1530
|
|
|
|
|
|
|
(L<Treex::Tool::Parser::MSTperl::TrainerLabelling>) |
1531
|
|
|
|
|
|
|
to adjust the scores to whatever seems to be |
1532
|
|
|
|
|
|
|
the best idea at the moment. Used only in MIRA training |
1533
|
|
|
|
|
|
|
(MLE uses C<add_unigram>, C<add_emission>, C<add_transition> |
1534
|
|
|
|
|
|
|
and C<compute_probs_from_counts> instead). |
1535
|
|
|
|
|
|
|
|
1536
|
|
|
|
|
|
|
=over 4 |
1537
|
|
|
|
|
|
|
|
1538
|
|
|
|
|
|
|
=item set_feature_score($feature, $score, $label, $label_prev) |
1539
|
|
|
|
|
|
|
|
1540
|
|
|
|
|
|
|
Sets the specified emission score (if label_prev is not set) |
1541
|
|
|
|
|
|
|
or transition score (if it is) |
1542
|
|
|
|
|
|
|
to the given value (C<$score>). |
1543
|
|
|
|
|
|
|
|
1544
|
|
|
|
|
|
|
=item update_feature_score($feature, $update, $label, $label_prev) |
1545
|
|
|
|
|
|
|
|
1546
|
|
|
|
|
|
|
Updates the specified emission score (if label_prev is not set) |
1547
|
|
|
|
|
|
|
or transition score (if it is) |
1548
|
|
|
|
|
|
|
by the given value (C<$update>), i.e. adds that value to the |
1549
|
|
|
|
|
|
|
current value. |
1550
|
|
|
|
|
|
|
|
1551
|
|
|
|
|
|
|
=back |
1552
|
|
|
|
|
|
|
|
1553
|
|
|
|
|
|
|
=head1 AUTHORS |
1554
|
|
|
|
|
|
|
|
1555
|
|
|
|
|
|
|
Rudolf Rosa <rosa@ufal.mff.cuni.cz> |
1556
|
|
|
|
|
|
|
|
1557
|
|
|
|
|
|
|
=head1 COPYRIGHT AND LICENSE |
1558
|
|
|
|
|
|
|
|
1559
|
|
|
|
|
|
|
Copyright © 2011 by Institute of Formal and Applied Linguistics, Charles |
1560
|
|
|
|
|
|
|
University in Prague |
1561
|
|
|
|
|
|
|
|
1562
|
|
|
|
|
|
|
This module is free software; you can redistribute it and/or modify it under |
1563
|
|
|
|
|
|
|
the same terms as Perl itself. |