confidence.cc
4.98 KB
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
107
108
109
110
111
112
113
114
115
116
117
// lat/confidence.cc
// Copyright 2013 Johns Hopkins University (Author: Daniel Povey)
// See ../../COPYING for clarification regarding multiple authors
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// THIS CODE IS PROVIDED *AS IS* BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
// KIND, EITHER EXPRESS OR IMPLIED, INCLUDING WITHOUT LIMITATION ANY IMPLIED
// WARRANTIES OR CONDITIONS OF TITLE, FITNESS FOR A PARTICULAR PURPOSE,
// MERCHANTABLITY OR NON-INFRINGEMENT.
// See the Apache 2 License for the specific language governing permissions and
// limitations under the License.
#include "lat/confidence.h"
#include "lat/lattice-functions.h"
#include "lat/determinize-lattice-pruned.h"
namespace kaldi {
BaseFloat SentenceLevelConfidence(const CompactLattice &clat,
int32 *num_paths,
std::vector<int32> *best_sentence,
std::vector<int32> *second_best_sentence) {
/* It may seem strange that the first thing we do is to convert the
CompactLattice to a Lattice, given that we may have just created the
CompactLattice by determinizing a Lattice. However, this is not just
a circular conversion; "lat" will have the property that distinct
paths have distinct word sequences.
Below, we could run NbestAsFsts on a CompactLattice, but the time
taken would be quadratic in the length in words of the CompactLattice,
because of the alignment information getting appended as vectors.
That's why we convert back to Lattice.
*/
Lattice lat;
ConvertLattice(clat, &lat);
std::vector<Lattice> lats;
NbestAsFsts(lat, 2, &lats);
int32 n = lats.size();
KALDI_ASSERT(n >= 0 && n <= 2);
if (num_paths != NULL) *num_paths = n;
if (best_sentence != NULL) best_sentence->clear();
if (second_best_sentence != NULL) second_best_sentence->clear();
LatticeWeight weight1, weight2;
if (n >= 1)
fst::GetLinearSymbolSequence<LatticeArc,int32>(lats[0], NULL,
best_sentence,
&weight1);
if (n >= 2)
fst::GetLinearSymbolSequence<LatticeArc,int32>(lats[1], NULL,
second_best_sentence,
&weight2);
if (n == 0) {
return 0; // this seems most appropriate because it will be interpreted as
// zero confidence, and something definitely went wrong for this
// to happen.
} else if (n == 1) {
// If there is only one sentence in the lattice, we interpret this as there
// being perfect confidence
return std::numeric_limits<BaseFloat>::infinity();
} else {
BaseFloat best_cost = ConvertToCost(weight1),
second_best_cost = ConvertToCost(weight2);
BaseFloat ans = second_best_cost - best_cost;
if (!(ans >= -0.001 * (fabs(best_cost) + fabs(second_best_cost)))) {
// Answer should be positive. Make sure it's at at least not
// substantially negative. This would be very strange.
KALDI_WARN << "Very negative difference." << ans;
}
if (ans < 0) ans = 0;
return ans;
}
}
BaseFloat SentenceLevelConfidence(const Lattice &lat,
int32 *num_paths,
std::vector<int32> *best_sentence,
std::vector<int32> *second_best_sentence) {
int32 max_sentence_length = LongestSentenceLength(lat);
fst::DeterminizeLatticePrunedOptions determinize_opts;
// The basic idea of expanding only up to "max_sentence_length * 2" arcs,
// is that that should be sufficient to get the best and second-best paths
// through the lattice, which is all we need for this particular application.
// "safety_term" is just in case there is some reason why we might need a few
// extra arcs, e.g. in case of a tie on the weights of the second-best path.
int32 safety_term = 4 + max_sentence_length;
determinize_opts.max_arcs = max_sentence_length * 2 + safety_term;
// set prune_beam to a large value... we don't really rely on the beam; we
// rely on the max_arcs variable to limit the size of the lattice.
double prune_beam = std::numeric_limits<double>::infinity();
CompactLattice clat;
// We ignore the return status of DeterminizeLatticePruned. It will likely
// return false, but this is expected because the expansion is limited
// by "max_arcs" not "prune_beam".
Lattice inverse_lat(lat);
fst::Invert(&inverse_lat); // Swap input and output symbols.
DeterminizeLatticePruned(inverse_lat, prune_beam, &clat, determinize_opts);
// Call the version of this function that takes a CompactLattice.
return SentenceLevelConfidence(clat, num_paths,
best_sentence, second_best_sentence);
}
} // namespace kaldi