1 /*
2 * Copyright (C) 2018 The Android Open Source Project
3 *
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at
7 *
8 * http://www.apache.org/licenses/LICENSE-2.0
9 *
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
15 */
16
17 #include "lang_id/features/char-ngram-feature.h"
18
19 #include <string>
20 #include <utility>
21 #include <vector>
22
23 #include "lang_id/common/fel/feature-types.h"
24 #include "lang_id/common/fel/task-context.h"
25 #include "lang_id/common/lite_base/logging.h"
26 #include "lang_id/common/math/hash.h"
27 #include "lang_id/common/utf8.h"
28
29 namespace libtextclassifier3 {
30 namespace mobile {
31 namespace lang_id {
32
Setup(TaskContext * context)33 bool ContinuousBagOfNgramsFunction::Setup(TaskContext *context) {
34 // Parameters in the feature function descriptor.
35 bool include_terminators = GetBoolParameter("include_terminators", false);
36 if (!include_terminators) {
37 SAFTM_LOG(ERROR) << "No support for include_terminators=true";
38 return false;
39 }
40
41 bool include_spaces = GetBoolParameter("include_spaces", false);
42 if (include_spaces) {
43 SAFTM_LOG(ERROR) << "No support for include_spaces=true";
44 return false;
45 }
46
47 bool use_equal_ngram_weight = GetBoolParameter("use_equal_weight", false);
48 if (use_equal_ngram_weight) {
49 SAFTM_LOG(ERROR) << "No support for use_equal_weight=true";
50 return false;
51 }
52
53 ngram_id_dimension_ = GetIntParameter("id_dim", 10000);
54 ngram_size_ = GetIntParameter("size", 3);
55
56 counts_.assign(ngram_id_dimension_, 0);
57 return true;
58 }
59
Init(TaskContext * context)60 bool ContinuousBagOfNgramsFunction::Init(TaskContext *context) {
61 set_feature_type(new NumericFeatureType(name(), ngram_id_dimension_));
62 return true;
63 }
64
ComputeNgramCounts(const LightSentence & sentence) const65 int ContinuousBagOfNgramsFunction::ComputeNgramCounts(
66 const LightSentence &sentence) const {
67 SAFTM_CHECK_EQ(counts_.size(), ngram_id_dimension_);
68 SAFTM_CHECK_EQ(non_zero_count_indices_.size(), 0);
69
70 int total_count = 0;
71
72 for (const std::string &word : sentence) {
73 const char *const word_end = word.data() + word.size();
74
75 // Set ngram_start at the start of the current token (word).
76 const char *ngram_start = word.data();
77
78 // Set ngram_end ngram_size UTF8 characters after ngram_start. Note: each
79 // UTF8 character contains between 1 and 4 bytes.
80 const char *ngram_end = ngram_start;
81 int num_utf8_chars = 0;
82 do {
83 ngram_end += utils::OneCharLen(ngram_end);
84 num_utf8_chars++;
85 } while ((num_utf8_chars < ngram_size_) && (ngram_end < word_end));
86
87 if (num_utf8_chars < ngram_size_) {
88 // Current token is so small, it does not contain a single ngram of
89 // ngram_size UTF8 characters. Not much we can do in this case ...
90 continue;
91 }
92
93 // At this point, [ngram_start, ngram_end) is the first ngram of ngram_size
94 // UTF8 characters from current token.
95 while (true) {
96 // Compute ngram id: hash(ngram) % ngram_id_dimension
97 int ngram_id = (
98 utils::Hash32WithDefaultSeed(ngram_start, ngram_end - ngram_start)
99 % ngram_id_dimension_);
100
101 // Use a reference to the actual count, such that we can both test whether
102 // the count was 0 and increment it without perfoming two lookups.
103 int &ref_to_count_for_ngram = counts_[ngram_id];
104 if (ref_to_count_for_ngram == 0) {
105 non_zero_count_indices_.push_back(ngram_id);
106 }
107 ref_to_count_for_ngram++;
108 total_count++;
109 if (ngram_end >= word_end) {
110 break;
111 }
112
113 // Advance both ngram_start and ngram_end by one UTF8 character. This
114 // way, the number of UTF8 characters between them remains constant
115 // (ngram_size).
116 ngram_start += utils::OneCharLen(ngram_start);
117 ngram_end += utils::OneCharLen(ngram_end);
118 }
119 } // end of loop over tokens.
120
121 return total_count;
122 }
123
Evaluate(const WorkspaceSet & workspaces,const LightSentence & sentence,FeatureVector * result) const124 void ContinuousBagOfNgramsFunction::Evaluate(const WorkspaceSet &workspaces,
125 const LightSentence &sentence,
126 FeatureVector *result) const {
127 // NOTE: we use std::* constructs (instead of absl::Mutex & co) to simplify
128 // porting to Android and to avoid pulling in absl (which increases our code
129 // size).
130 std::lock_guard<std::mutex> mlock(state_mutex_);
131
132 // Find the char ngram counts.
133 int total_count = ComputeNgramCounts(sentence);
134
135 // Populate the feature vector.
136 const float norm = static_cast<float>(total_count);
137
138 // TODO(salcianu): explore treating dense vectors (i.e., many non-zero
139 // elements) separately.
140 for (int ngram_id : non_zero_count_indices_) {
141 const float weight = counts_[ngram_id] / norm;
142 FloatFeatureValue value(ngram_id, weight);
143 result->add(feature_type(), value.discrete_value);
144
145 // Clear up counts_, for the next invocation of Evaluate().
146 counts_[ngram_id] = 0;
147 }
148
149 // Clear up non_zero_count_indices_, for the next invocation of Evaluate().
150 non_zero_count_indices_.clear();
151 }
152
153 SAFTM_STATIC_REGISTRATION(ContinuousBagOfNgramsFunction);
154
155 } // namespace lang_id
156 } // namespace mobile
157 } // namespace nlp_saft
158