• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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 #ifndef NLP_SAFT_COMPONENTS_LANG_ID_MOBILE_FEATURES_CHAR_NGRAM_FEATURE_H_
18 #define NLP_SAFT_COMPONENTS_LANG_ID_MOBILE_FEATURES_CHAR_NGRAM_FEATURE_H_
19 
20 #include <mutex>  // NOLINT: see comments for state_mutex_
21 #include <string>
22 #include <vector>
23 
24 #include "lang_id/common/fel/feature-extractor.h"
25 #include "lang_id/common/fel/task-context.h"
26 #include "lang_id/common/fel/workspace.h"
27 #include "lang_id/features/light-sentence-features.h"
28 #include "lang_id/light-sentence.h"
29 
30 // TODO(abakalov): Add a test.
31 namespace libtextclassifier3 {
32 namespace mobile {
33 namespace lang_id {
34 
35 // Class for computing continuous char ngram features.
36 //
37 // Feature function descriptor parameters:
38 //   include_terminators(bool, false):
39 //     If 'true', then splits the text based on spaces to get tokens, adds "^"
40 //     to the beginning of each token, and adds "$" to the end of each token.
41 //     NOTE: currently, we support only include_terminators=true.
42 //   include_spaces(bool, false):
43 //     If 'true', then includes char ngrams containing spaces.
44 //     NOTE: currently, we support only include_spaces=false.
45 //   use_equal_weight(bool, false):
46 //     If 'true', then weighs each unique ngram by 1.0 / (number of unique
47 //     ngrams in the input). Otherwise, weighs each unique ngram by (ngram
48 //     count) / (total number of ngrams).
49 //     NOTE: currently, we support only use_equal_weight=false.
50 //   id_dim(int, 10000):
51 //     The integer id of each char ngram is computed as follows:
52 //     Hash32WithDefault(char ngram) % id_dim.
53 //   size(int, 3):
54 //     Only ngrams of this size will be extracted.
55 //
56 // NOTE: this class is not thread-safe.  TODO(salcianu): make it thread-safe.
57 class ContinuousBagOfNgramsFunction : public LightSentenceFeature {
58  public:
59   bool Setup(TaskContext *context) override;
60   bool Init(TaskContext *context) override;
61 
62   // Appends the features computed from the sentence to the feature vector.
63   void Evaluate(const WorkspaceSet &workspaces, const LightSentence &sentence,
64                 FeatureVector *result) const override;
65 
66   SAFTM_DEFINE_REGISTRATION_METHOD("continuous-bag-of-ngrams",
67                                    ContinuousBagOfNgramsFunction);
68 
69  private:
70   // Auxiliary for Evaluate().  Fills counts_ and non_zero_count_indices_ (see
71   // below), and returns the total ngram count.
72   int ComputeNgramCounts(const LightSentence &sentence) const;
73 
74   // Guards counts_ and non_zero_count_indices_.  NOTE: we use std::* constructs
75   // (instead of absl::Mutex & co) to simplify porting to Android and to avoid
76   // pulling in absl (which increases our code size).
77   mutable std::mutex state_mutex_;
78 
79   // counts_[i] is the count of all ngrams with id i.  Work data for Evaluate().
80   // NOTE: we declare this vector as a field, such that its underlying capacity
81   // stays allocated in between calls to Evaluate().
82   mutable std::vector<int> counts_;
83 
84   // Indices of non-zero elements of counts_.  See comments for counts_.
85   mutable std::vector<int> non_zero_count_indices_;
86 
87   // The integer id of each char ngram is computed as follows:
88   // Hash32WithDefaultSeed(char_ngram) % ngram_id_dimension_.
89   int ngram_id_dimension_;
90 
91   // Only ngrams of size ngram_size_ will be extracted.
92   int ngram_size_;
93 };
94 
95 }  // namespace lang_id
96 }  // namespace mobile
97 }  // namespace nlp_saft
98 
99 #endif  // NLP_SAFT_COMPONENTS_LANG_ID_MOBILE_FEATURES_CHAR_NGRAM_FEATURE_H_
100