• 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 LIBTEXTCLASSIFIER_UTILS_TOKEN_FEATURE_EXTRACTOR_H_
18 #define LIBTEXTCLASSIFIER_UTILS_TOKEN_FEATURE_EXTRACTOR_H_
19 
20 #include <memory>
21 #include <unordered_set>
22 #include <vector>
23 
24 #include "annotator/types.h"
25 #include "utils/strings/stringpiece.h"
26 #include "utils/utf8/unilib.h"
27 
28 namespace libtextclassifier3 {
29 
30 struct TokenFeatureExtractorOptions {
31   // Number of buckets used for hashing charactergrams.
32   int num_buckets = 0;
33 
34   // Orders of charactergrams to extract. E.g., 2 means character bigrams, 3
35   // character trigrams etc.
36   std::vector<int> chargram_orders;
37 
38   // Whether to extract the token case feature.
39   bool extract_case_feature = false;
40 
41   // If true, will use the unicode-aware functionality for extracting features.
42   bool unicode_aware_features = false;
43 
44   // Whether to extract the selection mask feature.
45   bool extract_selection_mask_feature = false;
46 
47   // Regexp features to extract.
48   std::vector<std::string> regexp_features;
49 
50   // Whether to remap digits to a single number.
51   bool remap_digits = false;
52 
53   // Whether to lowercase all tokens.
54   bool lowercase_tokens = false;
55 
56   // Maximum length of a word.
57   int max_word_length = 20;
58 
59   // List of allowed charactergrams. The extracted charactergrams are filtered
60   // using this list, and charactergrams that are not present are interpreted as
61   // out-of-vocabulary.
62   // If no allowed_chargrams are specified, all charactergrams are allowed.
63   std::unordered_set<std::string> allowed_chargrams;
64 };
65 
66 class TokenFeatureExtractor {
67  public:
68   // Des not take ownership of unilib, which must refer to a valid unilib
69   // instance that outlives this feature extractor.
70   explicit TokenFeatureExtractor(const TokenFeatureExtractorOptions& options,
71                                  const UniLib* unilib);
72 
73   // Extracts both the sparse (charactergram) and the dense features from a
74   // token. is_in_span is a bool indicator whether the token is a part of the
75   // selection span (true) or not (false).
76   // The sparse_features output is optional. Fails and returns false if
77   // dense_fatures in a nullptr.
78   bool Extract(const Token& token, bool is_in_span,
79                std::vector<int>* sparse_features,
80                std::vector<float>* dense_features) const;
81 
82   // Extracts the sparse (charactergram) features from the token.
83   std::vector<int> ExtractCharactergramFeatures(const Token& token) const;
84 
85   // Extracts the dense features from the token. is_in_span is a bool indicator
86   // whether the token is a part of the selection span (true) or not (false).
87   std::vector<float> ExtractDenseFeatures(const Token& token,
88                                           bool is_in_span) const;
89 
DenseFeaturesCount()90   int DenseFeaturesCount() const {
91     int feature_count =
92         options_.extract_case_feature + options_.extract_selection_mask_feature;
93     feature_count += regex_patterns_.size();
94     return feature_count;
95   }
96 
97  protected:
98   // Hashes given token to given number of buckets.
99   int HashToken(StringPiece token) const;
100 
101   // Extracts the charactergram features from the token in a non-unicode-aware
102   // way.
103   std::vector<int> ExtractCharactergramFeaturesAscii(const Token& token) const;
104 
105   // Extracts the charactergram features from the token in a unicode-aware way.
106   std::vector<int> ExtractCharactergramFeaturesUnicode(
107       const Token& token) const;
108 
109  private:
110   TokenFeatureExtractorOptions options_;
111   std::vector<std::unique_ptr<UniLib::RegexPattern>> regex_patterns_;
112   const UniLib& unilib_;
113 };
114 
115 }  // namespace libtextclassifier3
116 
117 #endif  // LIBTEXTCLASSIFIER_UTILS_TOKEN_FEATURE_EXTRACTOR_H_
118