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_SENTENCEPIECE_ENCODER_H_ 18 #define LIBTEXTCLASSIFIER_UTILS_SENTENCEPIECE_ENCODER_H_ 19 20 #include <vector> 21 22 #include "utils/base/logging.h" 23 #include "utils/sentencepiece/matcher.h" 24 #include "utils/strings/stringpiece.h" 25 26 namespace libtextclassifier3 { 27 28 // Encoder to segment/tokenize strings into pieces such that the sum of the 29 // scores of the pieces used is maximized. 30 class Encoder { 31 public: 32 // matcher: the list of valid sentence pieces represented as a matcher, e.g. 33 // a trie. 34 // num_pieces: the number of pieces in the trie. 35 // pieces_scores: the scores of the individual pieces. 36 // start_code: code that is used as encoding of the start of input. 37 // end_code: code that is used as encoding of the end of input. 38 // encoding_offset: value added to the sentence piece ids to make them 39 // not interesecting with start_code and end_code. 40 // unknown_code: code that is used for out-of-dictionary characters. 41 // unknown_score: the penality score associated with the unknown code. 42 Encoder(const SentencePieceMatcher* matcher, const int num_pieces, 43 const float* pieces_scores, int start_code = 0, int end_code = 1, 44 int encoding_offset = 2, int unknown_code = -1, 45 float unknown_score = 0.f) num_pieces_(num_pieces)46 : num_pieces_(num_pieces), 47 scores_(pieces_scores), 48 matcher_(matcher), 49 start_code_(start_code), 50 end_code_(end_code), 51 encoding_offset_(encoding_offset), 52 unknown_code_(unknown_code), 53 unknown_score_(unknown_score) {} 54 55 // Segment the input so that the total score of the pieces used is maximized. 56 // This is a simplified implementation of the general Viterbi algorithm, 57 // assuming independence between individual pieces. 58 bool Encode(StringPiece normalized_text, 59 std::vector<int>* encoded_text) const; 60 61 private: 62 // State in the dynamic programming algorithm. 63 struct SegmentationEntry { 64 // Accumulated score. 65 float score; 66 67 // Position before last piece. 68 int previous_pos; 69 70 // Last piece used. 71 int piece_id; 72 73 // Total number of pieces used. 74 int num_pieces; 75 }; 76 77 const int num_pieces_; 78 const float* scores_; 79 const SentencePieceMatcher* matcher_; 80 const int start_code_; 81 const int end_code_; 82 const int encoding_offset_; 83 const int unknown_code_; 84 const int unknown_score_; 85 }; 86 87 } // namespace libtextclassifier3 88 89 #endif // LIBTEXTCLASSIFIER_UTILS_SENTENCEPIECE_ENCODER_H_ 90