Initial commit: FunASR Speech Recognition Toolkit

Add complete FunASR codebase including models, runtime, and documentation.

runtime/onnxruntime/third_party/jieba/include/cppjieba/DictTrie.hpp

@@ -0,0 +1,274 @@
+#ifndef CPPJIEBA_DICT_TRIE_HPP
+#define CPPJIEBA_DICT_TRIE_HPP
+
+#include <iostream>
+#include <fstream>
+#include <map>
+#include <string>
+#include <cstring>
+#include <cstdlib>
+#include <stdint.h>
+#include <cmath>
+#include <limits>
+#include "limonp/StringUtil.hpp"
+#include "limonp/Logging.hpp"
+#include "Unicode.hpp"
+#include "Trie.hpp"
+
+namespace cppjieba {
+
+using namespace limonp;
+
+const double MIN_DOUBLE = -3.14e+100;
+const double MAX_DOUBLE = 3.14e+100;
+const size_t DICT_COLUMN_NUM = 3;
+const char* const UNKNOWN_TAG = "";
+
+class DictTrie {
+ public:
+  enum UserWordWeightOption {
+    WordWeightMin,
+    WordWeightMedian,
+    WordWeightMax,
+  }; // enum UserWordWeightOption
+
+  DictTrie(const string& dict_path, const string& user_dict_paths = "", UserWordWeightOption user_word_weight_opt = WordWeightMedian) {
+    Init(dict_path, user_dict_paths, user_word_weight_opt);
+  }
+  ~DictTrie() {
+    delete trie_;
+  }
+
+  bool InsertUserWord(const string& word, const string& tag = UNKNOWN_TAG) {
+    DictUnit node_info;
+    if (!MakeNodeInfo(node_info, word, user_word_default_weight_, tag)) {
+      return false;
+    }
+    active_node_infos_.push_back(node_info);
+    trie_->InsertNode(node_info.word, &active_node_infos_.back());
+    return true;
+  }
+
+  bool InsertUserWord(const string& word,int freq, const string& tag = UNKNOWN_TAG) {
+    DictUnit node_info;
+    double weight = freq ? log(1.0 * freq / freq_sum_) : user_word_default_weight_ ;
+    if (!MakeNodeInfo(node_info, word, weight , tag)) {
+      return false;
+    }
+    active_node_infos_.push_back(node_info);
+    trie_->InsertNode(node_info.word, &active_node_infos_.back());
+    return true;
+  }
+
+  bool DeleteUserWord(const string& word, const string& tag = UNKNOWN_TAG) {
+    DictUnit node_info;
+    if (!MakeNodeInfo(node_info, word, user_word_default_weight_, tag)) {
+      return false;
+    }
+    trie_->DeleteNode(node_info.word, &node_info);
+    return true;
+  }
+  
+  const DictUnit* Find(RuneStrArray::const_iterator begin, RuneStrArray::const_iterator end) const {
+    return trie_->Find(begin, end);
+  }
+
+  void Find(RuneStrArray::const_iterator begin, 
+        RuneStrArray::const_iterator end, 
+        vector<struct Dag>&res,
+        size_t max_word_len = MAX_WORD_LENGTH) const {
+    trie_->Find(begin, end, res, max_word_len);
+  }
+
+  bool Find(const string& word)
+  {
+    const DictUnit *tmp = NULL;
+    RuneStrArray runes;
+    if (!DecodeRunesInString(word, runes))
+    {
+      XLOG(ERROR) << "Decode failed.";
+    }
+    tmp = Find(runes.begin(), runes.end());
+    if (tmp == NULL)
+    {
+      return false;
+    }
+    else
+    {
+      return true;
+    }
+  }
+
+  bool IsUserDictSingleChineseWord(const Rune& word) const {
+    return IsIn(user_dict_single_chinese_word_, word);
+  }
+
+  double GetMinWeight() const {
+    return min_weight_;
+  }
+
+  void InserUserDictNode(const string& line) {
+    vector<string> buf;
+    DictUnit node_info;
+    Split(line, buf, " ");
+    if(buf.size() == 1){
+          MakeNodeInfo(node_info, 
+                buf[0], 
+                user_word_default_weight_,
+                UNKNOWN_TAG);
+        } else if (buf.size() == 2) {
+          MakeNodeInfo(node_info, 
+                buf[0], 
+                user_word_default_weight_,
+                buf[1]);
+        } else if (buf.size() == 3) {
+          int freq = atoi(buf[1].c_str());
+          assert(freq_sum_ > 0.0);
+          double weight = log(1.0 * freq / freq_sum_);
+          MakeNodeInfo(node_info, buf[0], weight, buf[2]);
+        }
+        static_node_infos_.push_back(node_info);
+        if (node_info.word.size() == 1) {
+          user_dict_single_chinese_word_.insert(node_info.word[0]);
+        }
+  }
+  
+  void LoadUserDict(const vector<string>& buf) {
+    for (size_t i = 0; i < buf.size(); i++) {
+      InserUserDictNode(buf[i]);
+    }
+  }
+
+   void LoadUserDict(const set<string>& buf) {
+    std::set<string>::const_iterator iter;
+    for (iter = buf.begin(); iter != buf.end(); iter++){
+      InserUserDictNode(*iter);
+    }
+  }
+
+  void LoadUserDict(const string& filePaths) {
+    vector<string> files = limonp::Split(filePaths, "|;");
+    size_t lineno = 0;
+    for (size_t i = 0; i < files.size(); i++) {
+      ifstream ifs(files[i].c_str());
+      XCHECK(ifs.is_open()) << "open " << files[i] << " failed"; 
+      string line;
+      
+      for (; getline(ifs, line); lineno++) {
+        if (line.size() == 0) {
+          continue;
+        }
+        InserUserDictNode(line);
+      }
+    }
+  }
+
+
+ private:
+  void Init(const string& dict_path, const string& user_dict_paths, UserWordWeightOption user_word_weight_opt) {
+    LoadDict(dict_path);
+    Shrink(static_node_infos_);
+    CreateTrie(static_node_infos_);
+  }
+  
+  void CreateTrie(const vector<DictUnit>& dictUnits) {
+    assert(dictUnits.size());
+    vector<Unicode> words;
+    vector<const DictUnit*> valuePointers;
+    for (size_t i = 0 ; i < dictUnits.size(); i ++) {
+      words.push_back(dictUnits[i].word);
+      valuePointers.push_back(&dictUnits[i]);
+    }
+    trie_ = new Trie(words, valuePointers);
+  }
+
+  bool MakeNodeInfo(DictUnit& node_info,
+        const string& word, 
+        double weight, 
+        const string& tag) {
+    if (!DecodeRunesInString(word, node_info.word)) {
+      XLOG(ERROR) << "Decode " << word << " failed.";
+      return false;
+    }
+    node_info.weight = weight;
+    node_info.tag = tag;
+    return true;
+  }
+
+  void LoadDict(const string& filePath) {
+    ifstream ifs(filePath.c_str());
+    XCHECK(ifs.is_open()) << "open " << filePath << " failed.";
+    string line;
+    vector<string> buf;
+
+    DictUnit node_info;
+    for (size_t lineno = 0; getline(ifs, line); lineno++) {
+      Split(line, buf, " ");
+      XCHECK(buf.size() == DICT_COLUMN_NUM) << "split result illegal, line:" << line;
+      MakeNodeInfo(node_info, 
+            buf[0], 
+            atof(buf[1].c_str()), 
+            buf[2]);
+      static_node_infos_.push_back(node_info);
+    }
+  }
+
+  static bool WeightCompare(const DictUnit& lhs, const DictUnit& rhs) {
+    return lhs.weight < rhs.weight;
+  }
+
+  void SetStaticWordWeights(UserWordWeightOption option) {
+    XCHECK(!static_node_infos_.empty());
+    vector<DictUnit> x = static_node_infos_;
+    sort(x.begin(), x.end(), WeightCompare);
+    min_weight_ = x[0].weight;
+    max_weight_ = x[x.size() - 1].weight;
+    median_weight_ = x[x.size() / 2].weight;
+    switch (option) {
+     case WordWeightMin:
+       user_word_default_weight_ = min_weight_;
+       break;
+     case WordWeightMedian:
+       user_word_default_weight_ = median_weight_;
+       break;
+     default:
+       user_word_default_weight_ = max_weight_;
+       break;
+    }
+  }
+
+  double CalcFreqSum(const vector<DictUnit>& node_infos) const {
+    double sum = 0.0;
+    for (size_t i = 0; i < node_infos.size(); i++) {
+      sum += node_infos[i].weight;
+    }
+    return sum;
+  }
+
+  void CalculateWeight(vector<DictUnit>& node_infos, double sum) const {
+    assert(sum > 0.0);
+    for (size_t i = 0; i < node_infos.size(); i++) {
+      DictUnit& node_info = node_infos[i];
+      assert(node_info.weight > 0.0);
+      node_info.weight = log(double(node_info.weight)/sum);
+    }
+  }
+
+  void Shrink(vector<DictUnit>& units) const {
+    vector<DictUnit>(units.begin(), units.end()).swap(units);
+  }
+
+  vector<DictUnit> static_node_infos_;
+  deque<DictUnit> active_node_infos_; // must not be vector
+  Trie * trie_;
+
+  double freq_sum_;
+  double min_weight_;
+  double max_weight_;
+  double median_weight_;
+  double user_word_default_weight_;
+  unordered_set<Rune> user_dict_single_chinese_word_;
+};
+}
+
+#endif

runtime/onnxruntime/third_party/jieba/include/cppjieba/FullSegment.hpp

@@ -0,0 +1,102 @@
+#ifndef CPPJIEBA_FULLSEGMENT_H
+#define CPPJIEBA_FULLSEGMENT_H
+
+#include <algorithm>
+#include <set>
+#include <cassert>
+#include "limonp/Logging.hpp"
+#include "DictTrie.hpp"
+#include "SegmentBase.hpp"
+#include "Unicode.hpp"
+
+namespace cppjieba {
+class FullSegment: public SegmentBase {
+ public:
+  FullSegment(const string& dictPath) {
+    dictTrie_ = new DictTrie(dictPath);
+    isNeedDestroy_ = true;
+  }
+  FullSegment(const DictTrie* dictTrie)
+    : dictTrie_(dictTrie), isNeedDestroy_(false) {
+    assert(dictTrie_);
+  }
+  FullSegment() {
+    dictTrie_ = NULL;
+  }
+  ~FullSegment() {
+    if (isNeedDestroy_) {
+      delete dictTrie_;
+    }
+  }
+
+  void setRes(DictTrie *&dictTrie) {
+    dictTrie_ = dictTrie;
+    isNeedDestroy_ = false;
+    assert(dictTrie_);
+  }
+  void Cut(const string& sentence, 
+        vector<string>& words) const {
+    vector<Word> tmp;
+    Cut(sentence, tmp);
+    GetStringsFromWords(tmp, words);
+  }
+  void Cut(const string& sentence, 
+        vector<Word>& words) const {
+    PreFilter pre_filter(symbols_, sentence);
+    PreFilter::Range range;
+    vector<WordRange> wrs;
+    wrs.reserve(sentence.size()/2);
+    while (pre_filter.HasNext()) {
+      range = pre_filter.Next();
+      Cut(range.begin, range.end, wrs);
+    }
+    words.clear();
+    words.reserve(wrs.size());
+    GetWordsFromWordRanges(sentence, wrs, words);
+  }
+  void Cut(RuneStrArray::const_iterator begin, 
+        RuneStrArray::const_iterator end, 
+        vector<WordRange>& res) const {
+    // result of searching in trie tree
+    LocalVector<pair<size_t, const DictUnit*> > tRes;
+
+    // max index of res's words
+    size_t maxIdx = 0;
+
+    // always equals to (uItr - begin)
+    size_t uIdx = 0;
+
+    // tmp variables
+    size_t wordLen = 0;
+    assert(dictTrie_);
+    vector<struct Dag> dags;
+    dictTrie_->Find(begin, end, dags);
+    for (size_t i = 0; i < dags.size(); i++) {
+      for (size_t j = 0; j < dags[i].nexts.size(); j++) {
+        size_t nextoffset = dags[i].nexts[j].first;
+        assert(nextoffset < dags.size());
+        const DictUnit* du = dags[i].nexts[j].second;
+        if (du == NULL) {
+          if (dags[i].nexts.size() == 1 && maxIdx <= uIdx) {
+            WordRange wr(begin + i, begin + nextoffset);
+            res.push_back(wr);
+          }
+        } else {
+          wordLen = du->word.size();
+          if (wordLen >= 2 || (dags[i].nexts.size() == 1 && maxIdx <= uIdx)) {
+            WordRange wr(begin + i, begin + nextoffset);
+            res.push_back(wr);
+          }
+        }
+        maxIdx = uIdx + wordLen > maxIdx ? uIdx + wordLen : maxIdx;
+      }
+      uIdx++;
+    }
+  }
+ private:
+  const DictTrie* dictTrie_;
+  bool isNeedDestroy_;
+};
+}
+
+#endif

runtime/onnxruntime/third_party/jieba/include/cppjieba/HMMModel.hpp

@@ -0,0 +1,129 @@
+#ifndef CPPJIEBA_HMMMODEL_H
+#define CPPJIEBA_HMMMODEL_H
+
+#include "limonp/StringUtil.hpp"
+#include "Trie.hpp"
+
+namespace cppjieba {
+
+using namespace limonp;
+typedef unordered_map<Rune, double> EmitProbMap;
+
+struct HMMModel {
+  /*
+   * STATUS:
+   * 0: HMMModel::B, 1: HMMModel::E, 2: HMMModel::M, 3:HMMModel::S
+   * */
+  enum {B = 0, E = 1, M = 2, S = 3, STATUS_SUM = 4};
+
+  HMMModel(const string& modelPath) {
+    memset(startProb, 0, sizeof(startProb));
+    memset(transProb, 0, sizeof(transProb));
+    statMap[0] = 'B';
+    statMap[1] = 'E';
+    statMap[2] = 'M';
+    statMap[3] = 'S';
+    emitProbVec.push_back(&emitProbB);
+    emitProbVec.push_back(&emitProbE);
+    emitProbVec.push_back(&emitProbM);
+    emitProbVec.push_back(&emitProbS);
+    LoadModel(modelPath);
+  }
+  ~HMMModel() {
+  }
+  void LoadModel(const string& filePath) {
+    ifstream ifile(filePath.c_str());
+    XCHECK(ifile.is_open()) << "open " << filePath << " failed";
+    string line;
+    vector<string> tmp;
+    vector<string> tmp2;
+    //Load startProb
+    XCHECK(GetLine(ifile, line));
+    Split(line, tmp, " ");
+    XCHECK(tmp.size() == STATUS_SUM);
+    for (size_t j = 0; j< tmp.size(); j++) {
+      startProb[j] = atof(tmp[j].c_str());
+    }
+
+    //Load transProb
+    for (size_t i = 0; i < STATUS_SUM; i++) {
+      XCHECK(GetLine(ifile, line));
+      Split(line, tmp, " ");
+      XCHECK(tmp.size() == STATUS_SUM);
+      for (size_t j =0; j < STATUS_SUM; j++) {
+        transProb[i][j] = atof(tmp[j].c_str());
+      }
+    }
+
+    //Load emitProbB
+    XCHECK(GetLine(ifile, line));
+    XCHECK(LoadEmitProb(line, emitProbB));
+
+    //Load emitProbE
+    XCHECK(GetLine(ifile, line));
+    XCHECK(LoadEmitProb(line, emitProbE));
+
+    //Load emitProbM
+    XCHECK(GetLine(ifile, line));
+    XCHECK(LoadEmitProb(line, emitProbM));
+
+    //Load emitProbS
+    XCHECK(GetLine(ifile, line));
+    XCHECK(LoadEmitProb(line, emitProbS));
+  }
+  double GetEmitProb(const EmitProbMap* ptMp, Rune key, 
+        double defVal)const {
+    EmitProbMap::const_iterator cit = ptMp->find(key);
+    if (cit == ptMp->end()) {
+      return defVal;
+    }
+    return cit->second;
+  }
+  bool GetLine(ifstream& ifile, string& line) {
+    while (getline(ifile, line)) {
+      Trim(line);
+      if (line.empty()) {
+        continue;
+      }
+      if (StartsWith(line, "#")) {
+        continue;
+      }
+      return true;
+    }
+    return false;
+  }
+  bool LoadEmitProb(const string& line, EmitProbMap& mp) {
+    if (line.empty()) {
+      return false;
+    }
+    vector<string> tmp, tmp2;
+    Unicode unicode;
+    Split(line, tmp, ",");
+    for (size_t i = 0; i < tmp.size(); i++) {
+      Split(tmp[i], tmp2, ":");
+      if (2 != tmp2.size()) {
+        XLOG(ERROR) << "emitProb illegal.";
+        return false;
+      }
+      if (!DecodeRunesInString(tmp2[0], unicode) || unicode.size() != 1) {
+        XLOG(ERROR) << "TransCode failed.";
+        return false;
+      }
+      mp[unicode[0]] = atof(tmp2[1].c_str());
+    }
+    return true;
+  }
+
+  char statMap[STATUS_SUM];
+  double startProb[STATUS_SUM];
+  double transProb[STATUS_SUM][STATUS_SUM];
+  EmitProbMap emitProbB;
+  EmitProbMap emitProbE;
+  EmitProbMap emitProbM;
+  EmitProbMap emitProbS;
+  vector<EmitProbMap* > emitProbVec;
+}; // struct HMMModel
+
+} // namespace cppjieba
+
+#endif

runtime/onnxruntime/third_party/jieba/include/cppjieba/HMMSegment.hpp

@@ -0,0 +1,197 @@
+#ifndef CPPJIBEA_HMMSEGMENT_H
+#define CPPJIBEA_HMMSEGMENT_H
+
+#include <iostream>
+#include <fstream>
+#include <memory.h>
+#include <cassert>
+#include "HMMModel.hpp"
+#include "SegmentBase.hpp"
+
+namespace cppjieba {
+class HMMSegment: public SegmentBase {
+ public:
+  HMMSegment(const string& filePath)
+  : model_(new HMMModel(filePath)), isNeedDestroy_(true) {
+  }
+  HMMSegment(const HMMModel* model) 
+  : model_(model), isNeedDestroy_(false) {
+  }
+  HMMSegment() {
+    model_ = NULL;
+  }
+
+  ~HMMSegment() {
+    if (isNeedDestroy_) {
+      delete model_;
+    }
+  }
+  void setRes(HMMModel *&model) {
+    model_ = model;
+    isNeedDestroy_ = false;
+  }
+  void Cut(const string& sentence, 
+        vector<string>& words) const {
+    vector<Word> tmp;
+    Cut(sentence, tmp);
+    GetStringsFromWords(tmp, words);
+  }
+  void Cut(const string& sentence, 
+        vector<Word>& words) const {
+    PreFilter pre_filter(symbols_, sentence);
+    PreFilter::Range range;
+    vector<WordRange> wrs;
+    wrs.reserve(sentence.size()/2);
+    while (pre_filter.HasNext()) {
+      range = pre_filter.Next();
+      Cut(range.begin, range.end, wrs);
+    }
+    words.clear();
+    words.reserve(wrs.size());
+    GetWordsFromWordRanges(sentence, wrs, words);
+  }
+  void Cut(RuneStrArray::const_iterator begin, RuneStrArray::const_iterator end, vector<WordRange>& res) const {
+    RuneStrArray::const_iterator left = begin;
+    RuneStrArray::const_iterator right = begin;
+    while (right != end) {
+      if (right->rune < 0x80) {
+        if (left != right) {
+          InternalCut(left, right, res);
+        }
+        left = right;
+        do {
+          right = SequentialLetterRule(left, end);
+          if (right != left) {
+            break;
+          }
+          right = NumbersRule(left, end);
+          if (right != left) {
+            break;
+          }
+          right ++;
+        } while (false);
+        WordRange wr(left, right - 1);
+        res.push_back(wr);
+        left = right;
+      } else {
+        right++;
+      }
+    }
+    if (left != right) {
+      InternalCut(left, right, res);
+    }
+  }
+ private:
+  // sequential letters rule
+  RuneStrArray::const_iterator SequentialLetterRule(RuneStrArray::const_iterator begin, RuneStrArray::const_iterator end) const {
+    Rune x = begin->rune;
+    if (('a' <= x && x <= 'z') || ('A' <= x && x <= 'Z')) {
+      begin ++;
+    } else {
+      return begin;
+    }
+    while (begin != end) {
+      x = begin->rune;
+      if (('a' <= x && x <= 'z') || ('A' <= x && x <= 'Z') || ('0' <= x && x <= '9')) {
+        begin ++;
+      } else {
+        break;
+      }
+    }
+    return begin;
+  }
+  //
+  RuneStrArray::const_iterator NumbersRule(RuneStrArray::const_iterator begin, RuneStrArray::const_iterator end) const {
+    Rune x = begin->rune;
+    if ('0' <= x && x <= '9') {
+      begin ++;
+    } else {
+      return begin;
+    }
+    while (begin != end) {
+      x = begin->rune;
+      if ( ('0' <= x && x <= '9') || x == '.') {
+        begin++;
+      } else {
+        break;
+      }
+    }
+    return begin;
+  }
+  void InternalCut(RuneStrArray::const_iterator begin, RuneStrArray::const_iterator end, vector<WordRange>& res) const {
+    vector<size_t> status;
+    Viterbi(begin, end, status);
+
+    RuneStrArray::const_iterator left = begin;
+    RuneStrArray::const_iterator right;
+    for (size_t i = 0; i < status.size(); i++) {
+      if (status[i] % 2) { //if (HMMModel::E == status[i] || HMMModel::S == status[i])
+        right = begin + i + 1;
+        WordRange wr(left, right - 1);
+        res.push_back(wr);
+        left = right;
+      }
+    }
+  }
+
+  void Viterbi(RuneStrArray::const_iterator begin, 
+        RuneStrArray::const_iterator end, 
+        vector<size_t>& status) const {
+    size_t Y = HMMModel::STATUS_SUM;
+    size_t X = end - begin;
+
+    size_t XYSize = X * Y;
+    size_t now, old, stat;
+    double tmp, endE, endS;
+
+    vector<int> path(XYSize);
+    vector<double> weight(XYSize);
+
+    //start
+    for (size_t y = 0; y < Y; y++) {
+      weight[0 + y * X] = model_->startProb[y] + model_->GetEmitProb(model_->emitProbVec[y], begin->rune, MIN_DOUBLE);
+      path[0 + y * X] = -1;
+    }
+
+    double emitProb;
+
+    for (size_t x = 1; x < X; x++) {
+      for (size_t y = 0; y < Y; y++) {
+        now = x + y*X;
+        weight[now] = MIN_DOUBLE;
+        path[now] = HMMModel::E; // warning
+        emitProb = model_->GetEmitProb(model_->emitProbVec[y], (begin+x)->rune, MIN_DOUBLE);
+        for (size_t preY = 0; preY < Y; preY++) {
+          old = x - 1 + preY * X;
+          tmp = weight[old] + model_->transProb[preY][y] + emitProb;
+          if (tmp > weight[now]) {
+            weight[now] = tmp;
+            path[now] = preY;
+          }
+        }
+      }
+    }
+
+    endE = weight[X-1+HMMModel::E*X];
+    endS = weight[X-1+HMMModel::S*X];
+    stat = 0;
+    if (endE >= endS) {
+      stat = HMMModel::E;
+    } else {
+      stat = HMMModel::S;
+    }
+
+    status.resize(X);
+    for (int x = X -1 ; x >= 0; x--) {
+      status[x] = stat;
+      stat = path[x + stat*X];
+    }
+  }
+
+  const HMMModel* model_;
+  bool isNeedDestroy_;
+}; // class HMMSegment
+
+} // namespace cppjieba
+
+#endif

runtime/onnxruntime/third_party/jieba/include/cppjieba/Jieba.hpp

@@ -0,0 +1,141 @@
+#ifndef CPPJIEAB_JIEBA_H
+#define CPPJIEAB_JIEBA_H
+
+#include "QuerySegment.hpp"
+#include "KeywordExtractor.hpp"
+
+namespace cppjieba {
+
+class Jieba {
+ public:
+  Jieba(DictTrie *jieba_dict_trie, 
+        HMMModel *jieba_model)
+    : dict_trie_(jieba_dict_trie),
+      model_(jieba_model),
+      mp_seg_(dict_trie_),
+      hmm_seg_(model_),
+      mix_seg_(dict_trie_, model_),
+      full_seg_(dict_trie_),
+      query_seg_(dict_trie_, model_) {
+  }
+  Jieba() {
+    dict_trie_ = NULL;
+    model_ = NULL;
+  }
+  ~Jieba() {
+  }
+
+  struct LocWord {
+    string word;
+    size_t begin;
+    size_t end;
+  }; // struct LocWord
+  void SetJiebaRes(cppjieba::DictTrie *&dict, cppjieba::HMMModel *&hmm) {
+    dict_trie_ = dict;
+    model_ = hmm;
+    mp_seg_.setRes(dict);
+    hmm_seg_.setRes(hmm);
+    mix_seg_.setRes(dict, hmm);
+    full_seg_.setRes(dict);
+    query_seg_.setRes(dict, hmm);
+  }
+  void Cut(const string& sentence, vector<string>& words, bool hmm = true) const {
+    mix_seg_.Cut(sentence, words, hmm);
+  }
+  void Cut(const string& sentence, vector<Word>& words, bool hmm = true) const {
+    mix_seg_.Cut(sentence, words, hmm);
+  }
+  void CutAll(const string& sentence, vector<string>& words) const {
+    full_seg_.Cut(sentence, words);
+  }
+  void CutAll(const string& sentence, vector<Word>& words) const {
+    full_seg_.Cut(sentence, words);
+  }
+  void CutForSearch(const string& sentence, vector<string>& words, bool hmm = true) const {
+    query_seg_.Cut(sentence, words, hmm);
+  }
+  void CutForSearch(const string& sentence, vector<Word>& words, bool hmm = true) const {
+    query_seg_.Cut(sentence, words, hmm);
+  }
+  void CutHMM(const string& sentence, vector<string>& words) const {
+    hmm_seg_.Cut(sentence, words);
+  }
+  void CutHMM(const string& sentence, vector<Word>& words) const {
+    hmm_seg_.Cut(sentence, words);
+  }
+  void CutSmall(const string& sentence, vector<string>& words, size_t max_word_len) const {
+    mp_seg_.Cut(sentence, words, max_word_len);
+  }
+  void CutSmall(const string& sentence, vector<Word>& words, size_t max_word_len) const {
+    mp_seg_.Cut(sentence, words, max_word_len);
+  }
+  
+  void Tag(const string& sentence, vector<pair<string, string> >& words) const {
+    mix_seg_.Tag(sentence, words);
+  }
+  string LookupTag(const string &str) const {
+    return mix_seg_.LookupTag(str);
+  }
+  bool InsertUserWord(const string& word, const string& tag = UNKNOWN_TAG) {
+    return dict_trie_->InsertUserWord(word, tag);
+  }
+
+  bool InsertUserWord(const string& word,int freq, const string& tag = UNKNOWN_TAG) {
+    return dict_trie_->InsertUserWord(word,freq, tag);
+  }
+
+  bool DeleteUserWord(const string& word, const string& tag = UNKNOWN_TAG) {
+    return dict_trie_->DeleteUserWord(word, tag);
+  }
+  
+  bool Find(const string& word)
+  {
+    return dict_trie_->Find(word);
+  }
+
+  void ResetSeparators(const string& s) {
+    //TODO
+    mp_seg_.ResetSeparators(s);
+    hmm_seg_.ResetSeparators(s);
+    mix_seg_.ResetSeparators(s);
+    full_seg_.ResetSeparators(s);
+    query_seg_.ResetSeparators(s);
+  }
+
+  const DictTrie* GetDictTrie() const {
+    return dict_trie_;
+  } 
+  
+  const HMMModel* GetHMMModel() const {
+    return model_;
+  }
+
+  void LoadUserDict(const vector<string>& buf)  {
+    dict_trie_->LoadUserDict(buf);
+  }
+
+  void LoadUserDict(const set<string>& buf)  {
+    dict_trie_->LoadUserDict(buf);
+  }
+
+  void LoadUserDict(const string& path)  {
+    dict_trie_->LoadUserDict(path);
+  }
+
+ private:
+  DictTrie *dict_trie_;
+  HMMModel *model_;
+  
+  // They share the same dict trie and model
+  MPSegment mp_seg_;
+  HMMSegment hmm_seg_;
+  MixSegment mix_seg_;
+  FullSegment full_seg_;
+  QuerySegment query_seg_;
+
+ public:
+}; // class Jieba
+
+} // namespace cppjieba
+
+#endif // CPPJIEAB_JIEBA_H

runtime/onnxruntime/third_party/jieba/include/cppjieba/KeywordExtractor.hpp

@@ -0,0 +1,154 @@
+#ifndef CPPJIEBA_KEYWORD_EXTRACTOR_H
+#define CPPJIEBA_KEYWORD_EXTRACTOR_H
+
+#include <cmath>
+#include <set>
+#include "MixSegment.hpp"
+
+namespace cppjieba {
+
+using namespace limonp;
+using namespace std;
+
+/*utf8*/
+class KeywordExtractor {
+ public:
+  struct Word {
+    string word;
+    vector<size_t> offsets;
+    double weight;
+  }; // struct Word
+
+  KeywordExtractor(const string& dictPath, 
+        const string& hmmFilePath, 
+        const string& idfPath, 
+        const string& stopWordPath, 
+        const string& userDict = "") 
+    : segment_(dictPath, hmmFilePath, userDict) {
+    LoadIdfDict(idfPath);
+    LoadStopWordDict(stopWordPath);
+  }
+  KeywordExtractor(const DictTrie* dictTrie, 
+        const HMMModel* model,
+        const string& idfPath, 
+        const string& stopWordPath) 
+    : segment_(dictTrie, model) {
+    LoadIdfDict(idfPath);
+    LoadStopWordDict(stopWordPath);
+  }
+  KeywordExtractor() {}
+  ~KeywordExtractor() {
+  }
+
+  void Extract(const string& sentence, vector<string>& keywords, size_t topN) const {
+    vector<Word> topWords;
+    Extract(sentence, topWords, topN);
+    for (size_t i = 0; i < topWords.size(); i++) {
+      keywords.push_back(topWords[i].word);
+    }
+  }
+
+  void Extract(const string& sentence, vector<pair<string, double> >& keywords, size_t topN) const {
+    vector<Word> topWords;
+    Extract(sentence, topWords, topN);
+    for (size_t i = 0; i < topWords.size(); i++) {
+      keywords.push_back(pair<string, double>(topWords[i].word, topWords[i].weight));
+    }
+  }
+
+  void Extract(const string& sentence, vector<Word>& keywords, size_t topN) const {
+    vector<string> words;
+    segment_.Cut(sentence, words);
+
+    map<string, Word> wordmap;
+    size_t offset = 0;
+    for (size_t i = 0; i < words.size(); ++i) {
+      size_t t = offset;
+      offset += words[i].size();
+      if (IsSingleWord(words[i]) || stopWords_.find(words[i]) != stopWords_.end()) {
+        continue;
+      }
+      wordmap[words[i]].offsets.push_back(t);
+      wordmap[words[i]].weight += 1.0;
+    }
+    if (offset != sentence.size()) {
+      XLOG(ERROR) << "words illegal";
+      return;
+    }
+
+    keywords.clear();
+    keywords.reserve(wordmap.size());
+    for (map<string, Word>::iterator itr = wordmap.begin(); itr != wordmap.end(); ++itr) {
+      unordered_map<string, double>::const_iterator cit = idfMap_.find(itr->first);
+      if (cit != idfMap_.end()) {
+        itr->second.weight *= cit->second;
+      } else {
+        itr->second.weight *= idfAverage_;
+      }
+      itr->second.word = itr->first;
+      keywords.push_back(itr->second);
+    }
+    topN = min(topN, keywords.size());
+    partial_sort(keywords.begin(), keywords.begin() + topN, keywords.end(), Compare);
+    keywords.resize(topN);
+  }
+ private:
+  void LoadIdfDict(const string& idfPath) {
+    ifstream ifs(idfPath.c_str());
+    XCHECK(ifs.is_open()) << "open " << idfPath << " failed";
+    string line ;
+    vector<string> buf;
+    double idf = 0.0;
+    double idfSum = 0.0;
+    size_t lineno = 0;
+    for (; getline(ifs, line); lineno++) {
+      buf.clear();
+      if (line.empty()) {
+        XLOG(ERROR) << "lineno: " << lineno << " empty. skipped.";
+        continue;
+      }
+      Split(line, buf, " ");
+      if (buf.size() != 2) {
+        XLOG(ERROR) << "line: " << line << ", lineno: " << lineno << " empty. skipped.";
+        continue;
+      }
+      idf = atof(buf[1].c_str());
+      idfMap_[buf[0]] = idf;
+      idfSum += idf;
+
+    }
+
+    assert(lineno);
+    idfAverage_ = idfSum / lineno;
+    assert(idfAverage_ > 0.0);
+  }
+  void LoadStopWordDict(const string& filePath) {
+    ifstream ifs(filePath.c_str());
+    XCHECK(ifs.is_open()) << "open " << filePath << " failed";
+    string line ;
+    while (getline(ifs, line)) {
+      stopWords_.insert(line);
+    }
+    assert(stopWords_.size());
+  }
+
+  static bool Compare(const Word& lhs, const Word& rhs) {
+    return lhs.weight > rhs.weight;
+  }
+
+  MixSegment segment_;
+  unordered_map<string, double> idfMap_;
+  double idfAverage_;
+
+  unordered_set<string> stopWords_;
+}; // class KeywordExtractor
+
+inline ostream& operator << (ostream& os, const KeywordExtractor::Word& word) {
+  return os << "{\"word\": \"" << word.word << "\", \"offset\": " << word.offsets << ", \"weight\": " << word.weight << "}"; 
+}
+
+} // namespace cppjieba
+
+#endif
+
+

runtime/onnxruntime/third_party/jieba/include/cppjieba/MPSegment.hpp

@@ -0,0 +1,144 @@
+#ifndef CPPJIEBA_MPSEGMENT_H
+#define CPPJIEBA_MPSEGMENT_H
+
+#include <algorithm>
+#include <set>
+#include <cassert>
+#include "limonp/Logging.hpp"
+#include "DictTrie.hpp"
+#include "SegmentTagged.hpp"
+#include "PosTagger.hpp"
+
+namespace cppjieba {
+
+class MPSegment: public SegmentTagged {
+ public:
+  MPSegment(const string& dictPath, const string& userDictPath = "")
+    : dictTrie_(new DictTrie(dictPath, userDictPath)), isNeedDestroy_(true) {
+  }
+  MPSegment(const DictTrie* dictTrie)
+    : dictTrie_(dictTrie), isNeedDestroy_(false) {
+    assert(dictTrie_);
+  }
+  MPSegment() {
+    dictTrie_ = NULL;
+  }
+  ~MPSegment() {
+    if (isNeedDestroy_) {
+      delete dictTrie_;
+    }
+  }
+  void setRes(DictTrie *&dictTrie) {
+    dictTrie_ = dictTrie;
+    isNeedDestroy_ = false;
+    assert(dictTrie_);
+  }
+  void Cut(const string& sentence, vector<string>& words) const {
+    Cut(sentence, words, MAX_WORD_LENGTH);
+  }
+
+  void Cut(const string& sentence,
+        vector<string>& words,
+        size_t max_word_len) const {
+    vector<Word> tmp;
+    Cut(sentence, tmp, max_word_len);
+    GetStringsFromWords(tmp, words);
+  }
+  void Cut(const string& sentence, 
+        vector<Word>& words, 
+        size_t max_word_len = MAX_WORD_LENGTH) const {
+    PreFilter pre_filter(symbols_, sentence);
+    PreFilter::Range range;
+    vector<WordRange> wrs;
+    wrs.reserve(sentence.size()/2);
+    while (pre_filter.HasNext()) {
+      range = pre_filter.Next();
+      Cut(range.begin, range.end, wrs, max_word_len);
+    }
+    words.clear();
+    words.reserve(wrs.size());
+    GetWordsFromWordRanges(sentence, wrs, words);
+  }
+  void Cut(RuneStrArray::const_iterator begin,
+           RuneStrArray::const_iterator end,
+           vector<WordRange>& words,
+           size_t max_word_len = MAX_WORD_LENGTH) const {
+    vector<Dag> dags;
+    dictTrie_->Find(begin, 
+          end, 
+          dags,
+          max_word_len);
+    CalcDP(dags);
+    CutByDag(begin, end, dags, words);
+  }
+
+  const DictTrie* GetDictTrie() const {
+    return dictTrie_;
+  }
+
+  bool Tag(const string& src, vector<pair<string, string> >& res) const {
+    return tagger_.Tag(src, res, *this);
+  }
+
+  bool IsUserDictSingleChineseWord(const Rune& value) const {
+    return dictTrie_->IsUserDictSingleChineseWord(value);
+  }
+ private:
+  void CalcDP(vector<Dag>& dags) const {
+    size_t nextPos;
+    const DictUnit* p;
+    double val;
+
+    for (vector<Dag>::reverse_iterator rit = dags.rbegin(); rit != dags.rend(); rit++) {
+      rit->pInfo = NULL;
+      rit->weight = MIN_DOUBLE;
+      assert(!rit->nexts.empty());
+      for (LocalVector<pair<size_t, const DictUnit*> >::const_iterator it = rit->nexts.begin(); it != rit->nexts.end(); it++) {
+        nextPos = it->first;
+        p = it->second;
+        val = 0.0;
+        if (nextPos + 1 < dags.size()) {
+          val += dags[nextPos + 1].weight;
+        }
+
+        if (p) {
+          val += p->weight;
+        } else {
+          val += dictTrie_->GetMinWeight();
+        }
+        if (val > rit->weight) {
+          rit->pInfo = p;
+          rit->weight = val;
+        }
+      }
+    }
+  }
+  void CutByDag(RuneStrArray::const_iterator begin, 
+        RuneStrArray::const_iterator end, 
+        const vector<Dag>& dags, 
+        vector<WordRange>& words) const {
+    size_t i = 0;
+    while (i < dags.size()) {
+      const DictUnit* p = dags[i].pInfo;
+      if (p) {
+        assert(p->word.size() >= 1);
+        WordRange wr(begin + i, begin + i + p->word.size() - 1);
+        words.push_back(wr);
+        i += p->word.size();
+      } else { //single chinese word
+        WordRange wr(begin + i, begin + i);
+        words.push_back(wr);
+        i++;
+      }
+    }
+  }
+
+  const DictTrie* dictTrie_;
+  bool isNeedDestroy_;
+  PosTagger tagger_;
+
+}; // class MPSegment
+
+} // namespace cppjieba
+
+#endif

runtime/onnxruntime/third_party/jieba/include/cppjieba/MixSegment.hpp

@@ -0,0 +1,113 @@
+#ifndef CPPJIEBA_MIXSEGMENT_H
+#define CPPJIEBA_MIXSEGMENT_H
+
+#include <cassert>
+#include "MPSegment.hpp"
+#include "HMMSegment.hpp"
+#include "limonp/StringUtil.hpp"
+#include "PosTagger.hpp"
+
+namespace cppjieba {
+class MixSegment: public SegmentTagged {
+ public:
+  MixSegment(const string& mpSegDict, const string& hmmSegDict, 
+        const string& userDict = "") 
+    : mpSeg_(mpSegDict, userDict), 
+      hmmSeg_(hmmSegDict) {
+  }
+  MixSegment(const DictTrie* dictTrie, const HMMModel* model) 
+    : mpSeg_(dictTrie), hmmSeg_(model) {
+  }
+  MixSegment() {}
+  ~MixSegment() {
+  }
+  void setRes(DictTrie *&dictTrie, HMMModel *&model) {
+    mpSeg_.setRes(dictTrie);
+    hmmSeg_.setRes(model);
+  }
+  void Cut(const string& sentence, vector<string>& words) const {
+    Cut(sentence, words, true);
+  }
+  void Cut(const string& sentence, vector<string>& words, bool hmm) const {
+    vector<Word> tmp;
+    Cut(sentence, tmp, hmm);
+    GetStringsFromWords(tmp, words);
+  }
+  void Cut(const string& sentence, vector<Word>& words, bool hmm = true) const {
+    PreFilter pre_filter(symbols_, sentence);
+    PreFilter::Range range;
+    vector<WordRange> wrs;
+    wrs.reserve(sentence.size() / 2);
+    while (pre_filter.HasNext()) {
+      range = pre_filter.Next();
+      Cut(range.begin, range.end, wrs, hmm);
+    }
+    words.clear();
+    words.reserve(wrs.size());
+    GetWordsFromWordRanges(sentence, wrs, words);
+  }
+
+  void Cut(RuneStrArray::const_iterator begin, RuneStrArray::const_iterator end, vector<WordRange>& res, bool hmm) const {
+    if (!hmm) {
+      mpSeg_.Cut(begin, end, res);
+      return;
+    }
+    vector<WordRange> words;
+    assert(end >= begin);
+    words.reserve(end - begin);
+    mpSeg_.Cut(begin, end, words);
+
+    vector<WordRange> hmmRes;
+    hmmRes.reserve(end - begin);
+    for (size_t i = 0; i < words.size(); i++) {
+      //if mp Get a word, it's ok, put it into result
+      if (words[i].left != words[i].right || (words[i].left == words[i].right && mpSeg_.IsUserDictSingleChineseWord(words[i].left->rune))) {
+        res.push_back(words[i]);
+        continue;
+      }
+
+      // if mp Get a single one and it is not in userdict, collect it in sequence
+      size_t j = i;
+      while (j < words.size() && words[j].left == words[j].right && !mpSeg_.IsUserDictSingleChineseWord(words[j].left->rune)) {
+        j++;
+      }
+
+      // Cut the sequence with hmm
+      assert(j - 1 >= i);
+      // TODO
+      hmmSeg_.Cut(words[i].left, words[j - 1].left + 1, hmmRes);
+      //put hmm result to result
+      for (size_t k = 0; k < hmmRes.size(); k++) {
+        res.push_back(hmmRes[k]);
+      }
+
+      //clear tmp vars
+      hmmRes.clear();
+
+      //let i jump over this piece
+      i = j - 1;
+    }
+  }
+
+  const DictTrie* GetDictTrie() const {
+    return mpSeg_.GetDictTrie();
+  }
+
+  bool Tag(const string& src, vector<pair<string, string> >& res) const {
+    return tagger_.Tag(src, res, *this);
+  }
+
+  string LookupTag(const string &str) const {
+    return tagger_.LookupTag(str, *this);
+  }
+
+ private:
+  MPSegment mpSeg_;
+  HMMSegment hmmSeg_;
+  PosTagger tagger_;
+
+}; // class MixSegment
+
+} // namespace cppjieba
+
+#endif

runtime/onnxruntime/third_party/jieba/include/cppjieba/PosTagger.hpp

@@ -0,0 +1,77 @@
+#ifndef CPPJIEBA_POS_TAGGING_H
+#define CPPJIEBA_POS_TAGGING_H
+
+#include "limonp/StringUtil.hpp"
+#include "SegmentTagged.hpp"
+#include "DictTrie.hpp"
+
+namespace cppjieba {
+using namespace limonp;
+
+static const char* const POS_M = "m";
+static const char* const POS_ENG = "eng";
+static const char* const POS_X = "x";
+
+class PosTagger {
+ public:
+  PosTagger() {
+  }
+  ~PosTagger() {
+  }
+
+  bool Tag(const string& src, vector<pair<string, string> >& res, const SegmentTagged& segment) const {
+    vector<string> CutRes;
+    segment.Cut(src, CutRes);
+
+    for (vector<string>::iterator itr = CutRes.begin(); itr != CutRes.end(); ++itr) {
+      res.push_back(make_pair(*itr, LookupTag(*itr, segment)));
+    }
+    return !res.empty();
+  }
+
+  string LookupTag(const string &str, const SegmentTagged& segment) const {
+    const DictUnit *tmp = NULL;
+    RuneStrArray runes;
+    const DictTrie * dict = segment.GetDictTrie();
+    assert(dict != NULL);
+      if (!DecodeRunesInString(str, runes)) {
+        XLOG(ERROR) << "Decode failed.";
+        return POS_X;
+      }
+      tmp = dict->Find(runes.begin(), runes.end());
+      if (tmp == NULL || tmp->tag.empty()) {
+        return SpecialRule(runes);
+      } else {
+        return tmp->tag;
+      }
+  }
+
+ private:
+  const char* SpecialRule(const RuneStrArray& unicode) const {
+    size_t m = 0;
+    size_t eng = 0;
+    for (size_t i = 0; i < unicode.size() && eng < unicode.size() / 2; i++) {
+      if (unicode[i].rune < 0x80) {
+        eng ++;
+        if ('0' <= unicode[i].rune && unicode[i].rune <= '9') {
+          m++;
+        }
+      }
+    }
+    // ascii char is not found
+    if (eng == 0) {
+      return POS_X;
+    }
+    // all the ascii is number char
+    if (m == eng) {
+      return POS_M;
+    }
+    // the ascii chars contain english letter
+    return POS_ENG;
+  }
+
+}; // class PosTagger
+
+} // namespace cppjieba
+
+#endif

runtime/onnxruntime/third_party/jieba/include/cppjieba/PreFilter.hpp

@@ -0,0 +1,54 @@
+#ifndef CPPJIEBA_PRE_FILTER_H
+#define CPPJIEBA_PRE_FILTER_H
+
+#include "Trie.hpp"
+#include "limonp/Logging.hpp"
+
+namespace cppjieba {
+
+class PreFilter {
+ public:
+  //TODO use WordRange instead of Range
+  struct Range {
+    RuneStrArray::const_iterator begin;
+    RuneStrArray::const_iterator end;
+  }; // struct Range
+
+  PreFilter(const unordered_set<Rune>& symbols, 
+        const string& sentence)
+    : symbols_(symbols) {
+    if (!DecodeRunesInString(sentence, sentence_)) {
+      XLOG(ERROR) << "decode failed. "; 
+    }
+    cursor_ = sentence_.begin();
+  }
+  ~PreFilter() {
+  }
+  bool HasNext() const {
+    return cursor_ != sentence_.end();
+  }
+  Range Next() {
+    Range range;
+    range.begin = cursor_;
+    while (cursor_ != sentence_.end()) {
+      if (IsIn(symbols_, cursor_->rune)) {
+        if (range.begin == cursor_) {
+          cursor_ ++;
+        }
+        range.end = cursor_;
+        return range;
+      }
+      cursor_ ++;
+    }
+    range.end = sentence_.end();
+    return range;
+  }
+ private:
+  RuneStrArray::const_iterator cursor_;
+  RuneStrArray sentence_;
+  const unordered_set<Rune>& symbols_;
+}; // class PreFilter
+
+} // namespace cppjieba
+
+#endif // CPPJIEBA_PRE_FILTER_H

runtime/onnxruntime/third_party/jieba/include/cppjieba/QuerySegment.hpp

@@ -0,0 +1,95 @@
+#ifndef CPPJIEBA_QUERYSEGMENT_H
+#define CPPJIEBA_QUERYSEGMENT_H
+
+#include <algorithm>
+#include <set>
+#include <cassert>
+#include "limonp/Logging.hpp"
+#include "DictTrie.hpp"
+#include "SegmentBase.hpp"
+#include "FullSegment.hpp"
+#include "MixSegment.hpp"
+#include "Unicode.hpp"
+
+namespace cppjieba {
+class QuerySegment: public SegmentBase {
+ public:
+  QuerySegment(const string& dict, const string& model, const string& userDict = "")
+    : mixSeg_(dict, model, userDict),
+      trie_(mixSeg_.GetDictTrie()) {
+  }
+  QuerySegment(const DictTrie* dictTrie, const HMMModel* model)
+    : mixSeg_(dictTrie, model), trie_(dictTrie) {
+  }
+  QuerySegment() {
+    trie_ = NULL;
+  }
+  ~QuerySegment() {
+  }
+  void setRes(DictTrie *&dictTrie, HMMModel *&model) {
+    mixSeg_.setRes(dictTrie, model);
+    trie_ = dictTrie;
+  }
+  void Cut(const string& sentence, vector<string>& words) const {
+    Cut(sentence, words, true);
+  }
+  void Cut(const string& sentence, vector<string>& words, bool hmm) const {
+    vector<Word> tmp;
+    Cut(sentence, tmp, hmm);
+    GetStringsFromWords(tmp, words);
+  }
+  void Cut(const string& sentence, vector<Word>& words, bool hmm = true) const {
+    PreFilter pre_filter(symbols_, sentence);
+    PreFilter::Range range;
+    vector<WordRange> wrs;
+    wrs.reserve(sentence.size()/2);
+    while (pre_filter.HasNext()) {
+      range = pre_filter.Next();
+      Cut(range.begin, range.end, wrs, hmm);
+    }
+    words.clear();
+    words.reserve(wrs.size());
+    GetWordsFromWordRanges(sentence, wrs, words);
+  }
+  void Cut(RuneStrArray::const_iterator begin, RuneStrArray::const_iterator end, vector<WordRange>& res, bool hmm) const {
+    //use mix Cut first
+    vector<WordRange> mixRes;
+    mixSeg_.Cut(begin, end, mixRes, hmm);
+
+    vector<WordRange> fullRes;
+    for (vector<WordRange>::const_iterator mixResItr = mixRes.begin(); mixResItr != mixRes.end(); mixResItr++) {
+      if (mixResItr->Length() > 2) {
+        for (size_t i = 0; i + 1 < mixResItr->Length(); i++) {
+          WordRange wr(mixResItr->left + i, mixResItr->left + i + 1);
+          if (trie_->Find(wr.left, wr.right + 1) != NULL) {
+            res.push_back(wr);
+          }
+        }
+      }
+      if (mixResItr->Length() > 3) {
+        for (size_t i = 0; i + 2 < mixResItr->Length(); i++) {
+          WordRange wr(mixResItr->left + i, mixResItr->left + i + 2);
+          if (trie_->Find(wr.left, wr.right + 1) != NULL) {
+            res.push_back(wr);
+          }
+        }
+      }
+      res.push_back(*mixResItr);
+    }
+  }
+ private:
+  bool IsAllAscii(const Unicode& s) const {
+   for(size_t i = 0; i < s.size(); i++) {
+     if (s[i] >= 0x80) {
+       return false;
+     }
+   }
+   return true;
+  }
+  MixSegment mixSeg_;
+  const DictTrie* trie_;
+}; // QuerySegment
+
+} // namespace cppjieba
+
+#endif

runtime/onnxruntime/third_party/jieba/include/cppjieba/SegmentBase.hpp

@@ -0,0 +1,46 @@
+#ifndef CPPJIEBA_SEGMENTBASE_H
+#define CPPJIEBA_SEGMENTBASE_H
+
+#include "limonp/Logging.hpp"
+#include "PreFilter.hpp"
+#include <cassert>
+
+
+namespace cppjieba {
+
+const char* const SPECIAL_SEPARATORS = " \t\n\xEF\xBC\x8C\xE3\x80\x82";
+
+using namespace limonp;
+
+class SegmentBase {
+ public:
+  SegmentBase() {
+    XCHECK(ResetSeparators(SPECIAL_SEPARATORS));
+  }
+  virtual ~SegmentBase() {
+  }
+
+  virtual void Cut(const string& sentence, vector<string>& words) const = 0;
+
+  bool ResetSeparators(const string& s) {
+    symbols_.clear();
+    RuneStrArray runes;
+    if (!DecodeRunesInString(s, runes)) {
+      XLOG(ERROR) << "decode " << s << " failed";
+      return false;
+    }
+    for (size_t i = 0; i < runes.size(); i++) {
+      if (!symbols_.insert(runes[i].rune).second) {
+        XLOG(ERROR) << s.substr(runes[i].offset, runes[i].len) << " already exists";
+        return false;
+      }
+    }
+    return true;
+  }
+ protected:
+  unordered_set<Rune> symbols_;
+}; // class SegmentBase
+
+} // cppjieba
+
+#endif

runtime/onnxruntime/third_party/jieba/include/cppjieba/SegmentTagged.hpp

@@ -0,0 +1,23 @@
+#ifndef CPPJIEBA_SEGMENTTAGGED_H
+#define CPPJIEBA_SEGMENTTAGGED_H
+
+#include "SegmentBase.hpp"
+
+namespace cppjieba {
+
+class SegmentTagged : public SegmentBase{
+ public:
+  SegmentTagged() {
+  }
+  virtual ~SegmentTagged() {
+  }
+
+  virtual bool Tag(const string& src, vector<pair<string, string> >& res) const = 0;
+
+  virtual const DictTrie* GetDictTrie() const = 0;
+
+}; // class SegmentTagged
+
+} // cppjieba
+
+#endif

runtime/onnxruntime/third_party/jieba/include/cppjieba/TextRankExtractor.hpp

@@ -0,0 +1,190 @@
+#ifndef CPPJIEBA_TEXTRANK_EXTRACTOR_H
+#define CPPJIEBA_TEXTRANK_EXTRACTOR_H
+
+#include <cmath>
+#include "Jieba.hpp"
+
+namespace cppjieba {
+  using namespace limonp;
+  using namespace std;
+
+  class TextRankExtractor {
+  public:
+    typedef struct _Word {string word;vector<size_t> offsets;double weight;}    Word; // struct Word
+  private:
+    typedef std::map<string,Word> WordMap;
+  
+    class WordGraph{
+    private:
+      typedef double Score;
+      typedef string Node;
+      typedef std::set<Node> NodeSet;
+
+      typedef std::map<Node,double> Edges;
+      typedef std::map<Node,Edges> Graph;
+      //typedef std::unordered_map<Node,double> Edges;
+      //typedef std::unordered_map<Node,Edges> Graph;
+
+      double d;
+      Graph graph;
+      NodeSet nodeSet;
+    public:
+      WordGraph(): d(0.85) {};
+      WordGraph(double in_d): d(in_d) {};
+
+      void addEdge(Node start,Node end,double weight){
+        Edges temp;
+        Edges::iterator gotEdges;
+        nodeSet.insert(start);
+        nodeSet.insert(end);
+        graph[start][end]+=weight;
+        graph[end][start]+=weight;
+      }
+
+      void rank(WordMap &ws,size_t rankTime=10){
+        WordMap outSum;
+        Score wsdef, min_rank, max_rank;
+
+        if( graph.size() == 0)
+          return;
+
+        wsdef = 1.0 / graph.size();
+
+        for(Graph::iterator edges=graph.begin();edges!=graph.end();++edges){
+          // edges->first start节点；edge->first end节点；edge->second 权重
+          ws[edges->first].word=edges->first;
+          ws[edges->first].weight=wsdef;
+          outSum[edges->first].weight=0;
+          for(Edges::iterator edge=edges->second.begin();edge!=edges->second.end();++edge){
+            outSum[edges->first].weight+=edge->second;
+          }
+        }
+        //sort(nodeSet.begin(),nodeSet.end()); 是否需要排序?
+        for( size_t i=0; i<rankTime; i++ ){
+          for(NodeSet::iterator node = nodeSet.begin(); node != nodeSet.end(); node++ ){
+            double s = 0;
+            for( Edges::iterator edge= graph[*node].begin(); edge != graph[*node].end(); edge++ )
+              // edge->first end节点；edge->second 权重
+              s += edge->second / outSum[edge->first].weight * ws[edge->first].weight;
+            ws[*node].weight = (1 - d) + d * s;
+          }
+        }
+
+        min_rank=max_rank=ws.begin()->second.weight;
+        for(WordMap::iterator i = ws.begin(); i != ws.end(); i ++){
+          if( i->second.weight < min_rank ){
+            min_rank = i->second.weight;
+          }
+          if( i->second.weight > max_rank ){
+            max_rank = i->second.weight;
+          }
+        }
+        for(WordMap::iterator i = ws.begin(); i != ws.end(); i ++){
+          ws[i->first].weight = (i->second.weight - min_rank / 10.0) / (max_rank - min_rank / 10.0);
+        }
+      }
+    };
+
+  public: 
+  TextRankExtractor(const string& dictPath, 
+        const string& hmmFilePath, 
+        const string& stopWordPath, 
+        const string& userDict = "") 
+    : segment_(dictPath, hmmFilePath, userDict) {
+    LoadStopWordDict(stopWordPath);
+  }
+  TextRankExtractor(const DictTrie* dictTrie, 
+        const HMMModel* model,
+        const string& stopWordPath) 
+    : segment_(dictTrie, model) {
+    LoadStopWordDict(stopWordPath);
+  }
+    TextRankExtractor(const Jieba& jieba, const string& stopWordPath) : segment_(jieba.GetDictTrie(), jieba.GetHMMModel()) {
+        LoadStopWordDict(stopWordPath);
+    }
+    ~TextRankExtractor() {
+    }
+
+    void Extract(const string& sentence, vector<string>& keywords, size_t topN) const {
+      vector<Word> topWords;
+      Extract(sentence, topWords, topN);
+      for (size_t i = 0; i < topWords.size(); i++) {
+        keywords.push_back(topWords[i].word);
+      }
+    }
+
+    void Extract(const string& sentence, vector<pair<string, double> >& keywords, size_t topN) const {
+      vector<Word> topWords;
+      Extract(sentence, topWords, topN);
+      for (size_t i = 0; i < topWords.size(); i++) {
+        keywords.push_back(pair<string, double>(topWords[i].word, topWords[i].weight));
+      }
+    }
+
+    void Extract(const string& sentence, vector<Word>& keywords, size_t topN, size_t span=5,size_t rankTime=10) const {
+      vector<string> words;
+      segment_.Cut(sentence, words);
+
+      TextRankExtractor::WordGraph graph;
+      WordMap wordmap;
+      size_t offset = 0;
+
+      for(size_t i=0; i < words.size(); i++){
+        size_t t = offset;
+        offset += words[i].size();
+        if (IsSingleWord(words[i]) || stopWords_.find(words[i]) != stopWords_.end()) {
+          continue;
+        }
+        for(size_t j=i+1,skip=0;j<i+span+skip && j<words.size();j++){
+          if (IsSingleWord(words[j]) || stopWords_.find(words[j]) != stopWords_.end()) {
+            skip++;
+            continue;
+          }
+          graph.addEdge(words[i],words[j],1);
+        }
+        wordmap[words[i]].offsets.push_back(t);
+      }
+      if (offset != sentence.size()) {
+        XLOG(ERROR) << "words illegal";
+        return;
+      }
+
+      graph.rank(wordmap,rankTime);
+      
+      keywords.clear();
+      keywords.reserve(wordmap.size());
+      for (WordMap::iterator itr = wordmap.begin(); itr != wordmap.end(); ++itr) {
+        keywords.push_back(itr->second);
+      }
+      
+      topN = min(topN, keywords.size());
+      partial_sort(keywords.begin(), keywords.begin() + topN, keywords.end(), Compare);
+      keywords.resize(topN);
+    }
+  private:
+    void LoadStopWordDict(const string& filePath) {
+      ifstream ifs(filePath.c_str());
+      XCHECK(ifs.is_open()) << "open " << filePath << " failed";
+      string line ;
+      while (getline(ifs, line)) {
+        stopWords_.insert(line);
+      }
+      assert(stopWords_.size());
+    }
+
+    static bool Compare(const Word &x,const Word &y){
+      return x.weight > y.weight;
+    }
+
+    MixSegment segment_;
+    unordered_set<string> stopWords_;
+  }; // class TextRankExtractor
+  
+  inline ostream& operator << (ostream& os, const TextRankExtractor::Word& word) {
+    return os << "{\"word\": \"" << word.word << "\", \"offset\": " << word.offsets << ", \"weight\": " << word.weight << "}"; 
+  }
+} // namespace cppjieba
+
+#endif
+
+

runtime/onnxruntime/third_party/jieba/include/cppjieba/Trie.hpp

@@ -0,0 +1,200 @@
+#ifndef CPPJIEBA_TRIE_HPP
+#define CPPJIEBA_TRIE_HPP
+
+#include <vector>
+#include <queue>
+#include "limonp/StdExtension.hpp"
+#include "Unicode.hpp"
+
+namespace cppjieba {
+
+using namespace std;
+
+const size_t MAX_WORD_LENGTH = 512;
+
+struct DictUnit {
+  Unicode word;
+  double weight;
+  string tag;
+}; // struct DictUnit
+
+// for debugging
+// inline ostream & operator << (ostream& os, const DictUnit& unit) {
+//   string s;
+//   s << unit.word;
+//   return os << StringFormat("%s %s %.3lf", s.c_str(), unit.tag.c_str(), unit.weight);
+// }
+
+struct Dag {
+  RuneStr runestr;
+  // [offset, nexts.first]
+  limonp::LocalVector<pair<size_t, const DictUnit*> > nexts;
+  const DictUnit * pInfo;
+  double weight;
+  size_t nextPos; // TODO
+  Dag():runestr(), pInfo(NULL), weight(0.0), nextPos(0) {
+  }
+}; // struct Dag
+
+typedef Rune TrieKey;
+
+class TrieNode {
+ public :
+  TrieNode(): next(NULL), ptValue(NULL) {
+  }
+ public:
+  typedef unordered_map<TrieKey, TrieNode*> NextMap;
+  NextMap *next;
+  const DictUnit *ptValue;
+};
+
+class Trie {
+ public:
+  Trie(const vector<Unicode>& keys, const vector<const DictUnit*>& valuePointers)
+   : root_(new TrieNode) {
+    CreateTrie(keys, valuePointers);
+  }
+  ~Trie() {
+    DeleteNode(root_);
+  }
+
+  const DictUnit* Find(RuneStrArray::const_iterator begin, RuneStrArray::const_iterator end) const {
+    if (begin == end) {
+      return NULL;
+    }
+
+    const TrieNode* ptNode = root_;
+    TrieNode::NextMap::const_iterator citer;
+    for (RuneStrArray::const_iterator it = begin; it != end; it++) {
+      if (NULL == ptNode->next) {
+        return NULL;
+      }
+      citer = ptNode->next->find(it->rune);
+      if (ptNode->next->end() == citer) {
+        return NULL;
+      }
+      ptNode = citer->second;
+    }
+    return ptNode->ptValue;
+  }
+
+  void Find(RuneStrArray::const_iterator begin, 
+        RuneStrArray::const_iterator end, 
+        vector<struct Dag>&res, 
+        size_t max_word_len = MAX_WORD_LENGTH) const {
+    assert(root_ != NULL);
+    res.resize(end - begin);
+
+    const TrieNode *ptNode = NULL;
+    TrieNode::NextMap::const_iterator citer;
+    for (size_t i = 0; i < size_t(end - begin); i++) {
+      res[i].runestr = *(begin + i);
+
+      if (root_->next != NULL && root_->next->end() != (citer = root_->next->find(res[i].runestr.rune))) {
+        ptNode = citer->second;
+      } else {
+        ptNode = NULL;
+      }
+      if (ptNode != NULL) {
+        res[i].nexts.push_back(pair<size_t, const DictUnit*>(i, ptNode->ptValue));
+      } else {
+        res[i].nexts.push_back(pair<size_t, const DictUnit*>(i, static_cast<const DictUnit*>(NULL)));
+      }
+
+      for (size_t j = i + 1; j < size_t(end - begin) && (j - i + 1) <= max_word_len; j++) {
+        if (ptNode == NULL || ptNode->next == NULL) {
+          break;
+        }
+        citer = ptNode->next->find((begin + j)->rune);
+        if (ptNode->next->end() == citer) {
+          break;
+        }
+        ptNode = citer->second;
+        if (NULL != ptNode->ptValue) {
+          res[i].nexts.push_back(pair<size_t, const DictUnit*>(j, ptNode->ptValue));
+        }
+      }
+    }
+  }
+
+  void InsertNode(const Unicode& key, const DictUnit* ptValue) {
+    if (key.begin() == key.end()) {
+      return;
+    }
+
+    TrieNode::NextMap::const_iterator kmIter;
+    TrieNode *ptNode = root_;
+    for (Unicode::const_iterator citer = key.begin(); citer != key.end(); ++citer) {
+      if (NULL == ptNode->next) {
+        ptNode->next = new TrieNode::NextMap;
+      }
+      kmIter = ptNode->next->find(*citer);
+      if (ptNode->next->end() == kmIter) {
+        TrieNode *nextNode = new TrieNode;
+
+        ptNode->next->insert(make_pair(*citer, nextNode));
+        ptNode = nextNode;
+      } else {
+        ptNode = kmIter->second;
+      }
+    }
+    assert(ptNode != NULL);
+    ptNode->ptValue = ptValue;
+  }
+  void DeleteNode(const Unicode& key, const DictUnit* ptValue) {
+      if (key.begin() == key.end()) {
+        return;
+      }
+      //定义一个NextMap迭代器
+      TrieNode::NextMap::const_iterator kmIter;
+      //定义一个指向root的TrieNode指针
+      TrieNode *ptNode = root_;
+      for (Unicode::const_iterator citer = key.begin(); citer != key.end(); ++citer) {
+        //链表不存在元素
+        if (NULL == ptNode->next) {
+          return;
+        }
+        kmIter = ptNode->next->find(*citer);
+        //如果map中不存在,跳出循环
+        if (ptNode->next->end() == kmIter) {
+              break;
+        }
+        //从unordered_map中擦除该项
+        ptNode->next->erase(*citer);
+        //删除该node
+        ptNode = kmIter->second;
+        delete ptNode;
+        break;
+      }
+      return;
+ }
+ private:
+  void CreateTrie(const vector<Unicode>& keys, const vector<const DictUnit*>& valuePointers) {
+    if (valuePointers.empty() || keys.empty()) {
+      return;
+    }
+    assert(keys.size() == valuePointers.size());
+
+    for (size_t i = 0; i < keys.size(); i++) {
+      InsertNode(keys[i], valuePointers[i]);
+    }
+  }
+
+  void DeleteNode(TrieNode* node) {
+    if (NULL == node) {
+      return;
+    }
+    if (NULL != node->next) {
+      for (TrieNode::NextMap::iterator it = node->next->begin(); it != node->next->end(); ++it) {
+        DeleteNode(it->second);
+      }
+      delete node->next;
+    }
+    delete node;
+  }
+
+  TrieNode* root_;
+}; // class Trie
+} // namespace cppjieba
+
+#endif // CPPJIEBA_TRIE_HPP

runtime/onnxruntime/third_party/jieba/include/cppjieba/Unicode.hpp

@@ -0,0 +1,227 @@
+#ifndef CPPJIEBA_UNICODE_H
+#define CPPJIEBA_UNICODE_H
+
+#include <stdint.h>
+#include <stdlib.h>
+#include <string>
+#include <vector>
+#include <ostream>
+#include "limonp/LocalVector.hpp"
+
+namespace cppjieba {
+
+using std::string;
+using std::vector;
+
+typedef uint32_t Rune;
+
+struct Word {
+  string word;
+  uint32_t offset;
+  uint32_t unicode_offset;
+  uint32_t unicode_length;
+  Word(const string& w, uint32_t o)
+   : word(w), offset(o) {
+  }
+  Word(const string& w, uint32_t o, uint32_t unicode_offset, uint32_t unicode_length)
+          : word(w), offset(o), unicode_offset(unicode_offset), unicode_length(unicode_length) {
+  }
+}; // struct Word
+
+inline std::ostream& operator << (std::ostream& os, const Word& w) {
+  return os << "{\"word\": \"" << w.word << "\", \"offset\": " << w.offset << "}";
+}
+
+struct RuneStr {
+  Rune rune;
+  uint32_t offset;
+  uint32_t len;
+  uint32_t unicode_offset;
+  uint32_t unicode_length;
+  RuneStr(): rune(0), offset(0), len(0), unicode_offset(0), unicode_length(0) {
+  }
+  RuneStr(Rune r, uint32_t o, uint32_t l)
+    : rune(r), offset(o), len(l), unicode_offset(0), unicode_length(0) {
+  }
+  RuneStr(Rune r, uint32_t o, uint32_t l, uint32_t unicode_offset, uint32_t unicode_length)
+          : rune(r), offset(o), len(l), unicode_offset(unicode_offset), unicode_length(unicode_length) {
+  }
+}; // struct RuneStr
+
+inline std::ostream& operator << (std::ostream& os, const RuneStr& r) {
+  return os << "{\"rune\": \"" << r.rune << "\", \"offset\": " << r.offset << ", \"len\": " << r.len << "}";
+}
+
+typedef limonp::LocalVector<Rune> Unicode;
+typedef limonp::LocalVector<struct RuneStr> RuneStrArray;
+
+// [left, right]
+struct WordRange {
+  RuneStrArray::const_iterator left;
+  RuneStrArray::const_iterator right;
+  WordRange(RuneStrArray::const_iterator l, RuneStrArray::const_iterator r)
+   : left(l), right(r) {
+  }
+  size_t Length() const {
+    return right - left + 1;
+  }
+  bool IsAllAscii() const {
+    for (RuneStrArray::const_iterator iter = left; iter <= right; ++iter) {
+      if (iter->rune >= 0x80) {
+        return false;
+      }
+    }
+    return true;
+  }
+}; // struct WordRange
+
+struct RuneStrLite {
+  uint32_t rune;
+  uint32_t len;
+  RuneStrLite(): rune(0), len(0) {
+  }
+  RuneStrLite(uint32_t r, uint32_t l): rune(r), len(l) {
+  }
+}; // struct RuneStrLite
+
+inline RuneStrLite DecodeRuneInString(const char* str, size_t len) {
+  RuneStrLite rp(0, 0);
+  if (str == NULL || len == 0) {
+    return rp;
+  }
+  if (!(str[0] & 0x80)) { // 0xxxxxxx
+    // 7bit, total 7bit
+    rp.rune = (uint8_t)(str[0]) & 0x7f;
+    rp.len = 1;
+  } else if ((uint8_t)str[0] <= 0xdf &&  1 < len) { 
+    // 110xxxxxx
+    // 5bit, total 5bit
+    rp.rune = (uint8_t)(str[0]) & 0x1f;
+
+    // 6bit, total 11bit
+    rp.rune <<= 6;
+    rp.rune |= (uint8_t)(str[1]) & 0x3f;
+    rp.len = 2;
+  } else if((uint8_t)str[0] <= 0xef && 2 < len) { // 1110xxxxxx
+    // 4bit, total 4bit
+    rp.rune = (uint8_t)(str[0]) & 0x0f;
+
+    // 6bit, total 10bit
+    rp.rune <<= 6;
+    rp.rune |= (uint8_t)(str[1]) & 0x3f;
+
+    // 6bit, total 16bit
+    rp.rune <<= 6;
+    rp.rune |= (uint8_t)(str[2]) & 0x3f;
+
+    rp.len = 3;
+  } else if((uint8_t)str[0] <= 0xf7 && 3 < len) { // 11110xxxx
+    // 3bit, total 3bit
+    rp.rune = (uint8_t)(str[0]) & 0x07;
+
+    // 6bit, total 9bit
+    rp.rune <<= 6;
+    rp.rune |= (uint8_t)(str[1]) & 0x3f;
+
+    // 6bit, total 15bit
+    rp.rune <<= 6;
+    rp.rune |= (uint8_t)(str[2]) & 0x3f;
+
+    // 6bit, total 21bit
+    rp.rune <<= 6;
+    rp.rune |= (uint8_t)(str[3]) & 0x3f;
+
+    rp.len = 4;
+  } else {
+    rp.rune = 0;
+    rp.len = 0;
+  }
+  return rp;
+}
+
+inline bool DecodeRunesInString(const char* s, size_t len, RuneStrArray& runes) {
+  runes.clear();
+  runes.reserve(len / 2);
+  for (uint32_t i = 0, j = 0; i < len;) {
+    RuneStrLite rp = DecodeRuneInString(s + i, len - i);
+    if (rp.len == 0) {
+      runes.clear();
+      return false;
+    }
+    RuneStr x(rp.rune, i, rp.len, j, 1);
+    runes.push_back(x);
+    i += rp.len;
+    ++j;
+  }
+  return true;
+}
+
+inline bool DecodeRunesInString(const string& s, RuneStrArray& runes) {
+  return DecodeRunesInString(s.c_str(), s.size(), runes);
+}
+
+inline bool DecodeRunesInString(const char* s, size_t len, Unicode& unicode) {
+  unicode.clear();
+  RuneStrArray runes;
+  if (!DecodeRunesInString(s, len, runes)) {
+    return false;
+  }
+  unicode.reserve(runes.size());
+  for (size_t i = 0; i < runes.size(); i++) {
+    unicode.push_back(runes[i].rune);
+  }
+  return true;
+}
+
+inline bool IsSingleWord(const string& str) {
+  RuneStrLite rp = DecodeRuneInString(str.c_str(), str.size());
+  return rp.len == str.size();
+}
+
+inline bool DecodeRunesInString(const string& s, Unicode& unicode) {
+  return DecodeRunesInString(s.c_str(), s.size(), unicode);
+}
+
+inline Unicode DecodeRunesInString(const string& s) {
+  Unicode result;
+  DecodeRunesInString(s, result);
+  return result;
+}
+
+
+// [left, right]
+inline Word GetWordFromRunes(const string& s, RuneStrArray::const_iterator left, RuneStrArray::const_iterator right) {
+  assert(right->offset >= left->offset);
+  uint32_t len = right->offset - left->offset + right->len;
+  uint32_t unicode_length = right->unicode_offset - left->unicode_offset + right->unicode_length;
+  return Word(s.substr(left->offset, len), left->offset, left->unicode_offset, unicode_length);
+}
+
+inline string GetStringFromRunes(const string& s, RuneStrArray::const_iterator left, RuneStrArray::const_iterator right) {
+  assert(right->offset >= left->offset);
+  uint32_t len = right->offset - left->offset + right->len;
+  return s.substr(left->offset, len);
+}
+
+inline void GetWordsFromWordRanges(const string& s, const vector<WordRange>& wrs, vector<Word>& words) {
+  for (size_t i = 0; i < wrs.size(); i++) {
+    words.push_back(GetWordFromRunes(s, wrs[i].left, wrs[i].right));
+  }
+}
+
+inline vector<Word> GetWordsFromWordRanges(const string& s, const vector<WordRange>& wrs) {
+  vector<Word> result;
+  GetWordsFromWordRanges(s, wrs, result);
+  return result;
+}
+
+inline void GetStringsFromWords(const vector<Word>& words, vector<string>& strs) {
+  strs.resize(words.size());
+  for (size_t i = 0; i < words.size(); ++i) {
+    strs[i] = words[i].word;
+  }
+}
+
+} // namespace cppjieba
+
+#endif // CPPJIEBA_UNICODE_H