libtextclassifier/smartselect/feature-processor.cc

/*
 * Copyright (C) 2017 The Android Open Source Project
 *
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
 * You may obtain a copy of the License at
 *
 *      http://www.apache.org/licenses/LICENSE-2.0
 *
 * Unless required by applicable law or agreed to in writing, software
 * distributed under the License is distributed on an "AS IS" BASIS,
 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 * See the License for the specific language governing permissions and
 * limitations under the License.
 */

#include "smartselect/feature-processor.h"

#include <iterator>
#include <set>
#include <vector>

#include "smartselect/text-classification-model.pb.h"
#include "util/base/logging.h"
#include "util/strings/utf8.h"
#include "util/utf8/unicodetext.h"
#include "unicode/brkiter.h"
#include "unicode/errorcode.h"
#include "unicode/uchar.h"

namespace libtextclassifier {

namespace internal {

TokenFeatureExtractorOptions BuildTokenFeatureExtractorOptions(
    const FeatureProcessorOptions& options) {
  TokenFeatureExtractorOptions extractor_options;

  extractor_options.num_buckets = options.num_buckets();
  for (int order : options.chargram_orders()) {
    extractor_options.chargram_orders.push_back(order);
  }
  extractor_options.max_word_length = options.max_word_length();
  extractor_options.extract_case_feature = options.extract_case_feature();
  extractor_options.unicode_aware_features = options.unicode_aware_features();
  extractor_options.extract_selection_mask_feature =
      options.extract_selection_mask_feature();
  for (int i = 0; i < options.regexp_feature_size(); ++i) {
    extractor_options.regexp_features.push_back(options.regexp_feature(i));
  }
  extractor_options.remap_digits = options.remap_digits();
  extractor_options.lowercase_tokens = options.lowercase_tokens();

  return extractor_options;
}

FeatureProcessorOptions ParseSerializedOptions(
    const std::string& serialized_options) {
  FeatureProcessorOptions options;
  options.ParseFromString(serialized_options);
  return options;
}

void SplitTokensOnSelectionBoundaries(CodepointSpan selection,
                                      std::vector<Token>* tokens) {
  for (auto it = tokens->begin(); it != tokens->end(); ++it) {
    const UnicodeText token_word =
        UTF8ToUnicodeText(it->value, /*do_copy=*/false);

    auto last_start = token_word.begin();
    int last_start_index = it->start;
    std::vector<UnicodeText::const_iterator> split_points;

    // Selection start split point.
    if (selection.first > it->start && selection.first < it->end) {
      std::advance(last_start, selection.first - last_start_index);
      split_points.push_back(last_start);
      last_start_index = selection.first;
    }

    // Selection end split point.
    if (selection.second > it->start && selection.second < it->end) {
      std::advance(last_start, selection.second - last_start_index);
      split_points.push_back(last_start);
    }

    if (!split_points.empty()) {
      // Add a final split for the rest of the token unless it's been all
      // consumed already.
      if (split_points.back() != token_word.end()) {
        split_points.push_back(token_word.end());
      }

      std::vector<Token> replacement_tokens;
      last_start = token_word.begin();
      int current_pos = it->start;
      for (const auto& split_point : split_points) {
        Token new_token(token_word.UTF8Substring(last_start, split_point),
                        current_pos,
                        current_pos + std::distance(last_start, split_point));

        last_start = split_point;
        current_pos = new_token.end;

        replacement_tokens.push_back(new_token);
      }

      it = tokens->erase(it);
      it = tokens->insert(it, replacement_tokens.begin(),
                          replacement_tokens.end());
      std::advance(it, replacement_tokens.size() - 1);
    }
  }
}

void FindSubstrings(const UnicodeText& t, const std::set<char32>& codepoints,
                    std::vector<UnicodeTextRange>* ranges) {
  UnicodeText::const_iterator start = t.begin();
  UnicodeText::const_iterator curr = start;
  UnicodeText::const_iterator end = t.end();
  for (; curr != end; ++curr) {
    if (codepoints.find(*curr) != codepoints.end()) {
      if (start != curr) {
        ranges->push_back(std::make_pair(start, curr));
      }
      start = curr;
      ++start;
    }
  }
  if (start != end) {
    ranges->push_back(std::make_pair(start, end));
  }
}

void StripTokensFromOtherLines(const std::string& context, CodepointSpan span,
                               std::vector<Token>* tokens) {
  const UnicodeText context_unicode = UTF8ToUnicodeText(context,
                                                        /*do_copy=*/false);
  std::vector<UnicodeTextRange> lines;
  std::set<char32> codepoints;
  codepoints.insert('\n');
  codepoints.insert('|');
  internal::FindSubstrings(context_unicode, codepoints, &lines);

  auto span_start = context_unicode.begin();
  if (span.first > 0) {
    std::advance(span_start, span.first);
  }
  auto span_end = context_unicode.begin();
  if (span.second > 0) {
    std::advance(span_end, span.second);
  }
  for (const UnicodeTextRange& line : lines) {
    // Find the line that completely contains the span.
    if (line.first <= span_start && line.second >= span_end) {
      const CodepointIndex last_line_begin_index =
          std::distance(context_unicode.begin(), line.first);
      const CodepointIndex last_line_end_index =
          last_line_begin_index + std::distance(line.first, line.second);

      for (auto token = tokens->begin(); token != tokens->end();) {
        if (token->start >= last_line_begin_index &&
            token->end <= last_line_end_index) {
          ++token;
        } else {
          token = tokens->erase(token);
        }
      }
    }
  }
}

}  // namespace internal

std::string FeatureProcessor::GetDefaultCollection() const {
  if (options_.default_collection() >= options_.collections_size()) {
    TC_LOG(ERROR) << "No collections specified. Returning empty string.";
    return "";
  }
  return options_.collections(options_.default_collection());
}

std::vector<Token> FeatureProcessor::Tokenize(
    const std::string& utf8_text) const {
  if (options_.tokenization_type() ==
      libtextclassifier::FeatureProcessorOptions::INTERNAL_TOKENIZER) {
    return tokenizer_.Tokenize(utf8_text);
  } else if (options_.tokenization_type() ==
                 libtextclassifier::FeatureProcessorOptions::ICU ||
             options_.tokenization_type() ==
                 libtextclassifier::FeatureProcessorOptions::MIXED) {
    std::vector<Token> result;
    if (!ICUTokenize(utf8_text, &result)) {
      return {};
    }
    if (options_.tokenization_type() ==
        libtextclassifier::FeatureProcessorOptions::MIXED) {
      InternalRetokenize(utf8_text, &result);
    }
    return result;
  } else {
    TC_LOG(ERROR) << "Unknown tokenization type specified. Using "
                     "internal.";
    return tokenizer_.Tokenize(utf8_text);
  }
}

bool FeatureProcessor::LabelToSpan(
    const int label, const VectorSpan<Token>& tokens,
    std::pair<CodepointIndex, CodepointIndex>* span) const {
  if (tokens.size() != GetNumContextTokens()) {
    return false;
  }

  TokenSpan token_span;
  if (!LabelToTokenSpan(label, &token_span)) {
    return false;
  }

  const int result_begin_token = token_span.first;
  const int result_begin_codepoint =
      tokens[options_.context_size() - result_begin_token].start;
  const int result_end_token = token_span.second;
  const int result_end_codepoint =
      tokens[options_.context_size() + result_end_token].end;

  if (result_begin_codepoint == kInvalidIndex ||
      result_end_codepoint == kInvalidIndex) {
    *span = CodepointSpan({kInvalidIndex, kInvalidIndex});
  } else {
    *span = CodepointSpan({result_begin_codepoint, result_end_codepoint});
  }
  return true;
}

bool FeatureProcessor::LabelToTokenSpan(const int label,
                                        TokenSpan* token_span) const {
  if (label >= 0 && label < label_to_selection_.size()) {
    *token_span = label_to_selection_[label];
    return true;
  } else {
    return false;
  }
}

bool FeatureProcessor::SpanToLabel(
    const std::pair<CodepointIndex, CodepointIndex>& span,
    const std::vector<Token>& tokens, int* label) const {
  if (tokens.size() != GetNumContextTokens()) {
    return false;
  }

  const int click_position =
      options_.context_size();  // Click is always in the middle.
  const int padding = options_.context_size() - options_.max_selection_span();

  int span_left = 0;
  for (int i = click_position - 1; i >= padding; i--) {
    if (tokens[i].start != kInvalidIndex && tokens[i].end > span.first) {
      ++span_left;
    } else {
      break;
    }
  }

  int span_right = 0;
  for (int i = click_position + 1; i < tokens.size() - padding; ++i) {
    if (tokens[i].end != kInvalidIndex && tokens[i].start < span.second) {
      ++span_right;
    } else {
      break;
    }
  }

  // Check that the spanned tokens cover the whole span.
  bool tokens_match_span;
  if (options_.snap_label_span_boundaries_to_containing_tokens()) {
    tokens_match_span =
        tokens[click_position - span_left].start <= span.first &&
        tokens[click_position + span_right].end >= span.second;
  } else {
    tokens_match_span =
        tokens[click_position - span_left].start == span.first &&
        tokens[click_position + span_right].end == span.second;
  }

  if (tokens_match_span) {
    *label = TokenSpanToLabel({span_left, span_right});
  } else {
    *label = kInvalidLabel;
  }

  return true;
}

int FeatureProcessor::TokenSpanToLabel(const TokenSpan& span) const {
  auto it = selection_to_label_.find(span);
  if (it != selection_to_label_.end()) {
    return it->second;
  } else {
    return kInvalidLabel;
  }
}

TokenSpan CodepointSpanToTokenSpan(const std::vector<Token>& selectable_tokens,
                                   CodepointSpan codepoint_span) {
  const int codepoint_start = std::get<0>(codepoint_span);
  const int codepoint_end = std::get<1>(codepoint_span);

  TokenIndex start_token = kInvalidIndex;
  TokenIndex end_token = kInvalidIndex;
  for (int i = 0; i < selectable_tokens.size(); ++i) {
    if (codepoint_start <= selectable_tokens[i].start &&
        codepoint_end >= selectable_tokens[i].end &&
        !selectable_tokens[i].is_padding) {
      if (start_token == kInvalidIndex) {
        start_token = i;
      }
      end_token = i + 1;
    }
  }
  return {start_token, end_token};
}

CodepointSpan TokenSpanToCodepointSpan(
    const std::vector<Token>& selectable_tokens, TokenSpan token_span) {
  return {selectable_tokens[token_span.first].start,
          selectable_tokens[token_span.second - 1].end};
}

namespace {

// Finds a single token that completely contains the given span.
int FindTokenThatContainsSpan(const std::vector<Token>& selectable_tokens,
                              CodepointSpan codepoint_span) {
  const int codepoint_start = std::get<0>(codepoint_span);
  const int codepoint_end = std::get<1>(codepoint_span);

  for (int i = 0; i < selectable_tokens.size(); ++i) {
    if (codepoint_start >= selectable_tokens[i].start &&
        codepoint_end <= selectable_tokens[i].end) {
      return i;
    }
  }
  return kInvalidIndex;
}

}  // namespace

namespace internal {

int CenterTokenFromClick(CodepointSpan span,
                         const std::vector<Token>& selectable_tokens) {
  int range_begin;
  int range_end;
  std::tie(range_begin, range_end) =
      CodepointSpanToTokenSpan(selectable_tokens, span);

  // If no exact match was found, try finding a token that completely contains
  // the click span. This is useful e.g. when Android builds the selection
  // using ICU tokenization, and ends up with only a portion of our space-
  // separated token. E.g. for "(857)" Android would select "857".
  if (range_begin == kInvalidIndex || range_end == kInvalidIndex) {
    int token_index = FindTokenThatContainsSpan(selectable_tokens, span);
    if (token_index != kInvalidIndex) {
      range_begin = token_index;
      range_end = token_index + 1;
    }
  }

  // We only allow clicks that are exactly 1 selectable token.
  if (range_end - range_begin == 1) {
    return range_begin;
  } else {
    return kInvalidIndex;
  }
}

int CenterTokenFromMiddleOfSelection(
    CodepointSpan span, const std::vector<Token>& selectable_tokens) {
  int range_begin;
  int range_end;
  std::tie(range_begin, range_end) =
      CodepointSpanToTokenSpan(selectable_tokens, span);

  // Center the clicked token in the selection range.
  if (range_begin != kInvalidIndex && range_end != kInvalidIndex) {
    return (range_begin + range_end - 1) / 2;
  } else {
    return kInvalidIndex;
  }
}

}  // namespace internal

int FeatureProcessor::FindCenterToken(CodepointSpan span,
                                      const std::vector<Token>& tokens) const {
  if (options_.center_token_selection_method() ==
      FeatureProcessorOptions::CENTER_TOKEN_FROM_CLICK) {
    return internal::CenterTokenFromClick(span, tokens);
  } else if (options_.center_token_selection_method() ==
             FeatureProcessorOptions::CENTER_TOKEN_MIDDLE_OF_SELECTION) {
    return internal::CenterTokenFromMiddleOfSelection(span, tokens);
  } else if (options_.center_token_selection_method() ==
             FeatureProcessorOptions::DEFAULT_CENTER_TOKEN_METHOD) {
    // TODO(zilka): Remove once we have new models on the device.
    // It uses the fact that sharing model use
    // split_tokens_on_selection_boundaries and selection not. So depending on
    // this we select the right way of finding the click location.
    if (!options_.split_tokens_on_selection_boundaries()) {
      // SmartSelection model.
      return internal::CenterTokenFromClick(span, tokens);
    } else {
      // SmartSharing model.
      return internal::CenterTokenFromMiddleOfSelection(span, tokens);
    }
  } else {
    TC_LOG(ERROR) << "Invalid center token selection method.";
    return kInvalidIndex;
  }
}

bool FeatureProcessor::SelectionLabelSpans(
    const VectorSpan<Token> tokens,
    std::vector<CodepointSpan>* selection_label_spans) const {
  for (int i = 0; i < label_to_selection_.size(); ++i) {
    CodepointSpan span;
    if (!LabelToSpan(i, tokens, &span)) {
      TC_LOG(ERROR) << "Could not convert label to span: " << i;
      return false;
    }
    selection_label_spans->push_back(span);
  }
  return true;
}

void FeatureProcessor::PrepareCodepointRanges(
    const std::vector<FeatureProcessorOptions::CodepointRange>&
        codepoint_ranges,
    std::vector<CodepointRange>* prepared_codepoint_ranges) {
  prepared_codepoint_ranges->clear();
  prepared_codepoint_ranges->reserve(codepoint_ranges.size());
  for (const FeatureProcessorOptions::CodepointRange& range :
       codepoint_ranges) {
    prepared_codepoint_ranges->push_back(
        CodepointRange(range.start(), range.end()));
  }

  std::sort(prepared_codepoint_ranges->begin(),
            prepared_codepoint_ranges->end(),
            [](const CodepointRange& a, const CodepointRange& b) {
              return a.start < b.start;
            });
}

float FeatureProcessor::SupportedCodepointsRatio(
    int click_pos, const std::vector<Token>& tokens) const {
  int num_supported = 0;
  int num_total = 0;
  for (int i = click_pos - options_.context_size();
       i <= click_pos + options_.context_size(); ++i) {
    const bool is_valid_token = i >= 0 && i < tokens.size();
    if (is_valid_token) {
      const UnicodeText value =
          UTF8ToUnicodeText(tokens[i].value, /*do_copy=*/false);
      for (auto codepoint : value) {
        if (IsCodepointInRanges(codepoint, supported_codepoint_ranges_)) {
          ++num_supported;
        }
        ++num_total;
      }
    }
  }
  return static_cast<float>(num_supported) / static_cast<float>(num_total);
}

bool FeatureProcessor::IsCodepointInRanges(
    int codepoint, const std::vector<CodepointRange>& codepoint_ranges) const {
  auto it = std::lower_bound(codepoint_ranges.begin(), codepoint_ranges.end(),
                             codepoint,
                             [](const CodepointRange& range, int codepoint) {
                               // This function compares range with the
                               // codepoint for the purpose of finding the first
                               // greater or equal range. Because of the use of
                               // std::lower_bound it needs to return true when
                               // range < codepoint; the first time it will
                               // return false the lower bound is found and
                               // returned.
                               //
                               // It might seem weird that the condition is
                               // range.end <= codepoint here but when codepoint
                               // == range.end it means it's actually just
                               // outside of the range, thus the range is less
                               // than the codepoint.
                               return range.end <= codepoint;
                             });
  if (it != codepoint_ranges.end() && it->start <= codepoint &&
      it->end > codepoint) {
    return true;
  } else {
    return false;
  }
}

int FeatureProcessor::CollectionToLabel(const std::string& collection) const {
  const auto it = collection_to_label_.find(collection);
  if (it == collection_to_label_.end()) {
    return options_.default_collection();
  } else {
    return it->second;
  }
}

std::string FeatureProcessor::LabelToCollection(int label) const {
  if (label >= 0 && label < collection_to_label_.size()) {
    return options_.collections(label);
  } else {
    return GetDefaultCollection();
  }
}

void FeatureProcessor::MakeLabelMaps() {
  for (int i = 0; i < options_.collections().size(); ++i) {
    collection_to_label_[options_.collections(i)] = i;
  }

  int selection_label_id = 0;
  for (int l = 0; l < (options_.max_selection_span() + 1); ++l) {
    for (int r = 0; r < (options_.max_selection_span() + 1); ++r) {
      if (!options_.selection_reduced_output_space() ||
          r + l <= options_.max_selection_span()) {
        TokenSpan token_span{l, r};
        selection_to_label_[token_span] = selection_label_id;
        label_to_selection_.push_back(token_span);
        ++selection_label_id;
      }
    }
  }
}

void FeatureProcessor::TokenizeAndFindClick(const std::string& context,
                                            CodepointSpan input_span,
                                            std::vector<Token>* tokens,
                                            int* click_pos) const {
  TC_CHECK(tokens != nullptr);
  *tokens = Tokenize(context);

  if (options_.split_tokens_on_selection_boundaries()) {
    internal::SplitTokensOnSelectionBoundaries(input_span, tokens);
  }

  if (options_.only_use_line_with_click()) {
    internal::StripTokensFromOtherLines(context, input_span, tokens);
  }

  int local_click_pos;
  if (click_pos == nullptr) {
    click_pos = &local_click_pos;
  }
  *click_pos = FindCenterToken(input_span, *tokens);
}

namespace internal {

void StripOrPadTokens(TokenSpan relative_click_span, int context_size,
                      std::vector<Token>* tokens, int* click_pos) {
  int right_context_needed = relative_click_span.second + context_size;
  if (*click_pos + right_context_needed + 1 >= tokens->size()) {
    // Pad max the context size.
    const int num_pad_tokens = std::min(
        context_size, static_cast<int>(*click_pos + right_context_needed + 1 -
                                       tokens->size()));
    std::vector<Token> pad_tokens(num_pad_tokens);
    tokens->insert(tokens->end(), pad_tokens.begin(), pad_tokens.end());
  } else if (*click_pos + right_context_needed + 1 < tokens->size() - 1) {
    // Strip unused tokens.
    auto it = tokens->begin();
    std::advance(it, *click_pos + right_context_needed + 1);
    tokens->erase(it, tokens->end());
  }

  int left_context_needed = relative_click_span.first + context_size;
  if (*click_pos < left_context_needed) {
    // Pad max the context size.
    const int num_pad_tokens =
        std::min(context_size, left_context_needed - *click_pos);
    std::vector<Token> pad_tokens(num_pad_tokens);
    tokens->insert(tokens->begin(), pad_tokens.begin(), pad_tokens.end());
    *click_pos += num_pad_tokens;
  } else if (*click_pos > left_context_needed) {
    // Strip unused tokens.
    auto it = tokens->begin();
    std::advance(it, *click_pos - left_context_needed);
    *click_pos -= it - tokens->begin();
    tokens->erase(tokens->begin(), it);
  }
}

}  // namespace internal

bool FeatureProcessor::ExtractFeatures(
    const std::string& context, CodepointSpan input_span,
    TokenSpan relative_click_span, const FeatureVectorFn& feature_vector_fn,
    int feature_vector_size, std::vector<Token>* tokens, int* click_pos,
    std::unique_ptr<CachedFeatures>* cached_features) const {
  TokenizeAndFindClick(context, input_span, tokens, click_pos);

  // If the default click method failed, let's try to do sub-token matching
  // before we fail.
  if (*click_pos == kInvalidIndex) {
    *click_pos = internal::CenterTokenFromClick(input_span, *tokens);
    if (*click_pos == kInvalidIndex) {
      return false;
    }
  }

  internal::StripOrPadTokens(relative_click_span, options_.context_size(),
                             tokens, click_pos);

  if (options_.min_supported_codepoint_ratio() > 0) {
    const float supported_codepoint_ratio =
        SupportedCodepointsRatio(*click_pos, *tokens);
    if (supported_codepoint_ratio < options_.min_supported_codepoint_ratio()) {
      TC_LOG(INFO) << "Not enough supported codepoints in the context: "
                   << supported_codepoint_ratio;
      return false;
    }
  }

  std::vector<std::vector<int>> sparse_features(tokens->size());
  std::vector<std::vector<float>> dense_features(tokens->size());
  for (int i = 0; i < tokens->size(); ++i) {
    const Token& token = (*tokens)[i];
    if (!feature_extractor_.Extract(token, token.IsContainedInSpan(input_span),
                                    &(sparse_features[i]),
                                    &(dense_features[i]))) {
      TC_LOG(ERROR) << "Could not extract token's features: " << token;
      return false;
    }
  }

  cached_features->reset(new CachedFeatures(
      *tokens, options_.context_size(), sparse_features, dense_features,
      feature_vector_fn, feature_vector_size));

  if (*cached_features == nullptr) {
    return false;
  }

  if (options_.feature_version() == 0) {
    (*cached_features)
        ->SetV0FeatureMode(feature_vector_size -
                           feature_extractor_.DenseFeaturesCount());
  }

  return true;
}

bool FeatureProcessor::ICUTokenize(const std::string& context,
                                   std::vector<Token>* result) const {
  icu::ErrorCode status;
  icu::UnicodeString unicode_text = icu::UnicodeString::fromUTF8(context);
  std::unique_ptr<icu::BreakIterator> break_iterator(
      icu::BreakIterator::createWordInstance(icu::Locale("en"), status));
  if (!status.isSuccess()) {
    TC_LOG(ERROR) << "Break iterator did not initialize properly: "
                  << status.errorName();
    return false;
  }

  break_iterator->setText(unicode_text);

  size_t last_break_index = 0;
  size_t break_index = 0;
  size_t last_unicode_index = 0;
  size_t unicode_index = 0;
  while ((break_index = break_iterator->next()) != icu::BreakIterator::DONE) {
    icu::UnicodeString token(unicode_text, last_break_index,
                             break_index - last_break_index);
    int token_length = token.countChar32();
    unicode_index = last_unicode_index + token_length;

    std::string token_utf8;
    token.toUTF8String(token_utf8);

    bool is_whitespace = true;
    for (int i = 0; i < token.length(); i++) {
      if (!u_isWhitespace(token.char32At(i))) {
        is_whitespace = false;
      }
    }

    if (!is_whitespace || options_.icu_preserve_whitespace_tokens()) {
      result->push_back(Token(token_utf8, last_unicode_index, unicode_index));
    }

    last_break_index = break_index;
    last_unicode_index = unicode_index;
  }

  return true;
}

void FeatureProcessor::InternalRetokenize(const std::string& context,
                                          std::vector<Token>* tokens) const {
  const UnicodeText unicode_text =
      UTF8ToUnicodeText(context, /*do_copy=*/false);

  std::vector<Token> result;
  CodepointSpan span(-1, -1);
  for (Token& token : *tokens) {
    const UnicodeText unicode_token_value =
        UTF8ToUnicodeText(token.value, /*do_copy=*/false);
    bool should_retokenize = true;
    for (const int codepoint : unicode_token_value) {
      if (!IsCodepointInRanges(codepoint,
                               internal_tokenizer_codepoint_ranges_)) {
        should_retokenize = false;
        break;
      }
    }

    if (should_retokenize) {
      if (span.first < 0) {
        span.first = token.start;
      }
      span.second = token.end;
    } else {
      TokenizeSubstring(unicode_text, span, &result);
      span.first = -1;
      result.emplace_back(std::move(token));
    }
  }
  TokenizeSubstring(unicode_text, span, &result);

  *tokens = std::move(result);
}

void FeatureProcessor::TokenizeSubstring(const UnicodeText& unicode_text,
                                         CodepointSpan span,
                                         std::vector<Token>* result) const {
  if (span.first < 0) {
    // There is no span to tokenize.
    return;
  }

  // Extract the substring.
  UnicodeText::const_iterator it_begin = unicode_text.begin();
  for (int i = 0; i < span.first; ++i) {
    ++it_begin;
  }
  UnicodeText::const_iterator it_end = unicode_text.begin();
  for (int i = 0; i < span.second; ++i) {
    ++it_end;
  }
  const std::string text = unicode_text.UTF8Substring(it_begin, it_end);

  // Run the tokenizer and update the token bounds to reflect the offset of the
  // substring.
  std::vector<Token> tokens = tokenizer_.Tokenize(text);
  for (Token& token : tokens) {
    token.start += span.first;
    token.end += span.first;
    result->emplace_back(std::move(token));
  }
}

}  // namespace libtextclassifier