Consider escaped characters as single characters in BPE (#322)

Author: guillaumekln

bindings/python/test/test.py

@@ -341,6 +341,18 @@ def test_bpe_learner_no_pairs(tmpdir):
         tokenizer = learner.learn(model_path)
 
 
+def test_bpe_learner_escaped_character(tmpdir):
+    text = "คุณอาจจะทำอย ่ างนั ้ นไปซักพัก จนคุณเริ ่ มจะรู ้ สึกถึงมันจริงๆ"
+
+    tokenizer = pyonmttok.Tokenizer("aggressive", joiner_annotate=True)
+    learner = pyonmttok.BPELearner(tokenizer=tokenizer, symbols=5, min_frequency=1)
+    learner.ingest(text)
+    tokenizer = learner.learn(str(tmpdir.join("bpe.model")))
+
+    tokens = tokenizer(text)
+    assert "￭％0020่" in tokens
+
+
 @pytest.mark.parametrize("keep_vocab", [False, True])
 def test_sp_learner(tmpdir, keep_vocab):
     learner = pyonmttok.SentencePieceLearner(

include/onmt/Tokenizer.h

@@ -70,6 +70,8 @@ namespace onmt
     static const std::string spacer_marker;
     static const std::string ph_marker_open;
     static const std::string ph_marker_close;
+    static const std::string escaped_character_prefix;
+    static const size_t escaped_character_width;
 
     Tokenizer(Options options,
               const std::shared_ptr<const SubwordEncoder>& subword_encoder = nullptr);

src/BPE.cc

@@ -138,9 +138,23 @@ namespace onmt
     std::vector<std::string> pieces;
     pieces.reserve(chars.size());
 
+    static const auto escaped_character_prefix = (
+      unicode::utf8_to_cp(Tokenizer::escaped_character_prefix.c_str()));
+    size_t escaped_character_length = 0;
+
     for (const auto& c : chars)
     {
-      if (c.char_type == unicode::CharType::Mark)
+      if (escaped_character_length > 0)
+      {
+        pieces.back().append(c.data, c.length);
+        escaped_character_length--;
+      }
+      else if (c.value == escaped_character_prefix)
+      {
+        pieces.emplace_back(c.data, c.length);
+        escaped_character_length = Tokenizer::escaped_character_width;
+      }
+      else if (c.char_type == unicode::CharType::Mark)
       {
         if (pieces.empty())
           pieces.emplace_back(c.data, c.length);

src/Tokenizer.cc

@@ -18,9 +18,10 @@ namespace onmt
   const std::string Tokenizer::spacer_marker = "▁";
   const std::string Tokenizer::ph_marker_open = "｟";
   const std::string Tokenizer::ph_marker_close = "｠";
+  const std::string Tokenizer::escaped_character_prefix = "％";
+  const size_t Tokenizer::escaped_character_width = 4;
   static const unicode::code_point_t ph_marker_open_cp = 0xFF5F;
   static const unicode::code_point_t ph_marker_close_cp = 0xFF60;
-  static const std::string protected_character = "％";
   static const std::vector<std::pair<unicode::code_point_t, std::string>> substitutes = {
     {0x2581 /* ▁ */, "_"},
     {0xFFED /* ￭ */, "■"},
@@ -32,7 +33,6 @@ namespace onmt
 
   static const int placeholder_alphabet = -2;
   static const int number_alphabet = -3;
-  static const int hex_value_width = 4;
 
   Tokenizer::Mode Tokenizer::str_to_mode(const std::string& mode)
   {
@@ -290,21 +290,23 @@ namespace onmt
 
   static inline void unescape_characters(std::string& str)
   {
+    const auto& prefix = Tokenizer::escaped_character_prefix;
+    const auto& width = Tokenizer::escaped_character_width;
+
     for (size_t offset = 0;;)
     {
-      const size_t index = str.find(protected_character, offset);
-      if (index == std::string::npos
-          || index + protected_character.size() + hex_value_width > str.size())
+      const size_t index = str.find(prefix, offset);
+      if (index == std::string::npos || index + prefix.size() + width > str.size())
         break;
 
-      const std::string code = str.substr(index + protected_character.size(), hex_value_width);
+      const std::string code = str.substr(index + prefix.size(), width);
       const int v = hex_to_int(code);
       const std::string c = unicode::cp_to_utf8(v);
-      if (c.empty() || !c[0] || int_to_hex(v, hex_value_width) != code)
-        offset = index + protected_character.size();
+      if (c.empty() || !c[0] || int_to_hex(v, width) != code)
+        offset = index + prefix.size();
       else
       {
-        str.replace(index, protected_character.size() + hex_value_width, c);
+        str.replace(index, prefix.size() + width, c);
         offset = index + 1;
       }
     }
@@ -635,7 +637,8 @@ namespace onmt
       if (_no_substitution)
         append(character);
       else
-        append(protected_character + int_to_hex(character.value, hex_value_width));
+        append(Tokenizer::escaped_character_prefix
+               + int_to_hex(character.value, Tokenizer::escaped_character_width));
     }
 
     void flush_feature()