Handle the punctuation definition mismatch between different Unicode versions.

PiperOrigin-RevId: 707239296

Author: tf-text-github-robot

tensorflow_text/core/kernels/BUILD

@@ -403,6 +403,8 @@ cc_test(
     srcs = ["fast_wordpiece_tokenizer_test.cc"],
     data = [
         "//tensorflow_text:python/ops/test_data/fast_wordpiece_tokenizer_model.fb",
+        "//tensorflow_text:python/ops/test_data/fast_wordpiece_tokenizer_model_ver_15_1.fb",
+        "//tensorflow_text:python/ops/test_data/fast_wordpiece_tokenizer_model_ver_16_0.fb",
     ],
     deps = [
         ":fast_wordpiece_tokenizer",

tensorflow_text/core/kernels/fast_wordpiece_tokenizer.cc

@@ -278,14 +278,16 @@ void FastWordpieceTokenizer::TokenizeTextImpl(
                         prev_unicode_char))) {
       // If the current Unicode character is a valid word boundary, collect the
       // remaining tokens stored on a path on the trie.
+      absl::string_view cur_str = absl::string_view(
+          input_substr.data(), cur_pos - input_word_offset_in_text);
       HandleTheRemainingStringOnTriePath<kGetPieces, kGetIds, kGetOffsets>(
-          absl::string_view(input_substr.data(),
-                            cur_pos - input_word_offset_in_text),
-          input_word_offset_in_text, cur_node, original_num_tokens,
+          cur_str, input_word_offset_in_text, cur_node, original_num_tokens,
           cur_offset_in_input_word, output_pieces, output_ids,
           output_start_offsets, output_end_offsets);
       // Skip the whitespace.
-      if (is_white_space) cur_pos = next_pos;
+      // If the remaining tokens are empty, it means we encountered an
+      // unmappable separator, so skip to the next token.
+      if (is_white_space || cur_str.empty()) cur_pos = next_pos;
       // Continue in the outer while loop to process the remaining input.
       continue;
     }

tensorflow_text/core/kernels/fast_wordpiece_tokenizer_test.cc

@@ -58,6 +58,84 @@ TEST(FastWordpieceTokenizerTest, LoadAndTokenize) {
   EXPECT_THAT(output_end_offsets, ElementsAre(3, 5, 6, 9));
 }
 
+TEST(FastWordpieceTokenizerTest, PunctuationVersionMismatch) {
+  // The config_flatbuffer used here is built from the following config:
+  //  * vocab = {"a", "abc", "abcdefghi", "##de", "##defgxy", "##deh", "##f",
+  //             "##ghz", "<unk>"}
+  //  * unk_token = "<unk>"
+  //  * suffix_indicator = "##"
+  //  * max_bytes_per_token = 100
+  //  * end_to_end = True
+
+  // Built with unicode ver 15.1
+  const char kTestConfigUnicode15_1Path[] =
+      "third_party/tensorflow_text/python/ops/test_data/"
+      "fast_wordpiece_tokenizer_model_ver_15_1.fb";
+
+  // Built with unicode ver 16.0
+  const char kTestConfigUnicode16_0Path[] =
+      "third_party/tensorflow_text/python/ops/test_data/"
+      "fast_wordpiece_tokenizer_model_ver_16_0.fb";
+
+  // We test the new punctuation symbol: \341\255\277, which was available in
+  // Unicode 16: https://www.fileformat.info/info/unicode/char//1b7f/index.htm,
+  // but not in 15.1.
+  // We also test an existing punctuation symbol ">".
+  std::string input = "abc>abc\341\255\277abc";
+
+  // Read 15.1 config.
+  {
+    std::string config_flatbuffer;
+    auto status = tensorflow::ReadFileToString(tensorflow::Env::Default(),
+                                               kTestConfigUnicode15_1Path,
+                                               &config_flatbuffer);
+    ASSERT_TRUE(status.ok());
+
+    ASSERT_OK_AND_ASSIGN(auto tokenizer, FastWordpieceTokenizer::Create(
+                                             config_flatbuffer.data()));
+
+    std::vector<std::string> output_tokens;
+    std::vector<int> output_ids;
+    std::vector<int> output_start_offsets;
+    std::vector<int> output_end_offsets;
+    tokenizer.Tokenize(input, &output_tokens, &output_ids,
+                       &output_start_offsets, &output_end_offsets);
+
+    // For 15.1, the flatbuffer does not have \341\255\277 as a punctuation.
+    EXPECT_THAT(output_tokens, ElementsAre("abc", "<unk>", "abc", "abc"));
+    EXPECT_THAT(output_ids, ElementsAre(1, 8, 1, 1));
+    // Note that the new-version punctuation symbol is ignored.
+    EXPECT_THAT(output_start_offsets, ElementsAre(0, 3, 4, 10));
+    EXPECT_THAT(output_end_offsets, ElementsAre(3, 4, 7, 13));
+  }
+
+  // Read 16.0 config.
+  {
+    std::string config_flatbuffer;
+    auto status = tensorflow::ReadFileToString(tensorflow::Env::Default(),
+                                               kTestConfigUnicode16_0Path,
+                                               &config_flatbuffer);
+    ASSERT_TRUE(status.ok());
+
+    ASSERT_OK_AND_ASSIGN(auto tokenizer, FastWordpieceTokenizer::Create(
+                                             config_flatbuffer.data()));
+
+    std::vector<std::string> output_tokens;
+    std::vector<int> output_ids;
+    std::vector<int> output_start_offsets;
+    std::vector<int> output_end_offsets;
+    tokenizer.Tokenize(input, &output_tokens, &output_ids,
+                       &output_start_offsets, &output_end_offsets);
+
+    // For 16.0, \341\255\277 is treated as a punctuation.
+    EXPECT_THAT(output_tokens,
+                ElementsAre("abc", "<unk>", "abc", "<unk>", "abc"));
+    EXPECT_THAT(output_ids, ElementsAre(1, 8, 1, 8, 1));
+    EXPECT_THAT(output_start_offsets, ElementsAre(0, 3, 4, 7, 10));
+    EXPECT_THAT(output_end_offsets, ElementsAre(3, 4, 7, 10, 13));
+  }
+}
+
 template <typename T>
 std::string ListToString(const std::vector<T>& list) {
   return absl::StrCat("[", absl::StrJoin(list, ", "), "]");