Work on making Japanese tokenization use MeCab consistently

Former-commit-id: 05cf94d1fd

wordfreq/__init__.py

@@ -18,15 +18,41 @@ DATA_PATH = pathlib.Path(resource_filename('wordfreq', 'data'))
 CACHE_SIZE = 100000
 
 
-def tokenize(text):
+def simple_tokenize(text):
     """
-    A simple tokenizer that can be applied to most languages. Strings that
+    A simple tokenizer that can be applied to most languages.
-    are looked up in wordfreq will be run through this tokenizer first,
+
-    so that they can be expected to match the data.
+    It considers a word to be made of a sequence of 'token characters', an
+    overly inclusive range that includes letters, Han characters, emoji, and a
+    bunch of miscellaneous whatnot, but excludes most punctuation and
+    whitespace.
+
+    The single complication for the sake of English is that apostrophes are not
+    considered part of the token if they appear on the edge of the character
+    sequence, but they are if they appear internally. "cats'" is not a token,
+    but "cat's" is.
     """
     return [token.lower() for token in TOKEN_RE.findall(text)]
 
 
+def tokenize(text, lang):
+    """
+    Tokenize this text in a way that's straightforward but appropriate for
+    the language.
+
+    So far, this means that Japanese is handled by mecab_tokenize, and
+    everything else is handled by simple_tokenize.
+
+    Strings that are looked up in wordfreq will be run through this function
+    first, so that they can be expected to match the data.
+    """
+    if lang == 'ja':
+        from wordfreq.mecab import mecab_tokenize
+        return mecab_tokenize(text)
+    else:
+        return simple_tokenize(text)
+
+
 def read_dBpack(filename):
     """
     Read a file from an idiosyncratic format that we use for storing
@@ -163,7 +189,7 @@ def word_frequency(word, lang, wordlist='combined', default=0.):
     """
     freqs = get_frequency_dict(lang, wordlist)
     combined_value = None
-    for token in tokenize(word):
+    for token in tokenize(word, lang):
         if token not in freqs:
             # If any word is missing, just return the default value
             return default

wordfreq/mecab.py

@@ -0,0 +1,21 @@
+import MeCab
+
+
+# Instantiate the MeCab analyzer, which the mecab-python3 interface calls a
+# Tagger.
+MECAB_ANALYZER = MeCab.Tagger()
+
+
+def mecab_tokenize(text):
+    """
+    Use the mecab-python3 package to tokenize the given Japanese text.
+
+    The simplest output from mecab-python3 is the single-string form, which
+    contains the same table that the command-line version of MeCab would output.
+    We find the tokens in the first column of this table.
+    """
+    parsed_str = MECAB_ANALYZER.parse(text.strip())
+    lines = [line for line in parsed_str.split('\n')
+             if line != '' and line != 'EOS']
+    tokens = [line.split('\t')[0] for line in lines]
+    return tokens