chardet/hebrewprober.py

   1 ######################## BEGIN LICENSE BLOCK ########################
   2 # The Original Code is Mozilla Universal charset detector code.
   3 #
   4 # The Initial Developer of the Original Code is
   5 #          Shy Shalom
   6 # Portions created by the Initial Developer are Copyright (C) 2005
   7 # the Initial Developer. All Rights Reserved.
   8 #
   9 # Contributor(s):
  10 #   Mark Pilgrim - port to Python
  11 #
  12 # This library is free software; you can redistribute it and/or
  13 # modify it under the terms of the GNU Lesser General Public
  14 # License as published by the Free Software Foundation; either
  15 # version 2.1 of the License, or (at your option) any later version.
  16 #
  17 # This library is distributed in the hope that it will be useful,
  18 # but WITHOUT ANY WARRANTY; without even the implied warranty of
  19 # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  20 # Lesser General Public License for more details.
  21 #
  22 # You should have received a copy of the GNU Lesser General Public
  23 # License along with this library; if not, write to the Free Software
  24 # Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA
  25 # 02110-1301  USA
  26 ######################### END LICENSE BLOCK #########################
  27
  28 from charsetprober import CharSetProber
  29 import constants
  30
  31 # This prober doesn't actually recognize a language or a charset.
  32 # It is a helper prober for the use of the Hebrew model probers
  33
  34 ### General ideas of the Hebrew charset recognition ###
  35 #
  36 # Four main charsets exist in Hebrew:
  37 # "ISO-8859-8" - Visual Hebrew
  38 # "windows-1255" - Logical Hebrew
  39 # "ISO-8859-8-I" - Logical Hebrew
  40 # "x-mac-hebrew" - ?? Logical Hebrew ??
  41 #
  42 # Both "ISO" charsets use a completely identical set of code points, whereas
  43 # "windows-1255" and "x-mac-hebrew" are two different proper supersets of
  44 # these code points. windows-1255 defines additional characters in the range
  45 # 0x80-0x9F as some misc punctuation marks as well as some Hebrew-specific
  46 # diacritics and additional 'Yiddish' ligature letters in the range 0xc0-0xd6.
  47 # x-mac-hebrew defines similar additional code points but with a different
  48 # mapping.
  49 #
  50 # As far as an average Hebrew text with no diacritics is concerned, all four
  51 # charsets are identical with respect to code points. Meaning that for the
  52 # main Hebrew alphabet, all four map the same values to all 27 Hebrew letters
  53 # (including final letters).
  54 #
  55 # The dominant difference between these charsets is their directionality.
  56 # "Visual" directionality means that the text is ordered as if the renderer is
  57 # not aware of a BIDI rendering algorithm. The renderer sees the text and
  58 # draws it from left to right. The text itself when ordered naturally is read
  59 # backwards. A buffer of Visual Hebrew generally looks like so:
  60 # "[last word of first line spelled backwards] [whole line ordered backwards
  61 # and spelled backwards] [first word of first line spelled backwards]
  62 # [end of line] [last word of second line] ... etc' "
  63 # adding punctuation marks, numbers and English text to visual text is
  64 # naturally also "visual" and from left to right.
  65 #
  66 # "Logical" directionality means the text is ordered "naturally" according to
  67 # the order it is read. It is the responsibility of the renderer to display
  68 # the text from right to left. A BIDI algorithm is used to place general
  69 # punctuation marks, numbers and English text in the text.
  70 #
  71 # Texts in x-mac-hebrew are almost impossible to find on the Internet. From
  72 # what little evidence I could find, it seems that its general directionality
  73 # is Logical.
  74 #
  75 # To sum up all of the above, the Hebrew probing mechanism knows about two
  76 # charsets:
  77 # Visual Hebrew - "ISO-8859-8" - backwards text - Words and sentences are
  78 #    backwards while line order is natural. For charset recognition purposes
  79 #    the line order is unimportant (In fact, for this implementation, even
  80 #    word order is unimportant).
  81 # Logical Hebrew - "windows-1255" - normal, naturally ordered text.
  82 #
  83 # "ISO-8859-8-I" is a subset of windows-1255 and doesn't need to be
  84 #    specifically identified.
  85 # "x-mac-hebrew" is also identified as windows-1255. A text in x-mac-hebrew
  86 #    that contain special punctuation marks or diacritics is displayed with
  87 #    some unconverted characters showing as question marks. This problem might
  88 #    be corrected using another model prober for x-mac-hebrew. Due to the fact
  89 #    that x-mac-hebrew texts are so rare, writing another model prober isn't
  90 #    worth the effort and performance hit.
  91 #
  92 #### The Prober ####
  93 #
  94 # The prober is divided between two SBCharSetProbers and a HebrewProber,
  95 # all of which are managed, created, fed data, inquired and deleted by the
  96 # SBCSGroupProber. The two SBCharSetProbers identify that the text is in
  97 # fact some kind of Hebrew, Logical or Visual. The final decision about which
  98 # one is it is made by the HebrewProber by combining final-letter scores
  99 # with the scores of the two SBCharSetProbers to produce a final answer.
 100 #
 101 # The SBCSGroupProber is responsible for stripping the original text of HTML
 102 # tags, English characters, numbers, low-ASCII punctuation characters, spaces
 103 # and new lines. It reduces any sequence of such characters to a single space.
 104 # The buffer fed to each prober in the SBCS group prober is pure text in
 105 # high-ASCII.
 106 # The two SBCharSetProbers (model probers) share the same language model:
 107 # Win1255Model.
 108 # The first SBCharSetProber uses the model normally as any other
 109 # SBCharSetProber does, to recognize windows-1255, upon which this model was
 110 # built. The second SBCharSetProber is told to make the pair-of-letter
 111 # lookup in the language model backwards. This in practice exactly simulates
 112 # a visual Hebrew model using the windows-1255 logical Hebrew model.
 113 #
 114 # The HebrewProber is not using any language model. All it does is look for
 115 # final-letter evidence suggesting the text is either logical Hebrew or visual
 116 # Hebrew. Disjointed from the model probers, the results of the HebrewProber
 117 # alone are meaningless. HebrewProber always returns 0.00 as confidence
 118 # since it never identifies a charset by itself. Instead, the pointer to the
 119 # HebrewProber is passed to the model probers as a helper "Name Prober".
 120 # When the Group prober receives a positive identification from any prober,
 121 # it asks for the name of the charset identified. If the prober queried is a
 122 # Hebrew model prober, the model prober forwards the call to the
 123 # HebrewProber to make the final decision. In the HebrewProber, the
 124 # decision is made according to the final-letters scores maintained and Both
 125 # model probers scores. The answer is returned in the form of the name of the
 126 # charset identified, either "windows-1255" or "ISO-8859-8".
 127
 128 # windows-1255 / ISO-8859-8 code points of interest
 129 FINAL_KAF = '\xea'
 130 NORMAL_KAF = '\xeb'
 131 FINAL_MEM = '\xed'
 132 NORMAL_MEM = '\xee'
 133 FINAL_NUN = '\xef'
 134 NORMAL_NUN = '\xf0'
 135 FINAL_PE = '\xf3'
 136 NORMAL_PE = '\xf4'
 137 FINAL_TSADI = '\xf5'
 138 NORMAL_TSADI = '\xf6'
 139
 140 # Minimum Visual vs Logical final letter score difference.
 141 # If the difference is below this, don't rely solely on the final letter score distance.
 142 MIN_FINAL_CHAR_DISTANCE = 5
 143
 144 # Minimum Visual vs Logical model score difference.
 145 # If the difference is below this, don't rely at all on the model score distance.
 146 MIN_MODEL_DISTANCE = 0.01
 147
 148 VISUAL_HEBREW_NAME = "ISO-8859-8"
 149 LOGICAL_HEBREW_NAME = "windows-1255"
 150
 151 class HebrewProber(CharSetProber):
 152     def __init__(self):
 153         CharSetProber.__init__(self)
 154         self._mLogicalProber = None
 155         self._mVisualProber = None
 156         self.reset()
 157
 158     def reset(self):
 159         self._mFinalCharLogicalScore = 0
 160         self._mFinalCharVisualScore = 0
 161         # The two last characters seen in the previous buffer,
 162         # mPrev and mBeforePrev are initialized to space in order to simulate a word
 163         # delimiter at the beginning of the data
 164         self._mPrev = ' '
 165         self._mBeforePrev = ' '
 166         # These probers are owned by the group prober.
 167
 168     def set_model_probers(self, logicalProber, visualProber):
 169         self._mLogicalProber = logicalProber
 170         self._mVisualProber = visualProber
 171
 172     def is_final(self, c):
 173         return c in [FINAL_KAF, FINAL_MEM, FINAL_NUN, FINAL_PE, FINAL_TSADI]
 174
 175     def is_non_final(self, c):
 176         # The normal Tsadi is not a good Non-Final letter due to words like
 177         # 'lechotet' (to chat) containing an apostrophe after the tsadi. This
 178         # apostrophe is converted to a space in FilterWithoutEnglishLetters causing
 179         # the Non-Final tsadi to appear at an end of a word even though this is not
 180         # the case in the original text.
 181         # The letters Pe and Kaf rarely display a related behavior of not being a
 182         # good Non-Final letter. Words like 'Pop', 'Winamp' and 'Mubarak' for
 183         # example legally end with a Non-Final Pe or Kaf. However, the benefit of
 184         # these letters as Non-Final letters outweighs the damage since these words
 185         # are quite rare.
 186         return c in [NORMAL_KAF, NORMAL_MEM, NORMAL_NUN, NORMAL_PE]
 187
 188     def feed(self, aBuf):
 189         # Final letter analysis for logical-visual decision.
 190         # Look for evidence that the received buffer is either logical Hebrew or
 191         # visual Hebrew.
 192         # The following cases are checked:
 193         # 1) A word longer than 1 letter, ending with a final letter. This is an
 194         #    indication that the text is laid out "naturally" since the final letter
 195         #    really appears at the end. +1 for logical score.
 196         # 2) A word longer than 1 letter, ending with a Non-Final letter. In normal
 197         #    Hebrew, words ending with Kaf, Mem, Nun, Pe or Tsadi, should not end with
 198         #    the Non-Final form of that letter. Exceptions to this rule are mentioned
 199         #    above in isNonFinal(). This is an indication that the text is laid out
 200         #    backwards. +1 for visual score
 201         # 3) A word longer than 1 letter, starting with a final letter. Final letters
 202         #    should not appear at the beginning of a word. This is an indication that
 203         #    the text is laid out backwards. +1 for visual score.
 204         #
 205         # The visual score and logical score are accumulated throughout the text and
 206         # are finally checked against each other in GetCharSetName().
 207         # No checking for final letters in the middle of words is done since that case
 208         # is not an indication for either Logical or Visual text.
 209         #
 210         # We automatically filter out all 7-bit characters (replace them with spaces)
 211         # so the word boundary detection works properly. [MAP]
 212
 213         if self.get_state() == constants.eNotMe:
 214             # Both model probers say it's not them. No reason to continue.
 215             return constants.eNotMe
 216
 217         aBuf = self.filter_high_bit_only(aBuf)
 218
 219         for cur in aBuf:
 220             if cur == ' ':
 221                 # We stand on a space - a word just ended
 222                 if self._mBeforePrev != ' ':
 223                     # next-to-last char was not a space so self._mPrev is not a 1 letter word
 224                     if self.is_final(self._mPrev):
 225                         # case (1) [-2:not space][-1:final letter][cur:space]
 226                         self._mFinalCharLogicalScore += 1
 227                     elif self.is_non_final(self._mPrev):
 228                         # case (2) [-2:not space][-1:Non-Final letter][cur:space]
 229                         self._mFinalCharVisualScore += 1
 230             else:
 231                 # Not standing on a space
 232                 if (self._mBeforePrev == ' ') and (self.is_final(self._mPrev)) and (cur != ' '):
 233                     # case (3) [-2:space][-1:final letter][cur:not space]
 234                     self._mFinalCharVisualScore += 1
 235             self._mBeforePrev = self._mPrev
 236             self._mPrev = cur
 237
 238         # Forever detecting, till the end or until both model probers return eNotMe (handled above)
 239         return constants.eDetecting
 240
 241     def get_charset_name(self):
 242         # Make the decision: is it Logical or Visual?
 243         # If the final letter score distance is dominant enough, rely on it.
 244         finalsub = self._mFinalCharLogicalScore - self._mFinalCharVisualScore
 245         if finalsub >= MIN_FINAL_CHAR_DISTANCE:
 246             return LOGICAL_HEBREW_NAME
 247         if finalsub <= -MIN_FINAL_CHAR_DISTANCE:
 248             return VISUAL_HEBREW_NAME
 249
 250         # It's not dominant enough, try to rely on the model scores instead.
 251         modelsub = self._mLogicalProber.get_confidence() - self._mVisualProber.get_confidence()
 252         if modelsub > MIN_MODEL_DISTANCE:
 253             return LOGICAL_HEBREW_NAME
 254         if modelsub < -MIN_MODEL_DISTANCE:
 255             return VISUAL_HEBREW_NAME
 256
 257         # Still no good, back to final letter distance, maybe it'll save the day.
 258         if finalsub < 0.0:
 259             return VISUAL_HEBREW_NAME
 260
 261         # (finalsub > 0 - Logical) or (don't know what to do) default to Logical.
 262         return LOGICAL_HEBREW_NAME
 263
 264     def get_state(self):
 265         # Remain active as long as any of the model probers are active.
 266         if (self._mLogicalProber.get_state() == constants.eNotMe) and \
 267            (self._mVisualProber.get_state() == constants.eNotMe):
 268             return constants.eNotMe
 269         return constants.eDetecting