hebrewprober.py 13 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283
  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. from .charsetprober import CharSetProber
  28. from .constants import eNotMe, eDetecting
  29. from .compat import wrap_ord
  30. # This prober doesn't actually recognize a language or a charset.
  31. # It is a helper prober for the use of the Hebrew model probers
  32. ### General ideas of the Hebrew charset recognition ###
  33. #
  34. # Four main charsets exist in Hebrew:
  35. # "ISO-8859-8" - Visual Hebrew
  36. # "windows-1255" - Logical Hebrew
  37. # "ISO-8859-8-I" - Logical Hebrew
  38. # "x-mac-hebrew" - ?? Logical Hebrew ??
  39. #
  40. # Both "ISO" charsets use a completely identical set of code points, whereas
  41. # "windows-1255" and "x-mac-hebrew" are two different proper supersets of
  42. # these code points. windows-1255 defines additional characters in the range
  43. # 0x80-0x9F as some misc punctuation marks as well as some Hebrew-specific
  44. # diacritics and additional 'Yiddish' ligature letters in the range 0xc0-0xd6.
  45. # x-mac-hebrew defines similar additional code points but with a different
  46. # mapping.
  47. #
  48. # As far as an average Hebrew text with no diacritics is concerned, all four
  49. # charsets are identical with respect to code points. Meaning that for the
  50. # main Hebrew alphabet, all four map the same values to all 27 Hebrew letters
  51. # (including final letters).
  52. #
  53. # The dominant difference between these charsets is their directionality.
  54. # "Visual" directionality means that the text is ordered as if the renderer is
  55. # not aware of a BIDI rendering algorithm. The renderer sees the text and
  56. # draws it from left to right. The text itself when ordered naturally is read
  57. # backwards. A buffer of Visual Hebrew generally looks like so:
  58. # "[last word of first line spelled backwards] [whole line ordered backwards
  59. # and spelled backwards] [first word of first line spelled backwards]
  60. # [end of line] [last word of second line] ... etc' "
  61. # adding punctuation marks, numbers and English text to visual text is
  62. # naturally also "visual" and from left to right.
  63. #
  64. # "Logical" directionality means the text is ordered "naturally" according to
  65. # the order it is read. It is the responsibility of the renderer to display
  66. # the text from right to left. A BIDI algorithm is used to place general
  67. # punctuation marks, numbers and English text in the text.
  68. #
  69. # Texts in x-mac-hebrew are almost impossible to find on the Internet. From
  70. # what little evidence I could find, it seems that its general directionality
  71. # is Logical.
  72. #
  73. # To sum up all of the above, the Hebrew probing mechanism knows about two
  74. # charsets:
  75. # Visual Hebrew - "ISO-8859-8" - backwards text - Words and sentences are
  76. # backwards while line order is natural. For charset recognition purposes
  77. # the line order is unimportant (In fact, for this implementation, even
  78. # word order is unimportant).
  79. # Logical Hebrew - "windows-1255" - normal, naturally ordered text.
  80. #
  81. # "ISO-8859-8-I" is a subset of windows-1255 and doesn't need to be
  82. # specifically identified.
  83. # "x-mac-hebrew" is also identified as windows-1255. A text in x-mac-hebrew
  84. # that contain special punctuation marks or diacritics is displayed with
  85. # some unconverted characters showing as question marks. This problem might
  86. # be corrected using another model prober for x-mac-hebrew. Due to the fact
  87. # that x-mac-hebrew texts are so rare, writing another model prober isn't
  88. # worth the effort and performance hit.
  89. #
  90. #### The Prober ####
  91. #
  92. # The prober is divided between two SBCharSetProbers and a HebrewProber,
  93. # all of which are managed, created, fed data, inquired and deleted by the
  94. # SBCSGroupProber. The two SBCharSetProbers identify that the text is in
  95. # fact some kind of Hebrew, Logical or Visual. The final decision about which
  96. # one is it is made by the HebrewProber by combining final-letter scores
  97. # with the scores of the two SBCharSetProbers to produce a final answer.
  98. #
  99. # The SBCSGroupProber is responsible for stripping the original text of HTML
  100. # tags, English characters, numbers, low-ASCII punctuation characters, spaces
  101. # and new lines. It reduces any sequence of such characters to a single space.
  102. # The buffer fed to each prober in the SBCS group prober is pure text in
  103. # high-ASCII.
  104. # The two SBCharSetProbers (model probers) share the same language model:
  105. # Win1255Model.
  106. # The first SBCharSetProber uses the model normally as any other
  107. # SBCharSetProber does, to recognize windows-1255, upon which this model was
  108. # built. The second SBCharSetProber is told to make the pair-of-letter
  109. # lookup in the language model backwards. This in practice exactly simulates
  110. # a visual Hebrew model using the windows-1255 logical Hebrew model.
  111. #
  112. # The HebrewProber is not using any language model. All it does is look for
  113. # final-letter evidence suggesting the text is either logical Hebrew or visual
  114. # Hebrew. Disjointed from the model probers, the results of the HebrewProber
  115. # alone are meaningless. HebrewProber always returns 0.00 as confidence
  116. # since it never identifies a charset by itself. Instead, the pointer to the
  117. # HebrewProber is passed to the model probers as a helper "Name Prober".
  118. # When the Group prober receives a positive identification from any prober,
  119. # it asks for the name of the charset identified. If the prober queried is a
  120. # Hebrew model prober, the model prober forwards the call to the
  121. # HebrewProber to make the final decision. In the HebrewProber, the
  122. # decision is made according to the final-letters scores maintained and Both
  123. # model probers scores. The answer is returned in the form of the name of the
  124. # charset identified, either "windows-1255" or "ISO-8859-8".
  125. # windows-1255 / ISO-8859-8 code points of interest
  126. FINAL_KAF = 0xea
  127. NORMAL_KAF = 0xeb
  128. FINAL_MEM = 0xed
  129. NORMAL_MEM = 0xee
  130. FINAL_NUN = 0xef
  131. NORMAL_NUN = 0xf0
  132. FINAL_PE = 0xf3
  133. NORMAL_PE = 0xf4
  134. FINAL_TSADI = 0xf5
  135. NORMAL_TSADI = 0xf6
  136. # Minimum Visual vs Logical final letter score difference.
  137. # If the difference is below this, don't rely solely on the final letter score
  138. # distance.
  139. MIN_FINAL_CHAR_DISTANCE = 5
  140. # Minimum Visual vs Logical model score difference.
  141. # If the difference is below this, don't rely at all on the model score
  142. # distance.
  143. MIN_MODEL_DISTANCE = 0.01
  144. VISUAL_HEBREW_NAME = "ISO-8859-8"
  145. LOGICAL_HEBREW_NAME = "windows-1255"
  146. class HebrewProber(CharSetProber):
  147. def __init__(self):
  148. CharSetProber.__init__(self)
  149. self._mLogicalProber = None
  150. self._mVisualProber = None
  151. self.reset()
  152. def reset(self):
  153. self._mFinalCharLogicalScore = 0
  154. self._mFinalCharVisualScore = 0
  155. # The two last characters seen in the previous buffer,
  156. # mPrev and mBeforePrev are initialized to space in order to simulate
  157. # a word delimiter at the beginning of the data
  158. self._mPrev = ' '
  159. self._mBeforePrev = ' '
  160. # These probers are owned by the group prober.
  161. def set_model_probers(self, logicalProber, visualProber):
  162. self._mLogicalProber = logicalProber
  163. self._mVisualProber = visualProber
  164. def is_final(self, c):
  165. return wrap_ord(c) in [FINAL_KAF, FINAL_MEM, FINAL_NUN, FINAL_PE,
  166. FINAL_TSADI]
  167. def is_non_final(self, c):
  168. # The normal Tsadi is not a good Non-Final letter due to words like
  169. # 'lechotet' (to chat) containing an apostrophe after the tsadi. This
  170. # apostrophe is converted to a space in FilterWithoutEnglishLetters
  171. # causing the Non-Final tsadi to appear at an end of a word even
  172. # though this is not the case in the original text.
  173. # The letters Pe and Kaf rarely display a related behavior of not being
  174. # a good Non-Final letter. Words like 'Pop', 'Winamp' and 'Mubarak'
  175. # for example legally end with a Non-Final Pe or Kaf. However, the
  176. # benefit of these letters as Non-Final letters outweighs the damage
  177. # since these words are quite rare.
  178. return wrap_ord(c) in [NORMAL_KAF, NORMAL_MEM, NORMAL_NUN, NORMAL_PE]
  179. def feed(self, aBuf):
  180. # Final letter analysis for logical-visual decision.
  181. # Look for evidence that the received buffer is either logical Hebrew
  182. # or visual Hebrew.
  183. # The following cases are checked:
  184. # 1) A word longer than 1 letter, ending with a final letter. This is
  185. # an indication that the text is laid out "naturally" since the
  186. # final letter really appears at the end. +1 for logical score.
  187. # 2) A word longer than 1 letter, ending with a Non-Final letter. In
  188. # normal Hebrew, words ending with Kaf, Mem, Nun, Pe or Tsadi,
  189. # should not end with the Non-Final form of that letter. Exceptions
  190. # to this rule are mentioned above in isNonFinal(). This is an
  191. # indication that the text is laid out backwards. +1 for visual
  192. # score
  193. # 3) A word longer than 1 letter, starting with a final letter. Final
  194. # letters should not appear at the beginning of a word. This is an
  195. # indication that the text is laid out backwards. +1 for visual
  196. # score.
  197. #
  198. # The visual score and logical score are accumulated throughout the
  199. # text and are finally checked against each other in GetCharSetName().
  200. # No checking for final letters in the middle of words is done since
  201. # that case is not an indication for either Logical or Visual text.
  202. #
  203. # We automatically filter out all 7-bit characters (replace them with
  204. # spaces) so the word boundary detection works properly. [MAP]
  205. if self.get_state() == eNotMe:
  206. # Both model probers say it's not them. No reason to continue.
  207. return eNotMe
  208. aBuf = self.filter_high_bit_only(aBuf)
  209. for cur in aBuf:
  210. if cur == ' ':
  211. # We stand on a space - a word just ended
  212. if self._mBeforePrev != ' ':
  213. # next-to-last char was not a space so self._mPrev is not a
  214. # 1 letter word
  215. if self.is_final(self._mPrev):
  216. # case (1) [-2:not space][-1:final letter][cur:space]
  217. self._mFinalCharLogicalScore += 1
  218. elif self.is_non_final(self._mPrev):
  219. # case (2) [-2:not space][-1:Non-Final letter][
  220. # cur:space]
  221. self._mFinalCharVisualScore += 1
  222. else:
  223. # Not standing on a space
  224. if ((self._mBeforePrev == ' ') and
  225. (self.is_final(self._mPrev)) and (cur != ' ')):
  226. # case (3) [-2:space][-1:final letter][cur:not space]
  227. self._mFinalCharVisualScore += 1
  228. self._mBeforePrev = self._mPrev
  229. self._mPrev = cur
  230. # Forever detecting, till the end or until both model probers return
  231. # eNotMe (handled above)
  232. return eDetecting
  233. def get_charset_name(self):
  234. # Make the decision: is it Logical or Visual?
  235. # If the final letter score distance is dominant enough, rely on it.
  236. finalsub = self._mFinalCharLogicalScore - self._mFinalCharVisualScore
  237. if finalsub >= MIN_FINAL_CHAR_DISTANCE:
  238. return LOGICAL_HEBREW_NAME
  239. if finalsub <= -MIN_FINAL_CHAR_DISTANCE:
  240. return VISUAL_HEBREW_NAME
  241. # It's not dominant enough, try to rely on the model scores instead.
  242. modelsub = (self._mLogicalProber.get_confidence()
  243. - self._mVisualProber.get_confidence())
  244. if modelsub > MIN_MODEL_DISTANCE:
  245. return LOGICAL_HEBREW_NAME
  246. if modelsub < -MIN_MODEL_DISTANCE:
  247. return VISUAL_HEBREW_NAME
  248. # Still no good, back to final letter distance, maybe it'll save the
  249. # day.
  250. if finalsub < 0.0:
  251. return VISUAL_HEBREW_NAME
  252. # (finalsub > 0 - Logical) or (don't know what to do) default to
  253. # Logical.
  254. return LOGICAL_HEBREW_NAME
  255. def get_state(self):
  256. # Remain active as long as any of the model probers are active.
  257. if (self._mLogicalProber.get_state() == eNotMe) and \
  258. (self._mVisualProber.get_state() == eNotMe):
  259. return eNotMe
  260. return eDetecting