Java > Open Source Codes > java > text > DictionaryBasedBreakIterator


1 /*
2  * @(#)DictionaryBasedBreakIterator.java 1.13 03/12/19
3  *
4  * Copyright 2004 Sun Microsystems, Inc. All rights reserved.
5  * SUN PROPRIETARY/CONFIDENTIAL. Use is subject to license terms.
6  */
7
8 /*
9  * @(#)DictionaryBasedBreakIterator.java 1.3 99/05/03
10  *
11  * (C) Copyright Taligent, Inc. 1996, 1997 - All Rights Reserved
12  * (C) Copyright IBM Corp. 1996 - 2002 - All Rights Reserved
13  *
14  * The original version of this source code and documentation
15  * is copyrighted and owned by Taligent, Inc., a wholly-owned
16  * subsidiary of IBM. These materials are provided under terms
17  * of a License Agreement between Taligent and Sun. This technology
18  * is protected by multiple US and International patents.
19  *
20  * This notice and attribution to Taligent may not be removed.
21  * Taligent is a registered trademark of Taligent, Inc.
22  */
23
24 package java.text;
25
26 import java.util.Vector ;
27 import java.util.Stack ;
28 import java.util.Hashtable ;
29 import java.text.CharacterIterator ;
30 import java.io.InputStream ;
31 import java.io.IOException ;
32
33 /**
34  * A subclass of RuleBasedBreakIterator that adds the ability to use a dictionary
35  * to further subdivide ranges of text beyond what is possible using just the
36  * state-table-based algorithm. This is necessary, for example, to handle
37  * word and line breaking in Thai, which doesn't use spaces between words. The
38  * state-table-based algorithm used by RuleBasedBreakIterator is used to divide
39  * up text as far as possible, and then contiguous ranges of letters are
40  * repeatedly compared against a list of known words (i.e., the dictionary)
41  * to divide them up into words.
42  *
43  * DictionaryBasedBreakIterator uses the same rule language as RuleBasedBreakIterator,
44  * but adds one more special substitution name: <dictionary>. This substitution
45  * name is used to identify characters in words in the dictionary. The idea is that
46  * if the iterator passes over a chunk of text that includes two or more characters
47  * in a row that are included in <dictionary>, it goes back through that range and
48  * derives additional break positions (if possible) using the dictionary.
49  *
50  * DictionaryBasedBreakIterator is also constructed with the filename of a dictionary
51  * file. It follows a prescribed search path to locate the dictionary (right now,
52  * it looks for it in /com/ibm/text/resources in each directory in the classpath,
53  * and won't find it in JAR files, but this location is likely to change). The
54  * dictionary file is in a serialized binary format. We have a very primitive (and
55  * slow) BuildDictionaryFile utility for creating dictionary files, but aren't
56  * currently making it public. Contact us for help.
57  */
58 class DictionaryBasedBreakIterator extends RuleBasedBreakIterator {
59
60     /**
61      * a list of known words that is used to divide up contiguous ranges of letters,
62      * stored in a compressed, indexed, format that offers fast access
63      */
64     private BreakDictionary dictionary;
65
66     /**
67      * a list of flags indicating which character categories are contained in
68      * the dictionary file (this is used to determine which ranges of characters
69      * to apply the dictionary to)
70      */
71     private boolean[] categoryFlags;
72
73     /**
74      * a temporary hiding place for the number of dictionary characters in the
75      * last range passed over by next()
76      */
77     private int dictionaryCharCount;
78
79     /**
80      * when a range of characters is divided up using the dictionary, the break
81      * positions that are discovered are stored here, preventing us from having
82      * to use either the dictionary or the state table again until the iterator
83      * leaves this range of text
84      */
85     private int[] cachedBreakPositions;
86
87     /**
88      * if cachedBreakPositions is not null, this indicates which item in the
89      * cache the current iteration position refers to
90      */
91     private int positionInCache;
92
93     /**
94      * Constructs a DictionaryBasedBreakIterator.
95      * @param description Same as the description parameter on RuleBasedBreakIterator,
96      * except for the special meaning of "<dictionary>". This parameter is just
97      * passed through to RuleBasedBreakIterator's constructor.
98      * @param dictionaryFilename The filename of the dictionary file to use
99      */
100     public DictionaryBasedBreakIterator(String dataFile, String dictionaryFile)
101                                         throws IOException {
102         super(dataFile);
103         byte[] tmp = super.getAdditionalData();
104         if (tmp != null) {
105             prepareCategoryFlags(tmp);
106             super.setAdditionalData(null);
107         }
108         dictionary = new BreakDictionary (dictionaryFile);
109     }
110
111     private void prepareCategoryFlags(byte[] data) {
112     categoryFlags = new boolean[data.length];
113         for (int i = 0; i < data.length; i++) {
114             categoryFlags[i] = (data[i] == (byte)1) ? true : false;
115         }
116     }
117
118     public void setText(CharacterIterator newText) {
119         super.setText(newText);
120         cachedBreakPositions = null;
121         dictionaryCharCount = 0;
122         positionInCache = 0;
123     }
124
125     /**
126      * Sets the current iteration position to the beginning of the text.
127      * (i.e., the CharacterIterator's starting offset).
128      * @return The offset of the beginning of the text.
129      */
130     public int first() {
131         cachedBreakPositions = null;
132         dictionaryCharCount = 0;
133         positionInCache = 0;
134         return super.first();
135     }
136
137     /**
138      * Sets the current iteration position to the end of the text.
139      * (i.e., the CharacterIterator's ending offset).
140      * @return The text's past-the-end offset.
141      */
142     public int last() {
143         cachedBreakPositions = null;
144         dictionaryCharCount = 0;
145         positionInCache = 0;
146         return super.last();
147     }
148
149     /**
150      * Advances the iterator one step backwards.
151      * @return The position of the last boundary position before the
152      * current iteration position
153      */
154     public int previous() {
155         CharacterIterator text = getText();
156
157         // if we have cached break positions and we're still in the range
158 // covered by them, just move one step backward in the cache
159 if (cachedBreakPositions != null && positionInCache > 0) {
160             --positionInCache;
161             text.setIndex(cachedBreakPositions[positionInCache]);
162             return cachedBreakPositions[positionInCache];
163         }
164
165         // otherwise, dump the cache and use the inherited previous() method to move
166 // backward. This may fill up the cache with new break positions, in which
167 // case we have to mark our position in the cache
168 else {
169             cachedBreakPositions = null;
170             int result = super.previous();
171             if (cachedBreakPositions != null) {
172                 positionInCache = cachedBreakPositions.length - 2;
173             }
174             return result;
175         }
176     }
177
178     /**
179      * Sets the current iteration position to the last boundary position
180      * before the specified position.
181      * @param offset The position to begin searching from
182      * @return The position of the last boundary before "offset"
183      */
184     public int preceding(int offset) {
185         CharacterIterator text = getText();
186         checkOffset(offset, text);
187
188         // if we have no cached break positions, or "offset" is outside the
189 // range covered by the cache, we can just call the inherited routine
190 // (which will eventually call other routines in this class that may
191 // refresh the cache)
192 if (cachedBreakPositions == null || offset <= cachedBreakPositions[0] ||
193                 offset > cachedBreakPositions[cachedBreakPositions.length - 1]) {
194             cachedBreakPositions = null;
195             return super.preceding(offset);
196         }
197
198         // on the other hand, if "offset" is within the range covered by the cache,
199 // then all we have to do is search the cache for the last break position
200 // before "offset"
201 else {
202             positionInCache = 0;
203             while (positionInCache < cachedBreakPositions.length
204                    && offset > cachedBreakPositions[positionInCache]) {
205                 ++positionInCache;
206             }
207             --positionInCache;
208             text.setIndex(cachedBreakPositions[positionInCache]);
209             return text.getIndex();
210         }
211     }
212
213     /**
214      * Sets the current iteration position to the first boundary position after
215      * the specified position.
216      * @param offset The position to begin searching forward from
217      * @return The position of the first boundary after "offset"
218      */
219     public int following(int offset) {
220         CharacterIterator text = getText();
221         checkOffset(offset, text);
222
223         // if we have no cached break positions, or if "offset" is outside the
224 // range covered by the cache, then dump the cache and call our
225 // inherited following() method. This will call other methods in this
226 // class that may refresh the cache.
227 if (cachedBreakPositions == null || offset < cachedBreakPositions[0] ||
228                 offset >= cachedBreakPositions[cachedBreakPositions.length - 1]) {
229             cachedBreakPositions = null;
230             return super.following(offset);
231         }
232
233         // on the other hand, if "offset" is within the range covered by the
234 // cache, then just search the cache for the first break position
235 // after "offset"
236 else {
237             positionInCache = 0;
238             while (positionInCache < cachedBreakPositions.length
239                    && offset >= cachedBreakPositions[positionInCache]) {
240                 ++positionInCache;
241             }
242             text.setIndex(cachedBreakPositions[positionInCache]);
243             return text.getIndex();
244         }
245     }
246
247     /**
248      * This is the implementation function for next().
249      */
250     protected int handleNext() {
251         CharacterIterator text = getText();
252
253         // if there are no cached break positions, or if we've just moved
254 // off the end of the range covered by the cache, we have to dump
255 // and possibly regenerate the cache
256 if (cachedBreakPositions == null ||
257         positionInCache == cachedBreakPositions.length - 1) {
258
259             // start by using the inherited handleNext() to find a tentative return
260 // value. dictionaryCharCount tells us how many dictionary characters
261 // we passed over on our way to the tentative return value
262 int startPos = text.getIndex();
263             dictionaryCharCount = 0;
264             int result = super.handleNext();
265
266             // if we passed over more than one dictionary character, then we use
267 // divideUpDictionaryRange() to regenerate the cached break positions
268 // for the new range
269 if (dictionaryCharCount > 1 && result - startPos > 1) {
270                 divideUpDictionaryRange(startPos, result);
271             }
272
273             // otherwise, the value we got back from the inherited fuction
274 // is our return value, and we can dump the cache
275 else {
276                 cachedBreakPositions = null;
277                 return result;
278             }
279         }
280
281         // if the cache of break positions has been regenerated (or existed all
282 // along), then just advance to the next break position in the cache
283 // and return it
284 if (cachedBreakPositions != null) {
285             ++positionInCache;
286             text.setIndex(cachedBreakPositions[positionInCache]);
287             return cachedBreakPositions[positionInCache];
288         }
289         return -9999; // SHOULD NEVER GET HERE!
290 }
291
292     /**
293      * Looks up a character category for a character.
294      */
295     protected int lookupCategory(int c) {
296         // this override of lookupCategory() exists only to keep track of whether we've
297 // passed over any dictionary characters. It calls the inherited lookupCategory()
298 // to do the real work, and then checks whether its return value is one of the
299 // categories represented in the dictionary. If it is, bump the dictionary-
300 // character count.
301 int result = super.lookupCategory(c);
302         if (result != RuleBasedBreakIterator.IGNORE && categoryFlags[result]) {
303             ++dictionaryCharCount;
304         }
305         return result;
306     }
307
308     /**
309      * This is the function that actually implements the dictionary-based
310      * algorithm. Given the endpoints of a range of text, it uses the
311      * dictionary to determine the positions of any boundaries in this
312      * range. It stores all the boundary positions it discovers in
313      * cachedBreakPositions so that we only have to do this work once
314      * for each time we enter the range.
315      */
316     private void divideUpDictionaryRange(int startPos, int endPos) {
317         CharacterIterator text = getText();
318
319         // the range we're dividing may begin or end with non-dictionary characters
320 // (i.e., for line breaking, we may have leading or trailing punctuation
321 // that needs to be kept with the word). Seek from the beginning of the
322 // range to the first dictionary character
323 text.setIndex(startPos);
324         int c = getCurrent();
325         int category = lookupCategory(c);
326         while (category == IGNORE || !categoryFlags[category]) {
327             c = getNext();
328             category = lookupCategory(c);
329         }
330
331         // initialize. We maintain two stacks: currentBreakPositions contains
332 // the list of break positions that will be returned if we successfully
333 // finish traversing the whole range now. possibleBreakPositions lists
334 // all other possible word ends we've passed along the way. (Whenever
335 // we reach an error [a sequence of characters that can't begin any word
336 // in the dictionary], we back up, possibly delete some breaks from
337 // currentBreakPositions, move a break from possibleBreakPositions
338 // to currentBreakPositions, and start over from there. This process
339 // continues in this way until we either successfully make it all the way
340 // across the range, or exhaust all of our combinations of break
341 // positions.)
342 Stack currentBreakPositions = new Stack ();
343         Stack possibleBreakPositions = new Stack ();
344         Vector wrongBreakPositions = new Vector ();
345
346         // the dictionary is implemented as a trie, which is treated as a state
347 // machine. -1 represents the end of a legal word. Every word in the
348 // dictionary is represented by a path from the root node to -1. A path
349 // that ends in state 0 is an illegal combination of characters.
350 int state = 0;
351
352         // these two variables are used for error handling. We keep track of the
353 // farthest we've gotten through the range being divided, and the combination
354 // of breaks that got us that far. If we use up all possible break
355 // combinations, the text contains an error or a word that's not in the
356 // dictionary. In this case, we "bless" the break positions that got us the
357 // farthest as real break positions, and then start over from scratch with
358 // the character where the error occurred.
359 int farthestEndPoint = text.getIndex();
360         Stack bestBreakPositions = null;
361
362         // initialize (we always exit the loop with a break statement)
363 c = getCurrent();
364         while (true) {
365
366             // if we can transition to state "-1" from our current state, we're
367 // on the last character of a legal word. Push that position onto
368 // the possible-break-positions stack
369 if (dictionary.getNextState(state, 0) == -1) {
370                 possibleBreakPositions.push(new Integer (text.getIndex()));
371             }
372
373             // look up the new state to transition to in the dictionary
374 state = dictionary.getNextStateFromCharacter(state, c);
375
376             // if the character we're sitting on causes us to transition to
377 // the "end of word" state, then it was a non-dictionary character
378 // and we've successfully traversed the whole range. Drop out
379 // of the loop.
380 if (state == -1) {
381                 currentBreakPositions.push(new Integer (text.getIndex()));
382                 break;
383             }
384
385             // if the character we're sitting on causes us to transition to
386 // the error state, or if we've gone off the end of the range
387 // without transitioning to the "end of word" state, we've hit
388 // an error...
389 else if (state == 0 || text.getIndex() >= endPos) {
390
391                 // if this is the farthest we've gotten, take note of it in
392 // case there's an error in the text
393 if (text.getIndex() > farthestEndPoint) {
394                     farthestEndPoint = text.getIndex();
395                     bestBreakPositions = (Stack )(currentBreakPositions.clone());
396                 }
397
398                 // wrongBreakPositions is a list of all break positions
399 // we've tried starting that didn't allow us to traverse
400 // all the way through the text. Every time we pop a
401 //break position off of currentBreakPositions, we put it
402 // into wrongBreakPositions to avoid trying it again later.
403 // If we make it to this spot, we're either going to back
404 // up to a break in possibleBreakPositions and try starting
405 // over from there, or we've exhausted all possible break
406 // positions and are going to do the fallback procedure.
407 // This loop prevents us from messing with anything in
408 // possibleBreakPositions that didn't work as a starting
409 // point the last time we tried it (this is to prevent a bunch of
410 // repetitive checks from slowing down some extreme cases)
411 Integer newStartingSpot = null;
412                 while (!possibleBreakPositions.isEmpty() && wrongBreakPositions.contains(
413                             possibleBreakPositions.peek())) {
414                     possibleBreakPositions.pop();
415                 }
416                 
417                 // if we've used up all possible break-position combinations, there's
418 // an error or an unknown word in the text. In this case, we start
419 // over, treating the farthest character we've reached as the beginning
420 // of the range, and "blessing" the break positions that got us that
421 // far as real break positions
422 if (possibleBreakPositions.isEmpty()) {
423                     if (bestBreakPositions != null) {
424                         currentBreakPositions = bestBreakPositions;
425                         if (farthestEndPoint < endPos) {
426                             text.setIndex(farthestEndPoint + 1);
427                         }
428                         else {
429                             break;
430                         }
431                     }
432                     else {
433                         if ((currentBreakPositions.size() == 0 ||
434                  ((Integer )(currentBreakPositions.peek())).intValue() != text.getIndex())
435                 && text.getIndex() != startPos) {
436                             currentBreakPositions.push(new Integer (text.getIndex()));
437                         }
438                         getNext();
439                         currentBreakPositions.push(new Integer (text.getIndex()));
440                     }
441                 }
442
443                 // if we still have more break positions we can try, then promote the
444 // last break in possibleBreakPositions into currentBreakPositions,
445 // and get rid of all entries in currentBreakPositions that come after
446 // it. Then back up to that position and start over from there (i.e.,
447 // treat that position as the beginning of a new word)
448 else {
449                     Integer temp = (Integer )possibleBreakPositions.pop();
450                     Object temp2 = null;
451                     while (!currentBreakPositions.isEmpty() && temp.intValue() <
452                            ((Integer )currentBreakPositions.peek()).intValue()) {
453                         temp2 = currentBreakPositions.pop();
454                         wrongBreakPositions.addElement(temp2);
455                     }
456                     currentBreakPositions.push(temp);
457                     text.setIndex(((Integer )currentBreakPositions.peek()).intValue());
458                 }
459
460                 // re-sync "c" for the next go-round, and drop out of the loop if
461 // we've made it off the end of the range
462 c = getCurrent();
463                 if (text.getIndex() >= endPos) {
464                     break;
465                 }
466             }
467
468             // if we didn't hit any exceptional conditions on this last iteration,
469 // just advance to the next character and loop
470 else {
471                 c = getNext();
472             }
473         }
474
475         // dump the last break position in the list, and replace it with the actual
476 // end of the range (which may be the same character, or may be further on
477 // because the range actually ended with non-dictionary characters we want to
478 // keep with the word)
479 if (!currentBreakPositions.isEmpty()) {
480             currentBreakPositions.pop();
481         }
482         currentBreakPositions.push(new Integer (endPos));
483
484         // create a regular array to hold the break positions and copy
485 // the break positions from the stack to the array (in addition,
486 // our starting position goes into this array as a break position).
487 // This array becomes the cache of break positions used by next()
488 // and previous(), so this is where we actually refresh the cache.
489 cachedBreakPositions = new int[currentBreakPositions.size() + 1];
490         cachedBreakPositions[0] = startPos;
491
492         for (int i = 0; i < currentBreakPositions.size(); i++) {
493             cachedBreakPositions[i + 1] = ((Integer )currentBreakPositions.elementAt(i)).intValue();
494         }
495         positionInCache = 0;
496     }
497 }
498

A to Z: JavaDoc & Examples

---

Daily Java News & Articles

---

Open Source Projects

---

Open Source Codes

---

Free Computer Books

---

Remove Frame

---

Popular Tags