1 <!DOCTYPE html PUBLIC "-//W3C//DTD XHTML 1.0 Strict//EN"
2 "http://www.w3.org/TR/xhtml1/DTD/xhtml1-strict.dtd">
3 <html xmlns="http://www.w3.org/1999/xhtml"
4 lang="en" xml:lang="en">
7 <meta http-equiv="Content-Type" content="text/html;charset=utf-8"/>
8 <meta name="generator" content="Org-mode"/>
9 <meta name="generated" content="2008/06/04 14:25:59"/>
10 <meta name="author" content="Kevin Brubeck Unhammer"/>
11 <link rel="stylesheet" type="text/css" href="http://www.student.uib.no/~kun041/org.css">
13 <h1 class="title">DMV/CCM</h1>
14 <div id="table-of-contents">
15 <h2>Table of Contents</h2>
17 <li><a href="#sec-1">1 dmvccm report and project</a></li>
18 <li><a href="#sec-2">2 Adjacency and combining it with inner()</a></li>
19 <li><a href="#sec-3">3 [#A] P_STOP for IO/EM</a></li>
20 <li><a href="#sec-4">4 P_CHOOSE for IO/EM</a></li>
21 <li><a href="#sec-5">5 Initialization </a></li>
22 <li><a href="#sec-6">6 [#C] Deferred</a></li>
23 <li><a href="#sec-7">7 Expectation Maximation in IO/DMV-terms</a></li>
24 <li><a href="#sec-8">8 Python-stuff</a></li>
28 <div class="outline-2">
29 <h2 id="sec-1">1 dmvccm report and project</h2>
31 <p><span class="timestamp-kwd">DEADLINE: </span> <span class="timestamp">2008-06-30 Mon</span><br/>
32 But absolute, extended, really-quite-dead-now deadline: August 31…
35 <a href="src/dmv.py">dmv.py</a>
38 <a href="src/io.py">io.py</a>
41 <a href="src/harmonic.py">harmonic.py</a>
45 <div class="outline-2">
46 <h2 id="sec-2">2 <span class="todo">TODO</span> Adjacency and combining it with inner()</h2>
48 <p>Each DMV_Rule now has both a probN and a probA, for
49 adjacencies. inner() needs the correct one in each case.
52 Adjacency gives a problem with duplicate words/tags, eg. in the
53 sentence "a a b". If this has the dependency structure b->a<sub>0</sub>->a<sub>1</sub>,
54 then b is non-adjacent to a<sub>0</sub> and should use probN (for the LRStop and
55 the attachment of a<sub>0</sub>), while the other rules should all use
56 probA. But within the e(0,2,b) we can't just say "oh, a has index 0
57 so it's not adjacent to 2", since there's also an a at index 1, and
58 there's also a dependency structure b->a<sub>1</sub>->a<sub>0</sub> for that. We want
59 both. And in possibly much more complex versions.
65 I first thought of decorating the individual words/tags in a
66 sentence with their indices, and perhaps just duplicating the
67 relevant rules (one for each index of the duplicate tags). But this
68 gives an explosion in attachment rules (although a contained
69 explosion, within the rules used in a sentence; but most sentences
70 will have at least two NN's so it will be a problem).
73 Then, I had a <i>brilliant</i> idea. Just let e(), the helper function of
74 inner(), parametrize for an extra pair of boolean values for whether
75 or not we've attached anything to the left or right yet ("yet"
76 meaning "below"). So now, e() has a chart of the form [s, t, LHS,
77 Lattach, Rattach], and of course e(s,t,LHS) is the sum of the four
78 possible values for (Lattach,Rattach). This makes e() lots more
79 complex and DMV-specific though, so it's been rewritten in
80 inner_dmv() in dmv.py.
82 <li><span class="todo">TODO</span> document this adjacency stuff better<br/>
84 <li><span class="todo">TODO</span> test and debug my brilliant idea<br/>
86 <li><span class="done">DONE</span> implement my brilliant idea.<br/>
87 <span class="timestamp-kwd">CLOSED: </span> <span class="timestamp">2008-06-01 Sun 17:19</span><br/>
88 <a href="src/dmv.py">e(sti) in dmv.py</a>
91 <li><span class="done">DONE</span> [#A] test inner() on sentences with duplicate words<br/>
92 Works with eg. the sentence "h h h"
99 <div class="outline-2">
100 <h2 id="sec-3">3 <span class="todo">TODO</span> [#A] P_STOP for IO/EM</h2>
102 <p><a href="src/dmv.py">dmv-P_STOP</a>
103 Remember: The P<sub>STOP</sub> formula is upside-down (left-to-right also).
104 (In the article..not the thesis)
107 Remember: Initialization makes some "short-cut" rules, these will also
108 have to be updated along with the other P<sub>STOP</sub> updates:
111 b[(NOBAR, n<sub>h</sub>), 'h'] = 1.0 # always
114 b[(RBAR, n<sub>h</sub>), 'h'] = h_.probA # h_ is RBAR stop rule
117 b[(LRBAR, n<sub>h</sub>), 'h'] = h_.probA * _ h_.probA
120 <li>How is the P_STOP formula different given other values for dir and adj?<br/>
121 (Presumably, the P<sub>STOP</sub> formula where STOP is True is just the
122 rule-probability of _ h_ -> STOP h_ or h_ -> h STOP, but how does
123 adjacency fit in here?)
126 (And P<sub>STOP</sub>(-STOP|…) = 1 - P<sub>STOP</sub>(STOP|…) )
131 <div class="outline-2">
132 <h2 id="sec-4">4 <span class="todo">TODO</span> P_CHOOSE for IO/EM</h2>
134 <p>Write the formulas! should be easy?
137 <div class="outline-2">
138 <h2 id="sec-5">5 Initialization </h2>
140 <p><a href="/Users/kiwibird/Documents/Skole/V08/Probability/dmvccm/src/dmv.py">dmv-inits</a>
143 We do have to go through the corpus, since the probabilities are based
144 on how far away in the sentence arguments are from their heads.
146 <li><span class="todo">TODO</span> Separate initialization to another file? <span class="tag">PRETTIER</span><br/>
149 <li><span class="todo">TOGROK</span> CCM Initialization <br/>
150 P<sub>SPLIT</sub> used here… how, again?
152 <li><span class="done">DONE</span> DMV Initialization probabilities<br/>
153 (from initialization frequency)
155 <li><span class="done">DONE</span> DMV Initialization frequencies <br/>
156 <span class="timestamp-kwd">CLOSED: </span> <span class="timestamp">2008-05-27 Tue 20:04</span><br/>
159 P<sub>STOP</sub> is not well defined by K&M. One possible interpretation given
160 the sentence [det nn vb nn] is
162 f_{STOP}( STOP|det, L, adj) +1
163 f_{STOP}(-STOP|det, L, adj) +0
164 f_{STOP}( STOP|det, L, non_adj) +1
165 f_{STOP}(-STOP|det, L, non_adj) +0
166 f_{STOP}( STOP|det, R, adj) +0
167 f_{STOP}(-STOP|det, R, adj) +1
169 f_{STOP}( STOP|nn, L, adj) +0
170 f_{STOP}(-STOP|nn, L, adj) +1
171 f_{STOP}( STOP|nn, L, non_adj) +1 # since there's at least one to the left
172 f_{STOP}(-STOP|nn, L, non_adj) +0
175 <li><span class="todo">TODO</span> tweak<br/>
176 <a name="pstoptweak"> </a>
178 f[head, 'STOP', 'LN'] += (i_h <= 1) # first two words
179 f[head, '-STOP', 'LN'] += (not i_h <= 1)
180 f[head, 'STOP', 'LA'] += (i_h == 0) # very first word
181 f[head, '-STOP', 'LA'] += (not i_h == 0)
182 f[head, 'STOP', 'RN'] += (i_h >= n - 2) # last two words
183 f[head, '-STOP', 'RN'] += (not i_h >= n - 2)
184 f[head, 'STOP', 'RA'] += (i_h == n - 1) # very last word
185 f[head, '-STOP', 'RA'] += (not i_h == n - 1)
189 # this one requires some additional rewriting since it
190 # introduces divisions by zero
191 f[head, 'STOP', 'LN'] += (i_h == 1) # second word
192 f[head, '-STOP', 'LN'] += (not i_h <= 1) # not first two
193 f[head, 'STOP', 'LA'] += (i_h == 0) # first word
194 f[head, '-STOP', 'LA'] += (not i_h == 0) # not first
195 f[head, 'STOP', 'RN'] += (i_h == n - 2) # second-to-last
196 f[head, '-STOP', 'RN'] += (not i_h >= n - 2) # not last two
197 f[head, 'STOP', 'RA'] += (i_h == n - 1) # last word
198 f[head, '-STOP', 'RA'] += (not i_h == n - 1) # not last
202 f[head, 'STOP', 'LN'] += (i_h == 1) # second word
203 f[head, '-STOP', 'LN'] += (not i_h == 1) # not second
204 f[head, 'STOP', 'LA'] += (i_h == 0) # first word
205 f[head, '-STOP', 'LA'] += (not i_h == 0) # not first
206 f[head, 'STOP', 'RN'] += (i_h == n - 2) # second-to-last
207 f[head, '-STOP', 'RN'] += (not i_h == n - 2) # not second-to-last
208 f[head, 'STOP', 'RA'] += (i_h == n - 1) # last word
209 f[head, '-STOP', 'RA'] += (not i_h == n - 1) # not last
212 "all words take the same number of arguments" interpreted as
215 p_STOP(head, 'STOP', 'LN') = 0.3
216 p_STOP(head, 'STOP', 'LA') = 0.5
217 p_STOP(head, 'STOP', 'RN') = 0.4
218 p_STOP(head, 'STOP', 'RA') = 0.7
220 (which we easily may tweak in init_zeros())
225 Go through the corpus, counting distances between heads and
226 arguments. In [det nn vb nn], we give
229 f<sub>CHOOSE</sub>(nn|det, R) +1/1 + C
232 f<sub>CHOOSE</sub>(vb|det, R) +1/2 + C
235 f<sub>CHOOSE</sub>(nn|det, R) +1/3 + C
238 If this were the full corpus, P<sub>CHOOSE</sub>(nn|det, R) would have
239 (1+1/3+2C) / sum_a f<sub>CHOOSE</sub>(a|det, R)
244 <p>The ROOT gets "each argument with equal probability", so in a sentence
245 of three words, 1/3 for each (in [nn vb nn], 'nn' gets 2/3). Basically
246 just a frequency count of the corpus…
253 <div class="outline-2">
254 <h2 id="sec-6">6 [#C] Deferred</h2>
257 <li><span class="todo">TODO</span> inner_dmv() should disregard rules with heads not in sent <span class="tag">OPTIMIZE</span><br/>
258 If the sentence is "nn vbd det nn", we should not even look at rules
261 rule.head() not in "nn vbd det nn".split()
263 This is ruled out by getting rules from g.rules(LHS, sent).
266 Also, we optimize this further by saying we don't even recurse into
267 attachment rules where
269 rule.head() not in sent[ s :r+1]
270 rule.head() not in sent[r+1:t+1]
272 meaning, if we're looking at the span "vbd det", we only use
273 attachment rules where both daughters are members of ['vbd','det']
274 (although we don't (yet) care about removing rules that rewrite to the
275 same tag if there are no duplicate tags in the span, etc., that would
276 be a lot of trouble for little potential gain).
278 <li><span class="todo">TODO</span> when reestimating P_STOP etc, remove rules with p < epsilon <span class="tag">OPTIMIZE</span><br/>
280 <li><span class="todo">TODO</span> inner_dmv, short ranges and impossible attachment <span class="tag">OPTIMIZE</span><br/>
281 If s-t <= 2, there can be only one attachment below, so don't recurse
282 with both Lattach=True and Rattach=True.
285 If s-t <= 1, there can be no attachment below, so only recurse with
286 Lattach=False, Rattach=False.
289 Put this in the loop under rewrite rules (could also do it in the STOP
290 section, but that would only have an effect on very short sentences).
292 <li><span class="todo">TODO</span> clean up the module files <span class="tag">PRETTIER</span><br/>
293 Is there better way to divide dmv and harmonic? There's a two-way
294 dependency between the modules. Guess there could be a third file that
295 imports both the initialization and the actual EM stuff, while a file
296 containing constants and classes could be imported by all others:
298 dmv.py imports dmv_EM.py imports dmv_classes.py
299 dmv.py imports dmv_inits.py imports dmv_classes.py
303 <li><span class="todo">TOGROK</span> Some (tagged) sentences are bound to come twice <span class="tag">OPTIMIZE</span><br/>
304 Eg, first sort and count, so that the corpus
305 [['nn','vbd','det','nn'],
306 ['vbd','nn','det','nn'],
307 ['nn','vbd','det','nn']]
309 [(['nn','vbd','det','nn'],2),
310 (['vbd','nn','det','nn'],1)]
311 and then in each loop through sentences, make sure we handle the
315 Is there much to gain here?
318 <li><span class="todo">TOGROK</span> tags as numbers or tags as strings? <span class="tag">OPTIMIZE</span><br/>
319 Need to clean up the representation.
322 Stick with tag-strings in initialization then switch to numbers for
323 IO-algorithm perhaps? Can probably afford more string-matching in
329 <div class="outline-2">
330 <h2 id="sec-7">7 Expectation Maximation in IO/DMV-terms</h2>
332 <p>inner(s,t,LHS) calculates the expected number of trees headed by LHS
333 from s to t (sentence positions). This uses the P_STOP and P_CHOOSE
334 values, which have been conveniently distributed into CNF rules as
335 probN and probA (non-adjacent and adjacent probabilites).
338 When re-estimating, we use the expected values from inner() to get new
339 values for P_STOP and P_CHOOSE. When we've re-estimated for the entire
340 corpus, we distribute P_STOP and P_CHOOSE into the CNF rules again, so
341 that in the next round we use new probN and probA to find
345 The distribution of P_STOP and P_CHOOSE into CNF rules also happens in
346 init_normalize() (here along with the creation of P_STOP and
347 P_CHOOSE); P_STOP is used to create CNF rules where one branch of the
348 rule is STOP, P_CHOOSE is used to create rules of the form
355 Since "adjacency" is not captured in regular CNF rules, we need two
356 probabilites for each rule, and inner() has to know when to use which.
359 <li><span class="todo">TODO</span> Corpus access<br/>
361 <li><span class="todo">TOGROK</span> sentences or rules as the "outer loop"? <span class="tag">OPTIMIZE</span><br/>
362 In regard to the E/M-step, finding P<sub>STOP</sub>, P<sub>CHOOSE</sub>.
369 <div class="outline-2">
370 <h2 id="sec-8">8 Python-stuff</h2>
374 <a href="src/pseudo.py">pseudo.py</a>
377 <a href="http://nltk.org/doc/en/structured-programming.html">http://nltk.org/doc/en/structured-programming.html</a> recursive dynamic
380 <a href="http://nltk.org/doc/en/advanced-parsing.html">http://nltk.org/doc/en/advanced-parsing.html</a>
383 <a href="http://jaynes.colorado.edu/PythonIdioms.html">http://jaynes.colorado.edu/PythonIdioms.html</a>
389 <div id="postamble"><p class="author"> Author: Kevin Brubeck Unhammer
390 <a href="mailto:K.BrubeckUnhammer at student uva nl "><K.BrubeckUnhammer at student uva nl ></a>
392 <p class="date"> Date: 2008/06/04 14:25:59</p>
393 </div><p class="postamble">Skrive vha. emacs + <a href='http://orgmode.org/'>org-mode</a></p></body>