src/loc_h_dmv.py

   1 # loc_h_dmv.py
   2 #
   3 # dmv reestimation and inside-outside probabilities using loc_h, and
   4 # no CNF-style rules
   5 #
   6 # Table of Contents:
   7 # 1. Grammar-class and related functions
   8 # 2. P_INSIDE / inner() and inner_sent()
   9 # 3. P_OUTSIDE / outer()
  10 # 4. Reestimation v.1: sentences as outer loop
  11 # 5. Reestimation v.2: head-types as outer loop
  12 # 6. Most Probable Parse
  13 # 7. Testing functions
  14
  15 import io
  16 from common_dmv import *
  17
  18 ### todo: debug with @accepts once in a while, but it's SLOW
  19 # from typecheck import accepts, Any
  20
  21 if __name__ == "__main__":
  22     print "loc_h_dmv module tests:"
  23
  24 def adj(middle, loc_h):
  25     "middle is eg. k when rewriting for i<k<j (inside probabilities)."
  26     return middle == loc_h or middle == loc_h+1 # ADJ == True
  27
  28 def make_GO_AT(p_STOP,p_ATTACH):
  29     p_GO_AT = {}
  30     for (a,h,dir), p_ah in p_ATTACH.iteritems():
  31         p_GO_AT[a,h,dir, NON] = p_ah * (1-p_STOP[h, dir, NON])
  32         p_GO_AT[a,h,dir, ADJ] = p_ah * (1-p_STOP[h, dir, ADJ])
  33     return p_GO_AT
  34
  35 class DMV_Grammar(io.Grammar):
  36     def __str__(self):
  37         LJUST = 47
  38         def t(n):
  39             return "%d=%s" % (n, self.numtag(n))
  40         def p(dict,key):
  41             if key in dict: return dict[key]
  42             else: return 0.0
  43         def no_zeroL(str,tagstr,prob):
  44             if prob > 0.0: return (str%(tagstr,prob)).ljust(LJUST)
  45             else: return "".ljust(LJUST)
  46         def no_zeroR(str,tagstr,prob):
  47             if prob > 0.0: return str%(tagstr,prob)
  48             else: return ""
  49         def p_a(a,h):
  50             p_L = p(self.p_ATTACH,(a,h,LEFT))
  51             p_R = p(self.p_ATTACH,(a,h,RIGHT))
  52             if p_L == 0.0 and p_R == 0.0:
  53                 return ''
  54             else:
  55                 if p_L > 0.0:
  56                     str = "p_ATTACH[%s|%s,L] = %s" % (t(a), t(h), p_L)
  57                     str = str.ljust(LJUST)
  58                 else:
  59                     str = ''
  60                 if p_R > 0.0:
  61                     str = str.ljust(LJUST)
  62                     str += "p_ATTACH[%s|%s,R] = %s" % (t(a), t(h), p_R)
  63             return '\n'+str
  64
  65         root, stop, att, ord = "","","",""
  66         for h in self.headnums():
  67             root += no_zeroL("\np_ROOT[%s] = %s", t(h), p(self.p_ROOT, (h)))
  68             stop += '\n'
  69             stop += no_zeroL("p_STOP[stop|%s,L,adj] = %s", t(h), p(self.p_STOP, (h,LEFT,ADJ)))
  70             stop += no_zeroR("p_STOP[stop|%s,R,adj] = %s", t(h), p(self.p_STOP, (h,RIGHT,ADJ)))
  71             stop += '\n'
  72             stop += no_zeroL("p_STOP[stop|%s,L,non] = %s", t(h), p(self.p_STOP, (h,LEFT,NON)))
  73             stop += no_zeroR("p_STOP[stop|%s,R,non] = %s", t(h), p(self.p_STOP, (h,RIGHT,NON)))
  74             att  += ''.join([p_a(a,h) for a in self.headnums()])
  75             ord  += '\n'
  76             ord  += no_zeroL("p_ORDER[ left-first|%s ] = %s", t(h), p(self.p_ORDER, (GOL,h)))
  77             ord  += no_zeroR("p_ORDER[right-first|%s ] = %s", t(h), p(self.p_ORDER, (GOR,h)))
  78         return root + stop + att + ord
  79
  80     def __init__(self, numtag, tagnum, p_ROOT, p_STOP, p_ATTACH, p_ORDER):
  81         io.Grammar.__init__(self, numtag, tagnum)
  82         self.p_ROOT = p_ROOT     # p_ROOT[w] = p
  83         self.p_ORDER = p_ORDER   # p_ORDER[seals, w] = p
  84         self.p_STOP = p_STOP     # p_STOP[w, LEFT, NON] = p     (etc. for LA,RN,RA)
  85         self.p_ATTACH = p_ATTACH # p_ATTACH[a, h, LEFT] = p (etc. for R)
  86                                  # p_GO_AT[a, h, LEFT, NON] = p (etc. for LA,RN,RA)
  87         self.p_GO_AT = make_GO_AT(self.p_STOP, self.p_ATTACH)
  88         # these are used in reestimate2():
  89         self.reset_iocharts()
  90
  91     def get_iochart(self, sent_nums):
  92         ch_key = tuple(sent_nums)
  93         try:
  94             ichart = self._icharts[ch_key]
  95         except KeyError:
  96             ichart = {}
  97         try:
  98             ochart = self._ocharts[ch_key]
  99         except KeyError:
 100             ochart = {}
 101         return (ichart, ochart)
 102
 103     def set_iochart(self, sent_nums, ichart, ochart):
 104         self._icharts[tuple(sent_nums)] = ichart
 105         self._ocharts[tuple(sent_nums)] = ochart
 106
 107     def reset_iocharts(self):
 108         self._icharts = {}
 109         self._ocharts = {}
 110
 111     def p_GO_AT_or0(self, a, h, dir, adj):
 112         try:
 113             return self.p_GO_AT[a, h, dir, adj]
 114         except KeyError:
 115             return 0.0
 116
 117
 118 def locs(sent_nums, start, stop):
 119     '''Return the between-word locations of all words in some fragment of
 120     sent. We make sure to offset the locations correctly so that for
 121     any w in the returned list, sent[w]==loc_w.
 122
 123     start is inclusive, stop is exclusive, as in klein-thesis and
 124     Python's list-slicing.'''
 125     for i0,w in enumerate(sent_nums[start:stop]):
 126         loc_w = i0+start
 127         yield (loc_w, w)
 128
 129 ###################################################
 130 #            P_INSIDE (dmv-specific)              #
 131 ###################################################
 132
 133 #@accepts(int, int, (int, int), int, Any(), [str], {tuple:float}, IsOneOf(None,{}))
 134 def inner(i, j, node, loc_h, g, sent, ichart, mpptree=None):
 135     ''' The ichart is of this form:
 136     ichart[i,j,LHS, loc_h]
 137     where i and j are between-word positions.
 138
 139     loc_h gives adjacency (along with k for attachment rules), and is
 140     needed in P_STOP reestimation.
 141     '''
 142     sent_nums = g.sent_nums(sent)
 143
 144     def terminal(i,j,node, loc_h, tabs):
 145         if not i <= loc_h < j:
 146             if 'INNER' in DEBUG:
 147                 print "%s*= 0.0 (wrong loc_h)" % tabs
 148             return 0.0
 149         elif POS(node) == sent_nums[i] and node in g.p_ORDER:
 150             # todo: add to ichart perhaps? Although, it _is_ simple lookup..
 151             prob = g.p_ORDER[node]
 152         else:
 153             if 'INNER' in DEBUG:
 154                 print "%sLACKING TERMINAL:" % tabs
 155             prob = 0.0
 156         if 'INNER' in DEBUG:
 157             print "%s*= %.4f (terminal: %s -> %s_%d)" % (tabs,prob, node_str(node), sent[i], loc_h)
 158         return prob
 159
 160     def e(i,j, (s_h,h), loc_h, n_t):
 161         def to_mpp(p, L, R):
 162             if mpptree:
 163                 key = (i,j, (s_h,h), loc_h)
 164                 if key not in mpptree:
 165                     mpptree[key] = (p, L, R)
 166                 elif mpptree[key][0] < p:
 167                     mpptree[key] = (p, L, R)
 168
 169         def tab():
 170             "Tabs for debug output"
 171             return "\t"*n_t
 172
 173         if (i, j, (s_h,h), loc_h) in ichart:
 174             if 'INNER' in DEBUG:
 175                 print "%s*= %.4f in ichart: i:%d j:%d node:%s loc:%s" % (tab(),ichart[i, j, (s_h,h), loc_h], i, j,
 176                                                                        node_str((s_h,h)), loc_h)
 177             return ichart[i, j, (s_h,h), loc_h]
 178         else:
 179             # Either terminal rewrites, using p_ORDER:
 180             if i+1 == j and (s_h == GOR or s_h == GOL):
 181                 return terminal(i, j, (s_h,h), loc_h, tab())
 182             else: # Or not at terminal level yet:
 183                 if 'INNER' in DEBUG:
 184                     print "%s%s (%.1f) from %d to %d" % (tab(),node_str((s_h,h)),loc_h,i,j)
 185                 if s_h == SEAL:
 186                     if h == POS(ROOT): # only used in testing, o/w we use inner_sent
 187                         h = sent_nums[loc_h]
 188                         if i != 0 or j != len(sent): raise ValueError
 189                         else: return g.p_ROOT[h] * e(i,j,(SEAL,h),loc_h,n_t+1)
 190                     p_RGOL = g.p_STOP[h,  LEFT, adj(i,loc_h)] * e(i,j,(RGOL,h),loc_h,n_t+1)
 191                     p_LGOR = g.p_STOP[h, RIGHT, adj(j,loc_h)] * e(i,j,(LGOR,h),loc_h,n_t+1)
 192                     p = p_RGOL + p_LGOR
 193                     to_mpp(p_RGOL,       STOPKEY,         (i,j, (RGOL,h),loc_h))
 194                     to_mpp(p_LGOR, (i,j, (RGOL,h),loc_h),       STOPKEY )
 195                     if 'INNER' in DEBUG:
 196                         print "%sp= %.4f (STOP)" % (tab(), p)
 197                 elif s_h == RGOL or s_h == GOL:
 198                     p = 0.0
 199                     if s_h == RGOL:
 200                         p = g.p_STOP[h, RIGHT, adj(j,loc_h)] * e(i,j, (GOR,h),loc_h,n_t+1)
 201                         to_mpp(p, (i,j, (GOR,h),loc_h), STOPKEY)
 202                     for k in xgo_left(i, loc_h): # i < k <= loc_l(h)
 203                         p_R = e(k, j, ( s_h,h), loc_h, n_t+1)
 204                         if p_R > 0.0:
 205                             for loc_a,a in locs(sent_nums, i, k):
 206                                 p_ah = g.p_GO_AT_or0(a, h, LEFT, adj(k,loc_h))
 207                                 if p_ah > 0.0:
 208                                     p_L = e(i, k, (SEAL,a), loc_a, n_t+1)
 209                                     p_add = p_L * p_ah * p_R
 210                                     p += p_add
 211                                     to_mpp(p_add,
 212                                           (i, k, (SEAL,a), loc_a),
 213                                           (k, j, ( s_h,h), loc_h))
 214                     if 'INNER' in DEBUG:
 215                         print "%sp= %.4f (ATTACH)" % (tab(), p)
 216                 elif s_h == GOR or s_h == LGOR:
 217                     p = 0.0
 218                     if s_h == LGOR:
 219                         p = g.p_STOP[h, LEFT, adj(i,loc_h)] * e(i,j, (GOL,h),loc_h,n_t+1)
 220                         to_mpp(p, (i,j, (GOL,h),loc_h), STOPKEY)
 221                     for k in xgo_right(loc_h, j): # loc_l(h) < k < j
 222                         p_L = e(i, k, ( s_h,h), loc_h, n_t+1)
 223                         if p_L > 0.0:
 224                             for loc_a,a in locs(sent_nums,k,j):
 225                                 p_ah = g.p_GO_AT_or0(a, h, RIGHT, adj(k,loc_h))
 226                                 p_R = e(k, j, (SEAL,a), loc_a, n_t+1)
 227                                 p_add = p_L * p_ah * p_R
 228                                 p += p_add
 229                                 to_mpp(p_add,
 230                                       (i, k, ( s_h,h), loc_h),
 231                                       (k, j, (SEAL,a), loc_a))
 232
 233                     if 'INNER' in DEBUG:
 234                         print "%sp= %.4f (ATTACH)" % (tab(), p)
 235                 # elif s_h == GOL: # todo
 236
 237                 ichart[i, j, (s_h,h), loc_h] = p
 238                 return p
 239     # end of e-function
 240
 241     inner_prob = e(i,j,node,loc_h, 0)
 242     if 'INNER' in DEBUG:
 243         print debug_ichart(g,sent,ichart)
 244     return inner_prob
 245 # end of dmv.inner(i, j, node, loc_h, g, sent, ichart,mpptree)
 246
 247
 248 def debug_ichart(g,sent,ichart):
 249     str = "---ICHART:---\n"
 250     for (s,t,LHS,loc_h),v in ichart.iteritems():
 251         str += "%s -> %s_%d ... %s_%d (loc_h:%s):\t%s\n" % (node_str(LHS,g.numtag),
 252                                                               sent[s], s, sent[s], t, loc_h, v)
 253     str += "---ICHART:end---\n"
 254     return str
 255
 256
 257 def inner_sent(g, sent, ichart):
 258     return sum([g.p_ROOT[w] * inner(0, len(sent), (SEAL,w), loc_w, g, sent, ichart)
 259                 for loc_w,w in locs(g.sent_nums(sent),0,len(sent))])
 260
 261
 262
 263
 264
 265 ###################################################
 266 #           P_OUTSIDE (dmv-specific)              #
 267 ###################################################
 268
 269 #@accepts(int, int, (int, int), int, Any(), [str], {tuple:float}, {tuple:float})
 270 def outer(i,j,w_node,loc_w, g, sent, ichart, ochart):
 271     ''' http://www.student.uib.no/~kun041/dmvccm/DMVCCM.html#outer
 272
 273     w_node is a pair (seals,POS); the w in klein-thesis is made up of
 274     POS(w) and loc_w
 275     '''
 276     sent_nums = g.sent_nums(sent)
 277     if POS(w_node) not in sent_nums[i:j]:
 278         # sanity check, w must be able to dominate sent[i:j]
 279         return 0.0
 280
 281     # local functions:
 282     def e(i,j,LHS,loc_h): # P_{INSIDE}
 283         try:
 284             return ichart[i,j,LHS,loc_h]
 285         except KeyError:
 286             return inner(i,j,LHS,loc_h,g,sent,ichart)
 287
 288     def f(i,j,w_node,loc_w):
 289         if not (i <= loc_w < j):
 290             return 0.0
 291         if (i,j,w_node,loc_w) in ochart:
 292             return ochart[i,j, w_node,loc_w]
 293         if w_node == ROOT:
 294             if i == 0 and j == len(sent):
 295                 return 1.0
 296             else: # ROOT may only be used on full sentence
 297                 return 0.0
 298         # but we may have non-ROOTs (stops) over full sentence too:
 299         w = POS(w_node)
 300         s_w = seals(w_node)
 301
 302         # todo: try either if p_M > 0.0: or sum(), and speed-test them
 303
 304         if s_w == SEAL: # w == a
 305             # todo: do the i<sent<j check here to save on calls?
 306             p = g.p_ROOT[w] * f(i,j,ROOT,loc_w)
 307             # left attach
 308             for k in xgt(j, sent): # j<k<len(sent)+1
 309                 for loc_h,h in locs(sent_nums,j,k):
 310                     p_wh = g.p_GO_AT_or0(w, h, LEFT, adj(j, loc_h))
 311                     for s_h in [RGOL, GOL]:
 312                         p += f(i,k,(s_h,h),loc_h) * p_wh * e(j,k,(s_h,h),loc_h)
 313             # right attach
 314             for k in xlt(i): # k<i
 315                 for loc_h,h in locs(sent_nums,k,i):
 316                     p_wh = g.p_GO_AT_or0(w, h, RIGHT, adj(i, loc_h))
 317                     for s_h in [LGOR, GOR]:
 318                         p += e(k,i,(s_h,h), loc_h) * p_wh * f(k,j,(s_h,h), loc_h)
 319
 320         elif s_w == RGOL or s_w == GOL: # w == h, left stop + left attach
 321             if s_w == RGOL:
 322                 s_h = SEAL
 323             else: # s_w == GOL
 324                 s_h = LGOR
 325             p = g.p_STOP[w, LEFT,  adj(i,loc_w)] * f(i,j,( s_h,w),loc_w)
 326             for k in xlt(i): # k<i
 327                 for loc_a,a in locs(sent_nums,k,i):
 328                     p_aw = g.p_GO_AT_or0(a, w, LEFT, adj(i, loc_w))
 329                     p += e(k,i, (SEAL,a),loc_a) * p_aw * f(k,j,w_node,loc_w)
 330
 331         elif s_w == GOR or s_w == LGOR: # w == h, right stop + right attach
 332             if s_w == GOR:
 333                 s_h = RGOL
 334             else: # s_w == LGOR
 335                 s_h = SEAL
 336             p = g.p_STOP[w, RIGHT, adj(j,loc_w)] * f(i,j,( s_h,w),loc_w)
 337             for k in xgt(j, sent): # j<k<len(sent)+1
 338                 for loc_a,a in locs(sent_nums,j,k):
 339                     p_ah = g.p_GO_AT_or0(a, w, RIGHT, adj(j, loc_w))
 340                     p += f(i,k,w_node,loc_w) * p_ah * e(j,k,(SEAL,a),loc_a)
 341
 342         ochart[i,j,w_node,loc_w] = p
 343         return p
 344     # end outer.f()
 345
 346     return f(i,j,w_node,loc_w)
 347 # end outer(i,j,w_node,loc_w, g,sent, ichart,ochart)
 348
 349
 350
 351
 352 ###################################################
 353 #              Reestimation v.1:                  #
 354 #           Sentences as outer loop               #
 355 ###################################################
 356
 357 def reest_zeros(h_nums):
 358     '''A dict to hold numerators and denominators for our 6+ reestimation
 359 formulas. '''
 360     # todo: p_ORDER?
 361     fr = { ('ROOT','den'):0.0 } # holds sum over f_sent!! not p_sent...
 362     for h in h_nums:
 363         fr['ROOT','num',h] = 0.0
 364         for s_h in [GOR,GOL,RGOL,LGOR]:
 365             x = (s_h,h)
 366             fr['hat_a','den',x] = 0.0 # = c()
 367             # not all arguments are attached to, so we just initialize
 368             # fr['hat_a','num',a,(s_h,h)] as they show up, in reest_freq
 369             for adj in [NON, ADJ]:
 370                 for nd in ['num','den']:
 371                     fr['STOP',nd,x,adj] = 0.0
 372     return fr
 373
 374
 375 def reest_freq(g, corpus):
 376     fr = reest_zeros(g.headnums())
 377     ichart = {}
 378     ochart = {}
 379     p_sent = None # 50 % speed increase on storing this locally
 380
 381     # local functions altogether 2x faster than global
 382     def c(i,j,LHS,loc_h,sent):
 383         if not p_sent > 0.0:
 384             return p_sent
 385
 386         p_in = e(i,j, LHS,loc_h,sent)
 387         if not p_in > 0.0:
 388             return p_in
 389
 390         p_out = f(i,j, LHS,loc_h,sent)
 391         return p_in * p_out / p_sent
 392     # end reest_freq.c()
 393
 394     def f(i,j,LHS,loc_h,sent): # P_{OUTSIDE}
 395         try:
 396             return ochart[i,j,LHS,loc_h]
 397         except KeyError:
 398             return outer(i,j,LHS,loc_h,g,sent,ichart,ochart)
 399     # end reest_freq.f()
 400
 401     def e(i,j,LHS,loc_h,sent): # P_{INSIDE}
 402         try:
 403             return ichart[i,j,LHS,loc_h]
 404         except KeyError:
 405             return inner(i,j,LHS,loc_h,g,sent,ichart)
 406     # end reest_freq.e()
 407
 408     def w_left(i,j, x,loc_h,sent,sent_nums):
 409         if not p_sent > 0.0: return
 410
 411         h = POS(x)
 412         a_k = {}
 413         for k in xtween(i, j):
 414             p_out = f(i,j, x,loc_h, sent)
 415             if not p_out > 0.0:
 416                 continue
 417             p_R =   e(k,j, x,loc_h, sent)
 418             if not p_R > 0.0:
 419                 continue
 420
 421             for loc_a,a in locs(sent_nums, i,k): # i<=loc_l(a)<k
 422                 p_rule = g.p_GO_AT_or0(a, h, LEFT, adj(k, loc_h))
 423                 p_L = e(i,k, (SEAL,a), loc_a, sent)
 424                 p = p_L * p_out * p_R * p_rule
 425                 try:    a_k[a] += p
 426                 except KeyError: a_k[a]  = p
 427
 428         for a,p in a_k.iteritems():
 429             try:    fr['hat_a','num',a,x] += p / p_sent
 430             except KeyError: fr['hat_a','num',a,x]  = p / p_sent
 431     # end reest_freq.w_left()
 432
 433     def w_right(i,j, x,loc_h,sent,sent_nums):
 434         if not p_sent > 0.0: return
 435
 436         h = POS(x)
 437         a_k = {}
 438         for k in xtween(i, j):
 439             p_out = f(i,j, x,loc_h, sent)
 440             if not p_out > 0.0:
 441                 continue
 442             p_L =   e(i,k, x,loc_h, sent)
 443             if not p_L > 0.0:
 444                 continue
 445
 446             for loc_a,a in locs(sent_nums, k,j): # k<=loc_l(a)<j
 447                 p_rule = g.p_GO_AT_or0(a, h, RIGHT, adj(k, loc_h))
 448                 p_R = e(k,j, (SEAL,a),loc_a, sent)
 449                 p = p_L * p_out * p_R * p_rule
 450                 try:    a_k[a] += p
 451                 except KeyError: a_k[a]  = p
 452
 453         for a,p in a_k.iteritems():
 454             try:    fr['hat_a','num',a,x] += p / p_sent
 455             except KeyError: fr['hat_a','num',a,x]  = p / p_sent
 456     # end reest_freq.w_right()
 457
 458     # in reest_freq:
 459     for sent in corpus:
 460         if 'REEST' in DEBUG:
 461             print sent
 462         ichart = {}
 463         ochart = {}
 464         p_sent = inner_sent(g, sent, ichart)
 465         fr['ROOT','den'] += 1 # divide by p_sent per h!
 466
 467         sent_nums = g.sent_nums(sent)
 468
 469         for loc_h,h in locs(sent_nums,0,len(sent)+1): # locs-stop is exclusive, thus +1
 470             # root:
 471             fr['ROOT','num',h] += g.p_ROOT[h] * e(0,len(sent), (SEAL,h),loc_h, sent) \
 472                 / p_sent
 473
 474             loc_l_h = loc_h
 475             loc_r_h = loc_l_h+1
 476
 477             # left non-adjacent stop:
 478             for i in xlt(loc_l_h):
 479                 fr['STOP','num',(GOL,h),NON] += c(loc_l_h, j, (LGOR, h),loc_h, sent)
 480                 fr['STOP','den',(GOL,h),NON] += c(loc_l_h, j, (GOL, h),loc_h, sent)
 481                 for j in xgteq(loc_r_h, sent):
 482                     fr['STOP','num',(RGOL,h),NON] += c(i, j, (SEAL, h),loc_h, sent)
 483                     fr['STOP','den',(RGOL,h),NON] += c(i, j, (RGOL, h),loc_h, sent)
 484             # left adjacent stop, i = loc_l_h
 485             fr['STOP','num',(GOL,h),ADJ] += c(loc_l_h, loc_r_h, (LGOR, h),loc_h, sent)
 486             fr['STOP','den',(GOL,h),ADJ] += c(loc_l_h, loc_r_h, (GOL, h),loc_h, sent)
 487             for j in xgteq(loc_r_h, sent):
 488                 fr['STOP','num',(RGOL,h),ADJ] += c(loc_l_h, j, (SEAL, h),loc_h, sent)
 489                 fr['STOP','den',(RGOL,h),ADJ] += c(loc_l_h, j, (RGOL, h),loc_h, sent)
 490             # right non-adjacent stop:
 491             for j in xgt(loc_r_h, sent):
 492                 fr['STOP','num',(GOR,h),NON] += c(loc_l_h, j, (RGOL, h),loc_h, sent)
 493                 fr['STOP','den',(GOR,h),NON] += c(loc_l_h, j, (GOR,  h),loc_h, sent)
 494                 for i in xlteq(loc_l_h):
 495                     fr['STOP','num',(LGOR,h),NON] += c(loc_l_h, j, (SEAL, h),loc_h, sent)
 496                     fr['STOP','den',(LGOR,h),NON] += c(loc_l_h, j, (LGOR,  h),loc_h, sent)
 497             # right adjacent stop, j = loc_r_h
 498             fr['STOP','num',(GOR,h),ADJ] += c(loc_l_h, loc_r_h, (RGOL, h),loc_h, sent)
 499             fr['STOP','den',(GOR,h),ADJ] += c(loc_l_h, loc_r_h, (GOR,  h),loc_h, sent)
 500             for i in xlteq(loc_l_h):
 501                 fr['STOP','num',(LGOR,h),ADJ] += c(loc_l_h, j, (SEAL, h),loc_h, sent)
 502                 fr['STOP','den',(LGOR,h),ADJ] += c(loc_l_h, j, (LGOR,  h),loc_h, sent)
 503
 504             # left attachment:
 505             if 'REEST_ATTACH' in DEBUG:
 506                 print "Lattach %s: for i < %s"%(g.numtag(h),sent[0:loc_h+1])
 507             for s_h in [RGOL, GOL]:
 508                 x = (s_h, h)
 509                 for i in xlt(loc_l_h):  # i < loc_l(h)
 510                     if 'REEST_ATTACH' in DEBUG:
 511                         print "\tfor j >= %s"%sent[loc_h:len(sent)]
 512                     for j in xgteq(loc_r_h, sent): # j >= loc_r(h)
 513                         fr['hat_a','den',x] += c(i,j, x,loc_h, sent) # v_q in L&Y
 514                         if 'REEST_ATTACH' in DEBUG:
 515                             print "\t\tc( %d , %d, %s, %s, sent)=%.4f"%(i,j,node_str(x),loc_h,fr['hat_a','den',x])
 516                         w_left(i, j, x,loc_h, sent,sent_nums) # compute w for all a in sent
 517
 518             # right attachment:
 519             if 'REEST_ATTACH' in DEBUG:
 520                 print "Rattach %s: for i <= %s"%(g.numtag(h),sent[0:loc_h+1])
 521             for s_h in [GOR, LGOR]:
 522                 x = (s_h, h)
 523                 for i in xlteq(loc_l_h):  # i <= loc_l(h)
 524                     if 'REEST_ATTACH' in DEBUG:
 525                         print "\tfor j > %s"%sent[loc_h:len(sent)]
 526                     for j in xgt(loc_r_h, sent): # j > loc_r(h)
 527                         fr['hat_a','den',x] += c(i,j, x,loc_h, sent) # v_q in L&Y
 528                         if 'REEST_ATTACH' in DEBUG:
 529                             print "\t\tc( %d , %d, %s, %s, sent)=%.4f"%(loc_h,j,node_str(x),loc_h,fr['hat_a','den',x])
 530                         w_right(i,j, x,loc_h, sent,sent_nums) # compute w for all a in sent
 531         # end for loc_h,h
 532     # end for sent
 533
 534     return fr
 535
 536 def reestimate(old_g, corpus):
 537     fr = reest_freq(old_g, corpus)
 538     p_ROOT, p_STOP, p_ATTACH = {},{},{}
 539
 540     for h in old_g.headnums():
 541         # reest_head changes p_ROOT, p_STOP, p_ATTACH
 542         reest_head(h, fr, old_g, p_ROOT, p_STOP, p_ATTACH)
 543     p_ORDER = old_g.p_ORDER
 544     numtag, tagnum = old_g.get_nums_tags()
 545
 546     new_g = DMV_Grammar(numtag, tagnum, p_ROOT, p_STOP, p_ATTACH, p_ORDER)
 547     return new_g
 548
 549
 550 def reest_head(h, fr, g, p_ROOT, p_STOP, p_ATTACH):
 551     "Given a single head, update g with the reestimated probability."
 552     # remove 0-prob stuff? todo
 553     try:
 554         p_ROOT[h] = fr['ROOT','num',h] / fr['ROOT','den']
 555     except KeyError:
 556         p_ROOT[h] = 0.0
 557
 558     for dir in [LEFT,RIGHT]:
 559         for adj in [ADJ, NON]: # p_STOP
 560             p_STOP[h, dir, adj] = 0.0
 561             for s_h in dirseal(dir):
 562                 x = (s_h,h)
 563                 p = fr['STOP','den', x, adj]
 564                 if p > 0.0:
 565                     p = fr['STOP', 'num', x, adj] / p
 566                 p_STOP[h, dir, adj] += p
 567
 568         for s_h in dirseal(dir): # make hat_a for p_ATTACH
 569             x = (s_h,h)
 570             p_c = fr['hat_a','den',x]
 571
 572             for a in g.headnums():
 573                 if (a,h,dir) not in p_ATTACH:
 574                     p_ATTACH[a,h,dir] = 0.0
 575                 try: # (a,x) might not be in hat_a
 576                     p_ATTACH[a,h,dir] += fr['hat_a','num',a,x] / p_c
 577                 except KeyError: pass
 578                 except ZeroDivisionError: pass
 579
 580
 581
 582
 583
 584 ###################################################
 585 #              Reestimation v.2:                  #
 586 #            Heads as outer loop                  #
 587 ###################################################
 588
 589 def locs_h(h, sent_nums):
 590     '''Return the between-word locations of all tokens of h in sent.'''
 591     return [loc_w for loc_w,w in locs(sent_nums, 0, len(sent_nums))
 592             if w == h]
 593
 594 def locs_a(a, sent_nums, start, stop):
 595     '''Return the between-word locations of all tokens of h in some
 596     fragment of sent. We make sure to offset the locations correctly
 597     so that for any w in the returned list, sent[w]==loc_w.
 598
 599     start is inclusive, stop is exclusive, as in klein-thesis and
 600     Python's list-slicing (eg. return left-loc).'''
 601     return [loc_w for loc_w,w in locs(sent_nums, start, stop)
 602             if w == a]
 603
 604 def inner2(i, j, node, loc_h, g, sent):
 605     ichart,ochart = g.get_iochart(s_n)
 606     try: p = ichart[i,j,x,loc_h]
 607     except KeyError: p = inner(i,j,x,loc_h,g,sent,ichart)
 608     g.set_iochart(s_n,ichart,ochart)
 609     return p
 610
 611 def inner_sent2(g, sent):
 612     ichart,ochart = g.get_iochart(s_n)
 613     p = inner_sent(g,sent,ichart)
 614     g.set_iochart(s_n,ichart,ochart)
 615     return p
 616
 617 def outer2(i, j,w_node,loc_w, g, sent):
 618     ichart,ochart = g.get_iochart(s_n)
 619     try: p = ochart[i,j,w_node,loc_w]
 620     except KeyError: p = inner(i,j,w_node,loc_w,g,sent,ichart,ochart)
 621     g.set_iochart(s_n,ichart,ochart)
 622     return p
 623
 624 def reestimate2(old_g, corpus):
 625     p_ROOT, p_STOP, p_ATTACH = {},{},{}
 626
 627     for h in old_g.headnums():
 628         # reest_head changes p_ROOT, p_STOP, p_ATTACH
 629         reest_head2(h, old_g, corpus, p_ROOT, p_STOP, p_ATTACH)
 630     p_ORDER = old_g.p_ORDER
 631     numtag, tagnum = old_g.get_nums_tags()
 632
 633     new_g = DMV_Grammar(numtag, tagnum, p_ROOT, p_STOP, p_ATTACH, p_ORDER)
 634     return new_g
 635
 636 def hat_d2(xbar, x, xi, xj, g, corpus): # stop helper
 637     def c(x,loc_x,i,j): return c2(x,loc_x,i,j,g,s_n,sent)
 638
 639     h = POS(x)
 640     if h != POS(xbar): raise ValueError
 641
 642     num, den = 0.0, 0.0
 643     for s_n,sent in [(g.sent_nums(sent),sent) for sent in corpus]:
 644         for loc_h in locs_h(h,s_n):
 645             loc_l_h, loc_r_h = loc_h, loc_h + 1
 646             for i in xi(loc_l_h):
 647                 for j in xj(loc_r_h, s_n):
 648 #                     print "s:%s %d,%d"%(sent,i,j)
 649                     num += c(xbar,loc_h,i,j)
 650                     den += c(x,loc_h,i,j)
 651     if den == 0.0:
 652         return den
 653     return num/den # eg. SEAL/RGOL, xbar/x
 654
 655
 656 def c2(x,loc_h,i,j,g,s_n,sent):
 657     ichart,ochart = g.get_iochart(s_n)
 658
 659     def f(i,j,x,loc_h): # P_{OUTSIDE}
 660         try: return ochart[i,j,x,loc_h]
 661         except KeyError: return outer(i,j,x,loc_h,g,sent,ichart,ochart)
 662     def e(i,j,x,loc_h): # P_{INSIDE}
 663         try: return ichart[i,j,x,loc_h]
 664         except KeyError: return inner(i,j,x,loc_h,g,sent,ichart)
 665
 666     p_sent = inner_sent(g, sent, ichart)
 667     if not p_sent > 0.0:
 668         return p_sent
 669
 670     p_in = e(i,j, x,loc_h)
 671     if not p_in > 0.0:
 672         return p_in
 673
 674     p_out = f(i,j, x,loc_h)
 675
 676     g.set_iochart(s_n,ichart,ochart)
 677     return p_in * p_out / p_sent
 678
 679 def w2(a, x,loc_h, dir, i, j, g, s_n,sent):
 680     ichart,ochart = g.get_iochart(s_n)
 681
 682     def f(i,j,x,loc_h): # P_{OUTSIDE}
 683         try: return ochart[i,j,x,loc_h]
 684         except KeyError: return outer(i,j,x,loc_h,g,sent,ichart,ochart)
 685     def e(i,j,x,loc_h): # P_{INSIDE}
 686         try: return ichart[i,j,x,loc_h]
 687         except KeyError: return inner(i,j,x,loc_h,g,sent,ichart)
 688
 689     h = POS(x)
 690     p_sent = inner_sent(g, sent, ichart)
 691
 692     if dir == LEFT:
 693         L, R = (SEAL,a),x
 694     else:
 695         L, R = x,(SEAL,a)
 696     w_sum = 0.0
 697
 698     for k in xtween(i,j):
 699         if dir == LEFT:
 700             start, stop = i, k
 701         else:
 702             start, stop = k, j
 703         for loc_a in locs_a(a, s_n, start, stop):
 704             if dir == LEFT:
 705                 loc_L, loc_R = loc_a, loc_h
 706             else:
 707                 loc_L, loc_R = loc_h, loc_a
 708             p = g.p_GO_AT_or0(a,h,dir,adj(k,loc_h))
 709             in_L = e(i,k,L,loc_L)
 710             in_R = e(k,j,R,loc_R)
 711             out =  f(i,j,x,loc_h)
 712             w_sum += p * in_L * in_R * out
 713
 714     g.set_iochart(s_n,ichart,ochart)
 715     return w_sum/p_sent
 716
 717 def hat_a2(a, x, dir, g, corpus): # attachment helper
 718     def w(a,x,loc_x,dir,i,j): return w2(a,x,loc_x,dir,i,j,g,s_n,sent)
 719     def c(x,loc_x,i,j): return c2(x,loc_x,i,j,g,s_n,sent)
 720
 721     h = POS(x)
 722     if dir == LEFT:
 723         xi, xj = xlt, xgteq
 724     else:
 725         xi, xj = xlteq, xgt
 726     den, num = 0.0, 0.0
 727
 728     for s_n,sent in [(g.sent_nums(sent),sent) for sent in corpus]:
 729         for loc_h in locs_h(h,s_n):
 730             loc_l_h, loc_r_h = loc_h, loc_h + 1
 731             for i in xi(loc_l_h):
 732                 for j in xj(loc_r_h,sent):
 733                     num += w(a, x,loc_h, dir, i,j)
 734                     den += c(x,loc_h, i,j)
 735     if den == 0.0:
 736         return den
 737     return num/den
 738
 739 def reest_root2(h,g,corpus):
 740     sum = 0.0
 741     corpus_size = 0.0
 742     for s_n,sent in [(g.sent_nums(sent),sent) for sent in corpus]:
 743         num, den = 0.0, 0.0
 744         corpus_size += 1.0
 745         ichart, ochart = g.get_iochart(s_n)
 746         den += inner_sent(g, sent, ichart)
 747         for loc_h in locs_h(h,s_n):
 748             num += \
 749                 g.p_ROOT[h] * \
 750                 inner(0, len(s_n), (SEAL,h), loc_h, g, sent, ichart)
 751         g.set_iochart(s_n, ichart, ochart)
 752         sum += num / den
 753     return sum / corpus_size
 754
 755 def reest_head2(h, g, corpus, p_ROOT, p_STOP, p_ATTACH):
 756     print "h: %d=%s ..."%(h,g.numtag(h)),
 757     def hat_d(xbar,x,xi,xj): return hat_d2(xbar,x,xi,xj, g, corpus)
 758     def hat_a(a, x, dir  ): return hat_a2(a, x, dir,   g, corpus)
 759
 760     p_STOP[h, LEFT,NON] = \
 761         hat_d((SEAL,h),(RGOL,h),xlt,  xgteq) + \
 762         hat_d((LGOR,h),( GOL,h),xlt,  xeq)
 763     p_STOP[h, LEFT,ADJ] = \
 764         hat_d((SEAL,h),(RGOL,h),xeq,  xgteq) + \
 765         hat_d((LGOR,h),( GOL,h),xeq,  xeq)
 766     p_STOP[h,RIGHT,NON] = \
 767         hat_d((RGOL,h),( GOR,h),xeq,  xgt) + \
 768         hat_d((SEAL,h),(LGOR,h),xlteq,xgt)
 769     p_STOP[h,RIGHT,ADJ] = \
 770         hat_d((RGOL,h),( GOR,h),xeq,  xeq) + \
 771         hat_d((SEAL,h),(LGOR,h),xlteq,xeq)
 772     print "stops done...",
 773
 774     p_ROOT[h] = reest_root2(h,g,corpus)
 775     print "root done...",
 776
 777     for a in g.headnums():
 778         p_ATTACH[a,h,LEFT] = \
 779             hat_a(a, (GOL,h),LEFT) + \
 780             hat_a(a,(RGOL,h),LEFT)
 781         p_ATTACH[a,h,RIGHT] = \
 782             hat_a(a, (GOR,h),RIGHT) + \
 783             hat_a(a,(LGOR,h),RIGHT)
 784
 785     print "attachment done"
 786
 787
 788
 789 ###################################################
 790 #             Most Probable Parse:                #
 791 ###################################################
 792
 793 STOPKEY = (-1,-1,STOP,-1)
 794 ROOTKEY = (-1,-1,ROOT,-1)
 795
 796 def make_mpptree(g, sent):
 797     '''Tell inner() to make an mpptree, connect ROOT to this. (Logically,
 798     this should be part of inner_sent though...)'''
 799     ichart = {}
 800     mpptree = { ROOTKEY:(0.0, ROOTKEY, None) }
 801     for loc_w,w in locs(g.sent_nums(sent),0,len(sent)):
 802         p = g.p_ROOT[w] * inner(0, len(sent), (SEAL,w), loc_w, g, sent, ichart, mpptree)
 803         L = ROOTKEY
 804         R = (0,len(sent), (SEAL,w), loc_w)
 805         if mpptree[ROOTKEY][0] < p:
 806             mpptree[ROOTKEY] = (p, L, R)
 807     return mpptree
 808
 809 def parse_mpptree(mpptree, sent):
 810     '''mpptree is a dict of the form {k:(p,L,R),...}; where k, L and R
 811     are `keys' of the form (i,j,node,loc).
 812
 813     returns an mpp of the form [((head, loc_h),(arg, loc_a)), ...],
 814     where head and arg are tags.'''
 815     # local functions for clear access to mpptree:
 816     def k_node(key):
 817         return key[2]
 818     def k_POS(key):
 819         return POS(k_node(key))
 820     def k_seals(key):
 821         return seals(k_node(key))
 822     def k_locnode(key):
 823         return (k_node(key),key[3])
 824     def k_locPOS(key):
 825         return (k_POS(key),key[3])
 826     def k_terminal(key):
 827         s_k = k_seals(key) # i+1 == j
 828         return key[0] + 1 == key[1] and (s_k == GOR or s_k == GOL)
 829     def t_L(tree_entry):
 830         return tree_entry[1]
 831     def t_R(tree_entry):
 832         return tree_entry[2]
 833
 834     # arbitrarily, "ROOT attaches to right". We add it here to
 835     # avoid further complications:
 836     firstkey = t_R(mpptree[ROOTKEY])
 837     deps = set([ (k_locPOS(ROOTKEY), k_locPOS(firstkey), RIGHT) ])
 838     q = [firstkey]
 839
 840     while len(q) > 0:
 841         k = q.pop()
 842         if k_terminal(k):
 843             continue
 844         else:
 845             L = t_L( mpptree[k] )
 846             R = t_R( mpptree[k] )
 847             if k_locnode( k ) == k_locnode( L ): # Rattach
 848                 deps.add((k_locPOS( k ), k_locPOS( R ), LEFT))
 849                 q.extend( [L, R] )
 850             elif k_locnode( k ) == k_locnode( R ): # Lattach
 851                 deps.add((k_locPOS( k ), k_locPOS( L ), RIGHT))
 852                 q.extend( [L, R] )
 853             elif R == STOPKEY:
 854                 q.append( L )
 855             elif L == STOPKEY:
 856                 q.append( R )
 857     return deps
 858
 859 def mpp(g, sent):
 860     tagf = g.numtag # localized function, todo: speed-test
 861     mpptree = make_mpptree(g, sent)
 862     return set([((tagf(h), loc_h), (tagf(a), loc_a))
 863                 for (h, loc_h),(a,loc_a),dir in parse_mpptree(mpptree,sent)])
 864
 865
 866 ########################################################################
 867 #     testing functions:                                               #
 868 ########################################################################
 869
 870 testcorpus = [s.split() for s in ['det nn vbd c vbd','vbd nn c vbd',
 871                                   'det nn vbd',      'det nn vbd c pp',
 872                                   'det nn vbd',      'det vbd vbd c pp',
 873                                   'det nn vbd',      'det nn vbd c vbd',
 874                                   'det nn vbd',      'det nn vbd c vbd',
 875                                   'det nn vbd',      'det nn vbd c vbd',
 876                                   'det nn vbd',      'det nn vbd c pp',
 877                                   'det nn vbd pp',   'det nn vbd', ]]
 878
 879 def testgrammar():
 880     import loc_h_harmonic
 881     reload(loc_h_harmonic)
 882
 883     # make sure these are the way they were when setting up the tests:
 884     loc_h_harmonic.HARMONIC_C = 0.0
 885     loc_h_harmonic.FNONSTOP_MIN = 25
 886     loc_h_harmonic.FSTOP_MIN = 5
 887     loc_h_harmonic.RIGHT_FIRST = 1.0
 888     loc_h_harmonic.OTHER_STOP_CALC = False
 889
 890     return loc_h_harmonic.initialize(testcorpus)
 891
 892 def testreestimation2():
 893     g2 = testgrammar()
 894     reestimate2(g2, testcorpus)
 895     return g2
 896
 897 def testreestimation():
 898     g = testgrammar()
 899     g = reestimate(g, testcorpus)
 900     return g
 901
 902
 903 def testmpp_regression(mpptree,k_n):
 904     mpp = {ROOTKEY: (2.877072116829971e-05, STOPKEY, (0, 3, (2, 3), 1)),
 905  (0, 1, (1, 1), 0): (0.1111111111111111, (0, 1, (0, 1), 0), STOPKEY),
 906  (0, 1, (2, 1), 0): (0.049382716049382713, STOPKEY, (0, 1, (1, 1), 0)),
 907  (0, 3, (1, 3), 1): (0.00027619892321567721,
 908                        (0, 1, (2, 1), 0),
 909                        (1, 3, (1, 3), 1)),
 910  (0, 3, (2, 3), 1): (0.00012275507698474543, STOPKEY, (0, 3, (1, 3), 1)),
 911  (1, 3, (0, 3), 1): (0.025280986819448362,
 912                        (1, 2, (0, 3), 1),
 913                        (2, 3, (2, 4), 2)),
 914  (1, 3, (1, 3), 1): (0.0067415964851862296, (1, 3, (0, 3), 1), STOPKEY),
 915  (2, 3, (1, 4), 2): (0.32692307692307693, (2, 3, (0, 4), 2), STOPKEY),
 916  (2, 3, (2, 4), 2): (0.037721893491124266, STOPKEY, (2, 3, (1, 4), 2))}
 917     for k,(v,L,R) in mpp.iteritems():
 918         k2 = k[0:k_n] # 3 if the new does not check loc_h
 919         if type(k)==str:
 920             k2 = k
 921         if k2 not in mpptree:
 922             print "mpp regression, %s missing"%(k2,)
 923         else:
 924             vnew = mpptree[k2][0]
 925             if not "%.10f"%vnew == "%.10f"%v:
 926                 print "mpp regression, wanted %s=%.5f, got %.5f"%(k2,v,vnew)
 927
 928
 929 def testgrammar_a():
 930     h, a = 0, 1
 931     p_ROOT, p_STOP, p_ATTACH, p_ORDER = {},{},{},{}
 932     p_ROOT[h] = 0.9
 933     p_ROOT[a] = 0.1
 934     p_STOP[h,LEFT,NON] = 1.0
 935     p_STOP[h,LEFT,ADJ] = 1.0
 936     p_STOP[h,RIGHT,NON] = 0.4 # RSTOP
 937     p_STOP[h,RIGHT,ADJ] = 0.3 # RSTOP
 938     p_STOP[a,LEFT,NON] = 1.0
 939     p_STOP[a,LEFT,ADJ] = 1.0
 940     p_STOP[a,RIGHT,NON] = 0.4 # RSTOP
 941     p_STOP[a,RIGHT,ADJ] = 0.3 # RSTOP
 942     p_ATTACH[a,h,LEFT] = 1.0  # not used
 943     p_ATTACH[a,h,RIGHT] = 1.0 # not used
 944     p_ATTACH[h,a,LEFT] = 1.0  # not used
 945     p_ATTACH[h,a,RIGHT] = 1.0 # not used
 946     p_ATTACH[h,h,LEFT] = 1.0  # not used
 947     p_ATTACH[h,h,RIGHT] = 1.0 # not used
 948     p_ORDER[(GOR, h)] = 1.0
 949     p_ORDER[(GOL, h)] = 0.0
 950     p_ORDER[(GOR, a)] = 1.0
 951     p_ORDER[(GOL, a)] = 0.0
 952     g = DMV_Grammar({h:'h',a:'a'}, {'h':h,'a':a}, p_ROOT, p_STOP, p_ATTACH, p_ORDER)
 953     # these probabilities are impossible so add them manually:
 954     g.p_GO_AT[a,a,LEFT,NON] = 0.4  # Lattach
 955     g.p_GO_AT[a,a,LEFT,ADJ] = 0.6  # Lattach
 956     g.p_GO_AT[h,a,LEFT,NON] = 0.2  # Lattach to h
 957     g.p_GO_AT[h,a,LEFT,ADJ] = 0.1  # Lattach to h
 958     g.p_GO_AT[a,a,RIGHT,NON] = 1.0 # Rattach
 959     g.p_GO_AT[a,a,RIGHT,ADJ] = 1.0 # Rattach
 960     g.p_GO_AT[h,a,RIGHT,NON] = 1.0 # Rattach to h
 961     g.p_GO_AT[h,a,RIGHT,ADJ] = 1.0 # Rattach to h
 962     g.p_GO_AT[h,h,LEFT,NON] = 0.2 # Lattach
 963     g.p_GO_AT[h,h,LEFT,ADJ] = 0.1 # Lattach
 964     g.p_GO_AT[a,h,LEFT,NON] = 0.4 # Lattach to a
 965     g.p_GO_AT[a,h,LEFT,ADJ] = 0.6 # Lattach to a
 966     g.p_GO_AT[h,h,RIGHT,NON] = 1.0 # Rattach
 967     g.p_GO_AT[h,h,RIGHT,ADJ] = 1.0 # Rattach
 968     g.p_GO_AT[a,h,RIGHT,NON] = 1.0 # Rattach to a
 969     g.p_GO_AT[a,h,RIGHT,ADJ] = 1.0 # Rattach to a
 970     return g
 971
 972
 973 def testgrammar_h():
 974     h = 0
 975     p_ROOT, p_STOP, p_ATTACH, p_ORDER = {},{},{},{}
 976     p_ROOT[h] = 1.0
 977     p_STOP[h,LEFT,NON] = 1.0
 978     p_STOP[h,LEFT,ADJ] = 1.0
 979     p_STOP[h,RIGHT,NON] = 0.4
 980     p_STOP[h,RIGHT,ADJ] = 0.3
 981     p_ATTACH[h,h,LEFT] = 1.0  # not used
 982     p_ATTACH[h,h,RIGHT] = 1.0 # not used
 983     p_ORDER[(GOR, h)] = 1.0
 984     p_ORDER[(GOL, h)] = 0.0
 985     g = DMV_Grammar({h:'h'}, {'h':h}, p_ROOT, p_STOP, p_ATTACH, p_ORDER)
 986     g.p_GO_AT[h,h,LEFT,NON] = 0.6 # these probabilities are impossible
 987     g.p_GO_AT[h,h,LEFT,ADJ] = 0.7 # so add them manually...
 988     g.p_GO_AT[h,h,RIGHT,NON] = 1.0
 989     g.p_GO_AT[h,h,RIGHT,ADJ] = 1.0
 990     return g
 991
 992
 993
 994 def testreestimation_h():
 995     DEBUG.add('REEST')
 996     g = testgrammar_h()
 997     reestimate(g,['h h h'.split()])
 998
 999
1000 def test(wanted, got):
1001     if not wanted == got:
1002         raise Warning, "Regression! Should be %s: %s" % (wanted, got)
1003
1004 def regression_tests():
1005     testmpp_regression(make_mpptree(testgrammar(), testcorpus[2]),4)
1006     h = 0
1007
1008     test("0.120",
1009          "%.3f" % inner(0, 2, (SEAL,h), 0, testgrammar_h(), 'h h'.split(),{}))
1010     test("0.063",
1011          "%.3f" % inner(0, 2, (SEAL,h), 1, testgrammar_h(), 'h h'.split(),{}))
1012     test("0.1842",
1013          "%.4f" % inner_sent(testgrammar_h(), 'h h h'.split(),{}))
1014
1015     test("0.1092",
1016          "%.4f" % inner(0, 3, (SEAL,0), 0, testgrammar_h(), 'h h h'.split(),{}))
1017     test("0.0252",
1018          "%.4f" % inner(0, 3, (SEAL,0), 1, testgrammar_h(), 'h h h'.split(),{}))
1019     test("0.0498",
1020          "%.4f" % inner(0, 3, (SEAL,h), 2, testgrammar_h(), 'h h h'.split(),{}))
1021
1022     test("0.58" ,
1023          "%.2f" % outer(1, 3, (RGOL,h), 2, testgrammar_h(),'h h h'.split(),{},{}))
1024     test("0.61" , # ftw? can't be right... there's an 0.4 shared between these two...
1025          "%.2f" % outer(1, 3, (RGOL,h), 1, testgrammar_h(),'h h h'.split(),{},{}))
1026
1027     test("0.00" ,
1028          "%.2f" % outer(1, 3, (RGOL,h), 0, testgrammar_h(),'h h h'.split(),{},{}))
1029     test("0.00" ,
1030          "%.2f" % outer(1, 3, (RGOL,h), 3, testgrammar_h(),'h h h'.split(),{},{}))
1031
1032     test("0.1089" ,
1033          "%.4f" % outer(0, 1, (GOR,h),  0,testgrammar_a(),'h a'.split(),{},{}))
1034     test("0.3600" ,
1035          "%.4f" % outer(0, 2, (GOR,h),  0,testgrammar_a(),'h a'.split(),{},{}))
1036     test("0.0000" ,
1037          "%.4f" % outer(0, 3, (GOR,h),  0,testgrammar_a(),'h a'.split(),{},{}))
1038
1039     # todo: add more of these tests...
1040
1041
1042
1043 def compare_grammars(g1,g2):
1044     result = ""
1045     for d1,d2 in [(g1.p_ATTACH,g2.p_ATTACH),(g1.p_STOP,g2.p_STOP),
1046                   (g1.p_ORDER, g2.p_ORDER), (g1.p_ROOT,g2.p_ROOT) ]:
1047         for k,v in d1.iteritems():
1048             if k not in d2:
1049                 result += "\nreestimate1[%s]=%s missing from reestimate2"%(k,v)
1050             elif "%s"%d2[k] != "%s"%v:
1051                 result += "\nreestimate1[%s]=%s while \nreestimate2[%s]=%s."%(k,v,k,d2[k])
1052         for k,v in d2.iteritems():
1053             if k not in d1:
1054                 result += "\nreestimate2[%s]=%s missing from reestimate1"%(k,v)
1055     return result
1056
1057
1058 def testNVNgrammar():
1059     from loc_h_harmonic import initialize
1060     g = initialize(['n v n'.split()])
1061     return g # todo
1062
1063 def testIO():
1064     g = testgrammar()
1065     inners = [(sent, inner_sent(g, sent, {})) for sent in testcorpus]
1066     return inners
1067
1068 if __name__ == "__main__":
1069     DEBUG.clear()
1070     regression_tests()
1071
1072 #    import profile
1073 #    profile.run('testreestimation()')
1074
1075 #     import timeit
1076 #     print timeit.Timer("loc_h_dmv.testreestimation()",'''import loc_h_dmv
1077 # reload(loc_h_dmv)''').timeit(1)
1078
1079
1080 #     print "mpp-test:"
1081 #     import pprint
1082 #     for s in testcorpus:
1083 #         print "sent:%s\nparse:set(\n%s)"%(s,pprint.pformat(list(mpp(testgrammar(), s)),
1084 #                                                     width=40))
1085
1086 #     g1 = testreestimation()
1087 #     g2 = testreestimation2()
1088 #     print compare_grammars(g1,g2)
1089     g = testNVNgrammar()
1090     q_sent = inner_sent(g,'n v n'.split(),{})
1091     q_tree = {}
1092     q_tree[1] = 2.7213e-06 # n_0 -> v, n_0 -> n_2
1093     q_tree[2] = 9.738e-06   # n -> v -> n
1094     q_tree[3] = 2.268e-06   # n_0 -> n_2 -> v
1095     q_tree[4] = 2.7213e-06 # same as 1-3
1096     q_tree[5] = 9.738e-06
1097     q_tree[6] = 2.268e-06
1098     q_tree[7] = 1.086e-05   # n <- v -> n (e-05!!!)
1099     f_T_q = {}
1100     for i,q_t in q_tree.iteritems():
1101         f_T_q[i] = q_t / q_sent
1102     import pprint
1103     pprint.pprint(q_tree)
1104     pprint.pprint(f_T_q)
1105     print sum([f for f in f_T_q.values()])
1106
1107     def treediv(num,den):
1108         return \
1109             sum([f_T_q[i] for i in num ]) / \
1110             sum([f_T_q[i] for i in den ])
1111     g2 = {}
1112 #     g2['root --> _n_'] = treediv( (1,2,3,4,5,6), (1,2,3,4,5,6,7) )
1113 #     g2['root --> _v_'] = treediv( (7,), (1,2,3,4,5,6,7) )
1114 #     g2['_n_ --> STOP n><'] = treediv( (1,2,3,4,5,6,7,1,2,3,4,5,6,7),
1115 #                                       (1,2,3,4,5,6,7,1,2,3,4,5,6,7))
1116
1117 #     g2['_n_ --> STOP n>< NON'] = treediv( (3,4,5,6),
1118 #                                           (3,4,5,6,4) )
1119
1120 #     g2['_v_ --> STOP v><'] = treediv( (1,2,3,4,5,6,7),
1121 #                                       (1,2,3,4,5,6,7) )
1122 #     nlrtrees = (1,2,3,4,5,6,7,1,2,3,4,5,6,7,
1123 #                 3,4,4,5,6)
1124 #     g2['n>< --> _n_ n><'] = treediv( (  4,  6), nlrtrees )
1125 #     g2['n>< --> _v_ n><'] = treediv( (3,4,5), nlrtrees )
1126 #     g2['n>< --> n> STOP'] = treediv( (1,2,3,4,5,6,7,1,2,3,4,5,6,7),
1127 #                                      nlrtrees )
1128
1129 #     g2['n>< --> n> STOP ADJ'] = treediv( (      4,5,  7,1,2,3,4,5,6,7),
1130 #                                          nlrtrees )
1131 #     g2['n>< --> n> STOP NON'] = treediv( (1,2,3,    6),
1132 #                                          nlrtrees )
1133
1134 #     vlrtrees = (1,2,3,4,5,6,7,
1135 #                 7,5)
1136 #     g2['v>< --> _n_ v><'] = treediv( (5,7), vlrtrees )
1137 #     g2['v>< --> v> STOP'] = treediv( (1,2,3,4,5,6,7), vlrtrees )
1138 #     nrtrees = (1,2,3,4,5,6,7,1,2,3,4,5,6,7,
1139 #                1,1,2,3,6)
1140 #     g2['n> --> n> _n_'] = treediv( (1,3), nrtrees )
1141 #     g2['n> --> n> _v_'] = treediv( (1,2,6), nrtrees )
1142
1143 #     g2['n> --> n> _n_ NON'] = treediv( (1,), nrtrees )
1144 #     g2['n> --> n> _n_ ADJ'] = treediv( (      3,), nrtrees )
1145 #     g2['n> --> n> _v_ ADJ'] = treediv( (  1,2,  6), nrtrees )
1146
1147 #     vrtrees = (1,2,3,4,5,6,7,
1148 #                7,2)
1149 #     g2['v> --> v> _n_'] = treediv( (2,7), vrtrees )
1150
1151 #     g2[' v|n,R '] = treediv( (1,  2,  6),
1152 #                              (1,1,2,3,6) )
1153 #     g2[' n|n,R '] = treediv( (1,    3),
1154 #                              (1,1,2,3,6) )
1155
1156     g2[' stop|n,R,non '] = treediv( (  1,2,3,6),
1157                                     (1,1,2,3,6) )
1158     g2[' v|n,left '] = treediv( (    3,4,5),
1159                                 (6,4,3,4,5) )
1160     g2[' n|n,left '] = treediv( (6,4),
1161                                 (6,4,3,4,5) )
1162
1163     pprint.pprint(g2)
1164     g3 = reestimate2(g, ['n v n'.split()])
1165     print g3
1166     g4 = reestimate2(g, ['n v n'.split()])
1167     print g4
1168
1169