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 #import numpy # numpy provides Fast Arrays, for future optimization
   7 import io
   8 from common_dmv import *
   9
  10 ### todo: debug with @accepts once in a while, but it's SLOW
  11 # from typecheck import accepts, Any
  12
  13 if __name__ == "__main__":
  14     print "loc_h_dmv module tests:"
  15
  16 def adj(middle, loc_h):
  17     "middle is eg. k when rewriting for i<k<j (inside probabilities)."
  18     return middle == loc_h or middle == loc_h+1 # ADJ == True
  19
  20 def make_GO_AT(p_STOP,p_ATTACH):
  21     p_GO_AT = {}
  22     for (a,h,dir), p_ah in p_ATTACH.iteritems():
  23         p_GO_AT[a,h,dir, NON] = p_ah * (1-p_STOP[h, dir, NON])
  24         p_GO_AT[a,h,dir, ADJ] = p_ah * (1-p_STOP[h, dir, ADJ])
  25     return p_GO_AT
  26
  27 class DMV_Grammar(io.Grammar):
  28     def __str__(self):
  29         def t(n):
  30             return "%d=%s" % (n, self.numtag(n))
  31         def p(dict,key):
  32             if key in dict:
  33                 return dict[key]
  34             else:
  35                 return 0.0
  36         def p_a(a,h):
  37             p_L = p(self.p_ATTACH,(a,h,LEFT))
  38             p_R = p(self.p_ATTACH,(a,h,RIGHT))
  39             if p_L == 0.0 and p_R == 0.0:
  40                 return ''
  41             else:
  42                 if p_L > 0.0:
  43                     str = "p_ATTACH[ %s|%s,L] = %.4f" % (t(a), t(h), p_L)
  44                 else:
  45                     str = ''
  46                 if p_R > 0.0:
  47                     str = str.ljust(40)
  48                     str += "p_ATTACH[ %s|%s,R] = %.4f" % (t(a), t(h), p_R)
  49             return str+'\n'
  50
  51         root, stop, att, ord = "","","",""
  52         for h in self.headnums():
  53             root += "p_ROOT[%s] = %.4f\n" % (t(h), p(self.p_ROOT, (h)))
  54             stop += "p_STOP[stop|%s,L,adj] = %.4f\t" % (t(h), p(self.p_STOP, (h,LEFT,ADJ)))
  55             stop += "p_STOP[stop|%s,R,adj] = %.4f\n" % (t(h), p(self.p_STOP, (h,RIGHT,ADJ)))
  56             stop += "p_STOP[stop|%s,L,non] = %.4f\t" % (t(h), p(self.p_STOP, (h,LEFT,NON)))
  57             stop += "p_STOP[stop|%s,R,non] = %.4f\n" % (t(h), p(self.p_STOP, (h,RIGHT,NON)))
  58             att  += ''.join([p_a(a,h) for a in self.headnums()])
  59             ord  += "p_ORDER[ left-first|%s ] = %.4f\t" % (t(h), p(self.p_ORDER, (GOL,h)))
  60             ord  += "p_ORDER[right-first|%s ] = %.4f\n" % (t(h), p(self.p_ORDER, (GOR,h)))
  61         return root + stop + att + ord
  62
  63     def __init__(self, numtag, tagnum, p_ROOT, p_STOP, p_ATTACH, p_ORDER):
  64         io.Grammar.__init__(self, numtag, tagnum)
  65         self.p_ROOT = p_ROOT     # p_ROOT[w] = p
  66         self.p_ORDER = p_ORDER   # p_ORDER[seals, w] = p
  67         self.p_STOP = p_STOP     # p_STOP[w, LEFT, NON] = p     (etc. for LA,RN,RA)
  68         self.p_ATTACH = p_ATTACH # p_ATTACH[a, h, LEFT] = p (etc. for R)
  69                                  # p_GO_AT[a, h, LEFT, NON] = p (etc. for LA,RN,RA)
  70         self.p_GO_AT = make_GO_AT(self.p_STOP, self.p_ATTACH)
  71
  72     def p_GO_AT_or0(self, a, h, dir, adj):
  73         try:
  74             return self.p_GO_AT[a, h, dir, adj]
  75         except:
  76             return 0.0
  77
  78
  79 def locs(sent_nums, start, stop):
  80     '''Return the between-word locations of all words in some fragment of
  81     sent. We make sure to offset the locations correctly so that for
  82     any w in the returned list, sent[w]==loc_w.
  83
  84     start is inclusive, stop is exclusive, as in klein-thesis and
  85     Python's list-slicing.'''
  86     for i0,w in enumerate(sent_nums[start:stop]):
  87         loc_w = i0+start
  88         yield (loc_w, w)
  89
  90 ###################################################
  91 #            P_INSIDE (dmv-specific)              #
  92 ###################################################
  93
  94 #@accepts(int, int, (int, int), int, Any(), [str], {tuple:float}, IsOneOf(None,{}))
  95 def inner(i, j, node, loc_h, g, sent, ichart={}, mpptree=None):
  96     ''' The ichart is of this form:
  97     ichart[i,j,LHS, loc_h]
  98     where i and j are between-word positions.
  99
 100     loc_h gives adjacency (along with k for attachment rules), and is
 101     needed in P_STOP reestimation.
 102     '''
 103     sent_nums = g.sent_nums(sent)
 104
 105     def terminal(i,j,node, loc_h, tabs):
 106         if not i <= loc_h < j:
 107             if 'INNER' in DEBUG:
 108                 print "%s*= 0.0 (wrong loc_h)" % tabs
 109             return 0.0
 110         elif POS(node) == sent_nums[i] and node in g.p_ORDER:
 111             # todo: add to ichart perhaps? Although, it _is_ simple lookup..
 112             prob = g.p_ORDER[node]
 113         else:
 114             if 'INNER' in DEBUG:
 115                 print "%sLACKING TERMINAL:" % tabs
 116             prob = 0.0
 117         if 'INNER' in DEBUG:
 118             print "%s*= %.4f (terminal: %s -> %s_%d)" % (tabs,prob, node_str(node), sent[i], loc_h)
 119         return prob
 120
 121     def e(i,j, (s_h,h), loc_h, n_t):
 122         def to_mpp(p, L, R):
 123             if mpptree:
 124                 key = (i,j, (s_h,h), loc_h)
 125                 if key not in mpptree:
 126                     mpptree[key] = (p, L, R)
 127                 elif mpptree[key][0] < p:
 128                     mpptree[key] = (p, L, R)
 129
 130         def tab():
 131             "Tabs for debug output"
 132             return "\t"*n_t
 133
 134         if (i, j, (s_h,h), loc_h) in ichart:
 135             if 'INNER' in DEBUG:
 136                 print "%s*= %.4f in ichart: i:%d j:%d node:%s loc:%s" % (tab(),ichart[i, j, (s_h,h), loc_h], i, j,
 137                                                                        node_str((s_h,h)), loc_h)
 138             return ichart[i, j, (s_h,h), loc_h]
 139         else:
 140             # Either terminal rewrites, using p_ORDER:
 141             if i+1 == j and (s_h == GOR or s_h == GOL):
 142                 return terminal(i, j, (s_h,h), loc_h, tab())
 143             else: # Or not at terminal level yet:
 144                 if 'INNER' in DEBUG:
 145                     print "%s%s (%.1f) from %d to %d" % (tab(),node_str((s_h,h)),loc_h,i,j)
 146                 if s_h == SEAL:
 147                     p_RGOL = g.p_STOP[h,  LEFT, adj(i,loc_h)] * e(i,j,(RGOL,h),loc_h,n_t+1)
 148                     p_LGOR = g.p_STOP[h, RIGHT, adj(j,loc_h)] * e(i,j,(LGOR,h),loc_h,n_t+1)
 149                     p = p_RGOL + p_LGOR
 150                     to_mpp(p_RGOL,       STOPKEY,         (i,j, (RGOL,h),loc_h))
 151                     to_mpp(p_LGOR, (i,j, (RGOL,h),loc_h),       STOPKEY )
 152                     if 'INNER' in DEBUG:
 153                         print "%sp= %.4f (STOP)" % (tab(), p)
 154                 elif s_h == RGOL or s_h == GOL:
 155                     p = 0.0
 156                     if s_h == RGOL:
 157                         p = g.p_STOP[h, RIGHT, adj(j,loc_h)] * e(i,j, (GOR,h),loc_h,n_t+1)
 158                         to_mpp(p, (i,j, (GOR,h),loc_h), STOPKEY)
 159                     for k in xgo_left(i, loc_h): # i < k <= loc_l(h)
 160                         p_R = e(k, j, ( s_h,h), loc_h, n_t+1)
 161                         if p_R > 0.0:
 162                             for loc_a,a in locs(sent_nums, i, k):
 163                                 p_ah = g.p_GO_AT_or0(a, h, LEFT, adj(k,loc_h))
 164                                 if p_ah > 0.0:
 165                                     p_L = e(i, k, (SEAL,a), loc_a, n_t+1)
 166                                     p_add = p_L * p_ah * p_R
 167                                     p += p_add
 168                                     to_mpp(p_add,
 169                                           (i, k, (SEAL,a), loc_a),
 170                                           (k, j, ( s_h,h), loc_h))
 171                     if 'INNER' in DEBUG:
 172                         print "%sp= %.4f (ATTACH)" % (tab(), p)
 173                 elif s_h == GOR or s_h == LGOR:
 174                     p = 0.0
 175                     if s_h == LGOR:
 176                         p = g.p_STOP[h, LEFT, adj(i,loc_h)] * e(i,j, (GOL,h),loc_h,n_t+1)
 177                         to_mpp(p, (i,j, (GOL,h),loc_h), STOPKEY)
 178                     for k in xgo_right(loc_h, j): # loc_l(h) < k < j
 179                         p_L = e(i, k, ( s_h,h), loc_h, n_t+1)
 180                         if p_L > 0.0:
 181                             for loc_a,a in locs(sent_nums,k,j):
 182                                 p_ah = g.p_GO_AT_or0(a, h, RIGHT, adj(k,loc_h))
 183                                 p_R = e(k, j, (SEAL,a), loc_a, n_t+1)
 184                                 p_add = p_L * p_ah * p_R
 185                                 p += p_add
 186                                 to_mpp(p_add,
 187                                       (i, k, ( s_h,h), loc_h),
 188                                       (k, j, (SEAL,a), loc_a))
 189
 190                     if 'INNER' in DEBUG:
 191                         print "%sp= %.4f (ATTACH)" % (tab(), p)
 192                 # elif s_h == GOL: # todo
 193
 194                 ichart[i, j, (s_h,h), loc_h] = p
 195                 return p
 196     # end of e-function
 197
 198     inner_prob = e(i,j,node,loc_h, 0)
 199     if 'INNER' in DEBUG:
 200         print debug_ichart(g,sent,ichart)
 201     return inner_prob
 202 # end of dmv.inner(i, j, node, loc_h, g, sent, ichart={})
 203
 204
 205 def debug_ichart(g,sent,ichart):
 206     str = "---ICHART:---\n"
 207     for (s,t,LHS,loc_h),v in ichart.iteritems():
 208         str += "%s -> %s_%d ... %s_%d (loc_h:%s):\t%.4f\n" % (node_str(LHS,g.numtag),
 209                                                               sent[s], s, sent[s], t, loc_h, v)
 210     str += "---ICHART:end---\n"
 211     return str
 212
 213
 214 def inner_sent(g, sent, ichart={}):
 215     return sum([g.p_ROOT[w] * inner(0, len(sent), (SEAL,w), loc_w, g, sent, ichart)
 216                 for loc_w,w in locs(g.sent_nums(sent),0,len(sent))])
 217
 218
 219
 220
 221
 222 ###################################################
 223 #           P_OUTSIDE (dmv-specific)              #
 224 ###################################################
 225
 226 #@accepts(int, int, (int, int), int, Any(), [str], {tuple:float}, {tuple:float})
 227 def outer(i,j,w_node,loc_w, g, sent, ichart={}, ochart={}):
 228     ''' http://www.student.uib.no/~kun041/dmvccm/DMVCCM.html#outer
 229
 230     w_node is a pair (seals,POS); the w in klein-thesis is made up of
 231     POS(w) and loc_w
 232     '''
 233     sent_nums = g.sent_nums(sent)
 234
 235     # local functions:
 236     def e(i,j,LHS,loc_h): # P_{INSIDE}
 237         try:
 238             return ichart[i,j,LHS,loc_h]
 239         except:
 240             return inner(i,j,LHS,loc_h,g,sent,ichart)
 241
 242     def f(i,j,w_node,loc_w):
 243         if (i,j,w_node,loc_w) in ochart:
 244             return ochart[i,j, w_node,loc_w]
 245         if w_node == ROOT:
 246             if i == 0 and j == len(sent):
 247                 return 1.0
 248             else: # ROOT may only be used on full sentence
 249                 return 0.0
 250         # but we may have non-ROOTs (stops) over full sentence too:
 251         w = POS(w_node)
 252         s_w = seals(w_node)
 253
 254         # todo: try either if p_M > 0.0: or sum(), and speed-test them
 255
 256         if s_w == SEAL: # w == a
 257             # todo: do the i<sent<j check here to save on calls?
 258             p = g.p_ROOT[w] * f(i,j,ROOT,loc_w)
 259             # left attach
 260             for k in xgt(j, sent): # j<k<len(sent)+1
 261                 for loc_h,h in locs(sent_nums,j,k):
 262                     p_wh = g.p_GO_AT_or0(w, h, LEFT, adj(j, loc_h))
 263                     for s_h in [RGOL, GOL]:
 264                         p += f(i,k,(s_h,h),loc_h) * p_wh * e(j,k,(s_h,h),loc_h)
 265             # right attach
 266             for k in xlt(i): # k<i
 267                 for loc_h,h in locs(sent_nums,k,i):
 268                     p_wh = g.p_GO_AT_or0(w, h, RIGHT, adj(i, loc_h))
 269                     for s_h in [LGOR, GOR]:
 270                         p += e(k,i,(s_h,h), loc_h) * p_wh * f(k,j,(s_h,h), loc_h)
 271
 272         elif s_w == RGOL or s_w == GOL: # w == h, left stop + left attach
 273             if s_w == RGOL:
 274                 s_h = SEAL
 275             else: # s_w == GOL
 276                 s_h = LGOR
 277             p = g.p_STOP[w, LEFT,  adj(i,loc_w)] * f(i,j,( s_h,w),loc_w)
 278             for k in xlt(i): # k<i
 279                 for loc_a,a in locs(sent_nums,k,i):
 280                     p_aw = g.p_GO_AT_or0(a, w, LEFT, adj(i, loc_w))
 281                     p += e(k,i, (SEAL,a),loc_a) * p_aw * f(k,j,w_node,loc_w)
 282
 283         elif s_w == GOR or s_w == LGOR: # w == h, right stop + right attach
 284             if s_w == GOR:
 285                 s_h = RGOL
 286             else: # s_w == LGOR
 287                 s_h = SEAL
 288             p = g.p_STOP[w, RIGHT, adj(j,loc_w)] * f(i,j,( s_h,w),loc_w)
 289             for k in xgt(j, sent): # j<k<len(sent)+1
 290                 for loc_a,a in locs(sent_nums,j,k):
 291                     p_ah = g.p_GO_AT_or0(a, w, RIGHT, adj(j, loc_w))
 292                     p += f(i,k,w_node,loc_w) * p_ah * e(j,k,(SEAL,a),loc_a)
 293
 294         ochart[i,j,w_node,loc_w] = p
 295         return p
 296     # end outer.f()
 297
 298     return f(i,j,w_node,loc_w)
 299 # end outer(i,j,w_node,loc_w, g,sent, ichart,ochart)
 300
 301
 302
 303
 304 ###################################################
 305 #                  Reestimation:                  #
 306 ###################################################
 307
 308 def reest_zeros(h_nums):
 309     '''A dict to hold numerators and denominators for our 6+ reestimation
 310 formulas. '''
 311     # todo: p_ORDER?
 312     fr = { ('ROOT','den'):0.0 } # holds sum over p_sent
 313     for h in h_nums:
 314         fr['ROOT','num',h] = 0.0
 315         for s_h in [GOR,GOL,RGOL,LGOR]:
 316             x = (s_h,h)
 317             fr['hat_a','den',x] = 0.0 # = c()
 318             # not all arguments are attached to, so we just initialize
 319             # fr['hat_a','num',a,(s_h,h)] as they show up, in reest_freq
 320             for adj in [NON, ADJ]:
 321                 for nd in ['num','den']:
 322                     fr['STOP',nd,x,adj] = 0.0
 323     return fr
 324
 325 def reest_freq(g, corpus):
 326     fr = reest_zeros(g.headnums())
 327     ichart = {}
 328     ochart = {}
 329     p_sent = None # 50 % speed increase on storing this locally
 330
 331     # local functions altogether 2x faster than global
 332     def c(i,j,LHS,loc_h,sent):
 333         if not p_sent > 0.0:
 334             return p_sent
 335
 336         p_in = e(i,j, LHS,loc_h,sent)
 337         if not p_in > 0.0:
 338             return p_in
 339
 340         p_out = f(i,j, LHS,loc_h,sent)
 341         return p_in * p_out / p_sent
 342     # end reest_freq.c()
 343
 344     def f(i,j,LHS,loc_h,sent): # P_{OUTSIDE}
 345         try:
 346             return ochart[i,j,LHS,loc_h]
 347         except:
 348             return outer(i,j,LHS,loc_h,g,sent,ichart,ochart)
 349     # end reest_freq.f()
 350
 351     def e(i,j,LHS,loc_h,sent): # P_{INSIDE}
 352         try:
 353             return ichart[i,j,LHS,loc_h]
 354         except:
 355             return inner(i,j,LHS,loc_h,g,sent,ichart)
 356     # end reest_freq.e()
 357
 358     def w_left(i,j, x,loc_h,sent,sent_nums):
 359         h = POS(x)
 360         if not p_sent > 0.0:
 361             return p_sent
 362
 363         for k in xtween(i, j):
 364             p_out = f(i,j, x,loc_h, sent)
 365             if not p_out > 0.0:
 366                 continue
 367             p_R =   e(k,j, x,loc_h, sent)
 368             if not p_R > 0.0:
 369                 continue
 370
 371             for loc_a,a in locs(sent_nums, i,k): # i<=loc_l(a)<k
 372                 p_rule = g.p_GO_AT_or0(a, h, LEFT, adj(k, loc_h))
 373                 p_L = e(i,k, (SEAL,a), loc_a, sent)
 374                 p = p_L * p_out * p_R * p_rule / p_sent
 375                 try:
 376                     fr['hat_a','num',a,x] += p
 377                 except:
 378                     fr['hat_a','num',a,x] = p
 379     # end reest_freq.w_left()
 380
 381     def w_right(i,j, x,loc_h,sent,sent_nums):
 382         h = POS(x)
 383         if not p_sent > 0.0:
 384             return p_sent
 385
 386         for k in xtween(i, j):
 387             p_out = f(i,j, x,loc_h, sent)
 388             if not p_out > 0.0:
 389                 continue
 390             p_L =   e(i,k, x,loc_h, sent)
 391             if not p_L > 0.0:
 392                 continue
 393
 394             for loc_a,a in locs(sent_nums, k,j): # k<=loc_l(a)<j
 395                 p_rule = g.p_GO_AT_or0(a, h, RIGHT, adj(k, loc_h))
 396                 p_R = e(k,j, (SEAL,a),loc_a, sent)
 397                 p = p_L * p_out * p_R * p_rule / p_sent
 398                 try:
 399                     fr['hat_a','num',a,x] += p
 400                 except:
 401                     fr['hat_a','num',a,x] = p
 402     # end reest_freq.w_right()
 403
 404     # in reest_freq:
 405     for sent in corpus:
 406         if 'REEST' in DEBUG:
 407             print sent
 408         ichart = {}
 409         ochart = {}
 410         p_sent = inner_sent(g, sent, ichart)
 411         fr['ROOT','den'] += p_sent
 412
 413         sent_nums = g.sent_nums(sent)
 414
 415         for loc_h,h in locs(sent_nums,0,len(sent)+1): # locs-stop is exclusive, thus +1
 416             # root:
 417             fr['ROOT','num',h] += g.p_ROOT[h] * e(0,len(sent), (SEAL,h),loc_h, sent)
 418
 419             loc_l_h = loc_h
 420             loc_r_h = loc_l_h+1
 421
 422             # left non-adjacent stop:
 423             for i in xlt(loc_l_h):
 424                 fr['STOP','num',(GOL,h),NON] += c(loc_l_h, j, (LGOR, h),loc_h, sent)
 425                 fr['STOP','den',(GOL,h),NON] += c(loc_l_h, j, (GOL, h),loc_h, sent)
 426                 for j in xgteq(loc_r_h, sent):
 427                     fr['STOP','num',(RGOL,h),NON] += c(i, j, (SEAL, h),loc_h, sent)
 428                     fr['STOP','den',(RGOL,h),NON] += c(i, j, (RGOL, h),loc_h, sent)
 429             # left adjacent stop, i = loc_l_h
 430             fr['STOP','num',(GOL,h),ADJ] += c(loc_l_h, loc_r_h, (LGOR, h),loc_h, sent)
 431             fr['STOP','den',(GOL,h),ADJ] += c(loc_l_h, loc_r_h, (GOL, h),loc_h, sent)
 432             for j in xgteq(loc_r_h, sent):
 433                 fr['STOP','num',(RGOL,h),ADJ] += c(loc_l_h, j, (SEAL, h),loc_h, sent)
 434                 fr['STOP','den',(RGOL,h),ADJ] += c(loc_l_h, j, (RGOL, h),loc_h, sent)
 435             # right non-adjacent stop:
 436             for j in xgt(loc_r_h, sent):
 437                 fr['STOP','num',(GOR,h),NON] += c(loc_l_h, j, (RGOL, h),loc_h, sent)
 438                 fr['STOP','den',(GOR,h),NON] += c(loc_l_h, j, (GOR,  h),loc_h, sent)
 439                 for i in xlteq(loc_l_h):
 440                     fr['STOP','num',(LGOR,h),NON] += c(loc_l_h, j, (SEAL, h),loc_h, sent)
 441                     fr['STOP','den',(LGOR,h),NON] += c(loc_l_h, j, (LGOR,  h),loc_h, sent)
 442             # right adjacent stop, j = loc_r_h
 443             fr['STOP','num',(GOR,h),ADJ] += c(loc_l_h, loc_r_h, (RGOL, h),loc_h, sent)
 444             fr['STOP','den',(GOR,h),ADJ] += c(loc_l_h, loc_r_h, (GOR,  h),loc_h, sent)
 445             for i in xlteq(loc_l_h):
 446                 fr['STOP','num',(LGOR,h),ADJ] += c(loc_l_h, j, (SEAL, h),loc_h, sent)
 447                 fr['STOP','den',(LGOR,h),ADJ] += c(loc_l_h, j, (LGOR,  h),loc_h, sent)
 448
 449             # left attachment:
 450             if 'REEST_ATTACH' in DEBUG:
 451                 print "Lattach %s: for i < %s"%(g.numtag(h),sent[0:loc_h+1])
 452             for s_h in [RGOL, GOL]:
 453                 x = (s_h, h)
 454                 for i in xlt(loc_l_h):  # i < loc_l(h)
 455                     if 'REEST_ATTACH' in DEBUG:
 456                         print "\tfor j >= %s"%sent[loc_h:len(sent)]
 457                     for j in xgteq(loc_r_h, sent): # j >= loc_r(h)
 458                         fr['hat_a','den',x] += c(i,j, x,loc_h, sent) # v_q in L&Y
 459                         if 'REEST_ATTACH' in DEBUG:
 460                             print "\t\tc( %d , %d, %s, %s, sent)=%.4f"%(i,j,node_str(x),loc_h,fr['hat_a','den',x])
 461                         w_left(i, j, x,loc_h, sent,sent_nums) # compute w for all a in sent
 462
 463             # right attachment:
 464             if 'REEST_ATTACH' in DEBUG:
 465                 print "Rattach %s: for i <= %s"%(g.numtag(h),sent[0:loc_h+1])
 466             for s_h in [GOR, LGOR]:
 467                 x = (s_h, h)
 468                 for i in xlteq(loc_l_h):  # i <= loc_l(h)
 469                     if 'REEST_ATTACH' in DEBUG:
 470                         print "\tfor j > %s"%sent[loc_h:len(sent)]
 471                     for j in xgt(loc_r_h, sent): # j > loc_r(h)
 472                         fr['hat_a','den',x] += c(i,j, x,loc_h, sent) # v_q in L&Y
 473                         if 'REEST_ATTACH' in DEBUG:
 474                             print "\t\tc( %d , %d, %s, %s, sent)=%.4f"%(loc_h,j,node_str(x),loc_h,fr['hat_a','den',x])
 475                         w_right(loc_l_h,j, x,loc_h, sent,sent_nums) # compute w for all a in sent
 476
 477         # end for loc_h,h
 478     # end for sent
 479
 480     return fr
 481
 482 def reestimate(g, corpus):
 483     fr = reest_freq(g, corpus)
 484     p_ROOT, p_STOP, p_ATTACH = {},{},{}
 485
 486     for h in g.headnums():
 487         reest_head(h, fr, g, p_ROOT, p_STOP, p_ATTACH)
 488
 489     g.p_STOP = p_STOP
 490     g.p_ATTACH = p_ATTACH
 491     g.p_GO_AT = make_GO_AT(p_STOP,p_ATTACH)
 492     g.p_ROOT = p_ROOT
 493     return fr
 494
 495
 496 def reest_head(h, fr, g, p_ROOT, p_STOP, p_ATTACH):
 497     "Given a single head, update g with the reestimated probability."
 498     # remove 0-prob stuff? todo
 499     try:
 500         p_ROOT[h] = fr['ROOT','num',h] / fr['ROOT','den']
 501     except:
 502         p_ROOT[h] = fr['ROOT','den']
 503
 504     for dir in [LEFT,RIGHT]:
 505         for adj in [ADJ, NON]: # p_STOP
 506             p_STOP[h, dir, adj] = 0.0
 507             for s_h in dirseal(dir):
 508                 x = (s_h,h)
 509                 p = fr['STOP','den', x, adj]
 510                 if p > 0.0:
 511                     p = fr['STOP', 'num', x, adj] / p
 512                 p_STOP[h, dir, adj] += p
 513
 514         for s_h in dirseal(dir): # make hat_a for p_ATTACH
 515             x = (s_h,h)
 516             hat_a = {}
 517
 518             p_c = fr['hat_a','den',x]
 519             for a in g.headnums():
 520                 try:
 521                     hat_a[a,x] = fr['hat_a','num',a,x] / p_c
 522                 except:
 523                     pass
 524
 525             sum_hat_a = sum([hat_a[w,x] for w in g.headnums()
 526                              if (w,x) in hat_a])
 527
 528             for a in g.headnums():
 529                 if (a,h,dir) not in p_ATTACH:
 530                     p_ATTACH[a,h,dir] = 0.0
 531                 try: # (a,x) might not be in hat_a
 532                     p_ATTACH[a,h,dir] += hat_a[a,x] / sum_hat_a
 533                 except:
 534                     pass
 535
 536     # todo ROOT!!!
 537
 538
 539
 540
 541
 542
 543 ###################################################
 544 #             Most Probable Parse:                #
 545 ###################################################
 546
 547 STOPKEY = (-1,-1,STOP,-1)
 548 ROOTKEY = (-1,-1,ROOT,-1)
 549
 550 def make_mpptree(g, sent):
 551     '''Tell inner() to make an mpptree, connect ROOT to this. (Logically,
 552     this should be part of inner_sent though...)'''
 553     ichart = {}
 554     mpptree = { ROOTKEY:(0.0, ROOTKEY, None) }
 555     for loc_w,w in locs(g.sent_nums(sent),0,len(sent)):
 556         p = g.p_ROOT[w] * inner(0, len(sent), (SEAL,w), loc_w, g, sent, ichart, mpptree)
 557         L = ROOTKEY
 558         R = (0,len(sent), (SEAL,w), loc_w)
 559         if mpptree[ROOTKEY][0] < p:
 560             mpptree[ROOTKEY] = (p, L, R)
 561     return mpptree
 562
 563 def parse_mpptree(mpptree, sent):
 564     '''mpptree is a dict of the form {k:(p,L,R),...}; where k, L and R
 565     are `keys' of the form (i,j,node,loc).
 566
 567     returns an mpp of the form [((head, loc_h),(arg, loc_a)), ...],
 568     where head and arg are tags.'''
 569     # local functions for clear access to mpptree:
 570     def k_node(key):
 571         return key[2]
 572     def k_POS(key):
 573         return POS(k_node(key))
 574     def k_seals(key):
 575         return seals(k_node(key))
 576     def k_locnode(key):
 577         return (k_node(key),key[3])
 578     def k_locPOS(key):
 579         return (k_POS(key),key[3])
 580     def k_terminal(key):
 581         s_k = k_seals(key) # i+1 == j
 582         return key[0] + 1 == key[1] and (s_k == GOR or s_k == GOL)
 583     def t_L(tree_entry):
 584         return tree_entry[1]
 585     def t_R(tree_entry):
 586         return tree_entry[2]
 587
 588     # arbitrarily, "ROOT attaches to right". We add it here to
 589     # avoid further complications:
 590     firstkey = t_R(mpptree[ROOTKEY])
 591     deps = set([ (k_locPOS(ROOTKEY), k_locPOS(firstkey), RIGHT) ])
 592     q = [firstkey]
 593
 594     while len(q) > 0:
 595         k = q.pop()
 596         if k_terminal(k):
 597             continue
 598         else:
 599             L = t_L( mpptree[k] )
 600             R = t_R( mpptree[k] )
 601             if k_locnode( k ) == k_locnode( L ): # Rattach
 602                 deps.add((k_locPOS( k ), k_locPOS( R ), LEFT))
 603                 q.extend( [L, R] )
 604             elif k_locnode( k ) == k_locnode( R ): # Lattach
 605                 deps.add((k_locPOS( k ), k_locPOS( L ), RIGHT))
 606                 q.extend( [L, R] )
 607             elif R == STOPKEY:
 608                 q.append( L )
 609             elif L == STOPKEY:
 610                 q.append( R )
 611     return deps
 612
 613 def mpp(g, sent):
 614     tagf = g.numtag # localized function, todo: speed-test
 615     mpptree = make_mpptree(g, sent)
 616     return set([((tagf(h), loc_h), (tagf(a), loc_a))
 617                 for (h, loc_h),(a,loc_a),dir in parse_mpptree(mpptree,sent)])
 618
 619
 620 ########################################################################
 621 #     testing functions:                                               #
 622 ########################################################################
 623
 624 testcorpus = [s.split() for s in ['det nn vbd c vbd','vbd nn c vbd',
 625                                   'det nn vbd',      'det nn vbd c pp',
 626                                   'det nn vbd',      'det vbd vbd c pp',
 627                                   'det nn vbd',      'det nn vbd c vbd',
 628                                   'det nn vbd',      'det nn vbd c vbd',
 629                                   'det nn vbd',      'det nn vbd c vbd',
 630                                   'det nn vbd',      'det nn vbd c pp',
 631                                   'det nn vbd pp',   'det nn vbd', ]]
 632
 633 def testgrammar():
 634     import loc_h_harmonic
 635     reload(loc_h_harmonic)
 636     return loc_h_harmonic.initialize(testcorpus)
 637
 638 def ig(s,t,LHS,loc_h):
 639     return inner(s,t,LHS,loc_h,testgrammar(),'det nn vbd'.split(),{})
 640
 641 def testreestimation():
 642     g = testgrammar()
 643     print g
 644 #    DEBUG.add('REEST_ATTACH')
 645     f = reestimate(g, testcorpus)
 646     print g
 647     testreestimation_regression(f)
 648     return f
 649
 650 def testreestimation_regression(fr):
 651     f_stops = {('STOP', 'den', (RGOL,3),NON): 12.212773236178391, ('STOP', 'den', (GOR,2),ADJ): 4.0, ('STOP', 'num', (GOR,4),NON): 2.5553487221351365, ('STOP', 'den', (RGOL,2),NON): 1.274904052793207, ('STOP', 'num', (RGOL,1),ADJ): 14.999999999999995, ('STOP', 'den', (GOR,3),ADJ): 15.0, ('STOP', 'num', (RGOL,4),ADJ): 16.65701084787457, ('STOP', 'num', (RGOL,0),ADJ): 4.1600647714443468, ('STOP', 'den', (RGOL,4),NON): 6.0170669155897105, ('STOP', 'num', (RGOL,3),ADJ): 2.7872267638216113, ('STOP', 'num', (RGOL,2),ADJ): 2.9723139990470515, ('STOP', 'den', (RGOL,2),ADJ): 4.0, ('STOP', 'den', (GOR,3),NON): 12.945787931730905, ('STOP', 'den', (RGOL,3),ADJ): 14.999999999999996, ('STOP', 'den', (GOR,2),NON): 0.0, ('STOP', 'den', (RGOL,0),ADJ): 8.0, ('STOP', 'num', (GOR,4),ADJ): 19.44465127786486, ('STOP', 'den', (GOR,1),NON): 3.1966410324085777, ('STOP', 'den', (RGOL,1),ADJ): 14.999999999999995, ('STOP', 'num', (GOR,3),ADJ): 4.1061665495365558, ('STOP', 'den', (GOR,0),NON): 4.8282499043902476, ('STOP', 'num', (RGOL,4),NON): 5.3429891521254289, ('STOP', 'num', (GOR,2),ADJ): 4.0, ('STOP', 'den', (RGOL,4),ADJ): 22.0, ('STOP', 'num', (GOR,1),ADJ): 12.400273895299103, ('STOP', 'num', (RGOL,2),NON): 1.0276860009529487, ('STOP', 'num', (GOR,0),ADJ): 3.1717500956097533, ('STOP', 'num', (RGOL,3),NON): 12.212773236178391, ('STOP', 'den', (GOR,4),ADJ): 22.0, ('STOP', 'den', (GOR,4),NON): 2.8705211946979836, ('STOP', 'num', (RGOL,0),NON): 3.8399352285556518, ('STOP', 'num', (RGOL,1),NON): 0.0, ('STOP', 'num', (GOR,0),NON): 4.8282499043902476, ('STOP', 'num', (GOR,1),NON): 2.5997261047008959, ('STOP', 'den', (RGOL,1),NON): 0.0, ('STOP', 'den', (GOR,0),ADJ): 8.0, ('STOP', 'num', (GOR,2),NON): 0.0, ('STOP', 'den', (RGOL,0),NON): 4.6540557322109795, ('STOP', 'den', (GOR,1),ADJ): 15.0, ('STOP', 'num', (GOR,3),NON): 10.893833450463443}
 652     for k,v in f_stops.iteritems():
 653         if not k in fr:
 654             print '''Regression in P_STOP reestimation, should be fr[%s]=%.4f,
 655 but %s not in fr'''%(k,v,k)
 656         elif not "%.10f"%fr[k] == "%.10f"%v:
 657             print '''Regression in P_STOP reestimation, should be fr[%s]=%.4f,
 658 got fr[%s]=%.4f.'''%(k,v,k,fr[k])
 659
 660 def testmpp_regression(mpptree,k_n):
 661     mpp = {ROOTKEY: (2.877072116829971e-05, STOPKEY, (0, 3, (2, 3), 1)),
 662  (0, 1, (1, 1), 0): (0.1111111111111111, (0, 1, (0, 1), 0), STOPKEY),
 663  (0, 1, (2, 1), 0): (0.049382716049382713, STOPKEY, (0, 1, (1, 1), 0)),
 664  (0, 3, (1, 3), 1): (0.00027619892321567721,
 665                        (0, 1, (2, 1), 0),
 666                        (1, 3, (1, 3), 1)),
 667  (0, 3, (2, 3), 1): (0.00012275507698474543, STOPKEY, (0, 3, (1, 3), 1)),
 668  (1, 3, (0, 3), 1): (0.025280986819448362,
 669                        (1, 2, (0, 3), 1),
 670                        (2, 3, (2, 4), 2)),
 671  (1, 3, (1, 3), 1): (0.0067415964851862296, (1, 3, (0, 3), 1), STOPKEY),
 672  (2, 3, (1, 4), 2): (0.32692307692307693, (2, 3, (0, 4), 2), STOPKEY),
 673  (2, 3, (2, 4), 2): (0.037721893491124266, STOPKEY, (2, 3, (1, 4), 2))}
 674     for k,(v,L,R) in mpp.iteritems():
 675         k2 = k[0:k_n] # 3 if the new does not check loc_h
 676         if type(k)==str:
 677             k2 = k
 678         if k2 not in mpptree:
 679             print "mpp regression, %s missing"%(k2,)
 680         else:
 681             vnew = mpptree[k2][0]
 682             if not "%.10f"%vnew == "%.10f"%v:
 683                 print "mpp regression, wanted %s=%.5f, got %.5f"%(k2,v,vnew)
 684
 685
 686 def testgrammar_a2():
 687     h, a = 0, 1
 688     p_ROOT, p_STOP, p_ATTACH, p_ORDER = {},{},{},{}
 689     p_ROOT[h] = 0.9
 690     p_ROOT[a] = 0.1
 691     p_STOP[h,LEFT,NON] = 1.0
 692     p_STOP[h,LEFT,ADJ] = 1.0
 693     p_STOP[h,RIGHT,NON] = 0.4 # RSTOP
 694     p_STOP[h,RIGHT,ADJ] = 0.3 # RSTOP
 695     p_STOP[a,LEFT,NON] = 1.0
 696     p_STOP[a,LEFT,ADJ] = 1.0
 697     p_STOP[a,RIGHT,NON] = 0.4 # RSTOP
 698     p_STOP[a,RIGHT,ADJ] = 0.3 # RSTOP
 699     p_ATTACH[a,h,LEFT] = 1.0  # not used
 700     p_ATTACH[a,h,RIGHT] = 1.0 # not used
 701     p_ATTACH[h,a,LEFT] = 1.0  # not used
 702     p_ATTACH[h,a,RIGHT] = 1.0 # not used
 703     p_ATTACH[h,h,LEFT] = 1.0  # not used
 704     p_ATTACH[h,h,RIGHT] = 1.0 # not used
 705     p_ORDER[(GOR, h)] = 1.0
 706     p_ORDER[(GOL, h)] = 0.0
 707     p_ORDER[(GOR, a)] = 1.0
 708     p_ORDER[(GOL, a)] = 0.0
 709     g = DMV_Grammar({h:'h',a:'a'}, {'h':h,'a':a}, p_ROOT, p_STOP, p_ATTACH, p_ORDER)
 710     # these probabilities are impossible so add them manually:
 711     g.p_GO_AT[a,a,LEFT,NON] = 0.4  # Lattach
 712     g.p_GO_AT[a,a,LEFT,ADJ] = 0.6  # Lattach
 713     g.p_GO_AT[h,a,LEFT,NON] = 0.2  # Lattach to h
 714     g.p_GO_AT[h,a,LEFT,ADJ] = 0.1  # Lattach to h
 715     g.p_GO_AT[a,a,RIGHT,NON] = 1.0 # Rattach
 716     g.p_GO_AT[a,a,RIGHT,ADJ] = 1.0 # Rattach
 717     g.p_GO_AT[h,a,RIGHT,NON] = 1.0 # Rattach to h
 718     g.p_GO_AT[h,a,RIGHT,ADJ] = 1.0 # Rattach to h
 719     g.p_GO_AT[h,h,LEFT,NON] = 0.2 # Lattach
 720     g.p_GO_AT[h,h,LEFT,ADJ] = 0.1 # Lattach
 721     g.p_GO_AT[a,h,LEFT,NON] = 0.4 # Lattach to a
 722     g.p_GO_AT[a,h,LEFT,ADJ] = 0.6 # Lattach to a
 723     g.p_GO_AT[h,h,RIGHT,NON] = 1.0 # Rattach
 724     g.p_GO_AT[h,h,RIGHT,ADJ] = 1.0 # Rattach
 725     g.p_GO_AT[a,h,RIGHT,NON] = 1.0 # Rattach to a
 726     g.p_GO_AT[a,h,RIGHT,ADJ] = 1.0 # Rattach to a
 727     return g
 728
 729
 730 def testgrammar_h2():
 731     h = 0
 732     p_ROOT, p_STOP, p_ATTACH, p_ORDER = {},{},{},{}
 733     p_ROOT[h] = 1.0
 734     p_STOP[h,LEFT,NON] = 1.0
 735     p_STOP[h,LEFT,ADJ] = 1.0
 736     p_STOP[h,RIGHT,NON] = 0.4
 737     p_STOP[h,RIGHT,ADJ] = 0.3
 738     p_ATTACH[h,h,LEFT] = 1.0  # not used
 739     p_ATTACH[h,h,RIGHT] = 1.0 # not used
 740     p_ORDER[(GOR, h)] = 1.0
 741     p_ORDER[(GOL, h)] = 0.0
 742     g = DMV_Grammar({h:'h'}, {'h':h}, p_ROOT, p_STOP, p_ATTACH, p_ORDER)
 743     g.p_GO_AT[h,h,LEFT,NON] = 0.6 # these probabilities are impossible
 744     g.p_GO_AT[h,h,LEFT,ADJ] = 0.7 # so add them manually...
 745     g.p_GO_AT[h,h,RIGHT,NON] = 1.0
 746     g.p_GO_AT[h,h,RIGHT,ADJ] = 1.0
 747     return g
 748
 749
 750
 751 def testreestimation_h():
 752     DEBUG.add('REEST')
 753     g = testgrammar_h2()
 754     reestimate(g,['h h h'.split()])
 755
 756
 757 def regression_tests():
 758     testmpp_regression(make_mpptree(testgrammar(), testcorpus[2]),4)
 759
 760     def test(wanted, got):
 761         if not wanted == got:
 762             print "Regression! Should be %s: %s" % (wanted, got)
 763
 764     g_dup2 = testgrammar_h2()
 765     h = 0
 766
 767     test("0.120",
 768          "%.3f" % inner(0, 2, (SEAL,h), 0, g_dup2, 'h h'.split(),{}))
 769
 770     test("0.063",
 771          "%.3f" % inner(0, 2, (SEAL,h), 1, g_dup2, 'h h'.split(),{}))
 772
 773     test("0.0498",
 774          "%.4f" % inner(0, 3, (SEAL,h), 2, g_dup2, 'h h h'.split(),{}))
 775
 776     test("0.58" ,
 777          "%.2f" % outer(1, 3, (RGOL,h), 2, g_dup2,'h h h'.split(),{},{}))
 778
 779     test("0.1089" ,
 780          "%.4f" % outer(0, 1, (GOR,h),  0,testgrammar_a2(),'h a'.split(),{},{}))
 781     test("0.3600" ,
 782          "%.4f" % outer(0, 2, (GOR,h),  0,testgrammar_a2(),'h a'.split(),{},{}))
 783     test("0.0000" ,
 784          "%.4f" % outer(0, 3, (GOR,h),  0,testgrammar_a2(),'h a'.split(),{},{}))
 785
 786     # todo: add more of these tests...
 787
 788 if __name__ == "__main__":
 789     DEBUG.clear()
 790
 791 #    import profile
 792 #    profile.run('testreestimation()')
 793
 794     import timeit
 795     print timeit.Timer("loc_h_dmv.testreestimation()",'''import loc_h_dmv
 796 reload(loc_h_dmv)''').timeit(1)
 797
 798     regression_tests()
 799
 800 #     print "mpp-test:"
 801     import pprint
 802     for s in testcorpus:
 803         print "sent:%s\nparse:set(\n%s)"%(s,pprint.pformat(list(mpp(testgrammar(), s)),
 804                                                     width=40))
 805
 806
 807 #     import pprint
 808 #     pprint.pprint( testreestimation())
 809
 810
 811
 812 def testIO():
 813     g = testgrammar()
 814     inners = [(sent, inner_sent(g, sent, {})) for sent in testcorpus]
 815     return inners
 816