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 not (i <= loc_w < j):
 244             return 0.0
 245         if (i,j,w_node,loc_w) in ochart:
 246             return ochart[i,j, w_node,loc_w]
 247         if w_node == ROOT:
 248             if i == 0 and j == len(sent):
 249                 return 1.0
 250             else: # ROOT may only be used on full sentence
 251                 return 0.0
 252         # but we may have non-ROOTs (stops) over full sentence too:
 253         w = POS(w_node)
 254         s_w = seals(w_node)
 255
 256         # todo: try either if p_M > 0.0: or sum(), and speed-test them
 257
 258         if s_w == SEAL: # w == a
 259             # todo: do the i<sent<j check here to save on calls?
 260             p = g.p_ROOT[w] * f(i,j,ROOT,loc_w)
 261             # left attach
 262             for k in xgt(j, sent): # j<k<len(sent)+1
 263                 for loc_h,h in locs(sent_nums,j,k):
 264                     p_wh = g.p_GO_AT_or0(w, h, LEFT, adj(j, loc_h))
 265                     for s_h in [RGOL, GOL]:
 266                         p += f(i,k,(s_h,h),loc_h) * p_wh * e(j,k,(s_h,h),loc_h)
 267             # right attach
 268             for k in xlt(i): # k<i
 269                 for loc_h,h in locs(sent_nums,k,i):
 270                     p_wh = g.p_GO_AT_or0(w, h, RIGHT, adj(i, loc_h))
 271                     for s_h in [LGOR, GOR]:
 272                         p += e(k,i,(s_h,h), loc_h) * p_wh * f(k,j,(s_h,h), loc_h)
 273
 274         elif s_w == RGOL or s_w == GOL: # w == h, left stop + left attach
 275             if s_w == RGOL:
 276                 s_h = SEAL
 277             else: # s_w == GOL
 278                 s_h = LGOR
 279             p = g.p_STOP[w, LEFT,  adj(i,loc_w)] * f(i,j,( s_h,w),loc_w)
 280             for k in xlt(i): # k<i
 281                 for loc_a,a in locs(sent_nums,k,i):
 282                     p_aw = g.p_GO_AT_or0(a, w, LEFT, adj(i, loc_w))
 283                     p += e(k,i, (SEAL,a),loc_a) * p_aw * f(k,j,w_node,loc_w)
 284
 285         elif s_w == GOR or s_w == LGOR: # w == h, right stop + right attach
 286             if s_w == GOR:
 287                 s_h = RGOL
 288             else: # s_w == LGOR
 289                 s_h = SEAL
 290             p = g.p_STOP[w, RIGHT, adj(j,loc_w)] * f(i,j,( s_h,w),loc_w)
 291             for k in xgt(j, sent): # j<k<len(sent)+1
 292                 for loc_a,a in locs(sent_nums,j,k):
 293                     p_ah = g.p_GO_AT_or0(a, w, RIGHT, adj(j, loc_w))
 294                     p += f(i,k,w_node,loc_w) * p_ah * e(j,k,(SEAL,a),loc_a)
 295
 296         ochart[i,j,w_node,loc_w] = p
 297         return p
 298     # end outer.f()
 299
 300     return f(i,j,w_node,loc_w)
 301 # end outer(i,j,w_node,loc_w, g,sent, ichart,ochart)
 302
 303
 304
 305
 306 ###################################################
 307 #                  Reestimation:                  #
 308 ###################################################
 309
 310 def reest_zeros(h_nums):
 311     '''A dict to hold numerators and denominators for our 6+ reestimation
 312 formulas. '''
 313     # todo: p_ORDER?
 314     fr = { ('ROOT','den'):0.0 } # holds sum over p_sent
 315     for h in h_nums:
 316         fr['ROOT','num',h] = 0.0
 317         for s_h in [GOR,GOL,RGOL,LGOR]:
 318             x = (s_h,h)
 319             fr['hat_a','den',x] = 0.0 # = c()
 320             # not all arguments are attached to, so we just initialize
 321             # fr['hat_a','num',a,(s_h,h)] as they show up, in reest_freq
 322             for adj in [NON, ADJ]:
 323                 for nd in ['num','den']:
 324                     fr['STOP',nd,x,adj] = 0.0
 325     return fr
 326
 327 def reest_freq(g, corpus):
 328     fr = reest_zeros(g.headnums())
 329     ichart = {}
 330     ochart = {}
 331     p_sent = None # 50 % speed increase on storing this locally
 332
 333     # local functions altogether 2x faster than global
 334     def c(i,j,LHS,loc_h,sent):
 335         if not p_sent > 0.0:
 336             return p_sent
 337
 338         p_in = e(i,j, LHS,loc_h,sent)
 339         if not p_in > 0.0:
 340             return p_in
 341
 342         p_out = f(i,j, LHS,loc_h,sent)
 343         return p_in * p_out / p_sent
 344     # end reest_freq.c()
 345
 346     def f(i,j,LHS,loc_h,sent): # P_{OUTSIDE}
 347         try:
 348             return ochart[i,j,LHS,loc_h]
 349         except:
 350             return outer(i,j,LHS,loc_h,g,sent,ichart,ochart)
 351     # end reest_freq.f()
 352
 353     def e(i,j,LHS,loc_h,sent): # P_{INSIDE}
 354         try:
 355             return ichart[i,j,LHS,loc_h]
 356         except:
 357             return inner(i,j,LHS,loc_h,g,sent,ichart)
 358     # end reest_freq.e()
 359
 360     def w_left(i,j, x,loc_h,sent,sent_nums):
 361         h = POS(x)
 362         if not p_sent > 0.0:
 363             return p_sent
 364
 365         for k in xtween(i, j):
 366             p_out = f(i,j, x,loc_h, sent)
 367             if not p_out > 0.0:
 368                 continue
 369             p_R =   e(k,j, x,loc_h, sent)
 370             if not p_R > 0.0:
 371                 continue
 372
 373             for loc_a,a in locs(sent_nums, i,k): # i<=loc_l(a)<k
 374                 p_rule = g.p_GO_AT_or0(a, h, LEFT, adj(k, loc_h))
 375                 p_L = e(i,k, (SEAL,a), loc_a, sent)
 376                 p = p_L * p_out * p_R * p_rule / p_sent
 377                 try:
 378                     fr['hat_a','num',a,x] += p
 379                 except:
 380                     fr['hat_a','num',a,x] = p
 381     # end reest_freq.w_left()
 382
 383     def w_right(i,j, x,loc_h,sent,sent_nums):
 384         h = POS(x)
 385         if not p_sent > 0.0:
 386             return p_sent
 387
 388         for k in xtween(i, j):
 389             p_out = f(i,j, x,loc_h, sent)
 390             if not p_out > 0.0:
 391                 continue
 392             p_L =   e(i,k, x,loc_h, sent)
 393             if not p_L > 0.0:
 394                 continue
 395
 396             for loc_a,a in locs(sent_nums, k,j): # k<=loc_l(a)<j
 397                 p_rule = g.p_GO_AT_or0(a, h, RIGHT, adj(k, loc_h))
 398                 p_R = e(k,j, (SEAL,a),loc_a, sent)
 399                 p = p_L * p_out * p_R * p_rule / p_sent
 400                 try:
 401                     fr['hat_a','num',a,x] += p
 402                 except:
 403                     fr['hat_a','num',a,x] = p
 404     # end reest_freq.w_right()
 405
 406     # in reest_freq:
 407     for sent in corpus:
 408         if 'REEST' in DEBUG:
 409             print sent
 410         ichart = {}
 411         ochart = {}
 412         p_sent = inner_sent(g, sent, ichart)
 413         fr['ROOT','den'] += p_sent
 414
 415         sent_nums = g.sent_nums(sent)
 416
 417         for loc_h,h in locs(sent_nums,0,len(sent)+1): # locs-stop is exclusive, thus +1
 418             # root:
 419             fr['ROOT','num',h] += g.p_ROOT[h] * e(0,len(sent), (SEAL,h),loc_h, sent)
 420
 421             loc_l_h = loc_h
 422             loc_r_h = loc_l_h+1
 423
 424             # left non-adjacent stop:
 425             for i in xlt(loc_l_h):
 426                 fr['STOP','num',(GOL,h),NON] += c(loc_l_h, j, (LGOR, h),loc_h, sent)
 427                 fr['STOP','den',(GOL,h),NON] += c(loc_l_h, j, (GOL, h),loc_h, sent)
 428                 for j in xgteq(loc_r_h, sent):
 429                     fr['STOP','num',(RGOL,h),NON] += c(i, j, (SEAL, h),loc_h, sent)
 430                     fr['STOP','den',(RGOL,h),NON] += c(i, j, (RGOL, h),loc_h, sent)
 431             # left adjacent stop, i = loc_l_h
 432             fr['STOP','num',(GOL,h),ADJ] += c(loc_l_h, loc_r_h, (LGOR, h),loc_h, sent)
 433             fr['STOP','den',(GOL,h),ADJ] += c(loc_l_h, loc_r_h, (GOL, h),loc_h, sent)
 434             for j in xgteq(loc_r_h, sent):
 435                 fr['STOP','num',(RGOL,h),ADJ] += c(loc_l_h, j, (SEAL, h),loc_h, sent)
 436                 fr['STOP','den',(RGOL,h),ADJ] += c(loc_l_h, j, (RGOL, h),loc_h, sent)
 437             # right non-adjacent stop:
 438             for j in xgt(loc_r_h, sent):
 439                 fr['STOP','num',(GOR,h),NON] += c(loc_l_h, j, (RGOL, h),loc_h, sent)
 440                 fr['STOP','den',(GOR,h),NON] += c(loc_l_h, j, (GOR,  h),loc_h, sent)
 441                 for i in xlteq(loc_l_h):
 442                     fr['STOP','num',(LGOR,h),NON] += c(loc_l_h, j, (SEAL, h),loc_h, sent)
 443                     fr['STOP','den',(LGOR,h),NON] += c(loc_l_h, j, (LGOR,  h),loc_h, sent)
 444             # right adjacent stop, j = loc_r_h
 445             fr['STOP','num',(GOR,h),ADJ] += c(loc_l_h, loc_r_h, (RGOL, h),loc_h, sent)
 446             fr['STOP','den',(GOR,h),ADJ] += c(loc_l_h, loc_r_h, (GOR,  h),loc_h, sent)
 447             for i in xlteq(loc_l_h):
 448                 fr['STOP','num',(LGOR,h),ADJ] += c(loc_l_h, j, (SEAL, h),loc_h, sent)
 449                 fr['STOP','den',(LGOR,h),ADJ] += c(loc_l_h, j, (LGOR,  h),loc_h, sent)
 450
 451             # left attachment:
 452             if 'REEST_ATTACH' in DEBUG:
 453                 print "Lattach %s: for i < %s"%(g.numtag(h),sent[0:loc_h+1])
 454             for s_h in [RGOL, GOL]:
 455                 x = (s_h, h)
 456                 for i in xlt(loc_l_h):  # i < loc_l(h)
 457                     if 'REEST_ATTACH' in DEBUG:
 458                         print "\tfor j >= %s"%sent[loc_h:len(sent)]
 459                     for j in xgteq(loc_r_h, sent): # j >= loc_r(h)
 460                         fr['hat_a','den',x] += c(i,j, x,loc_h, sent) # v_q in L&Y
 461                         if 'REEST_ATTACH' in DEBUG:
 462                             print "\t\tc( %d , %d, %s, %s, sent)=%.4f"%(i,j,node_str(x),loc_h,fr['hat_a','den',x])
 463                         w_left(i, j, x,loc_h, sent,sent_nums) # compute w for all a in sent
 464
 465             # right attachment:
 466             if 'REEST_ATTACH' in DEBUG:
 467                 print "Rattach %s: for i <= %s"%(g.numtag(h),sent[0:loc_h+1])
 468             for s_h in [GOR, LGOR]:
 469                 x = (s_h, h)
 470                 for i in xlteq(loc_l_h):  # i <= loc_l(h)
 471                     if 'REEST_ATTACH' in DEBUG:
 472                         print "\tfor j > %s"%sent[loc_h:len(sent)]
 473                     for j in xgt(loc_r_h, sent): # j > loc_r(h)
 474                         fr['hat_a','den',x] += c(i,j, x,loc_h, sent) # v_q in L&Y
 475                         if 'REEST_ATTACH' in DEBUG:
 476                             print "\t\tc( %d , %d, %s, %s, sent)=%.4f"%(loc_h,j,node_str(x),loc_h,fr['hat_a','den',x])
 477                         w_right(loc_l_h,j, x,loc_h, sent,sent_nums) # compute w for all a in sent
 478
 479         # end for loc_h,h
 480     # end for sent
 481
 482     return fr
 483
 484 def reestimate(g, corpus):
 485     fr = reest_freq(g, corpus)
 486     p_ROOT, p_STOP, p_ATTACH = {},{},{}
 487
 488     for h in g.headnums():
 489         reest_head(h, fr, g, p_ROOT, p_STOP, p_ATTACH)
 490
 491     g.p_STOP = p_STOP
 492     g.p_ATTACH = p_ATTACH
 493     g.p_GO_AT = make_GO_AT(p_STOP,p_ATTACH)
 494     g.p_ROOT = p_ROOT
 495     return fr
 496
 497
 498 def reest_head(h, fr, g, p_ROOT, p_STOP, p_ATTACH):
 499     "Given a single head, update g with the reestimated probability."
 500     # remove 0-prob stuff? todo
 501     try:
 502         p_ROOT[h] = fr['ROOT','num',h] / fr['ROOT','den']
 503     except:
 504         p_ROOT[h] = fr['ROOT','den']
 505
 506     for dir in [LEFT,RIGHT]:
 507         for adj in [ADJ, NON]: # p_STOP
 508             p_STOP[h, dir, adj] = 0.0
 509             for s_h in dirseal(dir):
 510                 x = (s_h,h)
 511                 p = fr['STOP','den', x, adj]
 512                 if p > 0.0:
 513                     p = fr['STOP', 'num', x, adj] / p
 514                 p_STOP[h, dir, adj] += p
 515
 516         for s_h in dirseal(dir): # make hat_a for p_ATTACH
 517             x = (s_h,h)
 518             hat_a = {}
 519
 520             p_c = fr['hat_a','den',x]
 521             for a in g.headnums():
 522                 try:
 523                     hat_a[a,x] = fr['hat_a','num',a,x] / p_c
 524                 except:
 525                     pass
 526
 527             sum_hat_a = sum([hat_a[w,x] for w in g.headnums()
 528                              if (w,x) in hat_a])
 529
 530             for a in g.headnums():
 531                 if (a,h,dir) not in p_ATTACH:
 532                     p_ATTACH[a,h,dir] = 0.0
 533                 try: # (a,x) might not be in hat_a
 534                     p_ATTACH[a,h,dir] += hat_a[a,x] / sum_hat_a
 535                 except:
 536                     pass
 537
 538
 539
 540
 541
 542
 543
 544
 545 ###################################################
 546 #             Most Probable Parse:                #
 547 ###################################################
 548
 549 STOPKEY = (-1,-1,STOP,-1)
 550 ROOTKEY = (-1,-1,ROOT,-1)
 551
 552 def make_mpptree(g, sent):
 553     '''Tell inner() to make an mpptree, connect ROOT to this. (Logically,
 554     this should be part of inner_sent though...)'''
 555     ichart = {}
 556     mpptree = { ROOTKEY:(0.0, ROOTKEY, None) }
 557     for loc_w,w in locs(g.sent_nums(sent),0,len(sent)):
 558         p = g.p_ROOT[w] * inner(0, len(sent), (SEAL,w), loc_w, g, sent, ichart, mpptree)
 559         L = ROOTKEY
 560         R = (0,len(sent), (SEAL,w), loc_w)
 561         if mpptree[ROOTKEY][0] < p:
 562             mpptree[ROOTKEY] = (p, L, R)
 563     return mpptree
 564
 565 def parse_mpptree(mpptree, sent):
 566     '''mpptree is a dict of the form {k:(p,L,R),...}; where k, L and R
 567     are `keys' of the form (i,j,node,loc).
 568
 569     returns an mpp of the form [((head, loc_h),(arg, loc_a)), ...],
 570     where head and arg are tags.'''
 571     # local functions for clear access to mpptree:
 572     def k_node(key):
 573         return key[2]
 574     def k_POS(key):
 575         return POS(k_node(key))
 576     def k_seals(key):
 577         return seals(k_node(key))
 578     def k_locnode(key):
 579         return (k_node(key),key[3])
 580     def k_locPOS(key):
 581         return (k_POS(key),key[3])
 582     def k_terminal(key):
 583         s_k = k_seals(key) # i+1 == j
 584         return key[0] + 1 == key[1] and (s_k == GOR or s_k == GOL)
 585     def t_L(tree_entry):
 586         return tree_entry[1]
 587     def t_R(tree_entry):
 588         return tree_entry[2]
 589
 590     # arbitrarily, "ROOT attaches to right". We add it here to
 591     # avoid further complications:
 592     firstkey = t_R(mpptree[ROOTKEY])
 593     deps = set([ (k_locPOS(ROOTKEY), k_locPOS(firstkey), RIGHT) ])
 594     q = [firstkey]
 595
 596     while len(q) > 0:
 597         k = q.pop()
 598         if k_terminal(k):
 599             continue
 600         else:
 601             L = t_L( mpptree[k] )
 602             R = t_R( mpptree[k] )
 603             if k_locnode( k ) == k_locnode( L ): # Rattach
 604                 deps.add((k_locPOS( k ), k_locPOS( R ), LEFT))
 605                 q.extend( [L, R] )
 606             elif k_locnode( k ) == k_locnode( R ): # Lattach
 607                 deps.add((k_locPOS( k ), k_locPOS( L ), RIGHT))
 608                 q.extend( [L, R] )
 609             elif R == STOPKEY:
 610                 q.append( L )
 611             elif L == STOPKEY:
 612                 q.append( R )
 613     return deps
 614
 615 def mpp(g, sent):
 616     tagf = g.numtag # localized function, todo: speed-test
 617     mpptree = make_mpptree(g, sent)
 618     return set([((tagf(h), loc_h), (tagf(a), loc_a))
 619                 for (h, loc_h),(a,loc_a),dir in parse_mpptree(mpptree,sent)])
 620
 621
 622 ########################################################################
 623 #     testing functions:                                               #
 624 ########################################################################
 625
 626 testcorpus = [s.split() for s in ['det nn vbd c vbd','vbd nn c vbd',
 627                                   'det nn vbd',      'det nn vbd c pp',
 628                                   'det nn vbd',      'det vbd vbd c pp',
 629                                   'det nn vbd',      'det nn vbd c vbd',
 630                                   'det nn vbd',      'det nn vbd c vbd',
 631                                   'det nn vbd',      'det nn vbd c vbd',
 632                                   'det nn vbd',      'det nn vbd c pp',
 633                                   'det nn vbd pp',   'det nn vbd', ]]
 634
 635 def testgrammar():
 636     import loc_h_harmonic
 637     reload(loc_h_harmonic)
 638
 639     # make sure these are the way they were when setting up the tests:
 640     loc_h_harmonic.HARMONIC_C = 0.0
 641     loc_h_harmonic.FNONSTOP_MIN = 25
 642     loc_h_harmonic.FSTOP_MIN = 5
 643     loc_h_harmonic.RIGHT_FIRST = 1.0
 644
 645     return loc_h_harmonic.initialize(testcorpus)
 646
 647 def ig(s,t,LHS,loc_h):
 648     return inner(s,t,LHS,loc_h,testgrammar(),'det nn vbd'.split(),{})
 649
 650 def testreestimation():
 651     g = testgrammar()
 652     print g
 653 #    DEBUG.add('REEST_ATTACH')
 654     f = reestimate(g, testcorpus)
 655     print g
 656     testreestimation_regression(f)
 657     return f
 658
 659 def testreestimation_regression(fr):
 660     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}
 661     for k,v in f_stops.iteritems():
 662         if not k in fr:
 663             print '''Regression in P_STOP reestimation, should be fr[%s]=%.4f,
 664 but %s not in fr'''%(k,v,k)
 665         elif not "%.10f"%fr[k] == "%.10f"%v:
 666             print '''Regression in P_STOP reestimation, should be fr[%s]=%.4f,
 667 got fr[%s]=%.4f.'''%(k,v,k,fr[k])
 668
 669 def testmpp_regression(mpptree,k_n):
 670     mpp = {ROOTKEY: (2.877072116829971e-05, STOPKEY, (0, 3, (2, 3), 1)),
 671  (0, 1, (1, 1), 0): (0.1111111111111111, (0, 1, (0, 1), 0), STOPKEY),
 672  (0, 1, (2, 1), 0): (0.049382716049382713, STOPKEY, (0, 1, (1, 1), 0)),
 673  (0, 3, (1, 3), 1): (0.00027619892321567721,
 674                        (0, 1, (2, 1), 0),
 675                        (1, 3, (1, 3), 1)),
 676  (0, 3, (2, 3), 1): (0.00012275507698474543, STOPKEY, (0, 3, (1, 3), 1)),
 677  (1, 3, (0, 3), 1): (0.025280986819448362,
 678                        (1, 2, (0, 3), 1),
 679                        (2, 3, (2, 4), 2)),
 680  (1, 3, (1, 3), 1): (0.0067415964851862296, (1, 3, (0, 3), 1), STOPKEY),
 681  (2, 3, (1, 4), 2): (0.32692307692307693, (2, 3, (0, 4), 2), STOPKEY),
 682  (2, 3, (2, 4), 2): (0.037721893491124266, STOPKEY, (2, 3, (1, 4), 2))}
 683     for k,(v,L,R) in mpp.iteritems():
 684         k2 = k[0:k_n] # 3 if the new does not check loc_h
 685         if type(k)==str:
 686             k2 = k
 687         if k2 not in mpptree:
 688             print "mpp regression, %s missing"%(k2,)
 689         else:
 690             vnew = mpptree[k2][0]
 691             if not "%.10f"%vnew == "%.10f"%v:
 692                 print "mpp regression, wanted %s=%.5f, got %.5f"%(k2,v,vnew)
 693
 694
 695 def testgrammar_a():
 696     h, a = 0, 1
 697     p_ROOT, p_STOP, p_ATTACH, p_ORDER = {},{},{},{}
 698     p_ROOT[h] = 0.9
 699     p_ROOT[a] = 0.1
 700     p_STOP[h,LEFT,NON] = 1.0
 701     p_STOP[h,LEFT,ADJ] = 1.0
 702     p_STOP[h,RIGHT,NON] = 0.4 # RSTOP
 703     p_STOP[h,RIGHT,ADJ] = 0.3 # RSTOP
 704     p_STOP[a,LEFT,NON] = 1.0
 705     p_STOP[a,LEFT,ADJ] = 1.0
 706     p_STOP[a,RIGHT,NON] = 0.4 # RSTOP
 707     p_STOP[a,RIGHT,ADJ] = 0.3 # RSTOP
 708     p_ATTACH[a,h,LEFT] = 1.0  # not used
 709     p_ATTACH[a,h,RIGHT] = 1.0 # not used
 710     p_ATTACH[h,a,LEFT] = 1.0  # not used
 711     p_ATTACH[h,a,RIGHT] = 1.0 # not used
 712     p_ATTACH[h,h,LEFT] = 1.0  # not used
 713     p_ATTACH[h,h,RIGHT] = 1.0 # not used
 714     p_ORDER[(GOR, h)] = 1.0
 715     p_ORDER[(GOL, h)] = 0.0
 716     p_ORDER[(GOR, a)] = 1.0
 717     p_ORDER[(GOL, a)] = 0.0
 718     g = DMV_Grammar({h:'h',a:'a'}, {'h':h,'a':a}, p_ROOT, p_STOP, p_ATTACH, p_ORDER)
 719     # these probabilities are impossible so add them manually:
 720     g.p_GO_AT[a,a,LEFT,NON] = 0.4  # Lattach
 721     g.p_GO_AT[a,a,LEFT,ADJ] = 0.6  # Lattach
 722     g.p_GO_AT[h,a,LEFT,NON] = 0.2  # Lattach to h
 723     g.p_GO_AT[h,a,LEFT,ADJ] = 0.1  # Lattach to h
 724     g.p_GO_AT[a,a,RIGHT,NON] = 1.0 # Rattach
 725     g.p_GO_AT[a,a,RIGHT,ADJ] = 1.0 # Rattach
 726     g.p_GO_AT[h,a,RIGHT,NON] = 1.0 # Rattach to h
 727     g.p_GO_AT[h,a,RIGHT,ADJ] = 1.0 # Rattach to h
 728     g.p_GO_AT[h,h,LEFT,NON] = 0.2 # Lattach
 729     g.p_GO_AT[h,h,LEFT,ADJ] = 0.1 # Lattach
 730     g.p_GO_AT[a,h,LEFT,NON] = 0.4 # Lattach to a
 731     g.p_GO_AT[a,h,LEFT,ADJ] = 0.6 # Lattach to a
 732     g.p_GO_AT[h,h,RIGHT,NON] = 1.0 # Rattach
 733     g.p_GO_AT[h,h,RIGHT,ADJ] = 1.0 # Rattach
 734     g.p_GO_AT[a,h,RIGHT,NON] = 1.0 # Rattach to a
 735     g.p_GO_AT[a,h,RIGHT,ADJ] = 1.0 # Rattach to a
 736     return g
 737
 738
 739 def testgrammar_h():
 740     h = 0
 741     p_ROOT, p_STOP, p_ATTACH, p_ORDER = {},{},{},{}
 742     p_ROOT[h] = 1.0
 743     p_STOP[h,LEFT,NON] = 1.0
 744     p_STOP[h,LEFT,ADJ] = 1.0
 745     p_STOP[h,RIGHT,NON] = 0.4
 746     p_STOP[h,RIGHT,ADJ] = 0.3
 747     p_ATTACH[h,h,LEFT] = 1.0  # not used
 748     p_ATTACH[h,h,RIGHT] = 1.0 # not used
 749     p_ORDER[(GOR, h)] = 1.0
 750     p_ORDER[(GOL, h)] = 0.0
 751     g = DMV_Grammar({h:'h'}, {'h':h}, p_ROOT, p_STOP, p_ATTACH, p_ORDER)
 752     g.p_GO_AT[h,h,LEFT,NON] = 0.6 # these probabilities are impossible
 753     g.p_GO_AT[h,h,LEFT,ADJ] = 0.7 # so add them manually...
 754     g.p_GO_AT[h,h,RIGHT,NON] = 1.0
 755     g.p_GO_AT[h,h,RIGHT,ADJ] = 1.0
 756     return g
 757
 758
 759
 760 def testreestimation_h():
 761     DEBUG.add('REEST')
 762     g = testgrammar_h()
 763     reestimate(g,['h h h'.split()])
 764
 765
 766 def test(wanted, got):
 767     if not wanted == got:
 768         raise Warning, "Regression! Should be %s: %s" % (wanted, got)
 769
 770 def regression_tests():
 771     testmpp_regression(make_mpptree(testgrammar(), testcorpus[2]),4)
 772     h = 0
 773
 774     test("0.120",
 775          "%.3f" % inner(0, 2, (SEAL,h), 0, testgrammar_h(), 'h h'.split(),{}))
 776     test("0.063",
 777          "%.3f" % inner(0, 2, (SEAL,h), 1, testgrammar_h(), 'h h'.split(),{}))
 778
 779     test("0.1092",
 780          "%.4f" % inner(0, 3, (SEAL,0), 0, testgrammar_h(), 'h h h'.split(),{}))
 781     test("0.0252",
 782          "%.4f" % inner(0, 3, (SEAL,0), 1, testgrammar_h(), 'h h h'.split(),{}))
 783     test("0.0498",
 784          "%.4f" % inner(0, 3, (SEAL,h), 2, testgrammar_h(), 'h h h'.split(),{}))
 785
 786     test("0.58" ,
 787          "%.2f" % outer(1, 3, (RGOL,h), 2, testgrammar_h(),'h h h'.split(),{},{}))
 788     test("0.61" , # ftw? can't be right... there's an 0.4 shared between these two...
 789          "%.2f" % outer(1, 3, (RGOL,h), 1, testgrammar_h(),'h h h'.split(),{},{}))
 790
 791     test("0.00" ,
 792          "%.2f" % outer(1, 3, (RGOL,h), 0, testgrammar_h(),'h h h'.split(),{},{}))
 793     test("0.00" ,
 794          "%.2f" % outer(1, 3, (RGOL,h), 3, testgrammar_h(),'h h h'.split(),{},{}))
 795
 796     test("0.1089" ,
 797          "%.4f" % outer(0, 1, (GOR,h),  0,testgrammar_a(),'h a'.split(),{},{}))
 798     test("0.3600" ,
 799          "%.4f" % outer(0, 2, (GOR,h),  0,testgrammar_a(),'h a'.split(),{},{}))
 800     test("0.0000" ,
 801          "%.4f" % outer(0, 3, (GOR,h),  0,testgrammar_a(),'h a'.split(),{},{}))
 802
 803     # todo: add more of these tests...
 804
 805 if __name__ == "__main__":
 806     DEBUG.clear()
 807
 808 #    import profile
 809 #    profile.run('testreestimation()')
 810
 811 #     import timeit
 812 #     print timeit.Timer("loc_h_dmv.testreestimation()",'''import loc_h_dmv
 813 # reload(loc_h_dmv)''').timeit(1)
 814
 815     regression_tests()
 816
 817 #     print "mpp-test:"
 818 #     import pprint
 819 #     for s in testcorpus:
 820 #         print "sent:%s\nparse:set(\n%s)"%(s,pprint.pformat(list(mpp(testgrammar(), s)),
 821 #                                                     width=40))
 822
 823
 824 #     import pprint
 825 #     pprint.pprint( testreestimation())
 826
 827
 828
 829 def testIO():
 830     g = testgrammar()
 831     inners = [(sent, inner_sent(g, sent, {})) for sent in testcorpus]
 832     return inners
 833