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