src/loc_h_dmv.py

   1 # loc_h_dmv.py
   2 #
   3 # dmv reestimation and inside-outside probabilities using loc_h, and
   4 # no CNF-style rules
   5 #
   6 # Table of Contents:
   7 # 1. Grammar-class and related functions
   8 # 2. P_INSIDE / inner() and inner_sent()
   9 # 3. P_OUTSIDE / outer()
  10 # 4. Reestimation v.1: sentences as outer loop
  11 # 5. Reestimation v.2: head-types as outer loop
  12 # 6. Most Probable Parse
  13 # 7. Testing functions
  14
  15 import io
  16 from common_dmv import *
  17
  18 ### todo: debug with @accepts once in a while, but it's SLOW
  19 # from typecheck import accepts, Any
  20
  21 if __name__ == "__main__":
  22     print "loc_h_dmv module tests:"
  23
  24 def adj(middle, loc_h):
  25     "middle is eg. k when rewriting for i<k<j (inside probabilities)."
  26     return middle == loc_h or middle == loc_h+1 # ADJ == True
  27
  28 def make_GO_AT(p_STOP,p_ATTACH):
  29     p_GO_AT = {}
  30     for (a,h,dir), p_ah in p_ATTACH.iteritems():
  31         p_GO_AT[a,h,dir, NON] = p_ah * (1-p_STOP[h, dir, NON])
  32         p_GO_AT[a,h,dir, ADJ] = p_ah * (1-p_STOP[h, dir, ADJ])
  33     return p_GO_AT
  34
  35 class DMV_Grammar(io.Grammar):
  36     def __str__(self):
  37         LJUST = 47
  38         def t(n):
  39             return "%d=%s" % (n, self.numtag(n))
  40         def p(dict,key):
  41             if key in dict:
  42                 if dict[key] > 1.0: raise Exception, "probability > 1.0:%s"%key
  43                 return dict[key]
  44             else: return 0.0
  45         def no_zeroL(str,tagstr,prob):
  46             if prob > 0.0: return (str%(tagstr,prob)).ljust(LJUST)
  47             else: return "".ljust(LJUST)
  48         def no_zeroR(str,tagstr,prob):
  49             if prob > 0.0: return str%(tagstr,prob)
  50             else: return ""
  51         def p_a(a,h):
  52             p_L = p(self.p_ATTACH,(a,h,LEFT))
  53             p_R = p(self.p_ATTACH,(a,h,RIGHT))
  54             if p_L == 0.0 and p_R == 0.0:
  55                 return ''
  56             else:
  57                 if p_L > 0.0:
  58                     str = "p_ATTACH[%s|%s,L] = %s" % (t(a), t(h), p_L)
  59                     str = str.ljust(LJUST)
  60                 else:
  61                     str = ''
  62                 if p_R > 0.0:
  63                     str = str.ljust(LJUST)
  64                     str += "p_ATTACH[%s|%s,R] = %s" % (t(a), t(h), p_R)
  65             return '\n'+str
  66
  67         root, stop, att, ord = "","","",""
  68         for h in self.headnums():
  69             root += no_zeroL("\np_ROOT[%s] = %s", t(h), p(self.p_ROOT, (h)))
  70             stop += '\n'
  71             stop += no_zeroL("p_STOP[stop|%s,L,adj] = %s", t(h), p(self.p_STOP, (h,LEFT,ADJ)))
  72             stop += no_zeroR("p_STOP[stop|%s,R,adj] = %s", t(h), p(self.p_STOP, (h,RIGHT,ADJ)))
  73             stop += '\n'
  74             stop += no_zeroL("p_STOP[stop|%s,L,non] = %s", t(h), p(self.p_STOP, (h,LEFT,NON)))
  75             stop += no_zeroR("p_STOP[stop|%s,R,non] = %s", t(h), p(self.p_STOP, (h,RIGHT,NON)))
  76             att  += ''.join([p_a(a,h) for a in self.headnums()])
  77             ord  += '\n'
  78             ord  += no_zeroL("p_ORDER[ left-first|%s ] = %s", t(h), p(self.p_ORDER, (GOL,h)))
  79             ord  += no_zeroR("p_ORDER[right-first|%s ] = %s", t(h), p(self.p_ORDER, (GOR,h)))
  80         return root + stop + att + ord
  81
  82     def __init__(self, numtag, tagnum, p_ROOT, p_STOP, p_ATTACH, p_ORDER):
  83         io.Grammar.__init__(self, numtag, tagnum)
  84         self.p_ROOT = p_ROOT     # p_ROOT[w] = p
  85         self.p_ORDER = p_ORDER   # p_ORDER[seals, w] = p
  86         self.p_STOP = p_STOP     # p_STOP[w, LEFT, NON] = p     (etc. for LA,RN,RA)
  87         self.p_ATTACH = p_ATTACH # p_ATTACH[a, h, LEFT] = p (etc. for R)
  88                                  # p_GO_AT[a, h, LEFT, NON] = p (etc. for LA,RN,RA)
  89         self.p_GO_AT = make_GO_AT(self.p_STOP, self.p_ATTACH)
  90         # these are used in reestimate2():
  91         self.reset_iocharts()
  92
  93     def get_iochart(self, sent_nums):
  94         ch_key = tuple(sent_nums)
  95         try:
  96             ichart = self._icharts[ch_key]
  97         except KeyError:
  98             ichart = {}
  99         try:
 100             ochart = self._ocharts[ch_key]
 101         except KeyError:
 102             ochart = {}
 103         return (ichart, ochart)
 104
 105     def set_iochart(self, sent_nums, ichart, ochart):
 106         self._icharts[tuple(sent_nums)] = ichart
 107         self._ocharts[tuple(sent_nums)] = ochart
 108
 109     def reset_iocharts(self):
 110         self._icharts = {}
 111         self._ocharts = {}
 112
 113     def p_GO_AT_or0(self, a, h, dir, adj):
 114         try:
 115             return self.p_GO_AT[a, h, dir, adj]
 116         except KeyError:
 117             return 0.0
 118
 119
 120 def locs(sent_nums, start, stop):
 121     '''Return the between-word locations of all words in some fragment of
 122     sent. We make sure to offset the locations correctly so that for
 123     any w in the returned list, sent[w]==loc_w.
 124
 125     start is inclusive, stop is exclusive, as in klein-thesis and
 126     Python's list-slicing.'''
 127     for i0,w in enumerate(sent_nums[start:stop]):
 128         loc_w = i0+start
 129         yield (loc_w, w)
 130
 131 ###################################################
 132 #            P_INSIDE (dmv-specific)              #
 133 ###################################################
 134
 135 #@accepts(int, int, (int, int), int, Any(), [str], {tuple:float}, IsOneOf(None,{}))
 136 def inner(i, j, node, loc_h, g, sent, ichart, mpptree=None):
 137     ''' The ichart is of this form:
 138     ichart[i,j,LHS, loc_h]
 139     where i and j are between-word positions.
 140
 141     loc_h gives adjacency (along with k for attachment rules), and is
 142     needed in P_STOP reestimation.
 143     '''
 144     sent_nums = g.sent_nums(sent)
 145
 146     def terminal(i,j,node, loc_h, tabs):
 147         if not i <= loc_h < j:
 148             if 'INNER' in DEBUG:
 149                 print "%s*= 0.0 (wrong loc_h)" % tabs
 150             return 0.0
 151         elif POS(node) == sent_nums[i] and node in g.p_ORDER:
 152             # todo: add to ichart perhaps? Although, it _is_ simple lookup..
 153             prob = g.p_ORDER[node]
 154         else:
 155             if 'INNER' in DEBUG:
 156                 print "%sLACKING TERMINAL:" % tabs
 157             prob = 0.0
 158         if 'INNER' in DEBUG:
 159             print "%s*= %.4f (terminal: %s -> %s_%d)" % (tabs,prob, node_str(node), sent[i], loc_h)
 160         return prob
 161
 162     def e(i,j, (s_h,h), loc_h, n_t):
 163         def to_mpp(p, L, R):
 164             if mpptree:
 165                 key = (i,j, (s_h,h), loc_h)
 166                 if key not in mpptree:
 167                     mpptree[key] = (p, L, R)
 168                 elif mpptree[key][0] < p:
 169                     mpptree[key] = (p, L, R)
 170
 171         def tab():
 172             "Tabs for debug output"
 173             return "\t"*n_t
 174
 175         if (i, j, (s_h,h), loc_h) in ichart:
 176             if 'INNER' in DEBUG:
 177                 print "%s*= %.4f in ichart: i:%d j:%d node:%s loc:%s" % (tab(),ichart[i, j, (s_h,h), loc_h], i, j,
 178                                                                        node_str((s_h,h)), loc_h)
 179             return ichart[i, j, (s_h,h), loc_h]
 180         else:
 181             # Either terminal rewrites, using p_ORDER:
 182             if i+1 == j and (s_h == GOR or s_h == GOL):
 183                 return terminal(i, j, (s_h,h), loc_h, tab())
 184             else: # Or not at terminal level yet:
 185                 if 'INNER' in DEBUG:
 186                     print "%s%s (%.1f) from %d to %d" % (tab(),node_str((s_h,h)),loc_h,i,j)
 187                 if s_h == SEAL:
 188                     if h == POS(ROOT): # only used in testing, o/w we use inner_sent
 189                         h = sent_nums[loc_h]
 190                         if i != 0 or j != len(sent): raise ValueError
 191                         else: return g.p_ROOT[h] * e(i,j,(SEAL,h),loc_h,n_t+1)
 192                     p_RGOL = g.p_STOP[h,  LEFT, adj(i,loc_h)] * e(i,j,(RGOL,h),loc_h,n_t+1)
 193                     p_LGOR = g.p_STOP[h, RIGHT, adj(j,loc_h)] * e(i,j,(LGOR,h),loc_h,n_t+1)
 194                     p = p_RGOL + p_LGOR
 195                     to_mpp(p_RGOL,       STOPKEY,         (i,j, (RGOL,h),loc_h))
 196                     to_mpp(p_LGOR, (i,j, (RGOL,h),loc_h),       STOPKEY )
 197                     if 'INNER' in DEBUG:
 198                         print "%sp= %.4f (STOP)" % (tab(), p)
 199                 elif s_h == RGOL or s_h == GOL:
 200                     p = 0.0
 201                     if s_h == RGOL:
 202                         p = g.p_STOP[h, RIGHT, adj(j,loc_h)] * e(i,j, (GOR,h),loc_h,n_t+1)
 203                         to_mpp(p, (i,j, (GOR,h),loc_h), STOPKEY)
 204                     for k in xgo_left(i, loc_h): # i < k <= loc_l(h)
 205                         p_R = e(k, j, ( s_h,h), loc_h, n_t+1)
 206                         if p_R > 0.0:
 207                             for loc_a,a in locs(sent_nums, i, k):
 208                                 p_ah = g.p_GO_AT_or0(a, h, LEFT, adj(k,loc_h))
 209                                 if p_ah > 0.0:
 210                                     p_L = e(i, k, (SEAL,a), loc_a, n_t+1)
 211                                     p_add = p_L * p_ah * p_R
 212                                     p += p_add
 213                                     to_mpp(p_add,
 214                                           (i, k, (SEAL,a), loc_a),
 215                                           (k, j, ( s_h,h), loc_h))
 216                     if 'INNER' in DEBUG:
 217                         print "%sp= %.4f (ATTACH)" % (tab(), p)
 218                 elif s_h == GOR or s_h == LGOR:
 219                     p = 0.0
 220                     if s_h == LGOR:
 221                         p = g.p_STOP[h, LEFT, adj(i,loc_h)] * e(i,j, (GOL,h),loc_h,n_t+1)
 222                         to_mpp(p, (i,j, (GOL,h),loc_h), STOPKEY)
 223                     for k in xgo_right(loc_h, j): # loc_l(h) < k < j
 224                         p_L = e(i, k, ( s_h,h), loc_h, n_t+1)
 225                         if p_L > 0.0:
 226                             for loc_a,a in locs(sent_nums,k,j):
 227                                 p_ah = g.p_GO_AT_or0(a, h, RIGHT, adj(k,loc_h))
 228                                 p_R = e(k, j, (SEAL,a), loc_a, n_t+1)
 229                                 p_add = p_L * p_ah * p_R
 230                                 p += p_add
 231                                 to_mpp(p_add,
 232                                       (i, k, ( s_h,h), loc_h),
 233                                       (k, j, (SEAL,a), loc_a))
 234
 235                     if 'INNER' in DEBUG:
 236                         print "%sp= %.4f (ATTACH)" % (tab(), p)
 237                 # elif s_h == GOL: # todo
 238
 239                 ichart[i, j, (s_h,h), loc_h] = p
 240                 return p
 241     # end of e-function
 242
 243     inner_prob = e(i,j,node,loc_h, 0)
 244     if 'INNER' in DEBUG:
 245         print debug_ichart(g,sent,ichart)
 246     return inner_prob
 247 # end of dmv.inner(i, j, node, loc_h, g, sent, ichart,mpptree)
 248
 249
 250 def debug_ichart(g,sent,ichart):
 251     str = "---ICHART:---\n"
 252     for (s,t,LHS,loc_h),v in ichart.iteritems():
 253         str += "%s -> %s_%d ... %s_%d (loc_h:%s):\t%s\n" % (node_str(LHS,g.numtag),
 254                                                               sent[s], s, sent[s], t, loc_h, v)
 255     str += "---ICHART:end---\n"
 256     return str
 257
 258
 259 def inner_sent(g, sent, ichart):
 260     return sum([g.p_ROOT[w] * inner(0, len(sent), (SEAL,w), loc_w, g, sent, ichart)
 261                 for loc_w,w in locs(g.sent_nums(sent),0,len(sent))])
 262
 263
 264
 265
 266
 267 ###################################################
 268 #           P_OUTSIDE (dmv-specific)              #
 269 ###################################################
 270
 271 #@accepts(int, int, (int, int), int, Any(), [str], {tuple:float}, {tuple:float})
 272 def outer(i,j,w_node,loc_w, g, sent, ichart, ochart):
 273     ''' http://www.student.uib.no/~kun041/dmvccm/DMVCCM.html#outer
 274
 275     w_node is a pair (seals,POS); the w in klein-thesis is made up of
 276     POS(w) and loc_w
 277     '''
 278     sent_nums = g.sent_nums(sent)
 279     if POS(w_node) not in sent_nums[i:j]:
 280         # sanity check, w must be able to dominate sent[i:j]
 281         return 0.0
 282
 283     # local functions:
 284     def e(i,j,LHS,loc_h): # P_{INSIDE}
 285         try:
 286             return ichart[i,j,LHS,loc_h]
 287         except KeyError:
 288             return inner(i,j,LHS,loc_h,g,sent,ichart)
 289
 290     def f(i,j,w_node,loc_w):
 291         if not (i <= loc_w < j):
 292             return 0.0
 293         if (i,j,w_node,loc_w) in ochart:
 294             return ochart[i,j, w_node,loc_w]
 295         if w_node == ROOT:
 296             if i == 0 and j == len(sent):
 297                 return 1.0
 298             else: # ROOT may only be used on full sentence
 299                 return 0.0
 300         # but we may have non-ROOTs (stops) over full sentence too:
 301         w = POS(w_node)
 302         s_w = seals(w_node)
 303
 304         # todo: try either if p_M > 0.0: or sum(), and speed-test them
 305
 306         if s_w == SEAL: # w == a
 307             # todo: do the i<sent<j check here to save on calls?
 308             p = g.p_ROOT[w] * f(i,j,ROOT,loc_w)
 309             # left attach
 310             for k in xgt(j, sent): # j<k<len(sent)+1
 311                 for loc_h,h in locs(sent_nums,j,k):
 312                     p_wh = g.p_GO_AT_or0(w, h, LEFT, adj(j, loc_h))
 313                     for s_h in [RGOL, GOL]:
 314                         p += f(i,k,(s_h,h),loc_h) * p_wh * e(j,k,(s_h,h),loc_h)
 315             # right attach
 316             for k in xlt(i): # k<i
 317                 for loc_h,h in locs(sent_nums,k,i):
 318                     p_wh = g.p_GO_AT_or0(w, h, RIGHT, adj(i, loc_h))
 319                     for s_h in [LGOR, GOR]:
 320                         p += e(k,i,(s_h,h), loc_h) * p_wh * f(k,j,(s_h,h), loc_h)
 321
 322         elif s_w == RGOL or s_w == GOL: # w == h, left stop + left attach
 323             if s_w == RGOL:
 324                 s_h = SEAL
 325             else: # s_w == GOL
 326                 s_h = LGOR
 327             p = g.p_STOP[w, LEFT,  adj(i,loc_w)] * f(i,j,( s_h,w),loc_w)
 328             for k in xlt(i): # k<i
 329                 for loc_a,a in locs(sent_nums,k,i):
 330                     p_aw = g.p_GO_AT_or0(a, w, LEFT, adj(i, loc_w))
 331                     p += e(k,i, (SEAL,a),loc_a) * p_aw * f(k,j,w_node,loc_w)
 332
 333         elif s_w == GOR or s_w == LGOR: # w == h, right stop + right attach
 334             if s_w == GOR:
 335                 s_h = RGOL
 336             else: # s_w == LGOR
 337                 s_h = SEAL
 338             p = g.p_STOP[w, RIGHT, adj(j,loc_w)] * f(i,j,( s_h,w),loc_w)
 339             for k in xgt(j, sent): # j<k<len(sent)+1
 340                 for loc_a,a in locs(sent_nums,j,k):
 341                     p_ah = g.p_GO_AT_or0(a, w, RIGHT, adj(j, loc_w))
 342                     p += f(i,k,w_node,loc_w) * p_ah * e(j,k,(SEAL,a),loc_a)
 343
 344         ochart[i,j,w_node,loc_w] = p
 345         return p
 346     # end outer.f()
 347
 348     return f(i,j,w_node,loc_w)
 349 # end outer(i,j,w_node,loc_w, g,sent, ichart,ochart)
 350
 351
 352
 353
 354 ###################################################
 355 #              Reestimation v.1:                  #
 356 #           Sentences as outer loop               #
 357 ###################################################
 358
 359 def reest_zeros(h_nums):
 360     '''A dict to hold numerators and denominators for our 6+ reestimation
 361 formulas. '''
 362     # todo: p_ORDER?
 363     fr = { ('ROOT','den'):0.0 } # holds sum over f_sent!! not p_sent...
 364     for h in h_nums:
 365         fr['ROOT','num',h] = 0.0
 366         for s_h in [GOR,GOL,RGOL,LGOR]:
 367             x = (s_h,h)
 368             fr['hat_a','den',x] = 0.0 # = c()
 369             # not all arguments are attached to, so we just initialize
 370             # fr['hat_a','num',a,(s_h,h)] as they show up, in reest_freq
 371             for adj in [NON, ADJ]:
 372                 for nd in ['num','den']:
 373                     fr['STOP',nd,x,adj] = 0.0
 374     return fr
 375
 376
 377 def reest_freq(g, corpus):
 378     fr = reest_zeros(g.headnums())
 379     ichart = {}
 380     ochart = {}
 381     p_sent = None # 50 % speed increase on storing this locally
 382
 383     # local functions altogether 2x faster than global
 384     def c(i,j,LHS,loc_h,sent):
 385         if not p_sent > 0.0:
 386             return p_sent
 387
 388         p_in = e(i,j, LHS,loc_h,sent)
 389         if not p_in > 0.0:
 390             return p_in
 391
 392         p_out = f(i,j, LHS,loc_h,sent)
 393         return p_in * p_out / p_sent
 394     # end reest_freq.c()
 395
 396     def f(i,j,LHS,loc_h,sent): # P_{OUTSIDE}
 397         try:
 398             return ochart[i,j,LHS,loc_h]
 399         except KeyError:
 400             return outer(i,j,LHS,loc_h,g,sent,ichart,ochart)
 401     # end reest_freq.f()
 402
 403     def e(i,j,LHS,loc_h,sent): # P_{INSIDE}
 404         try:
 405             return ichart[i,j,LHS,loc_h]
 406         except KeyError:
 407             return inner(i,j,LHS,loc_h,g,sent,ichart)
 408     # end reest_freq.e()
 409
 410     def w_left(i,j, x,loc_h,sent,sent_nums):
 411         if not p_sent > 0.0: return
 412
 413         h = POS(x)
 414         a_k = {}
 415         for k in xtween(i, j):
 416             p_out = f(i,j, x,loc_h, sent)
 417             if not p_out > 0.0:
 418                 continue
 419             p_R =   e(k,j, x,loc_h, sent)
 420             if not p_R > 0.0:
 421                 continue
 422
 423             for loc_a,a in locs(sent_nums, i,k): # i<=loc_l(a)<k
 424                 p_rule = g.p_GO_AT_or0(a, h, LEFT, adj(k, loc_h))
 425                 p_L = e(i,k, (SEAL,a), loc_a, sent)
 426                 p = p_L * p_out * p_R * p_rule
 427                 try:    a_k[a] += p
 428                 except KeyError: a_k[a]  = p
 429
 430         for a,p in a_k.iteritems():
 431             try:    fr['hat_a','num',a,x] += p / p_sent
 432             except KeyError: fr['hat_a','num',a,x]  = p / p_sent
 433     # end reest_freq.w_left()
 434
 435     def w_right(i,j, x,loc_h,sent,sent_nums):
 436         if not p_sent > 0.0: return
 437
 438         h = POS(x)
 439         a_k = {}
 440         for k in xtween(i, j):
 441             p_out = f(i,j, x,loc_h, sent)
 442             if not p_out > 0.0:
 443                 continue
 444             p_L =   e(i,k, x,loc_h, sent)
 445             if not p_L > 0.0:
 446                 continue
 447
 448             for loc_a,a in locs(sent_nums, k,j): # k<=loc_l(a)<j
 449                 p_rule = g.p_GO_AT_or0(a, h, RIGHT, adj(k, loc_h))
 450                 p_R = e(k,j, (SEAL,a),loc_a, sent)
 451                 p = p_L * p_out * p_R * p_rule
 452                 try:    a_k[a] += p
 453                 except KeyError: a_k[a]  = p
 454
 455         for a,p in a_k.iteritems():
 456             try:    fr['hat_a','num',a,x] += p / p_sent
 457             except KeyError: fr['hat_a','num',a,x]  = p / p_sent
 458     # end reest_freq.w_right()
 459
 460     # in reest_freq:
 461     for sent in corpus:
 462         if 'REEST' in DEBUG:
 463             print sent
 464         ichart = {}
 465         ochart = {}
 466         p_sent = inner_sent(g, sent, ichart)
 467         fr['ROOT','den'] += 1 # divide by p_sent per h!
 468
 469         sent_nums = g.sent_nums(sent)
 470
 471         for loc_h,h in locs(sent_nums,0,len(sent)+1): # locs-stop is exclusive, thus +1
 472             # root:
 473             fr['ROOT','num',h] += g.p_ROOT[h] * e(0,len(sent), (SEAL,h),loc_h, sent) \
 474                 / p_sent
 475
 476             loc_l_h = loc_h
 477             loc_r_h = loc_l_h+1
 478
 479             # left non-adjacent stop:
 480             for i in xlt(loc_l_h):
 481                 fr['STOP','num',(GOL,h),NON] += c(loc_l_h, j, (LGOR, h),loc_h, sent)
 482                 fr['STOP','den',(GOL,h),NON] += c(loc_l_h, j, (GOL, h),loc_h, sent)
 483                 for j in xgteq(loc_r_h, sent):
 484                     fr['STOP','num',(RGOL,h),NON] += c(i, j, (SEAL, h),loc_h, sent)
 485                     fr['STOP','den',(RGOL,h),NON] += c(i, j, (RGOL, h),loc_h, sent)
 486             # left adjacent stop, i = loc_l_h
 487             fr['STOP','num',(GOL,h),ADJ] += c(loc_l_h, loc_r_h, (LGOR, h),loc_h, sent)
 488             fr['STOP','den',(GOL,h),ADJ] += c(loc_l_h, loc_r_h, (GOL, h),loc_h, sent)
 489             for j in xgteq(loc_r_h, sent):
 490                 fr['STOP','num',(RGOL,h),ADJ] += c(loc_l_h, j, (SEAL, h),loc_h, sent)
 491                 fr['STOP','den',(RGOL,h),ADJ] += c(loc_l_h, j, (RGOL, h),loc_h, sent)
 492             # right non-adjacent stop:
 493             for j in xgt(loc_r_h, sent):
 494                 fr['STOP','num',(GOR,h),NON] += c(loc_l_h, j, (RGOL, h),loc_h, sent)
 495                 fr['STOP','den',(GOR,h),NON] += c(loc_l_h, j, (GOR,  h),loc_h, sent)
 496                 for i in xlteq(loc_l_h):
 497                     fr['STOP','num',(LGOR,h),NON] += c(loc_l_h, j, (SEAL, h),loc_h, sent)
 498                     fr['STOP','den',(LGOR,h),NON] += c(loc_l_h, j, (LGOR,  h),loc_h, sent)
 499             # right adjacent stop, j = loc_r_h
 500             fr['STOP','num',(GOR,h),ADJ] += c(loc_l_h, loc_r_h, (RGOL, h),loc_h, sent)
 501             fr['STOP','den',(GOR,h),ADJ] += c(loc_l_h, loc_r_h, (GOR,  h),loc_h, sent)
 502             for i in xlteq(loc_l_h):
 503                 fr['STOP','num',(LGOR,h),ADJ] += c(loc_l_h, j, (SEAL, h),loc_h, sent)
 504                 fr['STOP','den',(LGOR,h),ADJ] += c(loc_l_h, j, (LGOR,  h),loc_h, sent)
 505
 506             # left attachment:
 507             if 'REEST_ATTACH' in DEBUG:
 508                 print "Lattach %s: for i < %s"%(g.numtag(h),sent[0:loc_h+1])
 509             for s_h in [RGOL, GOL]:
 510                 x = (s_h, h)
 511                 for i in xlt(loc_l_h):  # i < loc_l(h)
 512                     if 'REEST_ATTACH' in DEBUG:
 513                         print "\tfor j >= %s"%sent[loc_h:len(sent)]
 514                     for j in xgteq(loc_r_h, sent): # j >= loc_r(h)
 515                         fr['hat_a','den',x] += c(i,j, x,loc_h, sent) # v_q in L&Y
 516                         if 'REEST_ATTACH' in DEBUG:
 517                             print "\t\tc( %d , %d, %s, %s, sent)=%.4f"%(i,j,node_str(x),loc_h,fr['hat_a','den',x])
 518                         w_left(i, j, x,loc_h, sent,sent_nums) # compute w for all a in sent
 519
 520             # right attachment:
 521             if 'REEST_ATTACH' in DEBUG:
 522                 print "Rattach %s: for i <= %s"%(g.numtag(h),sent[0:loc_h+1])
 523             for s_h in [GOR, LGOR]:
 524                 x = (s_h, h)
 525                 for i in xlteq(loc_l_h):  # i <= loc_l(h)
 526                     if 'REEST_ATTACH' in DEBUG:
 527                         print "\tfor j > %s"%sent[loc_h:len(sent)]
 528                     for j in xgt(loc_r_h, sent): # j > loc_r(h)
 529                         fr['hat_a','den',x] += c(i,j, x,loc_h, sent) # v_q in L&Y
 530                         if 'REEST_ATTACH' in DEBUG:
 531                             print "\t\tc( %d , %d, %s, %s, sent)=%.4f"%(loc_h,j,node_str(x),loc_h,fr['hat_a','den',x])
 532                         w_right(i,j, x,loc_h, sent,sent_nums) # compute w for all a in sent
 533         # end for loc_h,h
 534     # end for sent
 535
 536     return fr
 537
 538 def reestimate(old_g, corpus):
 539     fr = reest_freq(old_g, corpus)
 540     p_ROOT, p_STOP, p_ATTACH = {},{},{}
 541
 542     for h in old_g.headnums():
 543         # reest_head changes p_ROOT, p_STOP, p_ATTACH
 544         reest_head(h, fr, old_g, p_ROOT, p_STOP, p_ATTACH)
 545     p_ORDER = old_g.p_ORDER
 546     numtag, tagnum = old_g.get_nums_tags()
 547
 548     new_g = DMV_Grammar(numtag, tagnum, p_ROOT, p_STOP, p_ATTACH, p_ORDER)
 549     return new_g
 550
 551
 552 def reest_head(h, fr, g, p_ROOT, p_STOP, p_ATTACH):
 553     "Given a single head, update g with the reestimated probability."
 554     # remove 0-prob stuff? todo
 555     try:
 556         p_ROOT[h] = fr['ROOT','num',h] / fr['ROOT','den']
 557     except KeyError:
 558         p_ROOT[h] = 0.0
 559
 560     for dir in [LEFT,RIGHT]:
 561         for adj in [ADJ, NON]: # p_STOP
 562             p_STOP[h, dir, adj] = 0.0
 563             for s_h in dirseal(dir):
 564                 x = (s_h,h)
 565                 p = fr['STOP','den', x, adj]
 566                 if p > 0.0:
 567                     p = fr['STOP', 'num', x, adj] / p
 568                 p_STOP[h, dir, adj] += p
 569
 570         for s_h in dirseal(dir): # make hat_a for p_ATTACH
 571             x = (s_h,h)
 572             p_c = fr['hat_a','den',x]
 573
 574             for a in g.headnums():
 575                 if (a,h,dir) not in p_ATTACH:
 576                     p_ATTACH[a,h,dir] = 0.0
 577                 try: # (a,x) might not be in hat_a
 578                     p_ATTACH[a,h,dir] += fr['hat_a','num',a,x] / p_c
 579                 except KeyError: pass
 580                 except ZeroDivisionError: pass
 581
 582
 583
 584
 585
 586 ###################################################
 587 #              Reestimation v.2:                  #
 588 #            Heads as outer loop                  #
 589 ###################################################
 590
 591 def locs_h(h, sent_nums):
 592     '''Return the between-word locations of all tokens of h in sent.'''
 593     return [loc_w for loc_w,w in locs(sent_nums, 0, len(sent_nums))
 594             if w == h]
 595
 596 def locs_a(a, sent_nums, start, stop):
 597     '''Return the between-word locations of all tokens of h in some
 598     fragment of sent. We make sure to offset the locations correctly
 599     so that for any w in the returned list, sent[w]==loc_w.
 600
 601     start is inclusive, stop is exclusive, as in klein-thesis and
 602     Python's list-slicing (eg. return left-loc).'''
 603     return [loc_w for loc_w,w in locs(sent_nums, start, stop)
 604             if w == a]
 605
 606 def inner2(i, j, node, loc_h, g, sent):
 607     ichart,ochart = g.get_iochart(s_n)
 608     try: p = ichart[i,j,x,loc_h]
 609     except KeyError: p = inner(i,j,x,loc_h,g,sent,ichart)
 610     g.set_iochart(s_n,ichart,ochart)
 611     return p
 612
 613 def inner_sent2(g, sent):
 614     ichart,ochart = g.get_iochart(s_n)
 615     p = inner_sent(g,sent,ichart)
 616     g.set_iochart(s_n,ichart,ochart)
 617     return p
 618
 619 def outer2(i, j,w_node,loc_w, g, sent):
 620     ichart,ochart = g.get_iochart(s_n)
 621     try: p = ochart[i,j,w_node,loc_w]
 622     except KeyError: p = inner(i,j,w_node,loc_w,g,sent,ichart,ochart)
 623     g.set_iochart(s_n,ichart,ochart)
 624     return p
 625
 626 def reestimate2(old_g, corpus):
 627     p_ROOT, p_STOP, p_ATTACH = {},{},{}
 628
 629     for h in old_g.headnums():
 630         # reest_head changes p_ROOT, p_STOP, p_ATTACH
 631         reest_head2(h, old_g, corpus, p_ROOT, p_STOP, p_ATTACH)
 632     p_ORDER = old_g.p_ORDER
 633     numtag, tagnum = old_g.get_nums_tags()
 634
 635     new_g = DMV_Grammar(numtag, tagnum, p_ROOT, p_STOP, p_ATTACH, p_ORDER)
 636     return new_g
 637
 638 def hat_d2(xbar, x, xi, xj, g, corpus): # stop helper
 639     def c(x,loc_x,i,j): return c2(x,loc_x,i,j,g,s_n,sent)
 640
 641     h = POS(x)
 642     if h != POS(xbar): raise ValueError
 643
 644     num, den = 0.0, 0.0
 645     for s_n,sent in [(g.sent_nums(sent),sent) for sent in corpus]:
 646         for loc_h in locs_h(h,s_n):
 647             loc_l_h, loc_r_h = loc_h, loc_h + 1
 648             for i in xi(loc_l_h):
 649                 for j in xj(loc_r_h, s_n):
 650 #                     print "s:%s %d,%d"%(sent,i,j)
 651                     num += c(xbar,loc_h,i,j)
 652                     den += c(x,loc_h,i,j)
 653     if den == 0.0:
 654         return den
 655     return num/den # eg. SEAL/RGOL, xbar/x
 656
 657
 658 def c2(x,loc_h,i,j,g,s_n,sent):
 659     ichart,ochart = g.get_iochart(s_n)
 660
 661     def f(i,j,x,loc_h): # P_{OUTSIDE}
 662         try: return ochart[i,j,x,loc_h]
 663         except KeyError: return outer(i,j,x,loc_h,g,sent,ichart,ochart)
 664     def e(i,j,x,loc_h): # P_{INSIDE}
 665         try: return ichart[i,j,x,loc_h]
 666         except KeyError: return inner(i,j,x,loc_h,g,sent,ichart)
 667
 668     p_sent = inner_sent(g, sent, ichart)
 669     if not p_sent > 0.0:
 670         return p_sent
 671
 672     p_in = e(i,j, x,loc_h)
 673     if not p_in > 0.0:
 674         return p_in
 675
 676     p_out = f(i,j, x,loc_h)
 677
 678     g.set_iochart(s_n,ichart,ochart)
 679     return p_in * p_out / p_sent
 680
 681 def w2(a, x,loc_h, dir, i, j, g, s_n,sent):
 682     ichart,ochart = g.get_iochart(s_n)
 683
 684     def f(i,j,x,loc_h): # P_{OUTSIDE}
 685         try: return ochart[i,j,x,loc_h]
 686         except KeyError: return outer(i,j,x,loc_h,g,sent,ichart,ochart)
 687     def e(i,j,x,loc_h): # P_{INSIDE}
 688         try: return ichart[i,j,x,loc_h]
 689         except KeyError: return inner(i,j,x,loc_h,g,sent,ichart)
 690
 691     h = POS(x)
 692     p_sent = inner_sent(g, sent, ichart)
 693
 694     if dir == LEFT:
 695         L, R = (SEAL,a),x
 696     else:
 697         L, R = x,(SEAL,a)
 698     w_sum = 0.0
 699
 700     for k in xtween(i,j):
 701         if dir == LEFT:
 702             start, stop = i, k
 703         else:
 704             start, stop = k, j
 705         for loc_a in locs_a(a, s_n, start, stop):
 706             if dir == LEFT:
 707                 loc_L, loc_R = loc_a, loc_h
 708             else:
 709                 loc_L, loc_R = loc_h, loc_a
 710             p = g.p_GO_AT_or0(a,h,dir,adj(k,loc_h))
 711             in_L = e(i,k,L,loc_L)
 712             in_R = e(k,j,R,loc_R)
 713             out =  f(i,j,x,loc_h)
 714             w_sum += p * in_L * in_R * out
 715
 716     g.set_iochart(s_n,ichart,ochart)
 717     return w_sum/p_sent
 718
 719 def hat_a2(a, x, dir, g, corpus): # attachment helper
 720     def w(a,x,loc_x,dir,i,j): return w2(a,x,loc_x,dir,i,j,g,s_n,sent)
 721     def c(x,loc_x,i,j): return c2(x,loc_x,i,j,g,s_n,sent)
 722
 723     h = POS(x)
 724     if dir == LEFT:
 725         xi, xj = xlt, xgteq
 726     else:
 727         xi, xj = xlteq, xgt
 728     den, num = 0.0, 0.0
 729
 730     for s_n,sent in [(g.sent_nums(sent),sent) for sent in corpus]:
 731         for loc_h in locs_h(h,s_n):
 732             loc_l_h, loc_r_h = loc_h, loc_h + 1
 733             for i in xi(loc_l_h):
 734                 for j in xj(loc_r_h,sent):
 735                     num += w(a, x,loc_h, dir, i,j)
 736                     den += c(x,loc_h, i,j)
 737     if den == 0.0:
 738         return den
 739     return num/den
 740
 741 def reest_root2(h,g,corpus):
 742     sum = 0.0
 743     corpus_size = 0.0
 744     for s_n,sent in [(g.sent_nums(sent),sent) for sent in corpus]:
 745         num, den = 0.0, 0.0
 746         corpus_size += 1.0
 747         ichart, ochart = g.get_iochart(s_n)
 748         den += inner_sent(g, sent, ichart)
 749         for loc_h in locs_h(h,s_n):
 750             num += \
 751                 g.p_ROOT[h] * \
 752                 inner(0, len(s_n), (SEAL,h), loc_h, g, sent, ichart)
 753         g.set_iochart(s_n, ichart, ochart)
 754         sum += num / den
 755     return sum / corpus_size
 756
 757 def reest_head2(h, g, corpus, p_ROOT, p_STOP, p_ATTACH):
 758     print "h: %d=%s ..."%(h,g.numtag(h)),
 759     def hat_d(xbar,x,xi,xj): return hat_d2(xbar,x,xi,xj, g, corpus)
 760     def hat_a(a, x, dir  ): return hat_a2(a, x, dir,   g, corpus)
 761
 762     p_STOP[h, LEFT,NON] = \
 763         hat_d((SEAL,h),(RGOL,h),xlt,  xgteq) + \
 764         hat_d((LGOR,h),( GOL,h),xlt,  xeq)
 765     p_STOP[h, LEFT,ADJ] = \
 766         hat_d((SEAL,h),(RGOL,h),xeq,  xgteq) + \
 767         hat_d((LGOR,h),( GOL,h),xeq,  xeq)
 768     p_STOP[h,RIGHT,NON] = \
 769         hat_d((RGOL,h),( GOR,h),xeq,  xgt) + \
 770         hat_d((SEAL,h),(LGOR,h),xlteq,xgt)
 771     p_STOP[h,RIGHT,ADJ] = \
 772         hat_d((RGOL,h),( GOR,h),xeq,  xeq) + \
 773         hat_d((SEAL,h),(LGOR,h),xlteq,xeq)
 774     print "stops done...",
 775
 776     p_ROOT[h] = reest_root2(h,g,corpus)
 777     print "root done...",
 778
 779     for a in g.headnums():
 780         p_ATTACH[a,h,LEFT] = \
 781             hat_a(a, (GOL,h),LEFT) + \
 782             hat_a(a,(RGOL,h),LEFT)
 783         p_ATTACH[a,h,RIGHT] = \
 784             hat_a(a, (GOR,h),RIGHT) + \
 785             hat_a(a,(LGOR,h),RIGHT)
 786
 787     print "attachment done"
 788
 789
 790
 791 ###################################################
 792 #             Most Probable Parse:                #
 793 ###################################################
 794
 795 STOPKEY = (-1,-1,STOP,-1)
 796 ROOTKEY = (-1,-1,ROOT,-1)
 797
 798 def make_mpptree(g, sent):
 799     '''Tell inner() to make an mpptree, connect ROOT to this. (Logically,
 800     this should be part of inner_sent though...)'''
 801     ichart = {}
 802     mpptree = { ROOTKEY:(0.0, ROOTKEY, None) }
 803     for loc_w,w in locs(g.sent_nums(sent),0,len(sent)):
 804         p = g.p_ROOT[w] * inner(0, len(sent), (SEAL,w), loc_w, g, sent, ichart, mpptree)
 805         L = ROOTKEY
 806         R = (0,len(sent), (SEAL,w), loc_w)
 807         if mpptree[ROOTKEY][0] < p:
 808             mpptree[ROOTKEY] = (p, L, R)
 809     return mpptree
 810
 811 def parse_mpptree(mpptree, sent):
 812     '''mpptree is a dict of the form {k:(p,L,R),...}; where k, L and R
 813     are `keys' of the form (i,j,node,loc).
 814
 815     returns an mpp of the form [((head, loc_h),(arg, loc_a)), ...],
 816     where head and arg are tags.'''
 817     # local functions for clear access to mpptree:
 818     def k_node(key):
 819         return key[2]
 820     def k_POS(key):
 821         return POS(k_node(key))
 822     def k_seals(key):
 823         return seals(k_node(key))
 824     def k_locnode(key):
 825         return (k_node(key),key[3])
 826     def k_locPOS(key):
 827         return (k_POS(key),key[3])
 828     def k_terminal(key):
 829         s_k = k_seals(key) # i+1 == j
 830         return key[0] + 1 == key[1] and (s_k == GOR or s_k == GOL)
 831     def t_L(tree_entry):
 832         return tree_entry[1]
 833     def t_R(tree_entry):
 834         return tree_entry[2]
 835
 836     # arbitrarily, "ROOT attaches to right". We add it here to
 837     # avoid further complications:
 838     firstkey = t_R(mpptree[ROOTKEY])
 839     deps = set([ (k_locPOS(ROOTKEY), k_locPOS(firstkey), RIGHT) ])
 840     q = [firstkey]
 841
 842     while len(q) > 0:
 843         k = q.pop()
 844         if k_terminal(k):
 845             continue
 846         else:
 847             L = t_L( mpptree[k] )
 848             R = t_R( mpptree[k] )
 849             if k_locnode( k ) == k_locnode( L ): # Rattach
 850                 deps.add((k_locPOS( k ), k_locPOS( R ), LEFT))
 851                 q.extend( [L, R] )
 852             elif k_locnode( k ) == k_locnode( R ): # Lattach
 853                 deps.add((k_locPOS( k ), k_locPOS( L ), RIGHT))
 854                 q.extend( [L, R] )
 855             elif R == STOPKEY:
 856                 q.append( L )
 857             elif L == STOPKEY:
 858                 q.append( R )
 859     return deps
 860
 861 def mpp(g, sent):
 862     tagf = g.numtag # localized function, todo: speed-test
 863     mpptree = make_mpptree(g, sent)
 864     return set([((tagf(h), loc_h), (tagf(a), loc_a))
 865                 for (h, loc_h),(a,loc_a),dir in parse_mpptree(mpptree,sent)])
 866
 867
 868 ########################################################################
 869 #     testing functions:                                               #
 870 ########################################################################
 871
 872 testcorpus = [s.split() for s in ['det nn vbd c vbd','vbd nn c vbd',
 873                                   'det nn vbd',      'det nn vbd c pp',
 874                                   'det nn vbd',      'det vbd vbd c pp',
 875                                   'det nn vbd',      'det nn vbd c vbd',
 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 pp',
 879                                   'det nn vbd pp',   'det nn vbd', ]]
 880
 881 def testgrammar():
 882     import loc_h_harmonic
 883     reload(loc_h_harmonic)
 884
 885     # make sure these are the way they were when setting up the tests:
 886     loc_h_harmonic.HARMONIC_C = 0.0
 887     loc_h_harmonic.FNONSTOP_MIN = 25
 888     loc_h_harmonic.FSTOP_MIN = 5
 889     loc_h_harmonic.RIGHT_FIRST = 1.0
 890     loc_h_harmonic.OLD_STOP_CALC = True
 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     import loc_h_harmonic
1062
1063     # make sure these are the way they were when setting up the tests:
1064     loc_h_harmonic.HARMONIC_C = 0.0
1065     loc_h_harmonic.FNONSTOP_MIN = 25
1066     loc_h_harmonic.FSTOP_MIN = 5
1067     loc_h_harmonic.RIGHT_FIRST = 1.0
1068     loc_h_harmonic.OLD_STOP_CALC = True
1069
1070     g = loc_h_harmonic.initialize(['n v n'.split()])
1071     return g # todo
1072
1073 def testIO():
1074     g = testgrammar()
1075     inners = [(sent, inner_sent(g, sent, {})) for sent in testcorpus]
1076     return inners
1077
1078 if __name__ == "__main__":
1079     DEBUG.clear()
1080     regression_tests()
1081
1082 #    import profile
1083 #    profile.run('testreestimation()')
1084
1085 #     import timeit
1086 #     print timeit.Timer("loc_h_dmv.testreestimation()",'''import loc_h_dmv
1087 # reload(loc_h_dmv)''').timeit(1)
1088
1089
1090 #     print "mpp-test:"
1091 #     import pprint
1092 #     for s in testcorpus:
1093 #         print "sent:%s\nparse:set(\n%s)"%(s,pprint.pformat(list(mpp(testgrammar(), s)),
1094 #                                                     width=40))
1095
1096 #     g1 = testreestimation()
1097 #     g2 = testreestimation2()
1098 #     print compare_grammars(g1,g2)
1099
1100
1101
1102
1103
1104
1105
1106 if False:
1107     g = testNVNgrammar()
1108     q_sent = inner_sent(g,'n v n'.split(),{})
1109     q_tree = {}
1110     q_tree[1] = 2.7213e-06 # n_0 -> v, n_0 -> n_2
1111     q_tree[2] = 9.738e-06   # n -> v -> n
1112     q_tree[3] = 2.268e-06   # n_0 -> n_2 -> v
1113     q_tree[4] = 2.7213e-06 # same as 1-3
1114     q_tree[5] = 9.738e-06
1115     q_tree[6] = 2.268e-06
1116     q_tree[7] = 1.086e-05   # n <- v -> n (e-05!!!)
1117     f_T_q = {}
1118     for i,q_t in q_tree.iteritems():
1119         f_T_q[i] = q_t / q_sent
1120     import pprint
1121     pprint.pprint(q_tree)
1122     pprint.pprint(f_T_q)
1123     print sum([f for f in f_T_q.values()])
1124
1125     def treediv(num,den):
1126         return \
1127             sum([f_T_q[i] for i in num ]) / \
1128             sum([f_T_q[i] for i in den ])
1129     g2 = {}
1130 #     g2['root --> _n_'] = treediv( (1,2,3,4,5,6), (1,2,3,4,5,6,7) )
1131 #     g2['root --> _v_'] = treediv( (7,), (1,2,3,4,5,6,7) )
1132 #     g2['_n_ --> STOP n><'] = treediv( (1,2,3,4,5,6,7,1,2,3,4,5,6,7),
1133 #                                       (1,2,3,4,5,6,7,1,2,3,4,5,6,7))
1134
1135 #     g2['_n_ --> STOP n>< NON'] = treediv( (3,4,5,6),
1136 #                                           (3,4,5,6,4) )
1137
1138 #     g2['_v_ --> STOP v><'] = treediv( (1,2,3,4,5,6,7),
1139 #                                       (1,2,3,4,5,6,7) )
1140 #     nlrtrees = (1,2,3,4,5,6,7,1,2,3,4,5,6,7,
1141 #                 3,4,4,5,6)
1142 #     g2['n>< --> _n_ n><'] = treediv( (  4,  6), nlrtrees )
1143 #     g2['n>< --> _v_ n><'] = treediv( (3,4,5), nlrtrees )
1144 #     g2['n>< --> n> STOP'] = treediv( (1,2,3,4,5,6,7,1,2,3,4,5,6,7),
1145 #                                      nlrtrees )
1146
1147 #     g2['n>< --> n> STOP ADJ'] = treediv( (      4,5,  7,1,2,3,4,5,6,7),
1148 #                                          nlrtrees )
1149 #     g2['n>< --> n> STOP NON'] = treediv( (1,2,3,    6),
1150 #                                          nlrtrees )
1151
1152 #     vlrtrees = (1,2,3,4,5,6,7,
1153 #                 7,5)
1154 #     g2['v>< --> _n_ v><'] = treediv( (5,7), vlrtrees )
1155 #     g2['v>< --> v> STOP'] = treediv( (1,2,3,4,5,6,7), vlrtrees )
1156 #     nrtrees = (1,2,3,4,5,6,7,1,2,3,4,5,6,7,
1157 #                1,1,2,3,6)
1158 #     g2['n> --> n> _n_'] = treediv( (1,3), nrtrees )
1159 #     g2['n> --> n> _v_'] = treediv( (1,2,6), nrtrees )
1160
1161 #     g2['n> --> n> _n_ NON'] = treediv( (1,), nrtrees )
1162 #     g2['n> --> n> _n_ ADJ'] = treediv( (      3,), nrtrees )
1163 #     g2['n> --> n> _v_ ADJ'] = treediv( (  1,2,  6), nrtrees )
1164
1165 #     vrtrees = (1,2,3,4,5,6,7,
1166 #                7,2)
1167 #     g2['v> --> v> _n_'] = treediv( (2,7), vrtrees )
1168
1169 #     g2[' v|n,R '] = treediv( (1,  2,  6),
1170 #                              (1,1,2,3,6) )
1171 #     g2[' n|n,R '] = treediv( (1,    3),
1172 #                              (1,1,2,3,6) )
1173
1174     g2[' stop|n,R,non '] = treediv( (  1,2,3,6),
1175                                     (1,1,2,3,6) )
1176     g2[' v|n,left '] = treediv( (    3,4,5),
1177                                 (6,4,3,4,5) )
1178     g2[' n|n,left '] = treediv( (6,4),
1179                                 (6,4,3,4,5) )
1180
1181     pprint.pprint(g2)
1182     g3 = reestimate2(g, ['n v n'.split()])
1183     print g3
1184     g4 = reestimate2(g, ['n v n'.split()])
1185     print g4
1186
1187