nominatim/api/search/token_assignment.py

   1 # SPDX-License-Identifier: GPL-3.0-or-later
   2 #
   3 # This file is part of Nominatim. (https://nominatim.org)
   4 #
   5 # Copyright (C) 2023 by the Nominatim developer community.
   6 # For a full list of authors see the git log.
   7 """
   8 Create query interpretations where each vertice in the query is assigned
   9 a specific function (expressed as a token type).
  10 """
  11 from typing import Optional, List, Iterator
  12 import dataclasses
  13
  14 import nominatim.api.search.query as qmod
  15 from nominatim.api.logging import log
  16
  17 # pylint: disable=too-many-return-statements,too-many-branches
  18
  19 @dataclasses.dataclass
  20 class TypedRange:
  21     """ A token range for a specific type of tokens.
  22     """
  23     ttype: qmod.TokenType
  24     trange: qmod.TokenRange
  25
  26
  27 PENALTY_TOKENCHANGE = {
  28     qmod.BreakType.START: 0.0,
  29     qmod.BreakType.END: 0.0,
  30     qmod.BreakType.PHRASE: 0.0,
  31     qmod.BreakType.WORD: 0.1,
  32     qmod.BreakType.PART: 0.2,
  33     qmod.BreakType.TOKEN: 0.4
  34 }
  35
  36 TypedRangeSeq = List[TypedRange]
  37
  38 @dataclasses.dataclass
  39 class TokenAssignment: # pylint: disable=too-many-instance-attributes
  40     """ Representation of a possible assignment of token types
  41         to the tokens in a tokenized query.
  42     """
  43     penalty: float = 0.0
  44     name: Optional[qmod.TokenRange] = None
  45     address: List[qmod.TokenRange] = dataclasses.field(default_factory=list)
  46     housenumber: Optional[qmod.TokenRange] = None
  47     postcode: Optional[qmod.TokenRange] = None
  48     country: Optional[qmod.TokenRange] = None
  49     category: Optional[qmod.TokenRange] = None
  50     qualifier: Optional[qmod.TokenRange] = None
  51
  52
  53     @staticmethod
  54     def from_ranges(ranges: TypedRangeSeq) -> 'TokenAssignment':
  55         """ Create a new token assignment from a sequence of typed spans.
  56         """
  57         out = TokenAssignment()
  58         for token in ranges:
  59             if token.ttype == qmod.TokenType.PARTIAL:
  60                 out.address.append(token.trange)
  61             elif token.ttype == qmod.TokenType.HOUSENUMBER:
  62                 out.housenumber = token.trange
  63             elif token.ttype == qmod.TokenType.POSTCODE:
  64                 out.postcode = token.trange
  65             elif token.ttype == qmod.TokenType.COUNTRY:
  66                 out.country = token.trange
  67             elif token.ttype == qmod.TokenType.CATEGORY:
  68                 out.category = token.trange
  69             elif token.ttype == qmod.TokenType.QUALIFIER:
  70                 out.qualifier = token.trange
  71         return out
  72
  73
  74 class _TokenSequence:
  75     """ Working state used to put together the token assignements.
  76
  77         Represents an intermediate state while traversing the tokenized
  78         query.
  79     """
  80     def __init__(self, seq: TypedRangeSeq,
  81                  direction: int = 0, penalty: float = 0.0) -> None:
  82         self.seq = seq
  83         self.direction = direction
  84         self.penalty = penalty
  85
  86
  87     def __str__(self) -> str:
  88         seq = ''.join(f'[{r.trange.start} - {r.trange.end}: {r.ttype.name}]' for r in self.seq)
  89         return f'{seq} (dir: {self.direction}, penalty: {self.penalty})'
  90
  91
  92     @property
  93     def end_pos(self) -> int:
  94         """ Return the index of the global end of the current sequence.
  95         """
  96         return self.seq[-1].trange.end if self.seq else 0
  97
  98
  99     def has_types(self, *ttypes: qmod.TokenType) -> bool:
 100         """ Check if the current sequence contains any typed ranges of
 101             the given types.
 102         """
 103         return any(s.ttype in ttypes for s in self.seq)
 104
 105
 106     def is_final(self) -> bool:
 107         """ Return true when the sequence cannot be extended by any
 108             form of token anymore.
 109         """
 110         # Country and category must be the final term for left-to-right
 111         return len(self.seq) > 1 and \
 112                self.seq[-1].ttype in (qmod.TokenType.COUNTRY, qmod.TokenType.CATEGORY)
 113
 114
 115     def appendable(self, ttype: qmod.TokenType) -> Optional[int]:
 116         """ Check if the give token type is appendable to the existing sequence.
 117
 118             Returns None if the token type is not appendable, otherwise the
 119             new direction of the sequence after adding such a type. The
 120             token is not added.
 121         """
 122         if ttype == qmod.TokenType.WORD:
 123             return None
 124
 125         if not self.seq:
 126             # Append unconditionally to the empty list
 127             if ttype == qmod.TokenType.COUNTRY:
 128                 return -1
 129             if ttype in (qmod.TokenType.HOUSENUMBER, qmod.TokenType.QUALIFIER):
 130                 return 1
 131             return self.direction
 132
 133         # Name tokens are always acceptable and don't change direction
 134         if ttype == qmod.TokenType.PARTIAL:
 135             return self.direction
 136
 137         # Other tokens may only appear once
 138         if self.has_types(ttype):
 139             return None
 140
 141         if ttype == qmod.TokenType.HOUSENUMBER:
 142             if self.direction == 1:
 143                 if len(self.seq) == 1 and self.seq[0].ttype == qmod.TokenType.QUALIFIER:
 144                     return None
 145                 if len(self.seq) > 2 \
 146                    or self.has_types(qmod.TokenType.POSTCODE, qmod.TokenType.COUNTRY):
 147                     return None # direction left-to-right: housenumber must come before anything
 148             elif self.direction == -1 \
 149                  or self.has_types(qmod.TokenType.POSTCODE, qmod.TokenType.COUNTRY):
 150                 return -1 # force direction right-to-left if after other terms
 151
 152             return self.direction
 153
 154         if ttype == qmod.TokenType.POSTCODE:
 155             if self.direction == -1:
 156                 if self.has_types(qmod.TokenType.HOUSENUMBER, qmod.TokenType.QUALIFIER):
 157                     return None
 158                 return -1
 159             if self.direction == 1:
 160                 return None if self.has_types(qmod.TokenType.COUNTRY) else 1
 161             if self.has_types(qmod.TokenType.HOUSENUMBER, qmod.TokenType.QUALIFIER):
 162                 return 1
 163             return self.direction
 164
 165         if ttype == qmod.TokenType.COUNTRY:
 166             return None if self.direction == -1 else 1
 167
 168         if ttype == qmod.TokenType.CATEGORY:
 169             return self.direction
 170
 171         if ttype == qmod.TokenType.QUALIFIER:
 172             if self.direction == 1:
 173                 if (len(self.seq) == 1
 174                     and self.seq[0].ttype in (qmod.TokenType.PARTIAL, qmod.TokenType.CATEGORY)) \
 175                    or (len(self.seq) == 2
 176                        and self.seq[0].ttype == qmod.TokenType.CATEGORY
 177                        and self.seq[1].ttype == qmod.TokenType.PARTIAL):
 178                     return 1
 179                 return None
 180             if self.direction == -1:
 181                 return -1
 182
 183             tempseq = self.seq[1:] if self.seq[0].ttype == qmod.TokenType.CATEGORY else self.seq
 184             if len(tempseq) == 0:
 185                 return 1
 186             if len(tempseq) == 1 and self.seq[0].ttype == qmod.TokenType.HOUSENUMBER:
 187                 return None
 188             if len(tempseq) > 1 or self.has_types(qmod.TokenType.POSTCODE, qmod.TokenType.COUNTRY):
 189                 return -1
 190             return 0
 191
 192         return None
 193
 194
 195     def advance(self, ttype: qmod.TokenType, end_pos: int,
 196                 btype: qmod.BreakType) -> Optional['_TokenSequence']:
 197         """ Return a new token sequence state with the given token type
 198             extended.
 199         """
 200         newdir = self.appendable(ttype)
 201         if newdir is None:
 202             return None
 203
 204         if not self.seq:
 205             newseq = [TypedRange(ttype, qmod.TokenRange(0, end_pos))]
 206             new_penalty = 0.0
 207         else:
 208             last = self.seq[-1]
 209             if btype != qmod.BreakType.PHRASE and last.ttype == ttype:
 210                 # extend the existing range
 211                 newseq = self.seq[:-1] + [TypedRange(ttype, last.trange.replace_end(end_pos))]
 212                 new_penalty = 0.0
 213             else:
 214                 # start a new range
 215                 newseq = list(self.seq) + [TypedRange(ttype,
 216                                                       qmod.TokenRange(last.trange.end, end_pos))]
 217                 new_penalty = PENALTY_TOKENCHANGE[btype]
 218
 219         return _TokenSequence(newseq, newdir, self.penalty + new_penalty)
 220
 221
 222     def _adapt_penalty_from_priors(self, priors: int, new_dir: int) -> bool:
 223         if priors == 2:
 224             self.penalty += 1.0
 225         elif priors > 2:
 226             if self.direction == 0:
 227                 self.direction = new_dir
 228             else:
 229                 return False
 230
 231         return True
 232
 233
 234     def recheck_sequence(self) -> bool:
 235         """ Check that the sequence is a fully valid token assignment
 236             and addapt direction and penalties further if necessary.
 237
 238             This function catches some impossible assignments that need
 239             forward context and can therefore not be exluded when building
 240             the assignment.
 241         """
 242         # housenumbers may not be further than 2 words from the beginning.
 243         # If there are two words in front, give it a penalty.
 244         hnrpos = next((i for i, tr in enumerate(self.seq)
 245                        if tr.ttype == qmod.TokenType.HOUSENUMBER),
 246                       None)
 247         if hnrpos is not None:
 248             if self.direction != -1:
 249                 priors = sum(1 for t in self.seq[:hnrpos] if t.ttype == qmod.TokenType.PARTIAL)
 250                 if not self._adapt_penalty_from_priors(priors, -1):
 251                     return False
 252             if self.direction != 1:
 253                 priors = sum(1 for t in self.seq[hnrpos+1:] if t.ttype == qmod.TokenType.PARTIAL)
 254                 if not self._adapt_penalty_from_priors(priors, 1):
 255                     return False
 256
 257         return True
 258
 259
 260     def get_assignments(self, query: qmod.QueryStruct) -> Iterator[TokenAssignment]:
 261         """ Yield possible assignments for the current sequence.
 262
 263             This function splits up general name assignments into name
 264             and address and yields all possible variants of that.
 265         """
 266         base = TokenAssignment.from_ranges(self.seq)
 267
 268         # Postcode search (postcode-only search is covered in next case)
 269         if base.postcode is not None and base.address:
 270             if (base.postcode.start == 0 and self.direction != -1)\
 271                or (base.postcode.end == query.num_token_slots() and self.direction != 1):
 272                 log().comment('postcode search')
 273                 yield dataclasses.replace(base, penalty=self.penalty)
 274
 275         # Postcode or country-only search
 276         if not base.address:
 277             if not base.housenumber and (base.postcode or base.country or base.category):
 278                 log().comment('postcode/country search')
 279                 yield dataclasses.replace(base, penalty=self.penalty)
 280         else:
 281             # Use entire first word as name
 282             if self.direction != -1:
 283                 log().comment('first word = name')
 284                 yield dataclasses.replace(base, name=base.address[0],
 285                                           penalty=self.penalty,
 286                                           address=base.address[1:])
 287
 288             # Use entire last word as name
 289             if self.direction == -1 or (self.direction == 0 and len(base.address) > 1):
 290                 log().comment('last word = name')
 291                 yield dataclasses.replace(base, name=base.address[-1],
 292                                           penalty=self.penalty,
 293                                           address=base.address[:-1])
 294
 295             # Use beginning of first word as name
 296             if self.direction != -1:
 297                 first = base.address[0]
 298                 if (not base.housenumber or first.end >= base.housenumber.start)\
 299                    and (not base.qualifier or first.start >= base.qualifier.end):
 300                     for i in range(first.start + 1, first.end):
 301                         name, addr = first.split(i)
 302                         penalty = self.penalty + PENALTY_TOKENCHANGE[query.nodes[i].btype]
 303                         log().comment(f'split first word = name ({i - first.start})')
 304                         yield dataclasses.replace(base, name=name, penalty=penalty,
 305                                                   address=[addr] + base.address[1:])
 306
 307             # Use end of last word as name
 308             if self.direction != 1:
 309                 last = base.address[-1]
 310                 if (not base.housenumber or last.start <= base.housenumber.end)\
 311                    and (not base.qualifier or last.end <= base.qualifier.start):
 312                     for i in range(last.start + 1, last.end):
 313                         addr, name = last.split(i)
 314                         penalty = self.penalty + PENALTY_TOKENCHANGE[query.nodes[i].btype]
 315                         log().comment(f'split last word = name ({i - last.start})')
 316                         yield dataclasses.replace(base, name=name, penalty=penalty,
 317                                                   address=base.address[:-1] + [addr])
 318
 319
 320
 321 def yield_token_assignments(query: qmod.QueryStruct) -> Iterator[TokenAssignment]:
 322     """ Return possible word type assignments to word positions.
 323
 324         The assignments are computed from the concrete tokens listed
 325         in the tokenized query.
 326
 327         The result includes the penalty for transitions from one word type to
 328         another. It does not include penalties for transitions within a
 329         type.
 330     """
 331     todo = [_TokenSequence([], direction=0 if query.source[0].ptype == qmod.PhraseType.NONE else 1)]
 332
 333     while todo:
 334         state = todo.pop()
 335         node = query.nodes[state.end_pos]
 336
 337         for tlist in node.starting:
 338             newstate = state.advance(tlist.ttype, tlist.end, node.btype)
 339             if newstate is not None:
 340                 if newstate.end_pos == query.num_token_slots():
 341                     if newstate.recheck_sequence():
 342                         log().var_dump('Assignment', newstate)
 343                         yield from newstate.get_assignments(query)
 344                 elif not newstate.is_final():
 345                     todo.append(newstate)