src/nominatim_api/search/icu_tokenizer.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) 2024 by the Nominatim developer community.
   6 # For a full list of authors see the git log.
   7 """
   8 Implementation of query analysis for the ICU tokenizer.
   9 """
  10 from typing import Tuple, Dict, List, Optional, Iterator, Any, cast
  11 import dataclasses
  12 import difflib
  13 import re
  14 from itertools import zip_longest
  15
  16 from icu import Transliterator
  17
  18 import sqlalchemy as sa
  19
  20 from ..errors import UsageError
  21 from ..typing import SaRow
  22 from ..sql.sqlalchemy_types import Json
  23 from ..connection import SearchConnection
  24 from ..logging import log
  25 from . import query as qmod
  26 from ..query_preprocessing.config import QueryConfig
  27 from .query_analyzer_factory import AbstractQueryAnalyzer
  28 from .postcode_parser import PostcodeParser
  29
  30
  31 DB_TO_TOKEN_TYPE = {
  32     'W': qmod.TOKEN_WORD,
  33     'w': qmod.TOKEN_PARTIAL,
  34     'H': qmod.TOKEN_HOUSENUMBER,
  35     'P': qmod.TOKEN_POSTCODE,
  36     'C': qmod.TOKEN_COUNTRY
  37 }
  38
  39 PENALTY_IN_TOKEN_BREAK = {
  40      qmod.BREAK_START: 0.5,
  41      qmod.BREAK_END: 0.5,
  42      qmod.BREAK_PHRASE: 0.5,
  43      qmod.BREAK_SOFT_PHRASE: 0.5,
  44      qmod.BREAK_WORD: 0.1,
  45      qmod.BREAK_PART: 0.0,
  46      qmod.BREAK_TOKEN: 0.0
  47 }
  48
  49
  50 @dataclasses.dataclass
  51 class ICUToken(qmod.Token):
  52     """ Specialised token for ICU tokenizer.
  53     """
  54     word_token: str
  55     info: Optional[Dict[str, Any]]
  56
  57     def get_category(self) -> Tuple[str, str]:
  58         assert self.info
  59         return self.info.get('class', ''), self.info.get('type', '')
  60
  61     def rematch(self, norm: str) -> None:
  62         """ Check how well the token matches the given normalized string
  63             and add a penalty, if necessary.
  64         """
  65         if not self.lookup_word:
  66             return
  67
  68         seq = difflib.SequenceMatcher(a=self.lookup_word, b=norm)
  69         distance = 0
  70         for tag, afrom, ato, bfrom, bto in seq.get_opcodes():
  71             if tag in ('delete', 'insert') and (afrom == 0 or ato == len(self.lookup_word)):
  72                 distance += 1
  73             elif tag == 'replace':
  74                 distance += max((ato-afrom), (bto-bfrom))
  75             elif tag != 'equal':
  76                 distance += abs((ato-afrom) - (bto-bfrom))
  77         self.penalty += (distance/len(self.lookup_word))
  78
  79     @staticmethod
  80     def from_db_row(row: SaRow, base_penalty: float = 0.0) -> 'ICUToken':
  81         """ Create a ICUToken from the row of the word table.
  82         """
  83         count = 1 if row.info is None else row.info.get('count', 1)
  84         addr_count = 1 if row.info is None else row.info.get('addr_count', 1)
  85
  86         penalty = base_penalty
  87         if row.type == 'w':
  88             penalty += 0.3
  89         elif row.type == 'W':
  90             if len(row.word_token) == 1 and row.word_token == row.word:
  91                 penalty += 0.2 if row.word.isdigit() else 0.3
  92         elif row.type == 'H':
  93             penalty += sum(0.1 for c in row.word_token if c != ' ' and not c.isdigit())
  94             if all(not c.isdigit() for c in row.word_token):
  95                 penalty += 0.2 * (len(row.word_token) - 1)
  96         elif row.type == 'C':
  97             if len(row.word_token) == 1:
  98                 penalty += 0.3
  99
 100         if row.info is None:
 101             lookup_word = row.word
 102         else:
 103             lookup_word = row.info.get('lookup', row.word)
 104         if lookup_word:
 105             lookup_word = lookup_word.split('@', 1)[0]
 106         else:
 107             lookup_word = row.word_token
 108
 109         return ICUToken(penalty=penalty, token=row.word_id, count=max(1, count),
 110                         lookup_word=lookup_word,
 111                         word_token=row.word_token, info=row.info,
 112                         addr_count=max(1, addr_count))
 113
 114
 115 class ICUQueryAnalyzer(AbstractQueryAnalyzer):
 116     """ Converter for query strings into a tokenized query
 117         using the tokens created by a ICU tokenizer.
 118     """
 119     def __init__(self, conn: SearchConnection) -> None:
 120         self.conn = conn
 121         self.postcode_parser = PostcodeParser(conn.config)
 122
 123     async def setup(self) -> None:
 124         """ Set up static data structures needed for the analysis.
 125         """
 126         async def _make_normalizer() -> Any:
 127             rules = await self.conn.get_property('tokenizer_import_normalisation')
 128             return Transliterator.createFromRules("normalization", rules)
 129
 130         self.normalizer = await self.conn.get_cached_value('ICUTOK', 'normalizer',
 131                                                            _make_normalizer)
 132
 133         async def _make_transliterator() -> Any:
 134             rules = await self.conn.get_property('tokenizer_import_transliteration')
 135             return Transliterator.createFromRules("transliteration", rules)
 136
 137         self.transliterator = await self.conn.get_cached_value('ICUTOK', 'transliterator',
 138                                                                _make_transliterator)
 139
 140         await self._setup_preprocessing()
 141
 142         if 'word' not in self.conn.t.meta.tables:
 143             sa.Table('word', self.conn.t.meta,
 144                      sa.Column('word_id', sa.Integer),
 145                      sa.Column('word_token', sa.Text, nullable=False),
 146                      sa.Column('type', sa.Text, nullable=False),
 147                      sa.Column('word', sa.Text),
 148                      sa.Column('info', Json))
 149
 150     async def _setup_preprocessing(self) -> None:
 151         """ Load the rules for preprocessing and set up the handlers.
 152         """
 153
 154         rules = self.conn.config.load_sub_configuration('icu_tokenizer.yaml',
 155                                                         config='TOKENIZER_CONFIG')
 156         preprocessing_rules = rules.get('query-preprocessing', [])
 157
 158         self.preprocessors = []
 159
 160         for func in preprocessing_rules:
 161             if 'step' not in func:
 162                 raise UsageError("Preprocessing rule is missing the 'step' attribute.")
 163             if not isinstance(func['step'], str):
 164                 raise UsageError("'step' attribute must be a simple string.")
 165
 166             module = self.conn.config.load_plugin_module(
 167                         func['step'], 'nominatim_api.query_preprocessing')
 168             self.preprocessors.append(
 169                 module.create(QueryConfig(func).set_normalizer(self.normalizer)))
 170
 171     async def analyze_query(self, phrases: List[qmod.Phrase]) -> qmod.QueryStruct:
 172         """ Analyze the given list of phrases and return the
 173             tokenized query.
 174         """
 175         log().section('Analyze query (using ICU tokenizer)')
 176         for func in self.preprocessors:
 177             phrases = func(phrases)
 178         query = qmod.QueryStruct(phrases)
 179
 180         log().var_dump('Normalized query', query.source)
 181         if not query.source:
 182             return query
 183
 184         self.split_query(query)
 185         log().var_dump('Transliterated query', lambda: query.get_transliterated_query())
 186         words = query.extract_words(base_penalty=PENALTY_IN_TOKEN_BREAK[qmod.BREAK_WORD])
 187
 188         for row in await self.lookup_in_db(list(words.keys())):
 189             for trange in words[row.word_token]:
 190                 token = ICUToken.from_db_row(row, trange.penalty or 0.0)
 191                 if row.type == 'S':
 192                     if row.info['op'] in ('in', 'near'):
 193                         if trange.start == 0:
 194                             query.add_token(trange, qmod.TOKEN_NEAR_ITEM, token)
 195                     else:
 196                         if trange.start == 0 and trange.end == query.num_token_slots():
 197                             query.add_token(trange, qmod.TOKEN_NEAR_ITEM, token)
 198                         else:
 199                             query.add_token(trange, qmod.TOKEN_QUALIFIER, token)
 200                 else:
 201                     query.add_token(trange, DB_TO_TOKEN_TYPE[row.type], token)
 202
 203         self.add_extra_tokens(query)
 204         for start, end, pc in self.postcode_parser.parse(query):
 205             query.add_token(qmod.TokenRange(start, end),
 206                             qmod.TOKEN_POSTCODE,
 207                             ICUToken(penalty=0.1, token=0, count=1, addr_count=1,
 208                                      lookup_word=pc, word_token=pc, info=None))
 209         self.rerank_tokens(query)
 210
 211         log().table_dump('Word tokens', _dump_word_tokens(query))
 212
 213         return query
 214
 215     def normalize_text(self, text: str) -> str:
 216         """ Bring the given text into a normalized form. That is the
 217             standardized form search will work with. All information removed
 218             at this stage is inevitably lost.
 219         """
 220         return cast(str, self.normalizer.transliterate(text)).strip('-: ')
 221
 222     def split_query(self, query: qmod.QueryStruct) -> None:
 223         """ Transliterate the phrases and split them into tokens.
 224         """
 225         for phrase in query.source:
 226             query.nodes[-1].ptype = phrase.ptype
 227             phrase_split = re.split('([ :-])', phrase.text)
 228             # The zip construct will give us the pairs of word/break from
 229             # the regular expression split. As the split array ends on the
 230             # final word, we simply use the fillvalue to even out the list and
 231             # add the phrase break at the end.
 232             for word, breakchar in zip_longest(*[iter(phrase_split)]*2, fillvalue=','):
 233                 if not word:
 234                     continue
 235                 trans = self.transliterator.transliterate(word)
 236                 if trans:
 237                     for term in trans.split(' '):
 238                         if term:
 239                             query.add_node(qmod.BREAK_TOKEN, phrase.ptype,
 240                                            PENALTY_IN_TOKEN_BREAK[qmod.BREAK_TOKEN],
 241                                            term, word)
 242                     query.nodes[-1].adjust_break(breakchar,
 243                                                  PENALTY_IN_TOKEN_BREAK[breakchar])
 244
 245         query.nodes[-1].adjust_break(qmod.BREAK_END, PENALTY_IN_TOKEN_BREAK[qmod.BREAK_END])
 246
 247     async def lookup_in_db(self, words: List[str]) -> 'sa.Result[Any]':
 248         """ Return the token information from the database for the
 249             given word tokens.
 250
 251             This function excludes postcode tokens
 252         """
 253         t = self.conn.t.meta.tables['word']
 254         return await self.conn.execute(t.select()
 255                                         .where(t.c.word_token.in_(words))
 256                                         .where(t.c.type != 'P'))
 257
 258     def add_extra_tokens(self, query: qmod.QueryStruct) -> None:
 259         """ Add tokens to query that are not saved in the database.
 260         """
 261         need_hnr = False
 262         for i, node in enumerate(query.nodes):
 263             is_full_token = node.btype not in (qmod.BREAK_TOKEN, qmod.BREAK_PART)
 264             if need_hnr and is_full_token \
 265                     and len(node.term_normalized) <= 4 and node.term_normalized.isdigit():
 266                 query.add_token(qmod.TokenRange(i-1, i), qmod.TOKEN_HOUSENUMBER,
 267                                 ICUToken(penalty=0.5, token=0,
 268                                          count=1, addr_count=1,
 269                                          lookup_word=node.term_lookup,
 270                                          word_token=node.term_lookup, info=None))
 271
 272             need_hnr = is_full_token and not node.has_tokens(i+1, qmod.TOKEN_HOUSENUMBER)
 273
 274     def rerank_tokens(self, query: qmod.QueryStruct) -> None:
 275         """ Add penalties to tokens that depend on presence of other token.
 276         """
 277         for i, node, tlist in query.iter_token_lists():
 278             if tlist.ttype == qmod.TOKEN_POSTCODE:
 279                 for repl in node.starting:
 280                     if repl.end == tlist.end and repl.ttype != qmod.TOKEN_POSTCODE \
 281                        and (repl.ttype != qmod.TOKEN_HOUSENUMBER
 282                             or len(tlist.tokens[0].lookup_word) > 4):
 283                         repl.add_penalty(0.39)
 284             elif (tlist.ttype == qmod.TOKEN_HOUSENUMBER
 285                   and len(tlist.tokens[0].lookup_word) <= 3):
 286                 if any(c.isdigit() for c in tlist.tokens[0].lookup_word):
 287                     for repl in node.starting:
 288                         if repl.end == tlist.end and repl.ttype != qmod.TOKEN_HOUSENUMBER:
 289                             repl.add_penalty(0.5 - tlist.tokens[0].penalty)
 290             elif tlist.ttype not in (qmod.TOKEN_COUNTRY, qmod.TOKEN_PARTIAL):
 291                 norm = ' '.join(n.term_normalized for n in query.nodes[i + 1:tlist.end + 1]
 292                                 if n.btype != qmod.BREAK_TOKEN)
 293                 if not norm:
 294                     # Can happen when the token only covers a partial term
 295                     norm = query.nodes[i + 1].term_normalized
 296                 for token in tlist.tokens:
 297                     cast(ICUToken, token).rematch(norm)
 298
 299
 300 def _dump_word_tokens(query: qmod.QueryStruct) -> Iterator[List[Any]]:
 301     yield ['type', 'from', 'to', 'token', 'word_token', 'lookup_word', 'penalty', 'count', 'info']
 302     for i, node in enumerate(query.nodes):
 303         for tlist in node.starting:
 304             for token in tlist.tokens:
 305                 t = cast(ICUToken, token)
 306                 yield [tlist.ttype, str(i), str(tlist.end), t.token, t.word_token or '',
 307                        t.lookup_word or '', t.penalty, t.count, t.info]
 308
 309
 310 async def create_query_analyzer(conn: SearchConnection) -> AbstractQueryAnalyzer:
 311     """ Create and set up a new query analyzer for a database based
 312         on the ICU tokenizer.
 313     """
 314     out = ICUQueryAnalyzer(conn)
 315     await out.setup()
 316
 317     return out