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