1 # SPDX-License-Identifier: GPL-3.0-or-later
3 # This file is part of Nominatim. (https://nominatim.org)
5 # Copyright (C) 2024 by the Nominatim developer community.
6 # For a full list of authors see the git log.
8 Implementation of query analysis for the ICU tokenizer.
10 from typing import Tuple, Dict, List, Optional, NamedTuple, Iterator, Any, cast
11 from collections import defaultdict
15 from icu import Transliterator
17 import sqlalchemy as sa
19 from ..typing import SaRow
20 from ..sql.sqlalchemy_types import Json
21 from ..connection import SearchConnection
22 from ..logging import log
23 from ..search import query as qmod
24 from ..search.query_analyzer_factory import AbstractQueryAnalyzer
28 'W': qmod.TokenType.WORD,
29 'w': qmod.TokenType.PARTIAL,
30 'H': qmod.TokenType.HOUSENUMBER,
31 'P': qmod.TokenType.POSTCODE,
32 'C': qmod.TokenType.COUNTRY
36 class QueryPart(NamedTuple):
37 """ Normalized and transliterated form of a single term in the query.
38 When the term came out of a split during the transliteration,
39 the normalized string is the full word before transliteration.
40 The word number keeps track of the word before transliteration
41 and can be used to identify partial transliterated terms.
48 QueryParts = List[QueryPart]
49 WordDict = Dict[str, List[qmod.TokenRange]]
51 def yield_words(terms: List[QueryPart], start: int) -> Iterator[Tuple[str, qmod.TokenRange]]:
52 """ Return all combinations of words in the terms list after the
56 for first in range(start, total):
57 word = terms[first].token
58 yield word, qmod.TokenRange(first, first + 1)
59 for last in range(first + 1, min(first + 20, total)):
60 word = ' '.join((word, terms[last].token))
61 yield word, qmod.TokenRange(first, last + 1)
64 @dataclasses.dataclass
65 class ICUToken(qmod.Token):
66 """ Specialised token for ICU tokenizer.
69 info: Optional[Dict[str, Any]]
71 def get_category(self) -> Tuple[str, str]:
73 return self.info.get('class', ''), self.info.get('type', '')
76 def rematch(self, norm: str) -> None:
77 """ Check how well the token matches the given normalized string
78 and add a penalty, if necessary.
80 if not self.lookup_word:
83 seq = difflib.SequenceMatcher(a=self.lookup_word, b=norm)
85 for tag, afrom, ato, bfrom, bto in seq.get_opcodes():
86 if tag in ('delete', 'insert') and (afrom == 0 or ato == len(self.lookup_word)):
88 elif tag == 'replace':
89 distance += max((ato-afrom), (bto-bfrom))
91 distance += abs((ato-afrom) - (bto-bfrom))
92 self.penalty += (distance/len(self.lookup_word))
96 def from_db_row(row: SaRow) -> 'ICUToken':
97 """ Create a ICUToken from the row of the word table.
99 count = 1 if row.info is None else row.info.get('count', 1)
100 addr_count = 1 if row.info is None else row.info.get('addr_count', 1)
105 elif row.type == 'W':
106 if len(row.word_token) == 1 and row.word_token == row.word:
107 penalty = 0.2 if row.word.isdigit() else 0.3
108 elif row.type == 'H':
109 penalty = sum(0.1 for c in row.word_token if c != ' ' and not c.isdigit())
110 if all(not c.isdigit() for c in row.word_token):
111 penalty += 0.2 * (len(row.word_token) - 1)
112 elif row.type == 'C':
113 if len(row.word_token) == 1:
117 lookup_word = row.word
119 lookup_word = row.info.get('lookup', row.word)
121 lookup_word = lookup_word.split('@', 1)[0]
123 lookup_word = row.word_token
125 return ICUToken(penalty=penalty, token=row.word_id, count=max(1, count),
126 lookup_word=lookup_word,
127 word_token=row.word_token, info=row.info,
128 addr_count=max(1, addr_count))
132 class ICUQueryAnalyzer(AbstractQueryAnalyzer):
133 """ Converter for query strings into a tokenized query
134 using the tokens created by a ICU tokenizer.
137 def __init__(self, conn: SearchConnection) -> None:
141 async def setup(self) -> None:
142 """ Set up static data structures needed for the analysis.
144 async def _make_normalizer() -> Any:
145 rules = await self.conn.get_property('tokenizer_import_normalisation')
146 return Transliterator.createFromRules("normalization", rules)
148 self.normalizer = await self.conn.get_cached_value('ICUTOK', 'normalizer',
151 async def _make_transliterator() -> Any:
152 rules = await self.conn.get_property('tokenizer_import_transliteration')
153 return Transliterator.createFromRules("transliteration", rules)
155 self.transliterator = await self.conn.get_cached_value('ICUTOK', 'transliterator',
156 _make_transliterator)
158 if 'word' not in self.conn.t.meta.tables:
159 sa.Table('word', self.conn.t.meta,
160 sa.Column('word_id', sa.Integer),
161 sa.Column('word_token', sa.Text, nullable=False),
162 sa.Column('type', sa.Text, nullable=False),
163 sa.Column('word', sa.Text),
164 sa.Column('info', Json))
167 async def analyze_query(self, phrases: List[qmod.Phrase]) -> qmod.QueryStruct:
168 """ Analyze the given list of phrases and return the
171 log().section('Analyze query (using ICU tokenizer)')
172 normalized = list(filter(lambda p: p.text,
173 (qmod.Phrase(p.ptype, self.normalize_text(p.text))
175 query = qmod.QueryStruct(normalized)
176 log().var_dump('Normalized query', query.source)
180 parts, words = self.split_query(query)
181 log().var_dump('Transliterated query', lambda: _dump_transliterated(query, parts))
183 for row in await self.lookup_in_db(list(words.keys())):
184 for trange in words[row.word_token]:
185 token = ICUToken.from_db_row(row)
187 if row.info['op'] in ('in', 'near'):
188 if trange.start == 0:
189 query.add_token(trange, qmod.TokenType.NEAR_ITEM, token)
191 if trange.start == 0 and trange.end == query.num_token_slots():
192 query.add_token(trange, qmod.TokenType.NEAR_ITEM, token)
194 query.add_token(trange, qmod.TokenType.QUALIFIER, token)
196 query.add_token(trange, DB_TO_TOKEN_TYPE[row.type], token)
198 self.add_extra_tokens(query, parts)
199 self.rerank_tokens(query, parts)
201 log().table_dump('Word tokens', _dump_word_tokens(query))
206 def normalize_text(self, text: str) -> str:
207 """ Bring the given text into a normalized form. That is the
208 standardized form search will work with. All information removed
209 at this stage is inevitably lost.
211 norm = cast(str, self.normalizer.transliterate(text))
212 numspaces = norm.count(' ')
213 if numspaces > 4 and len(norm) <= (numspaces + 1) * 3:
219 def split_query(self, query: qmod.QueryStruct) -> Tuple[QueryParts, WordDict]:
220 """ Transliterate the phrases and split them into tokens.
222 Returns the list of transliterated tokens together with their
223 normalized form and a dictionary of words for lookup together
226 parts: QueryParts = []
228 words = defaultdict(list)
230 for phrase in query.source:
231 query.nodes[-1].ptype = phrase.ptype
232 for word in phrase.text.split(' '):
233 trans = self.transliterator.transliterate(word)
235 for term in trans.split(' '):
237 parts.append(QueryPart(term, word, wordnr))
238 query.add_node(qmod.BreakType.TOKEN, phrase.ptype)
239 query.nodes[-1].btype = qmod.BreakType.WORD
241 query.nodes[-1].btype = qmod.BreakType.PHRASE
243 for word, wrange in yield_words(parts, phrase_start):
244 words[word].append(wrange)
246 phrase_start = len(parts)
247 query.nodes[-1].btype = qmod.BreakType.END
252 async def lookup_in_db(self, words: List[str]) -> 'sa.Result[Any]':
253 """ Return the token information from the database for the
256 t = self.conn.t.meta.tables['word']
257 return await self.conn.execute(t.select().where(t.c.word_token.in_(words)))
260 def add_extra_tokens(self, query: qmod.QueryStruct, parts: QueryParts) -> None:
261 """ Add tokens to query that are not saved in the database.
263 for part, node, i in zip(parts, query.nodes, range(1000)):
264 if len(part.token) <= 4 and part[0].isdigit()\
265 and not node.has_tokens(i+1, qmod.TokenType.HOUSENUMBER):
266 query.add_token(qmod.TokenRange(i, i+1), qmod.TokenType.HOUSENUMBER,
267 ICUToken(penalty=0.5, token=0,
268 count=1, addr_count=1, lookup_word=part.token,
269 word_token=part.token, info=None))
272 def rerank_tokens(self, query: qmod.QueryStruct, parts: QueryParts) -> None:
273 """ Add penalties to tokens that depend on presence of other token.
275 for i, node, tlist in query.iter_token_lists():
276 if tlist.ttype == qmod.TokenType.POSTCODE:
277 for repl in node.starting:
278 if repl.end == tlist.end and repl.ttype != qmod.TokenType.POSTCODE \
279 and (repl.ttype != qmod.TokenType.HOUSENUMBER
280 or len(tlist.tokens[0].lookup_word) > 4):
281 repl.add_penalty(0.39)
282 elif tlist.ttype == qmod.TokenType.HOUSENUMBER \
283 and len(tlist.tokens[0].lookup_word) <= 3:
284 if any(c.isdigit() for c in tlist.tokens[0].lookup_word):
285 for repl in node.starting:
286 if repl.end == tlist.end and repl.ttype != qmod.TokenType.HOUSENUMBER:
287 repl.add_penalty(0.5 - tlist.tokens[0].penalty)
288 elif tlist.ttype not in (qmod.TokenType.COUNTRY, qmod.TokenType.PARTIAL):
289 norm = parts[i].normalized
290 for j in range(i + 1, tlist.end):
291 if parts[j - 1].word_number != parts[j].word_number:
292 norm += ' ' + parts[j].normalized
293 for token in tlist.tokens:
294 cast(ICUToken, token).rematch(norm)
297 def _dump_transliterated(query: qmod.QueryStruct, parts: QueryParts) -> str:
298 out = query.nodes[0].btype.value
299 for node, part in zip(query.nodes[1:], parts):
300 out += part.token + node.btype.value
304 def _dump_word_tokens(query: qmod.QueryStruct) -> Iterator[List[Any]]:
305 yield ['type', 'token', 'word_token', 'lookup_word', 'penalty', 'count', 'info']
306 for node in query.nodes:
307 for tlist in node.starting:
308 for token in tlist.tokens:
309 t = cast(ICUToken, token)
310 yield [tlist.ttype.name, t.token, t.word_token or '',
311 t.lookup_word or '', t.penalty, t.count, t.info]
314 async def create_query_analyzer(conn: SearchConnection) -> AbstractQueryAnalyzer:
315 """ Create and set up a new query analyzer for a database based
316 on the ICU tokenizer.
318 out = ICUQueryAnalyzer(conn)