# SPDX-License-Identifier: GPL-3.0-or-later # # This file is part of Nominatim. (https://nominatim.org) # # Copyright (C) 2025 by the Nominatim developer community. # For a full list of authors see the git log. """ Implementation of query analysis for the ICU tokenizer. """ from typing import Tuple, Dict, List, Optional, Iterator, Any, cast import dataclasses import difflib import re from itertools import zip_longest from icu import Transliterator import sqlalchemy as sa from ..errors import UsageError from ..typing import SaRow from ..sql.sqlalchemy_types import Json from ..connection import SearchConnection from ..logging import log from . import query as qmod from ..query_preprocessing.config import QueryConfig from ..query_preprocessing.base import QueryProcessingFunc from .query_analyzer_factory import AbstractQueryAnalyzer from .postcode_parser import PostcodeParser DB_TO_TOKEN_TYPE = { 'W': qmod.TOKEN_WORD, 'w': qmod.TOKEN_PARTIAL, 'H': qmod.TOKEN_HOUSENUMBER, 'P': qmod.TOKEN_POSTCODE, 'C': qmod.TOKEN_COUNTRY } PENALTY_BREAK = { qmod.BREAK_START: -0.5, qmod.BREAK_END: -0.5, qmod.BREAK_PHRASE: -0.5, qmod.BREAK_SOFT_PHRASE: -0.5, qmod.BREAK_WORD: 0.1, qmod.BREAK_PART: 0.2, qmod.BREAK_TOKEN: 0.4 } @dataclasses.dataclass class ICUToken(qmod.Token): """ Specialised token for ICU tokenizer. """ word_token: str info: Optional[Dict[str, Any]] def get_category(self) -> Tuple[str, str]: assert self.info return self.info.get('class', ''), self.info.get('type', '') def get_country(self) -> str: assert self.info return cast(str, self.info.get('cc', '')) def match_penalty(self, norm: str) -> float: """ Check how well the token matches the given normalized string and add a penalty, if necessary. """ if not self.lookup_word: return 0.0 seq = difflib.SequenceMatcher(a=self.lookup_word, b=norm) distance = 0 for tag, afrom, ato, bfrom, bto in seq.get_opcodes(): if tag in ('delete', 'insert') and (afrom == 0 or ato == len(self.lookup_word)): distance += 1 elif tag == 'replace': distance += max((ato-afrom), (bto-bfrom)) elif tag != 'equal': distance += abs((ato-afrom) - (bto-bfrom)) return (distance/len(self.lookup_word)) @staticmethod def from_db_row(row: SaRow) -> 'ICUToken': """ Create a ICUToken from the row of the word table. """ count = 1 if row.info is None else row.info.get('count', 1) addr_count = 1 if row.info is None else row.info.get('addr_count', 1) penalty = 0.0 if row.type == 'w': penalty += 0.3 elif row.type == 'W': if len(row.word_token) == 1 and row.word_token == row.word: penalty += 0.2 if row.word.isdigit() else 0.3 elif row.type == 'H': penalty += sum(0.1 for c in row.word_token if c != ' ' and not c.isdigit()) if all(not c.isdigit() for c in row.word_token): penalty += 0.2 * (len(row.word_token) - 1) elif row.type == 'C': if len(row.word_token) == 1: penalty += 0.3 if row.info is None: lookup_word = row.word else: lookup_word = row.info.get('lookup', row.word) if lookup_word: lookup_word = lookup_word.split('@', 1)[0] else: lookup_word = row.word_token return ICUToken(penalty=penalty, token=row.word_id, count=max(1, count), lookup_word=lookup_word, word_token=row.word_token, info=row.info, addr_count=max(1, addr_count)) @dataclasses.dataclass class ICUAnalyzerConfig: postcode_parser: PostcodeParser normalizer: Transliterator transliterator: Transliterator preprocessors: List[QueryProcessingFunc] @staticmethod async def create(conn: SearchConnection) -> 'ICUAnalyzerConfig': rules = await conn.get_property('tokenizer_import_normalisation') normalizer = Transliterator.createFromRules("normalization", rules) rules = await conn.get_property('tokenizer_import_transliteration') transliterator = Transliterator.createFromRules("transliteration", rules) preprocessing_rules = conn.config.load_sub_configuration('icu_tokenizer.yaml', config='TOKENIZER_CONFIG')\ .get('query-preprocessing', []) preprocessors: List[QueryProcessingFunc] = [] for func in preprocessing_rules: if 'step' not in func: raise UsageError("Preprocessing rule is missing the 'step' attribute.") if not isinstance(func['step'], str): raise UsageError("'step' attribute must be a simple string.") module = conn.config.load_plugin_module( func['step'], 'nominatim_api.query_preprocessing') preprocessors.append( module.create(QueryConfig(func).set_normalizer(normalizer))) return ICUAnalyzerConfig(PostcodeParser(conn.config), normalizer, transliterator, preprocessors) class ICUQueryAnalyzer(AbstractQueryAnalyzer): """ Converter for query strings into a tokenized query using the tokens created by a ICU tokenizer. """ def __init__(self, conn: SearchConnection, config: ICUAnalyzerConfig) -> None: self.conn = conn self.postcode_parser = config.postcode_parser self.normalizer = config.normalizer self.transliterator = config.transliterator self.preprocessors = config.preprocessors async def analyze_query(self, phrases: List[qmod.Phrase]) -> qmod.QueryStruct: """ Analyze the given list of phrases and return the tokenized query. """ log().section('Analyze query (using ICU tokenizer)') for func in self.preprocessors: phrases = func(phrases) query = qmod.QueryStruct(phrases) log().var_dump('Normalized query', query.source) if not query.source: return query self.split_query(query) log().var_dump('Transliterated query', lambda: ''.join(f"{n.term_lookup}{n.btype}" for n in query.nodes) + ' / ' + ''.join(f"{n.term_normalized}{n.btype}" for n in query.nodes)) words = query.extract_words() for row in await self.lookup_in_db(list(words.keys())): for trange in words[row.word_token]: # Create a new token for each position because the token # penalty can vary depending on the position in the query. # (See rerank_tokens() below.) token = ICUToken.from_db_row(row) if row.type == 'S': if row.info['op'] in ('in', 'near'): if trange.start == 0: query.add_token(trange, qmod.TOKEN_NEAR_ITEM, token) else: if trange.start == 0 and trange.end == query.num_token_slots(): query.add_token(trange, qmod.TOKEN_NEAR_ITEM, token) else: query.add_token(trange, qmod.TOKEN_QUALIFIER, token) else: query.add_token(trange, DB_TO_TOKEN_TYPE[row.type], token) self.add_extra_tokens(query) for start, end, pc in self.postcode_parser.parse(query): term = ' '.join(n.term_lookup for n in query.nodes[start + 1:end + 1]) query.add_token(qmod.TokenRange(start, end), qmod.TOKEN_POSTCODE, ICUToken(penalty=0.0, token=0, count=1, addr_count=1, lookup_word=pc, word_token=term, info=None)) self.rerank_tokens(query) self.compute_break_penalties(query) log().table_dump('Word tokens', _dump_word_tokens(query)) return query def normalize_text(self, text: str) -> str: """ Bring the given text into a normalized form. That is the standardized form search will work with. All information removed at this stage is inevitably lost. """ return cast(str, self.normalizer.transliterate(text)).strip('-: ') def split_transliteration(self, trans: str, word: str) -> list[tuple[str, str]]: """ Split the given transliteration string into sub-words and return them together with the original part of the word. """ subwords = trans.split(' ') if len(subwords) == 1: return [(trans, word)] tlist = [] titer = filter(None, subwords) current_trans: Optional[str] = next(titer) assert current_trans current_word = '' for letter in word: current_word += letter if self.transliterator.transliterate(current_word).rstrip() == current_trans: tlist.append((current_trans, current_word)) current_trans = next(titer, None) if current_trans is None: return tlist current_word = '' if current_word: tlist.append((current_trans, current_word)) return tlist def split_query(self, query: qmod.QueryStruct) -> None: """ Transliterate the phrases and split them into tokens. """ for phrase in query.source: query.nodes[-1].ptype = phrase.ptype phrase_split = re.split('([ :-])', phrase.text) # The zip construct will give us the pairs of word/break from # the regular expression split. As the split array ends on the # final word, we simply use the fillvalue to even out the list and # add the phrase break at the end. for word, breakchar in zip_longest(*[iter(phrase_split)]*2, fillvalue=','): if not word: continue if trans := self.transliterator.transliterate(word): for term, term_word in self.split_transliteration(trans, word): if term: query.add_node(qmod.BREAK_TOKEN, phrase.ptype, term, term_word) query.nodes[-1].btype = breakchar query.nodes[-1].btype = qmod.BREAK_END async def lookup_in_db(self, words: List[str]) -> 'sa.Result[Any]': """ Return the token information from the database for the given word tokens. This function excludes postcode tokens """ t = self.conn.t.meta.tables['word'] return await self.conn.execute(t.select() .where(t.c.word_token.in_(words)) .where(t.c.type != 'P')) def add_extra_tokens(self, query: qmod.QueryStruct) -> None: """ Add tokens to query that are not saved in the database. """ need_hnr = False for i, node in enumerate(query.nodes): is_full_token = node.btype not in (qmod.BREAK_TOKEN, qmod.BREAK_PART) if need_hnr and is_full_token \ and len(node.term_normalized) <= 4 and node.term_normalized.isdigit(): query.add_token(qmod.TokenRange(i-1, i), qmod.TOKEN_HOUSENUMBER, ICUToken(penalty=0.2, token=0, count=1, addr_count=1, lookup_word=node.term_lookup, word_token=node.term_lookup, info=None)) need_hnr = is_full_token and not node.has_tokens(i+1, qmod.TOKEN_HOUSENUMBER) def rerank_tokens(self, query: qmod.QueryStruct) -> None: """ Add penalties to tokens that depend on presence of other token. """ for start, end, tlist in query.iter_tokens_by_edge(): if len(tlist) > 1: # If it looks like a Postcode, give preference. if qmod.TOKEN_POSTCODE in tlist: for ttype, tokens in tlist.items(): if ttype != qmod.TOKEN_POSTCODE and \ (ttype != qmod.TOKEN_HOUSENUMBER or start + 1 > end or len(query.nodes[end].term_lookup) > 4): for token in tokens: token.penalty += 0.39 if (start + 1 == end): if partial := query.nodes[start].partial: partial.penalty += 0.39 # If it looks like a simple housenumber, prefer that. if qmod.TOKEN_HOUSENUMBER in tlist: hnr_lookup = tlist[qmod.TOKEN_HOUSENUMBER][0].lookup_word if len(hnr_lookup) <= 3 and any(c.isdigit() for c in hnr_lookup): penalty = 0.5 - tlist[qmod.TOKEN_HOUSENUMBER][0].penalty for ttype, tokens in tlist.items(): if ttype != qmod.TOKEN_HOUSENUMBER: for token in tokens: token.penalty += penalty if (start + 1 == end): if partial := query.nodes[start].partial: partial.penalty += penalty # rerank tokens against the normalized form norm = ''.join(f"{n.term_normalized}{'' if n.btype == qmod.BREAK_TOKEN else ' '}" for n in query.nodes[start + 1:end + 1]).strip() for ttype, tokens in tlist.items(): for token in tokens: itok = cast(ICUToken, token) itok.penalty += itok.match_penalty(norm) * \ (1 if ttype in (qmod.TOKEN_WORD, qmod.TOKEN_PARTIAL) else 2) def compute_break_penalties(self, query: qmod.QueryStruct) -> None: """ Set the break penalties for the nodes in the query. """ for node in query.nodes: node.penalty = PENALTY_BREAK[node.btype] def _dump_word_tokens(query: qmod.QueryStruct) -> Iterator[List[Any]]: yield ['type', 'from', 'to', 'token', 'word_token', 'lookup_word', 'penalty', 'count', 'info'] for i, node in enumerate(query.nodes): if node.partial is not None: t = cast(ICUToken, node.partial) yield [qmod.TOKEN_PARTIAL, str(i), str(i + 1), t.token, t.word_token, t.lookup_word, t.penalty, t.count, t.info] for tlist in node.starting: for token in tlist.tokens: t = cast(ICUToken, token) yield [tlist.ttype, str(i), str(tlist.end), t.token, t.word_token or '', t.lookup_word or '', t.penalty, t.count, t.info] async def create_query_analyzer(conn: SearchConnection) -> AbstractQueryAnalyzer: """ Create and set up a new query analyzer for a database based on the ICU tokenizer. """ async def _get_config() -> ICUAnalyzerConfig: if 'word' not in conn.t.meta.tables: sa.Table('word', conn.t.meta, sa.Column('word_id', sa.Integer), sa.Column('word_token', sa.Text, nullable=False), sa.Column('type', sa.Text, nullable=False), sa.Column('word', sa.Text), sa.Column('info', Json)) return await ICUAnalyzerConfig.create(conn) config = await conn.get_cached_value('ICUTOK', 'config', _get_config) return ICUQueryAnalyzer(conn, config)