X-Git-Url: https://git.openstreetmap.org./nominatim.git/blobdiff_plain/d856788bf5cf7b4b3e2b4de10f37cf3002ae6826..d984100e23d28253b3659e7b628cde2c8b436cf8:/src/nominatim_api/search/icu_tokenizer.py diff --git a/src/nominatim_api/search/icu_tokenizer.py b/src/nominatim_api/search/icu_tokenizer.py index 1aadc97e..6f1dcf79 100644 --- a/src/nominatim_api/search/icu_tokenizer.py +++ b/src/nominatim_api/search/icu_tokenizer.py @@ -7,21 +7,25 @@ """ Implementation of query analysis for the ICU tokenizer. """ -from typing import Tuple, Dict, List, Optional, NamedTuple, Iterator, Any, cast +from typing import Tuple, Dict, List, Optional, Iterator, Any, cast from collections import defaultdict 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 ..search import query as qmod -from ..search.query_analyzer_factory import AbstractQueryAnalyzer +from . import query as qmod +from ..query_preprocessing.config import QueryConfig +from .query_analyzer_factory import AbstractQueryAnalyzer DB_TO_TOKEN_TYPE = { @@ -32,22 +36,36 @@ DB_TO_TOKEN_TYPE = { 'C': qmod.TokenType.COUNTRY } +PENALTY_IN_TOKEN_BREAK = { + qmod.BreakType.START: 0.5, + qmod.BreakType.END: 0.5, + qmod.BreakType.PHRASE: 0.5, + qmod.BreakType.SOFT_PHRASE: 0.5, + qmod.BreakType.WORD: 0.1, + qmod.BreakType.PART: 0.0, + qmod.BreakType.TOKEN: 0.0 +} + -class QueryPart(NamedTuple): +@dataclasses.dataclass +class QueryPart: """ Normalized and transliterated form of a single term in the query. When the term came out of a split during the transliteration, the normalized string is the full word before transliteration. The word number keeps track of the word before transliteration and can be used to identify partial transliterated terms. + Penalty is the break penalty for the break following the token. """ token: str normalized: str word_number: int + penalty: float QueryParts = List[QueryPart] WordDict = Dict[str, List[qmod.TokenRange]] + def yield_words(terms: List[QueryPart], start: int) -> Iterator[Tuple[str, qmod.TokenRange]]: """ Return all combinations of words in the terms list after the given position. @@ -55,10 +73,12 @@ def yield_words(terms: List[QueryPart], start: int) -> Iterator[Tuple[str, qmod. total = len(terms) for first in range(start, total): word = terms[first].token - yield word, qmod.TokenRange(first, first + 1) + penalty = PENALTY_IN_TOKEN_BREAK[qmod.BreakType.WORD] + yield word, qmod.TokenRange(first, first + 1, penalty=penalty) for last in range(first + 1, min(first + 20, total)): word = ' '.join((word, terms[last].token)) - yield word, qmod.TokenRange(first, last + 1) + penalty += terms[last - 1].penalty + yield word, qmod.TokenRange(first, last + 1, penalty=penalty) @dataclasses.dataclass @@ -72,7 +92,6 @@ class ICUToken(qmod.Token): assert self.info return self.info.get('class', ''), self.info.get('type', '') - def rematch(self, norm: str) -> None: """ Check how well the token matches the given normalized string and add a penalty, if necessary. @@ -91,27 +110,26 @@ class ICUToken(qmod.Token): distance += abs((ato-afrom) - (bto-bfrom)) self.penalty += (distance/len(self.lookup_word)) - @staticmethod - def from_db_row(row: SaRow) -> 'ICUToken': + def from_db_row(row: SaRow, base_penalty: float = 0.0) -> '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 + penalty = base_penalty if row.type == 'w': - penalty = 0.3 + 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 + 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()) + 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 + penalty += 0.3 if row.info is None: lookup_word = row.word @@ -128,16 +146,13 @@ class ICUToken(qmod.Token): addr_count=max(1, addr_count)) - class ICUQueryAnalyzer(AbstractQueryAnalyzer): """ Converter for query strings into a tokenized query using the tokens created by a ICU tokenizer. """ - def __init__(self, conn: SearchConnection) -> None: self.conn = conn - async def setup(self) -> None: """ Set up static data structures needed for the analysis. """ @@ -155,6 +170,8 @@ class ICUQueryAnalyzer(AbstractQueryAnalyzer): self.transliterator = await self.conn.get_cached_value('ICUTOK', 'transliterator', _make_transliterator) + await self._setup_preprocessing() + if 'word' not in self.conn.t.meta.tables: sa.Table('word', self.conn.t.meta, sa.Column('word_id', sa.Integer), @@ -163,16 +180,36 @@ class ICUQueryAnalyzer(AbstractQueryAnalyzer): sa.Column('word', sa.Text), sa.Column('info', Json)) + async def _setup_preprocessing(self) -> None: + """ Load the rules for preprocessing and set up the handlers. + """ + + rules = self.conn.config.load_sub_configuration('icu_tokenizer.yaml', + config='TOKENIZER_CONFIG') + preprocessing_rules = rules.get('query-preprocessing', []) + + self.preprocessors = [] + + 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 = self.conn.config.load_plugin_module( + func['step'], 'nominatim_api.query_preprocessing') + self.preprocessors.append( + module.create(QueryConfig(func).set_normalizer(self.normalizer))) 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)') - normalized = list(filter(lambda p: p.text, - (qmod.Phrase(p.ptype, self.normalize_text(p.text)) - for p in phrases))) - query = qmod.QueryStruct(normalized) + 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 @@ -182,7 +219,7 @@ class ICUQueryAnalyzer(AbstractQueryAnalyzer): for row in await self.lookup_in_db(list(words.keys())): for trange in words[row.word_token]: - token = ICUToken.from_db_row(row) + token = ICUToken.from_db_row(row, trange.penalty or 0.0) if row.type == 'S': if row.info['op'] in ('in', 'near'): if trange.start == 0: @@ -202,7 +239,6 @@ class ICUQueryAnalyzer(AbstractQueryAnalyzer): 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 @@ -210,7 +246,6 @@ class ICUQueryAnalyzer(AbstractQueryAnalyzer): """ return cast(str, self.normalizer.transliterate(text)) - def split_query(self, query: qmod.QueryStruct) -> Tuple[QueryParts, WordDict]: """ Transliterate the phrases and split them into tokens. @@ -224,16 +259,24 @@ class ICUQueryAnalyzer(AbstractQueryAnalyzer): wordnr = 0 for phrase in query.source: query.nodes[-1].ptype = phrase.ptype - for word in phrase.text.split(' '): + 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 trans = self.transliterator.transliterate(word) if trans: for term in trans.split(' '): if term: - parts.append(QueryPart(term, word, wordnr)) + parts.append(QueryPart(term, word, wordnr, + PENALTY_IN_TOKEN_BREAK[qmod.BreakType.TOKEN])) query.add_node(qmod.BreakType.TOKEN, phrase.ptype) - query.nodes[-1].btype = qmod.BreakType.WORD + query.nodes[-1].btype = qmod.BreakType(breakchar) + parts[-1].penalty = PENALTY_IN_TOKEN_BREAK[qmod.BreakType(breakchar)] wordnr += 1 - query.nodes[-1].btype = qmod.BreakType.PHRASE for word, wrange in yield_words(parts, phrase_start): words[word].append(wrange) @@ -243,7 +286,6 @@ class ICUQueryAnalyzer(AbstractQueryAnalyzer): return parts, words - async def lookup_in_db(self, words: List[str]) -> 'sa.Result[Any]': """ Return the token information from the database for the given word tokens. @@ -251,19 +293,17 @@ class ICUQueryAnalyzer(AbstractQueryAnalyzer): t = self.conn.t.meta.tables['word'] return await self.conn.execute(t.select().where(t.c.word_token.in_(words))) - def add_extra_tokens(self, query: qmod.QueryStruct, parts: QueryParts) -> None: """ Add tokens to query that are not saved in the database. """ for part, node, i in zip(parts, query.nodes, range(1000)): - if len(part.token) <= 4 and part[0].isdigit()\ + if len(part.token) <= 4 and part.token.isdigit()\ and not node.has_tokens(i+1, qmod.TokenType.HOUSENUMBER): query.add_token(qmod.TokenRange(i, i+1), qmod.TokenType.HOUSENUMBER, ICUToken(penalty=0.5, token=0, count=1, addr_count=1, lookup_word=part.token, word_token=part.token, info=None)) - def rerank_tokens(self, query: qmod.QueryStruct, parts: QueryParts) -> None: """ Add penalties to tokens that depend on presence of other token. """ @@ -274,8 +314,8 @@ class ICUQueryAnalyzer(AbstractQueryAnalyzer): and (repl.ttype != qmod.TokenType.HOUSENUMBER or len(tlist.tokens[0].lookup_word) > 4): repl.add_penalty(0.39) - elif tlist.ttype == qmod.TokenType.HOUSENUMBER \ - and len(tlist.tokens[0].lookup_word) <= 3: + elif (tlist.ttype == qmod.TokenType.HOUSENUMBER + and len(tlist.tokens[0].lookup_word) <= 3): if any(c.isdigit() for c in tlist.tokens[0].lookup_word): for repl in node.starting: if repl.end == tlist.end and repl.ttype != qmod.TokenType.HOUSENUMBER: