Merge remote-tracking branch 'upstream/master'

[nominatim.git] / nominatim / api / search / icu_tokenizer.py

diff --git a/nominatim/api/search/icu_tokenizer.py b/nominatim/api/search/icu_tokenizer.py
index 9bd16e1d85350de6d9879bd95af1c2d809f047b1..76a1a2e5d362688d5388044b511a9d3f0ae4a13c 100644 (file)
--- a/nominatim/api/search/icu_tokenizer.py
+++ b/nominatim/api/search/icu_tokenizer.py
@@ -8,7 +8,6 @@
 Implementation of query analysis for the ICU tokenizer.
 """
 from typing import Tuple, Dict, List, Optional, NamedTuple, Iterator, Any, cast
-from copy import copy
 from collections import defaultdict
 import dataclasses
 import difflib
@@ -22,6 +21,7 @@ from nominatim.api.connection import SearchConnection
 from nominatim.api.logging import log
 from nominatim.api.search import query as qmod
 from nominatim.api.search.query_analyzer_factory import AbstractQueryAnalyzer
+from nominatim.db.sqlalchemy_types import Json
 
 
 DB_TO_TOKEN_TYPE = {
@@ -83,7 +83,7 @@ class ICUToken(qmod.Token):
         seq = difflib.SequenceMatcher(a=self.lookup_word, b=norm)
         distance = 0
         for tag, afrom, ato, bfrom, bto in seq.get_opcodes():
-            if tag == 'delete' and (afrom == 0 or ato == len(self.lookup_word)):
+            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))
@@ -101,10 +101,16 @@ class ICUToken(qmod.Token):
         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
@@ -133,10 +139,19 @@ class ICUQueryAnalyzer(AbstractQueryAnalyzer):
     async def setup(self) -> None:
         """ Set up static data structures needed for the analysis.
         """
-        rules = await self.conn.get_property('tokenizer_import_normalisation')
-        self.normalizer = Transliterator.createFromRules("normalization", rules)
-        rules = await self.conn.get_property('tokenizer_import_transliteration')
-        self.transliterator = Transliterator.createFromRules("transliteration", rules)
+        async def _make_normalizer() -> Any:
+            rules = await self.conn.get_property('tokenizer_import_normalisation')
+            return Transliterator.createFromRules("normalization", rules)
+
+        self.normalizer = await self.conn.get_cached_value('ICUTOK', 'normalizer',
+                                                           _make_normalizer)
+
+        async def _make_transliterator() -> Any:
+            rules = await self.conn.get_property('tokenizer_import_transliteration')
+            return Transliterator.createFromRules("transliteration", rules)
+
+        self.transliterator = await self.conn.get_cached_value('ICUTOK', 'transliterator',
+                                                               _make_transliterator)
 
         if 'word' not in self.conn.t.meta.tables:
             sa.Table('word', self.conn.t.meta,
@@ -144,7 +159,7 @@ class ICUQueryAnalyzer(AbstractQueryAnalyzer):
                      sa.Column('word_token', sa.Text, nullable=False),
                      sa.Column('type', sa.Text, nullable=False),
                      sa.Column('word', sa.Text),
-                     sa.Column('info', self.conn.t.types.Json))
+                     sa.Column('info', Json))
 
 
     async def analyze_query(self, phrases: List[qmod.Phrase]) -> qmod.QueryStruct:
@@ -153,7 +168,7 @@ class ICUQueryAnalyzer(AbstractQueryAnalyzer):
         """
         log().section('Analyze query (using ICU tokenizer)')
         normalized = list(filter(lambda p: p.text,
-                                 (qmod.Phrase(p.ptype, self.normalizer.transliterate(p.text))
+                                 (qmod.Phrase(p.ptype, self.normalize_text(p.text))
                                   for p in phrases)))
         query = qmod.QueryStruct(normalized)
         log().var_dump('Normalized query', query.source)
@@ -169,13 +184,12 @@ class ICUQueryAnalyzer(AbstractQueryAnalyzer):
                 if row.type == 'S':
                     if row.info['op'] in ('in', 'near'):
                         if trange.start == 0:
-                            query.add_token(trange, qmod.TokenType.CATEGORY, token)
+                            query.add_token(trange, qmod.TokenType.NEAR_ITEM, token)
                     else:
-                        query.add_token(trange, qmod.TokenType.QUALIFIER, token)
-                        if trange.start == 0 or trange.end == query.num_token_slots():
-                            token = copy(token)
-                            token.penalty += 0.1 * (query.num_token_slots())
-                            query.add_token(trange, qmod.TokenType.CATEGORY, token)
+                        if trange.start == 0 and trange.end == query.num_token_slots():
+                            query.add_token(trange, qmod.TokenType.NEAR_ITEM, token)
+                        else:
+                            query.add_token(trange, qmod.TokenType.QUALIFIER, token)
                 else:
                     query.add_token(trange, DB_TO_TOKEN_TYPE[row.type], token)
 
@@ -187,6 +201,19 @@ 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
+            at this stage is inevitably lost.
+        """
+        norm = cast(str, self.normalizer.transliterate(text))
+        numspaces = norm.count(' ')
+        if numspaces > 4 and len(norm) <= (numspaces + 1) * 3:
+            return ''
+
+        return norm
+
+
     def split_query(self, query: qmod.QueryStruct) -> Tuple[QueryParts, WordDict]:
         """ Transliterate the phrases and split them into tokens.