add direction penalties

Direction penalties are estimated by getting the name to address
ratio usage for each partial term in the query and computing the
linear regression of that ratio over the entire phrase. Or to put
it in ither words: we try to determine if the terms at the beginning
or the end of the query are more likely to constitute a name.

Direction penalties are currently used only in classic name queries.
This commit is contained in:
Sarah Hoffmann
2025-04-11 20:41:06 +02:00
parent 2ef0e20a3f
commit 06e39e42d8
3 changed files with 58 additions and 8 deletions

View File

@@ -2,7 +2,7 @@
#
# This file is part of Nominatim. (https://nominatim.org)
#
# Copyright (C) 2024 by the Nominatim developer community.
# Copyright (C) 2025 by the Nominatim developer community.
# For a full list of authors see the git log.
"""
Public interface to the search code.
@@ -50,6 +50,9 @@ class ForwardGeocoder:
self.query_analyzer = await make_query_analyzer(self.conn)
query = await self.query_analyzer.analyze_query(phrases)
query.compute_direction_penalty()
log().var_dump('Query direction penalty',
lambda: f"[{'LR' if query.dir_penalty < 0 else 'RL'}] {query.dir_penalty}")
searches: List[AbstractSearch] = []
if query.num_token_slots() > 0: