When matching address parts from addr:* tags against place names,
the address names where so far converted to full names and compared
those to the place names. This can become problematic with the new
ICU tokenizer once we introduce creation of different variants
depending on the place name context. It wouldn't be clear which
variant to produce to get a match, so we would have to create all of
them. To work around this issue, switch to using the partial terms
for matching. This introduces a larger fuzziness between matches but
that shouldn't be a problem because matching is always geographically
restricted.
The search terms created for address parts have a different problem:
they are already created before we even know if they are going to be
used. This can lead to spurious entries in the word table, which slows
down searching. This problem can also be circumvented by using only
partial terms for the search terms. In terms of searching that means
that the address terms would not get the full-word boost, but given
that the case where an address part does not exist as an OSM object
should be the exception, this is likely acceptable.
Instead of requesting the match tokens from the tokenizer
when looking for parent streets/places and address parts,
hand in the saved tokens and ask if they match. This gives
the tokenizer more freedom to decide how name matching
should be done.
The new icu tokenizer is now no longer compatible with the old
legacy tokenizer in terms of data structures. Therefore there
is also no longer a need to refer to the legacy tokenizer in the
name.
Postgresql is very bad at creating statistics for jsonb
columns. The result is that the query planer tends to
use JIT for queries with a where over 'info' even when
there is an index.
Requires a second wrapper class for the word table with the new
layout. This class is interface-compatible, so that later when
the ICU tokenizer becomes the default, all tests that depend on
behaviour of the default tokenizer can be switched to the other
wrapper.
The table now directly reflects the different token types.
Extra information is saved in a json structure that may be
dynamically extended in the future without affecting the
table layout.
This adds precomputation of abbreviated terms for names and removes
abbreviation of terms in the query. Basic import works but still
needs some thorough testing as well as speed improvements during
import.
New dependency for python library datrie.
- only save partial words without internal spaces
- consider comma and semicolon a separator of full words
- consider parts before an opening bracket a full word
(but not the part after the bracket)
Fixes#244.
The BDD tests still use the old-style amenity creation scripts
because we don't have simple means to import a hand-crafted
test file of special phrases right now.
Normalization and token computation are now done in the tokenizer.
The tokenizer keeps a cache to the hundred most used house numbers
to keep the numbers of calls to the database low.
Creating and populating the word table is now the responsibility
of the tokenizer.
The get_maxwordfreq() function has been replaced with a
simple template parameter to the SQL during function installation.
The number is taken from the parameter list in the database to
ensure that it is not changed after installation.