This text documents the speller configuration that has turned out to be the optimal configuration for North Sámi (ISO code SME). There are several parts of the infrastructure that can be used for optimising the speller. They are:

• options to the ./configure command
• settings in Makefile.am files
• flag elimination

The optimisations described here relate to speed and file size. Fine tuning the error model is described in a separate document.

This document is up-to-date as of 28.11.2016.

configure options

The following configuration is what seems to produce the optimal speller:

./configure --with-hfst --without-xfst --enable-alignment --enable-spellers

Note specifically that the following option does not improve the SME speller (the default value seems good), although it could in certain cases:

 --disable-minimised-spellers

The following can be added to increase compilation speed, although it should not have an effect on the runtime speed or file size (but that has not been tested):

 --with-backend-format=foma --enable-reversed-intersect

Settings in Makefile.am files

The file tools/spellcheckers/fstbased/desktop/Makefile.am contains the following variables (with the settings used for SME):

ENABLE_CORPUS_WEIGHTS=yes
CORPUS_SIZE=

Enabling corpus weights does help improving suggestion quality quite a bit. And after experimenting, it seems there is no point in limiting the corpus size being used for frequency weighting — it does not increase the size of the speller fst very much, and there is some improvement in suggestion quality also for the last hapax entries in the generated frequency list.

Flag elimination

Eliminating flag diacritics can have a tremendeous effect on both speller speed and file size. Uncritical use of flag diacritics elimination can make the speller file size explode, so please use this optimisation tool carefully, and test the change in file size and speed after each new flag added to the elimination list.

The flag elimination is done in tools/spellcheckers/fstbased/Makefile.am. The following is used for SME:

eliminate flag CmpHyph
eliminate flag CmpN
eliminate flag Der1
eliminate flag Der2
eliminate flag Der3
eliminate flag Der4
eliminate flag Der5
eliminate flag Der_PassL
eliminate flag Der_PassS

There are more flags being used in SME, but eliminating them made the fst grow unreasonably big. The following flags are not removed for this reason:

eliminate flag NeedNoun
eliminate flag NeedsVowRed
eliminate flag Want_Left

NeedNoun and Want_Left crosses word boundaries, and will most likely cause a massive network size explosion if removed. NeedsVowRed also caused the fst to grow significantly in size, but that could probably be avoided with some more work on the lexc code - it should really be just a local constraint.