This is a free/libre open source morphology of Finnish: a database, tools and
APIs. This package is licenced under GNU GPL version
3, but not necessarily later. Licence can
also be found in the COPYING
file in the root directory of this package.
Other licences are possible by all the authors named in the AUTHORS
file.
Omorfi has been used for a number of tasks:
- morphological analysis
- morphological segmentation
- spell-checking and correction
- information retrieval
- ontologies
- statistical machine translation
- rule-based machine translation
- language modeling
- tokenisation and sentence boundary detection
- stemming, lemmatisation and shallow morph analysis
The lexical data of omorfi has been acquired from various sources with different original licences. The dictionaries used in omorfi are Nykysuomen sanalista (LGPL), Joukahainen (GPL) and FinnWordNet (Princeton Wordnet licence / GPL; relicenced with kind permission from University of Helsinki), and Finnish Wiktionary (Creative Commons Attribution–ShareAlike). Some words have also been collected by omorfi developers and contributors and are GPLv3 like the rest of the package.
These are the obligatory stamps of the day:
→ See also: Release policy
Omorfi is currently hosted at github. Omorfi's github repository contain most of the important information about omorfi: version control system for source codes, bug tracker for reporting bugs, and the stable releases as convenient packages. Omorfi's gh-pages site contain further information not found in this README.
Before you start: Apertium wiki has installation information for most dependencies on their Apertium installation pages, look at section called pre-requisites, e.g., if you are looking to build omorfi on Ubuntu, go to: Pre-requisites for Ubuntu.
Compilation of the morphological analyser, generation, lemmatisation or spell-checking requires HFST tools or compatible installed. For use, you will need the python bindings too, and a relatively recent version of python 3. Of course standard GNU build tools are needed as well. You should have versions no older than one or two years. The build is not guaranteed to work at all with all ancient versions of GNU build tools, HFST or python. The versions that should work are as follows:
- hfst-3.8 or greater, with python bindings
- python-3.2 or greater, with hfst python bindings available
- GNU autoconf-2.64 and automake-1.12
The use of certain automata also requires additional tools:
- hfst-ospell-0.2.0 or greater needed for spell-checking
APIs require:
- Python 3.2 for python API
- Java 7 for Java API
- Bash 3, coreutils for bash API
- The C++ API uses C++-11, this is basically available on all modern platforms.
Installation uses standard autotools system:
./configure && make && make install
The compiling may take forever or longer, depending on the hardware and settings used. You should be prepared with at least 4 gigs of RAM, however, it is possible to compile a limited vocabulary system with less. This is a limitation of the HFST system used to compile the language models, and it is only present at compile time, the final models use perhaps up to hundreds of megabytes in memory.
If configure cannot find HFST tools, you must tell it where to find them:
./configure --with-hfst=${HFSTPATH}
Autotools system supports installation to e.g. home directory:
./configure --prefix=${HOME}
With git version you must create the necessary autotools files in the host system once, after initial checkout:
./autogen.sh
For further instructions, see INSTALL
, the GNU standard install instructions
for autotools systems.
→ See also: man pages
Omorfi comes with several simple scripts for basic functionalities. These scripts cover the most basic usage with minimal amount of required extra tools, however, for advanced usage you may want to check the APIs or bindings for python and Java.
Following are basic shell scripts that only use HFST tools and GNU coreutils:
omorfi-analyse-text.sh
: analyse plain text into ambiguous word-form listsomorfi-analyse-tokenised.sh
: analyse pre-tokenised word-forms one per lineomorfi-generate.sh
: generate word-forms from omor descriptionsomorfi-segment.sh
: morphologically segment word-forms one per line
The following requires python and VISL CG 3
omorfi-disambiguate-text.sh
: analyse text and disambiguate using VISL CG-3
The following uses hfst-ospell:
omorfi-spell.sh
: spell-check and correct word-forms one per line
The following are python scripts: vi
omorfi-tokenise.py
: format raw text into tokens (words and puncts).omorfi-conllu.py
: analyse and generate CONLL-U formatted data (Universal Dependencies) formatomorfi-vislcg.py
: analyse raw texts into VISL CG 3 formatomorfi-factorise.py
: analyse raw texts into moses factored format
The following examples have been run in the omorfi source dir after succesful installation. The command lines look like this:
$
When testing the instructions, please do not copy/paste the dollar sign, it is not a part of the command, but a command-line prompt indicator!
For many tasks you need to tokenise text before using it, this may involve:
splitting punctuation from words, recasing, etc. Also formatting text into word
per line, space-separated or more advanced formats like CONLL-U. There's only
one tool omorfi-tokenise.py
for this.
$ omorfi-tokenise.py -v -i test/newstest2016-fien-ref.fi.text | tail
kokouksessa keskusteltiin myös lakimiesten ammattinharjoittamisoikeuksien takaamisesta sekä ammattituomareiden ja - syyttäjien tukemisesta .
kokoukseen osallistui myös pääministeri Li Keqiang , sekä vanhemmat johtajat Liu Yunshan ja Zhang Gaoli , kerrottiin kokouksen jälkeen julkaistussa lausunnossa .
Lines: 3000 Tokens: 47338 Ratio: 15.779333333333334 tokens/line
CPU time: 0.5221295579999996 Real time: 0.5221478860185016
Tokens per timeunit: 90660.1391436499 Lines per timeunit: 5745.498699373647
For CONLL-U and so Universal dependencies, you'd use -O conllu
:
$ omorfi-tokenise.py -v -i test/newstest2016-fien-ref.fi.text -O conllu| tail
# sentence-text: Kokoukseen osallistui myös pääministeri Li Keqiang, sekä vanhemmat johtajat Liu Yunshan ja Zhang Gaoli, kerrottiin kokouksen jälkeen julkaistussa lausunnossa.
1 kokoukseen _ _ _ _ _ _ _ LOWERCASED=Kokoukseen
2 osallistui _ _ _ _ _ _ _ ORIGINALCASE
3 myös _ _ _ _ _ _ _ ORIGINALCASE
4 pääministeri _ _ _ _ _ _ _ ORIGINALCASE
5 Li _ _ _ _ _ _ _ ORIGINALCASE
6 Keqiang _ _ _ _ _ _ _ SpaceBefore=No|SpaceAfter=No
7 , _ _ _ _ _ _ _ SpaceBefore=No
8 sekä _ _ _ _ _ _ _ ORIGINALCASE
9 vanhemmat _ _ _ _ _ _ _ ORIGINALCASE
10 johtajat _ _ _ _ _ _ _ ORIGINALCASE
11 Liu _ _ _ _ _ _ _ ORIGINALCASE
12 Yunshan _ _ _ _ _ _ _ SpaceBefore=No|SpaceAfter=No
13 ja _ _ _ _ _ _ _ ORIGINALCASE
14 Zhang _ _ _ _ _ _ _ ORIGINALCASE
15 Gaoli _ _ _ _ _ _ _ SpaceBefore=No|SpaceAfter=No
16 , _ _ _ _ _ _ _ SpaceBefore=No
17 kerrottiin _ _ _ _ _ _ _ ORIGINALCASE
18 kokouksen _ _ _ _ _ _ _ ORIGINALCASE
19 jälkeen _ _ _ _ _ _ _ ORIGINALCASE
20 julkaistussa _ _ _ _ _ _ _ ORIGINALCASE
21 lausunnossa _ _ _ _ _ _ _ ORIGINALCASE|SpaceAfter=No
22 . _ _ _ _ _ _ _ SpaceBefore=No
Lines: 3000 Tokens: 47338 Ratio: 15.779333333333334 tokens/line
CPU time: 0.9153448170000003 Real time: 1.1707193159963936
Tokens per timeunit: 40434.96964061865 Lines per timeunit: 2562.5271224355897
Different kinds of morphological analysis use cases: traditional linguistics with Xerox-style analysis, followed by constraitn grammars, or Universal dependency pre-parses, or factorised analysis for statistical machine ranslations.
The Xerox finite-State Morphology uses simple tab-separated analysis per line
format with tokens separated by empty lines, this is the output format of the
omorfi-analyse-text
tool.
$ omorfi-analyse-text.sh test/newstest2016-enfi-ref.fi.text |head
Tampereella [WORD_ID=Tampere][UPOS=PROPN][PROPER=GEO][NUM=SG][CASE=ADE]
karkuteillä [WORD_ID=karkuteillä][UPOS=ADV]
karkuteillä [WORD_ID=karkuteillä_2][UPOS=ADV]
ollut [WORD_ID=olla][UPOS=AUX][DRV=NUT][CMP=POS][NUM=SG][CASE=NOM]
ollut [WORD_ID=olla][UPOS=AUX][DRV=TU][CMP=POS][NUM=PL][CASE=NOM]
ollut [WORD_ID=olla][UPOS=AUX][VOICE=ACT][MOOD=INDV][TENSE=PAST][NUM=SG][NEG=CON]
ollut [WORD_ID=olla][UPOS=AUX][VOICE=ACT][PCP=NUT]
ollut [WORD_ID=olla][UPOS=AUX][VOICE=PSS][PCP=NUT][CMP=POS][CASE=NOM][NUM=PL]
If your text is already split into word-forms (one word-form per line), it can be analysed like this:
$ omorfi-analyse-tokenised.sh test/wordforms.list | head
. [WORD_ID=.][UPOS=PUNCT][BOUNDARY=SENTENCE] 0,000000
1 [WORD_ID=1][UPOS=NUM][NUMTYPE=CARD] 0,000000
1 [WORD_ID=1][UPOS=NUM][NUMTYPE=CARD][NUM=SG][CASE=NOM] 0,000000
10 [WORD_ID=10][UPOS=NUM][NUMTYPE=CARD] 0,000000
10 [WORD_ID=10][UPOS=NUM][NUMTYPE=CARD][NUM=SG][CASE=NOM] 0,000000
1 000 [WORD_ID=1 000][UPOS=NUM][NUMTYPE=CARD] 0,000000
1 000 [WORD_ID=1 000][UPOS=NUM][NUMTYPE=CARD][NUM=SG][CASE=NOM] 0,000000
A full pipeline for VISL CG 3 disambiguation is implemented as a convenience script that works like text analysis script:
$ omorfi-disambiguate-text.sh test/newstest2016-enfi-ref.fi.text | tail -n 40
"<kokoukseen>"
"kokouksi" NOUN SG ILL
"kokouksi_2" NOUN SG ILL
"kokous" NOUN SG ILL
"<osallistui>"
"osallistua" VERB ACT INDV PAST SG3
"<myös>"
"myödä" VERB <DIALECTAL> ACT IMPV SG2 S
"<pääministeri>"
"pääministeri" NOUN SG NOM
"<Li>"
"Li" NUM ROMAN
"Li" PROPN FIRST SG NOM
"Li_2" PROPN LAST SG NOM
"<Keqiang>"
"Keqiang" UNKNOWN <W=65536>
"<,>"
"," PUNCT CLAUSE COMMA CLB
",_2" SYM CLB
"<sekä>"
"sekä" CONJ
"<vanhemmat>"
"vanha" ADJ MPI CMP PL NOM
"vanhempi" NOUN PL NOM
"<johtajat>"
"johtaa" VERB JA PL NOM
"johtaja" NOUN PL NOM
"<Liu>"
"Liu" PROPN LAST SG NOM
"<Yunshan>"
"Yunshan" UNKNOWN <W=65536>
"<ja>"
"ja" CONJ
"<Zhang>"
"Zhang" PROPN LAST SG NOM
"<Gaoli>"
"Gaoli" UNKNOWN <W=65536>
"<,>"
"," PUNCT CLAUSE COMMA CLB
",_2" SYM CLB
"<kerrottiin>"
"kertoa" VERB PSS INDV PAST PE4
"<kokouksen>"
"kokouksi" NOUN SG GEN
"kokouksi_2" NOUN SG GEN
"kokous" NOUN SG GEN
"<jälkeen>"
"jälkeen_2" ADV PREP
"<julkaistussa>"
"julkaistu" ADJ POS SG INE
"<lausunnossa>"
"lausunto" NOUN SG INE
"<.>"
"." PUNCT SENTENCE
Tokens: 47338 Unknown: 1982 4.186911149604969 %
CPU time: 3.087152993 Real time: 3.1000033089949284
Tokens per timeunit: 15270.306280849665
CG style input format can be generated using python based analyser script
omorfi-vislcg.py
(this is the ambiguous morphological analysis, just formatted
for omorfi's CG ruleset):
$ omorfi-vislcg.py -i test/newstest2016-enfi-ref.fi.text | tail
"<,>"
"," PUNCT CLAUSE COMMA
",_2" SYM
"<kerrottiin>"
"kertoa" VERB PSS INDV PAST PE4
"<kokouksen>"
"kokouksi" NOUN SG GEN
"kokouksi_2" NOUN SG GEN
"kokous" NOUN SG GEN
"<jälkeen>"
"jälkeen" ADP POST
"jälkeen_2" ADV PREP
"jälki" NOUN SG ILL
"<julkaistussa>"
"julkaista" VERB TU POS SG INE
"julkaistu" ADJ POS SG INE
"<lausunnossa>"
"lausunto" NOUN SG INE
"<.>"
"." PUNCT SENTENCE
Tokens: 47338 Unknown: 1982 4.186911149604969 %
CPU time: 3.6671915069999996 Real time: 3.66774150999845
Tokens per timeunit: 12906.580213178655
Moses factored analysis format can be generated using python script:
$ omorfi-factorise.py -i test/newstest2016-enfi-ref.fi.text | tail
Kokouksessa|koko+uksi|UNK|NOUN.SG.INE|0 keskusteltiin|keskustella|UNK|VERB.PSS.INDV.PAST.PE4|0 myös|myödä|UNK|VERB.DIALECTAL.ACT.IMPV.SG2.S|0 lakimiesten|laki+mies|UNK|NOUN.PL.GEN|0 ammattinharjoittamisoikeuksien|ammattinharjoittamisoikeuksien|UNK|UNKNOWN|0 takaamisesta|taata_2|UNK|VERB.MINEN.SG.ELA|0 sekä|sekä|UNK|CONJ|0 ammattituomareiden|ammatti-+tuomari|UNK|NOUN.TITLE.PL.GEN|0 ja|ja|UNK|CONJ|0 -syyttäjien|syyttäjä|UNK|NOUN.TITLE.PL.GEN|0 tukemisesta.|tukemisesta.|UNK|UNKNOWN|0
Kokoukseen|koko+uksi|UNK|NOUN.SG.ILL|0 osallistui|osallistua|UNK|VERB.ACT.INDV.PAST.SG3|0 myös|myödä|UNK|VERB.DIALECTAL.ACT.IMPV.SG2.S|0 pääministeri|pää-+ministeri|UNK|NOUN.TITLE.SG.NOM|0 Li|Li|UNK|NUM.ROMAN|0 Keqiang,|Keqiang,|UNK|UNKNOWN|0 sekä|sekä|UNK|CONJ|0 vanhemmat|vanha|UNK|ADJ.MPI.CMP.PL.NOM|0 johtajat|johtaa|UNK|VERB.JA.PL.NOM|0 Liu|Liu|UNK|PROPN.LAST.SG.NOM|0 Yunshan|Yunshan|UNK|UNKNOWN|0 ja|ja|UNK|CONJ|0 Zhang|Zhang|UNK|PROPN.LAST.SG.NOM|0 Gaoli,|Gaoli,|UNK|UNKNOWN|0 kerrottiin|kertoa|UNK|VERB.PSS.INDV.PAST.PE4|0 kokouksen|koko+uksi|UNK|NOUN.SG.GEN|0 jälkeen|jälkeen|UNK|ADP.POST|0 julkaistussa|julkaistu|UNK|ADJ.POS.SG.INE|0 lausunnossa.|lausunnossa.|UNK|UNKNOWN|0
The input should be in format produced by moses's tokenizer.perl
(truecase or
clean-corpus-n not necessary). The output is readily usable by Moses train
model. If you don't use tokenizer.perl, the words next to punctuation will not
be analysed.
Some of the fields are produced by different automata, where automaton is missing, placeholder values can appear, such as UNK or 0.
Universal Dependencies are the up-and-coming standard for all your morpho-syntactic needs! Omorfi is currently scheduled to follow up on Universal dependencies relaeas schedules and analysis and design principles.
Universal dependencies parsing requires input in pre-tokenised, CONLL-U format, only fields INDEX, SURF and MISC are made use of in basic version.
$ omorfi-conllu.py -v -i test/UD_Finnish/fi-ud-dev.conllu | tail -n 40
# text = Luonto-Liitto ehdottaa Metsähallitukselle täydentämään suojelupäätöstä muuallekin Suomessa, kuten Kainuuseen.
# sent_id = wn074.11
1 Luonto-Liitto Luonto-Liitto PROPN N Case=Nom|Number=Sing _ _ _ _
2 ehdottaa ehdottaa VERB V InfForm=1|Number=Sing|VerbForm=Inf|Voice=Act _ _ __
3 Metsähallitukselle Metsähallitus PROPN N Case=All|Number=Sing _ _ _ _
4 täydentämään täydentää VERB V Case=Ill|InfForm=3|Number=Sing|VerbForm=Inf|Voice=Act _ __ _
5 suojelupäätöstä suojelu#päätös NOUN N Case=Par|Number=Sing _ _ _ _
6 muuallekin muualle ADV Adv Clitic=Kin _ _ _ _
7 Suomessa Suomi PROPN N Case=Ine|Number=Sing _ _ _ _
8 , , PUNCT Punct _ _ _ _ _
9 kuten kuten ADV Adv _ _ _ _ _
10 Kainuuseen Kainuu PROPN N Case=Ill|Number=Sing _ _ _ _
11 . . PUNCT Punct _ _ _ _ _
# text = Hakkuut vanhoissa metsissä saavat aina aikaan konfliktin.
# sent_id = wn074.12
1 Hakkuut Hakkuu PROPN N Case=Nom|Number=Plur _ _ _ _
2 vanhoissa vanha ADJ A Case=Ine|Degree=Pos|Number=Plur _ _ _ _
3 metsissä metsä NOUN N Case=Ine|Number=Plur _ _ _ _
4 saavat saada VERB V Mood=Ind|Number=Plur|Person=3|Tense=Pres|VerbForm=Fin|Voice=Act _ _ __
5 aina aina ADP Adp _ _ _ _ _
6 aikaan aika NOUN N Case=Ill|Number=Sing _ _ _ _
7 konfliktin konflikti NOUN N Case=Gen|Number=Sing _ _ _ _
8 . . PUNCT Punct _ _ _ _ _
This can be combined with tokenisation to analyse raw corpora:
$ omorfi-tokenise.py -O conllu -i test/newstest2016-enfi-ref.fi.text | omorfi-conllu.py | tail -n 40
# sentence-text: Kokoukseen osallistui myös pääministeri Li Keqiang, sekä vanhemmat johtajat Liu Yunshan ja Zhang Gaoli, kerrottiin kokouksen jälkeen julkaistussa lausunnossa.
1 kokoukseen koko#uksi NOUN N Case=Ill|Number=Sing _ _ _ _
2 osallistui osallistua VERB V Mood=Ind|Number=Sing|Person=3|Tense=Past|VerbForm=Fin|Voice=Act _ _ _ _
3 myös myödä VERB V Clitic=S|Mood=Imp|Number=Sing|Person=2|Style=Coll|VerbForm=Fin|Voice=Act __ _ _
4 pääministeri pää#ministeri NOUN N Case=Nom|Number=Sing _ _ _ _
5 Li Li NUM Num _ _ _ _ _
6 Keqiang PROPN N Case=Nom|Number=Sing _ _ _ GUESS=HEUR
7 , , PUNCT Punct _ _ _ _ _
8 sekä sekä CCONJ C _ _ _ _ _
9 vanhemmat vanha ADJ A Case=Nom|Degree=Cmp|Number=Plur _ _ _ _
10 johtajat johtaa VERB V Case=Nom|Derivation=Ja|Number=Plur _ _ _ _
11 Liu Liu PROPN N Case=Nom|Number=Sing _ _ _ _
12 Yunshan PROPN N Case=Nom|Number=Sing _ _ _ GUESS=HEUR
13 ja ja CCONJ C _ _ _ _ _
14 Zhang Zhang PROPN N Case=Nom|Number=Sing _ _ _ _
15 Gaoli PROPN N Case=Nom|Number=Sing _ _ _ GUESS=HEUR
16 , , PUNCT Punct _ _ _ _ _
17 kerrottiin kertoa VERB V Mood=Ind|Tense=Past|VerbForm=Fin|Voice=Pass _ _ _ _
18 kokouksen koko#uksi NOUN N Case=Gen|Number=Sing _ _ _ _
19 jälkeen jälkeen ADP Adp AdpType=Post _ _ _ _
20 julkaistussa julkaista VERB V Case=Ine|Degree=Pos|Number=Sing _ _ _ _
21 lausunnossa lausunto NOUN N Case=Ine|Number=Sing _ _ _ _
22 . . PUNCT Punct _ _ _ _ _
There's a cheat mode that can be used with UD training data to always select
the best match, for evaluation purposes: --oracle
. There's a debug mode to
print full n-best for each token: --debug
, this is pseudo CONLL-U.
The morphological segmentation can be done like this:
$ omorfi-segment.py -O segments -i test/newstest2016-enfi-ref.fi.text
lisäksi ulko→ ←maalais→ ←ten pysyv→ ←i→ ←en asukas→ ←lup→ ←i→ ←en , ” Green cardien , ” haku→ ←prosessi→ ←a helpote→ ←taan optimoima→ ←lla vaatimuks→ ←i→ ←a ja keventämä→ ←llä haku→ ←prosessi→ ←a .
kokoukse→ ←ssa keskustel→ ←tiin myös lakimies→ ←ten ammattinharjoittamisoikeuksien takaamise→ ←sta sekä ammatti→ ←tuomare→ ←i→ ←den ja - syyttäj→ ←i→ ←en tukemise→ ←sta .
kokoukse→ ←en osallistu→ ←i myös pää→ ←ministeri Li Keqiang , sekä vanhemma→ ←t johtaja→ ←t Liu Yunshan ja Zhang Gaoli , kerrot→ ←tiin kokoukse→ ←n jälkeen julkaistu→ ←ssa lausunno→ ←ssa .
Preliminary support for labeled segmentation is also available but not guaranteed to work.
Spelling correction may be done if hfst-ospell is installed:
$ omorfi-spell.sh test/wordforms.list | tail
"äyräässä" is in the lexicon...
"äänestys" is in the lexicon...
"äänioikeus" is in the lexicon...
"öykkärein" is in the lexicon...
"öykkäri" is in the lexicon...
"öykkärimpi" is in the lexicon...
"öykkäröi" is in the lexicon...
"öykkäröidä" is in the lexicon...
Generating word-forms can be done using:
$ omorfi-generate.sh
[WORD_ID=bisse][UPOS=NOUN][NUM=SG][CASE=NOM]
[WORD_ID=bisse][UPOS=NOUN][NUM=SG][CASE=NOM] bisse 0,000000
[WORD_ID=bisse][UPOS=NOUN][NUM=SG][CASE=INE]
[WORD_ID=bisse][UPOS=NOUN][NUM=SG][CASE=INE] bissessä 0,000000
The input for generator is simply the output of the raw analyser.
→ See also: API design, generated API docs ):
For serious business, the convenience shell-scripts are not usually sufficient. We offer bindings to several popular programming languages as well as low-level access to the automata either via command-line or the external programming libraries from the toolkit generating the automata.
Python interface:
$ python3
Python 3.4.0 (default, Mar 18 2014, 16:02:57)
[GCC 4.8.2] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> from omorfi.omorfi import Omorfi
>>> omorfi = Omorfi()
>>> omorfi.load_from_dir()
>>> omorfi.analyse("koira")
(('[WORD_ID=koira][UPOS=NOUN][NUM=SG][CASE=NOM]', 0.0),
('[WORD_ID=Koira][UPOS=PROPN][NUM=SG][CASE=NOM]', 0.0))
>>> for analysis in analyses:
... print(analysis[0])
...
[WORD_ID=alku][UPOS=NOUN][NUM=SG][CASE=ELA]
[WORD_ID=alunen][UPOS=NOUN][NUM=SG][CASE=PAR]
[WORD_ID=alus][UPOS=NOUN][NUM=SG][CASE=PAR]
[WORD_ID=alusta][UPOS=NOUN][NUM=SG][CASE=NOM]
[WORD_ID=alusta_2][UPOS=ADV]
[WORD_ID=alusta_3][UPOS=ADV]
[WORD_ID=alustaa][UPOS=VERB][VOICE=ACT][MOOD=IMPV][PERS=SG2]
[WORD_ID=alustaa][UPOS=VERB][VOICE=ACT][MOOD=INDV][TENSE=PRESENT][NEG=CON]
[WORD_ID=Alku][UPOS=PROPN]ERRORMACROPROPN-BLOCKING][NUM=SG][CASE=ELA]
[WORD_ID=Alku_2][UPOS=PROPN][NUM=SG][CASE=ELA]
[WORD_ID=Alku_3][UPOS=PROPN][NUM=SG][CASE=ELA]
[WORD_ID=Alus][UPOS=PROPN]ERRORMACROPROPN-BLOCKING][NUM=SG][CASE=PAR]
>>>
>>>
Java class (more details on java API pages):
$ CLASSPATH=$HOME/github/hfst/hfst-optimized-lookup/hfst-optimized-lookup-java/hfst-ol.jar:. java com.github.flammie.omorfi.Omorfi
...
Read all.
> talo
Analysing talo
{omorfi-omor_recased=net.sf.hfst.WeightedTransducer@63947c6b, omorfi-giella=net.sf.hfst.WeightedTransducer@2b193f2d, omorfi-ftb3=net.sf.hfst.WeightedTransducer@355da254, omorfi-omor=net.sf.hfst.WeightedTransducer@4dc63996}
Analysing talo with omorfi-omor
[WORD_ID=talo][UPOS=NOUN][NUM=SG][CASE=NOM][WEIGHT=0.0]
[WORD_ID=talo][UPOS=NOUN][NUM=SG][CASE=NOM][WEIGHT=0.0][CASECHANGE=LOWERCASED]
Especially loading all automata from system paths requires more memory than
java typically gives you, so use -Xmx
switch.
The installed files are in $prefix/share/omorfi
(my installation is in linux
default: /usr/local
)
$ ls /usr/local/share/omorfi/
master.tsv omorfi.describe.hfst omorfi_recased.analyse.hfst speller-omorfi.zhfst
omorfi.accept.hfst omorfi.generate.hfst omorfi_recased.describe.hfst
omorfi.analyse.hfst omorfi.labelsegment.hfst omorfi.segment.hfst
omorfi.cg3bin omorfi.lemmatise.hfst omorfi.tokenise.pmatchfst
The naming is probably not gonna be same forever.
You can directly access specific automata using finite-state tools from the HFST project (details can be found on their individual man pages and HFST wiki:
$ hfst-lookup /usr/local/share/omorfi/omorfi.segment.hfst
> talossani
talossani talo{DB}s{MB}sa{MB}ni 0,000000
talossani talo{MB}ssa{MB}ni 0,000000
> on
on on 0,000000
> hirveä
hirveä hirve{MB}ä 0,000000
hirveä hirveä 0,000000
> kissakoira-apina
kissakoira-apina kissa{hyph?}koira{hyph?}apina 0.000000
When using hfst-lookup
with large unclean material, it may get stuck at odd
looking long strings, consider using -t
switch to set timeout for individual
analyses; omorfi bash API sets this to 15 seconds.
Mac OS X may cause problems with its Unicode encoding (NFD), or with its non-GNU command-line tools.
This may happen when compiling the system with make:
hfst-lexc --Werror -o generated/omorfi-ftb3.lexc.hfst generated/omorfi-ftb3.lexc
Try ``/usr/local/bin/hfst-lexc --help'' for more information.
/usr/local/bin/hfst-lexc: Unknown option
It means your hfst-lexc is too old. You need at least version 3.8 to handle
--Werror
switch. You can workaround by removing --Werror
from
src/Makefile.am
, although this is not recommended, as the newer versions of
HFST have provided this option to ensure the data is not broken.
The same problem may re-appear with the --alignStrings option.
E.g. error message of form:
ImportError: No module named 'omorfi'
In order for python scripts to work you need to install them to same prefix as python, or define PYTHONPATH, e.g.:
$ PYTHONPATH=/usr/local/lib/python3.4/site-packages/ omorfi-disambiguate-text.sh kalevala.txt
The scripts require python3 as the system python, in case I forget to set the whole shebang right. You can work around this by, e.g.,:
python3 $(which omorfi-analyse.py )
This should not affect release versions but keep in mind if you are using a python2-based system and development versions.
When omorfi files are not where bash scripts are looking for them.
If you have moved your installation manually after make install, you may need to modify paths in omorfi.bash or set environment variable OMORFI_PATH.
If the file missing is omorfi.cg3bin
, it may mean that the vislcg3 was missing
at the time of the installation. Similarly may happen with omorfi-speller.zhfst,
it will only be created when hfst-ospell and it's dependencies and zip are all
available.
Occasionally some tokens yield very complicated analyses and take a lot of
memory or time. This happens especially with long strings that can be analysed
as combinations of interjections like ahahaha...ha (in theory, each ah, aha, ha
and hah within the string are ambiguous wrt. compounding), while current
versions have blacklisted most such combinations some may still exist. When
using hfst tools directly this can be solved using -t
option to set the
timeout. While these workarounds will slowly trickle to all parts of HFST and
omorfi, it is often a good idea to pre-process text to remove or normalise
offending strings as they will trip other NLP tools too.
Some operations of omorfi legitly take a lot of memory, and most tools are
suspectible to memory leaks. It may be often beneficial to split
your data
and process it in smaller chunks.
Compiling omorfi takes some memory, of course it is over 10 megabytes of words that need to be expanded and processed in a number of ways, it can momentarily take up more than 2 or 4 gigabytes of memory. If you have that or less RAM, depending on your setup it may either slow down the computer a lot, or just cause compilation to fail. The error message in that case will usually only read something like:
make[2]: Verzeichnis „/home/tpirinen/github/flammie/omorfi/src“ wird betreten
/usr/bin/hfst-lexc --alignStrings --Werror -o generated/omorfi.accept.lexc.hfst generated/omorfi.accept.lexc
/usr/bin/hfst-lexc: warning: Defaulting to OpenFst tropical type
... Getötet
make[2]: *** [generated/omorfi.lexc.hfst] Killed
it all depends on versions of make and language settings, but in general most of
the time it will say at least "killed" and usually nothing more as an error message
near the end. Current versions of omorfi support compiling a limited-size version with
high-quality dictionaries only, using ./configure --enable-small-lexicons
option.
If this works but normal compilation doesn't, you may need to acquire more RAM or
tweak some memory-related settings.
Omorfi code and data are free and libre open source, modifiable and redistributable by anyone. IRC channel #omorfi on Freenode is particularly good for immediate discussion about contributions. Any data or code contributed must be compatible with our licencing policy, i.e. GNU compatible free licence. In the github, you may use the "fork this project" button to contribute, read github's documentation for more information about this work-flow.
Python code should pass the flake8 style checker and imports should be sorted in accordance with isort. Ideally, you should integrate these into your editor, the development environment section of the python guide has instructions for a few editors. In addition, you can install a pre-commit hook to run the checks like so:
$ pip install pre-commit
$ pre-commit install