Omorfi Open morphology for Finnish

• Home
• Meta
• Articles
• Timeline
• TODO
• Releases
• News
• Contributing or helping
• AUTHORS
• Design
• Omorfi with moses
• Finnish morphosyntax
• Directory layout
• Database processing
• Statistics
• Speed tests
• Paradigms and stuffs
• Lexical statistics
• Inflection tables nominals
• Inflection paradigms by example
• Inflection paper dictionaries
• Faithfulness tests
• CoveragesTop100Deltas
• Coverages
• Automata sizes
• API Documentation
• Python API
• Java API
• Man pages
• omorfi-spell(1)
• omorfi-segment(1)
• omorfi-generate(1)
• omorfi-disambiguate-factorise(1)
• omorfi-disambiguate-text(1)
• omorfi-analyse-tokenised(1)
• omorfi-analyse-text(1)
• omorfi-analyse(1)

Introduction

One of the secondary goals for analysers built from omorfi data is to act as free/libre open source way to reproduce analyses in “gold” corpora for Finnish. Since most of the gold corpora have been made with commercial tools unavailable to us, without specs, we use automated tests to gauge how well can we even reproduce their results prior to disambiguation. This page lists automatic test results for those tests and discussion of whether we want to get 100 % recall or maintain correct readings in omorfi. We also provide ways to get to 100 % easily when needed

FTB 3.1 background

The standard versions of omorfi in 2013–2015 use so-called FTB3.1 standard as a format for analyses. In this format, morphological analyses are encoded as string composed of a reconstructed lemma, followed by analysis tags separated by spaces. This document describes the differences from the official FTB3 format, the full description of the format can be found from FTB web site.

The analysis of the first sentence of declaration of human rights in FTB3 mode in omorfi stable version 20141014:

Kaikki kaikki Pron Qnt Nom Sg

ihmiset ihminen N Nom Pl

syntyvät syntyä PrsPrc Act Pos Nom Pl
syntyvät syntyä PrsPrc Act Pl
syntyvät syntyä V Prs Act Pl3

vapaina vapaa A Pos Ess Pl
vapaina vapaa N Ess Pl

ja ja Adv
ja ja CC

tasavertaisina tasa#vertainen A Pos Ess Pl
tasavertaisina tasa#vertainen N Ess Pl

arvoltaan arvo N Abl Sg Px3

ja ja Adv
ja ja CC

oikeuksiltaan oikeus N Abl Pl Px3

. . Abbr Nom Sg
. . Punct

FTB3.1 tags

Extracted from our implementation:

 >>>,  A,  Abbr,  Abe,  Abl,  Acc,  Act,  Ade,  Adp,  Adv,  AgPrc,
  All,  CC,  CS,  Com,  Comp,  ConNeg,  Cond,  Dem,  Ela,  Ess,
  Foc_hAn,  Foc_kA,  Foc_kAAn,  Foc_kO,  Foc_kin,  Foc_pA,  Foc_s,
  Gen,  Ill,  Impv,  Indef,  Ine,  Inf1,  Inf2,  Inf3,  Inf5,  Ins,
  Interj,  Interr,  Lat,  Man,  N,  Neg,  Nom,  Num,  Par,  Pass,
  Pe4,  Pers,  Pl,  Pl1,  Pl2,  Pl3,  Po,  Pos,  Pot,  Pr,  PrfPrc,
  Pron,  Prop,  Prs,  PrsPrc,  Pst,  Px3,  PxPl1,  PxPl2,  PxSg1,
  PxSg2,  Qnt,  Quote,  Refl,  Rel,  Sg,  Sg1,  Sg2,  Sg3,  Superl,
  Tra,  TrunCo,  V

FTB3.1 Known differences

The releases of omorfi follow a regression test of 90 % recall for FTB3.1 tags. The results for 20141014 are (Work in Progress):

The errors come mainly from following classes (the order of popularity per word-form):

1. -sti adverbs are not analysed as ` Man` forms of their etymologically relevant adjectives
2. word-forms oli or olivat are not analysed as present tense. This should be done in the post-processing stage of the morphological analysis as needed. (Affects some other verbs too but in very small quantities)
3. eikä is either negation verb or conjunction, not both
4. compounds that lack the boundary data from boundaries.csv will lack it in analyses (fixable bug: just go and do it :-)
5. foreign languages are not analysed; this should be handled by the text processor at a higher level
6. mismatches from lexicalised vs. non-lexicalised (maybe fixable, consider if the lexicalisation is or isn’t reasonable)

Examples with counts (totals etc. as in table above)

These are the top word-forms of the known mismatches, and the single word-forms alone attribute for 2.7 % of the errors

Getting 100 % recall for ftb3

If you are doing some stuff that requires 100 % recall, please collect the strings from ftb3.1.conllx (with awk oneliner or whatever), compile them with hfst-fst2strings and hfst-disjunct that to your morphology .hfst. Omorfi will likely not support FTB3’s all quirks e.g. the ones mentioned above. If you find mismatch that can be reasonably fixed, report a bug preferably with git format-patch included or use a github pull request.

Automated test runs

2014-04-01 FTB 3.1 (manual test)

Some of the mismatches documented further down can be deducted from the scores: * Forgn: 285625 * Adv Pos Man: 517824 * oli V Prt Act: 86011

| Lines | Full matches | Lemma matches | Analyses match | Mismatched | No results | |:———-|:—————–|:——————|:——————-|:—————|:—————| | 75,479,979 | 68,000,879 | 2,298,104 | 1,552,020 | 812,694 | 2,816,282 | | 100.0 % | 90.1 % | 3.0 % | 2.0 % | 1.1 % | 3.7 % | ## 2015-03-26 faithfulness

ftb3.1

100 most common mismatches

2015-09-04 faithfulness

ftb3.1

100 most common mismatches