Migrating from pymorphy

pymorphy2 and pymorphy3 share one API: pymorphy3 continues pymorphy2, which does not work on Python 3.11 and newer. ruMorph uses the same OpenCorpora dictionary but has its own API: Universal Dependencies parts of speech and features instead of OpenCorpora grammemes, a probability instead of score, word forms as strings, exceptions instead of None.

pip install rumorph

The dictionary is included in the package, no separate dictionary package is needed.

pymorphy3ruMorph

>>> import pymorphy3
>>> morph = pymorphy3.MorphAnalyzer()
>>> p = morph.parse("стали")[0]
>>> p.normal_form, p.tag.POS, p.tag.number, round(p.score, 3)
('стать', 'VERB', 'plur', 0.975)
>>> cat = morph.parse("кошка")[0]
>>> cat.inflect({"gent", "plur"}).word
'кошек'
>>> morph.parse("рубль")[0].make_agree_with_number(5).word
'рублей'

>>> import rumorph
>>> morph = rumorph.Analyzer()
>>> a = morph.analyze("стали")[0]
>>> a.lemma, a.pos, a.feats["Number"], round(a.prob, 3)
('стать', 'VERB', 'Plur', 0.975)
>>> cat = morph.analyze("кошка")[0]
>>> cat.inflect(Case="Gen", Number="Plur")
'кошек'
>>> morph.analyze("рубль")[0].agree(5)
'рублей'

Call correspondence

Here morph is the analyzer, p is a pymorphy analysis, a is a ruMorph analysis.

pymorphy2, pymorphy3	ruMorph	Differences
MorphAnalyzer()	Analyzer()
MorphAnalyzer(path)	Analyzer(path)	the path to a ruMorph dictionary file; pymorphy dictionaries do not fit
morph.parse(word)	morph.analyze(word)	analyses that coincide in UD are merged
[morph.parse(w) for w in words]	morph.analyze_many(words)	runs without the GIL
[morph.parse(w)[0].normal_form for w in words]	morph.lemmatize(words)	runs without the GIL
morph.normal_forms(word)	list(dict.fromkeys(a.lemma for a in morph.analyze(word)))
morph.tag(word)	[(a.pos, a.feats) for a in morph.analyze(word)]
morph.word_is_known(word)	morph.is_known(word)
p.word	a.word	analyses by rules and guessed ones keep the case of the text
p.normal_form	a.lemma	lemmas follow the UD conventions
p.tag.POS	a.pos	UD parts of speech, see below
p.tag.case, p.tag.number and others	a.feats.get("Case"), a.feats.get("Number") and others	UD features, see below
"Infr" in p.tag	"Infr" in a.labels	labels holds the grammemes that have no UD feature
p.score	a.prob
p.methods_stack	a.source	"dictionary", "rule" or "guess"
p.lexeme	a.forms()
p.inflect({"gent", "plur"}).word	a.inflect(Case="Gen", Number="Plur")	a string; no form - FormNotFoundError instead of None
p.make_agree_with_number(n).word	a.agree(n)	a string; no form - FormNotFoundError instead of None

There is no counterpart for the Russian names of grammemes (cyr_repr, lat2cyr), the analysis of the initial form (p.normalized), iterating over the dictionary and configuring the predictor of unknown words.

Parts of speech

OpenCorpora	UD
NOUN	NOUN; given names, surnames, patronymics, places, organizations and trademarks - PROPN
ADJF	ADJ; pronominal adjectives - DET
ADJS	ADJ with Variant=Short
COMP	ADJ with Degree=Cmp
VERB	VERB with VerbForm=Fin; «быть» - AUX
INFN	VERB with VerbForm=Inf
PRTF	VERB with VerbForm=Part
PRTS	VERB with VerbForm=Part and Variant=Short
GRND	VERB with VerbForm=Conv
NUMR	NUM with NumType=Card
ADVB	ADV
NPRO	PRON
PRED	VERB without features
PREP	ADP
CONJ	CCONJ; subordinating conjunctions - SCONJ
PRCL	PART
INTJ	INTJ

Some words get their part of speech by the SynTagRus conventions: the parenthetical «однако», «например» are ADV, «бы» is AUX and so on (see Parts of speech and features). Tokens outside the dictionary: NUMB - NUM with NumForm=Digit, ROMN - ADJ with NumForm=Roman, LATN - X with Foreign=Yes, PNCT - PUNCT or SYM, UNKN - X without features (see Tokens outside the dictionary).

Grammemes

OpenCorpora	UD
nomn, gent, datv, accs, ablt, loct, voct	Case: Nom, Gen, Dat, Acc, Ins, Loc, Voc
gen1, acc2, loc1	Case: Gen, Acc, Loc
gen2 («стакан чаю»)	Case=Par
loc2 («в лесу»)	Case=Loc with the loc2 label
sing, plur	Number: Sing, Plur
masc, femn, neut	Gender: Masc, Fem, Neut
ms-f («сирота»)	Gender=Masc
anim, inan	Animacy: Anim, Inan
1per, 2per, 3per	Person: 1, 2, 3
pres, past, futr	Tense: Pres, Past, Fut; the future of «быть» is Tense=Pres with the futr label
perf, impf	Aspect: Perf, Imp
indc, impr	Mood: Ind, Imp
actv, pssv	Voice: Act, Pass; reflexive verbs - Mid
Name, Surn, Patr, Geox, Orgn, Trad	NameType: Giv, Sur, Pat, Geo, Com, Pro; labels for adjectives
Supr	Degree=Sup
Abbr	Abbr=Yes; a label in labels for declinable abbreviations («вуз»)
Abbr and Anum together («1-го», «90-х»)	NumType=Ord, NumForm=Combi, without Abbr
Poss	Poss=Yes; a label in labels for adjectives
Fixd	InflClass=Ind for nouns except abbreviations, a label in labels for other words
Apro	the DET part of speech
Dmns, Ques («там», «где»)	PronType: Dem, Rel, Int
Anum on words («первый»), GNdr	none
the rest: Infr, Slng, Arch, Qual, tran, intr, V-oy and others	labels

The Voice feature is present on all verb forms and Degree on all adverbs and on adjectives except ordinals written in Arabic or Roman numerals («1-го», «2026», «XX»), although in OpenCorpora only participles have voice grammemes and only superlatives and comparatives have degree.

Differences in behavior

• Lemmas follow SynTagRus: «Ивановой» - «иванова», not «иванов», «важнейшей» - «важнейший», not «важный», «денег» - «деньги», not «деньга».
• Analyses that coincide in UD (the conjunction and the adverb «так») are merged into one with the sum of their probabilities, so there may be fewer analyses than in pymorphy.
• Forms from inflect and agree are strings, not analyses; when there is no form, they raise FormNotFoundError, a subclass of LookupError.
• Tokens outside the dictionary are analyzed by rules with UD parts of speech and features: «2005-го» is an ordinal adjective in the genitive with forms («2005-му»), not an indeclinable noun as in pymorphy3.
• Speed. Analysis and lemmatization are 6-8 times faster than with pymorphy3[fast] and 30-60 times faster than with pymorphy3 without it, and lists of words are processed without the GIL, see Speed and memory.