Word analysis¶

An analysis is an Analysis object returned by the analyze and analyze_many methods of the analyzer: the lemma, the part of speech and the Universal Dependencies features, the probability and the source of the analysis. Analyses are immutable; they can be compared (the probability takes no part in the comparison: an analysis equals its form from forms()), put into sets and pickled. The forms of a word are given by the forms, inflect and agree methods - see Inflection.

CodeResult

from rumorph import Analyzer

morph = Analyzer()
analysis = morph.analyze("стали")[0]
analysis.lemma, analysis.pos, analysis.feats, analysis.labels

('стать',
 'VERB',
 {'Aspect': 'Perf',
  'Mood': 'Ind',
  'Number': 'Plur',
  'Tense': 'Past',
  'VerbForm': 'Fin',
  'Voice': 'Mid'},
 frozenset({'intr'}))

Attributes¶

Attribute	Type	Description
`word`	`str`	The word form as spelled in the dictionary: with «ё» if the dictionary has it; for analyses by rules and guessed ones - as in the text, but guessed ones take «ё» from the ending of the dictionary form («пингвиненка» - «пингвинёнка»)
`lemma`	`str`	Initial form
`pos`	`str`	UD part of speech: `NOUN`, `VERB`, `ADJ`, `DET`, `ADP` and so on
`feats`	`dict[str, str]`	UD features: `{"Case": "Gen", "Number": "Sing"}`; a new dictionary each time, changing it does not change the analysis
`labels`	`frozenset[str]`	Dictionary labels that have no UD feature: `Infr` (informal) and others
`prob`	`float`	Probability of the analysis; it sums to 1 over all analyses of the word form
`source`	`str`	Where the analysis comes from: `"dictionary"` - the word form is in the dictionary, `"rule"` - an analysis by rule, `"guess"` - a guessed analysis of an unknown word

Parts of speech and features¶

Parts of speech and features follow Universal Dependencies as annotated in the largest Russian treebank, UD_Russian-SynTagRus, so the output can be compared with spaCy, Stanza and other UD tools. Differences from the OpenCorpora parts of speech:

«который», «этот», «свой», «весь», «такой» and the other pronominal adjectives are DET;
«быть» and «бы» are AUX;
the parenthetical words «однако», «например», «конечно» are ADV;
«можно», «надо», «нельзя» are VERB without features;
prepositions are ADP, coordinating conjunctions CCONJ, subordinating ones SCONJ, particles PART.

Features: case, number, gender, animacy, person, tense, aspect, mood, voice, verb form, degree, the short form (Variant=Short), the type of a proper name (NameType), abbreviation (Abbr), possessiveness (Poss), the type of a numeral (NumType) and how it is written (NumForm), the type of a pronoun or pronominal adverb (PronType: «там» - Dem, «где» - Rel, «никогда» - Neg), reflexivity (Reflex), indeclinability (InflClass=Ind), negation (Polarity=Neg).

Lemmas follow SynTagRus too: feminine surnames and patronymics keep the feminine lemma («Ивановой» - «иванова»), superlatives stay superlative («важнейшей» - «важнейший»), «денег» - «деньги». When several OpenCorpora analyses coincide in UD (the conjunction and the adverb «так»), they are merged into one with the sum of their probabilities.

The labels are the OpenCorpora grammemes that have no UD feature: style (Infr, Slng, Arch), qualitative adjective (Qual), parenthetical word (Prnt), transitivity (tran, intr), whether the speaker is included in an imperative (incl - «пойдемте», excl - «иди»), the second locative (loc2 - «в лесу»), the future of «быть» (futr - «будет»: SynTagRus gives it Tense=Pres) and others.

Tokens outside the dictionary¶

Numbers, punctuation, Latin words and abbreviations are analyzed by rules, their source is "rule":

Token	Analysis
punctuation; symbols `%`, `$`, `№`, `−`, emoji	`PUNCT`; `SYM`
numbers in digits and ranges: «15», «3,14», «10 000», «17:04», «1990-2000»	`NUM` with `NumForm=Digit`, `NumType=Card`; an integer may also be an ordinal `ADJ` with `NumType=Ord` - for 1000-2100 (years), except round hundreds, this is the first analysis, and 1900 and 2000 are years too
numbers and ranges with an ending: «2005-го», «10-15-го», «5-ти»	an ordinal `ADJ` with the case and forms of «1-го», or a cardinal `NUM`
Roman numerals and ranges: «XX», «XIX-го», «II-III»	an ordinal `ADJ` with `NumForm=Roman`
Latin words: «iPhone»	`X` with `Foreign=Yes`
initials: «А.», «А»	`PROPN` with `Abbr=Yes`: a given name, a patronymic or a surname
abbreviations with a dot: «т.», «им.»	the analyses of the dictionary abbreviation without the dot
capitalized abbreviations: «НКО»	`PROPN` or `NOUN` with `Abbr=Yes`; a long word in capitals («ЧАТБОТАМИ») is also guessed as an ordinary word
adverbs «по-бурятски», «по-молодому»	`ADV` with `Degree=Pos`
other tokens: «км/ч», «РосПил»	`X` without features, the lemma is the token in lowercase

The shares of ambiguous analyses come from the frequencies in SynTagRus. The letters «й» and «ё» in decomposed Unicode and the non-breaking hyphen are converted to the usual ones.

CodeResult

[(a.lemma, a.pos, a.feats["NumType"], round(a.prob, 2)) for a in morph.analyze("2026")]
[(a.lemma, a.pos, a.feats["Case"]) for a in morph.analyze("2005-го")][:1]

[('2026', 'ADJ', 'Ord', 0.94), ('2026', 'NUM', 'Card', 0.06)]
[('2005-й', 'ADJ', 'Gen')]

Unknown words¶

The analyses of a Cyrillic word that is not in the dictionary are guessed by its ending: the analyzer takes the forms of dictionary words with the same ending - their part of speech, features and inflection. Their source is "guess". A lowercase word is guessed as a common word, a capitalized word mostly as a proper noun: a surname, a given name, a place, and a word in «-ский» mostly as an adjective. An abbreviation with an ending («МЭРТа») is guessed like a capitalized word.

A hyphenated word is analyzed by its last part: «крэш-тестов» - «крэш-тест», «IT-компании» - «it-компания»; if the last part is a particle («сколь-нибудь»), by the first one. If both parts inflect, the lemma is made of the lemmas of the parts: «дивана-кровати» - «диван-кровать», and such an analysis has no forms. A word with a known prefix - «эко», «анти», «супер», «микро» and others - is analyzed by the dictionary word without it: «экодомов» - «экодом». These analyses are guessed too.

A guessed analysis has forms, and forms, inflect and agree work with it; the forms keep the capital letters: «Мак-Киннона» in the dative is «Мак-Киннону», «ПИН-КОДОВ» is «ПИН-КОДАМ». A word with a particle inflects by its first part: «смотри-ка» in the plural is «смотрите-ка».

CodeResult

[(a.lemma, a.pos, a.feats["Case"], round(a.prob, 2)) for a in morph.analyze("чатботу")][:2]
guess = morph.analyze("Мартова")[0]
guess.lemma, guess.pos, guess.feats["NameType"], guess.inflect(Case="Dat")
morph.analyze("крэш-тестов")[0].lemma, morph.analyze("экодомов")[0].lemma

[('чатбот', 'NOUN', 'Dat', 0.44), ('чатбота', 'NOUN', 'Acc', 0.4)]
('мартов', 'PROPN', 'Sur', 'Мартову')
('крэш-тест', 'экодом')