Part of speech. Pronouns, determiners, predeterminers and articles are roofed by nouns and adjectives, and distinguished by values of other features. The num value is intended for cardinal numbers (numtype = card). Other types of numerals are roofed by syntactically defined parts of speech (adjectives or adverbs) and distinguished by values of numtype.

The difference between punctuation and symbols is that punctuation delimits parts of the sentence while symbols can be substituted for a word. For example, $ is not a punctuation, it is another form of writing the noun dollar. See also the definition of SYM for the Universal Dependencies.

Semantic classification of named entities and terms.

A deprecated feature. The only value that has not yet been moved elsewhere is pdt.

Although it reads as “pronoun type” (and we use the word “pronoun” for simplicity), it is also applied to words that are usually not considered pronouns, such as determiners, interrogative/indefinite adverbs (where, there, when, then, how, why) etc.

Subclasses of numerals. See also prontype (interrogative and indefinite numerals set it), numform and numvalue.

Is a number expressed by a word or by digits? Depending on tokenization and tagging scheme, this feature may be orthogonal to the distinction between cardinals and ordinals.

Low-value numerals display special behavior in some languages (e.g. Czech). This feature helps distinguish them. See also accommodability.

Note that this feature has not been approved so far, as it only occurs rarely in the old version of the Polish corpus, and it is not clear what exactly it encodes.

In Polish, special behavior of numerals (cf. numvalue) cannot be predicted from the value of the number. So unlike the tagset for Czech, Polish IPI PAN tagset (version 1 only; it disappeared from version 2) takes a different approach and encodes “accommodability”. The key question is: Does the numeral agree in case with the counted noun, or does it govern the noun and force it to genitive plural?

Příklady rec: trzech wileńskich i dwóch warszawskich; wyszło wielu znakomitych uczonych, pobierało dwunastu studentów, pięciu mych współpracowników; dziewięciu profesorów opowiedziało się

Příklady congr: obaj byli indywidualnościami; wszyci trzej prowadzili; dwaj synowie: Czesław; obaj synowie studiowali; czterej profesorowie głosowali

Řada výskytů číslovek nemá (ani v té verzi 1) vyznačenu ani jednu hodnotu, např.: o kilku artykułach, przedstawił trzy swoje dzieła, wystąpił z sześcioma prelekcjami, z których trzy zostały…

Subclasses of adverbs. See also prontype (pronominal adverbs set it).

Punctuation type.

Distinguishes between initial and final form of pairwise punctuation (brackets, quotation marks). Note that “initial” and “final” are better terms than “left” and “right”. The latter would be confusing in languages writing from right to left, like Arabic.

WARNING! This feature is deprecated. Most likely it will not be used in Interset version 2 drivers.

Does the pronoun or numeral behave syntactically as a noun, adjective, or adverb?

A word's morphological paradigm may behave like a different part of speech than the word is assigned to. For example, Slovak noun vstupné “admission (fee)” behaves syntactically as noun, is tagged as noun, but it originates from an adjective and retains adjectival paradigm. The paradigm feature of the sk::snk tagset maps to this Interset feature.

Is this a possessive adjective or pronoun?

Is this a reflexive pronoun?

See also the prontype feature.
In Arabic, definiteness is also called the state.
http://www.grammaticalfeatures.net/features/definiteness.html
http://de.wikipedia.org/wiki/Status_constructus
http://ufal.mff.cuni.cz/padt/PADT_1.0/docs/papers/2004-nemlar-padt.pdf

Possgender is possessor's gender.

Some languages distinguish only animate vs. inanimate, where the animate category includes humans, animals, fictious characters and sometimes also personified things. Some languages distinguish human vs. nonhuman, i.e. animals fall into the latter category. Some languages, e.g. Polish (see below) have a three-value system: human vs. animate non-human vs. inanimate. In that case we use the nhum value to denote the non-human animates, i.e. it excludes inanimates (while in hum-nhum systems, the nhum label includes inanimates).

The Polish word “który” (which) is an example of three-value animacy:

Pluralia tantum is a special case of plural, occurring e.g. in Czech. It applies to words that do not have singular forms. They use grammatical plural regardless of semantic number. Czech example: nůžky “scissors”: Papír rozstříhejte nůžkami. “Use scissors to cut the paper to pieces.” (semantic singular) vs. Koupil jsem si dvoje nůžky. “I bought two pairs of scissors.” (semantic plural)

Collective or mass or singularia tantum is a special case of singular. It applies to words that use grammatical singular to describe sets of objects, i.e. semantic plural. Although in theory they might be able to form plural, in practice it would be rarely semantically plausible. Sometimes, the plural form exists and means “several sorts of”. Czech example: lidstvo “mankind”.

Possnumber is possessor's number.

It applies e.g. to possessive pronouns and it can be different from their grammatical number, which is governed by agreement with the modified (possessed) noun phrase. Czech example: můj pes “my dog” (grammatical singular, possessor singular), mí psi “my dogs” (grammatical plural, possessor singular), náš pes “our dog” (grammatical singular, possessor plural), naši psi “our dogs” (grammatical plural, possessor plural).

Possednumber is the possessee's (possessed, owned noun phrase's) number.

In Hungarian, possession can be marked on the possessor or on the possessed. It is possible, though rare, that a noun has three distinct number features: its own grammatical number, number of its possessor and number of its possession. Examples from the Multext-East Hungarian lexicon:

• könnyedén (SSS)
  • könny = a tear (singular)
  • könnyed = your tear (singular owner)
  • könnyedé = (possession) of your tear (singular possession)
  • könnyedén = (on the possession) of your tear (superessive case)
• ellenfeleié (PSS)
  • ellenfél = an opponent (singular)
  • ellenfele = his/her/its opponent (singular owner)
  • ellenfelei = his/her/its opponents (core plural, singular owner)
  • ellenfeleié = (possession) of his/her/its opponents (singular possession)
• életeké (SPS)
  • él = point (singular)
  • élek = points (plural)
  • élén = his/her/its point (singular owner)
  • élünk = our point (plural owner)
  • életeké = (possession) of our point (singular possession)
• tárgyalópartnereinkét (PPS)
  • tárgyalópartner = negotiator (singular)
  • tárgyalópartnerei = his/her/its negotiators (plural, singular owner)
  • tárgyalópartnereinkét = (possession) of our negotiators (plural, plural owner, singular possession, accusative case)

Words marked for plural possessions are very rare, though. Note that in the following example from Multext-East, Columbus is marked for plural possession, but not for his own owner.

• Kolumbuszéinál
  • Kolumbusz = Columbus (singular)
  • Kolumbuszéi = (possessions) of Columbus (plural possession)
  • Kolumbuszéinál = (at the possessions) of Columbus (adessive case)

• Fine grained locative cases (Uralic languages)

http://en.wikipedia.org/wiki/List_of_grammatical_cases
http://www.hungarianreference.com/Nouns/-n-superessive.aspx

Ten příklad s knihou u partitivu máš blbě, ne? “Saanko lainata kirjaa?” je “Můžu si půjčit knihu?”, “tu knihu” by bylo “kirjan”, ne? — stepanek 30.10.2009 12:45

Copak já vím? Finsky neumím. Je to převzaté z anglické Wikipedie. — zeman 31.10.2009 16:29

Personal pronouns in some languages have different forms depending on whether they are objects of prepositions or not. For instance, Czech “on” (he) without prepositions has the forms jemu/DAT, jeho/ACC, jím/INS, while with a preposition it is němu/DAT, něho/ACC, ním/INS. Similarly, Portuguese pronouns in prepositional oblique case take forms different from oblique pronouns serving as direct objects of verbs: eu/NOM (I), me/ACC (give me that), mim/PREP-ACC (come to me).

Default empty value means that the word form is neutral w.r.t. a prepositional head.

Degree of comparison.

Note that this feature is used also for possessive pronouns, where it means the person of the possessor. E.g. “my” has person=1, “your” has person=2, “their” has person=3.

Possperson is possessor's person, marked e.g. in Hungarian. Don't use it for possessive pronouns (use person instead).

In quite a few languages, finite verb forms agree in person and number with the subject. In Basque, a subset of verbs agree with up to three arguments: one in the absolutive case, one in ergative and one in dative. To distinguish the different values of person, number (and politeness and rarely even gender), there are special features for each of the three arguments. Their names contain the three-letter code of the case of the argument: absperson, absnumber, ergperson, ergnumber etc. The value range is identical to the base features. That is, absnumber, ergnumber and datnumber may get the same values as number.

Adjectives in some languages (e.g. in Dutch) have different forms depending on how and where they are used. The same may hold for determiners (including possessive pronouns), quantifiers, numerals and participles.

Note: This feature has been introduced because of the nl::cgn tagset of Dutch. Similar features were previously observed in other tagsets (e.g. fa::conll) and they could be now re-implemented using this new Interset feature. We may also consider using this feature to distinguish the nominal (short) and pronominal (long) forms of Czech adjectives, the adverbial forms of German adjectives etc. (these are currently distinguished using the variant feature).

There are tag sets (e.g. Bulgarian CoNLL) that classify verbs as intransitive or transitive.

Note: In Interset 1.0, there were two separate features, tense and subtense, the latter consisting of the values aor, imp, nar and pqp. We used to avoid hierarchical feature values (e.g. aorist implies past; but many languages will only know past and their drivers will not check for the aor value). Hierarchical ordering of feature values is now normal and there are algorithms to exploit the hierarchy when a feature has to be replaced by another value, thus these two features have been merged.

Note that, unfortunately, imperfect tense is not always the same as past tense + imperfective aspect. For instance, in Bulgarian, there is lexical aspect, inherent in verb meaning, and grammatical aspect, which does not necessarily always match the lexical one. In main clauses, imperfective verbs can have imperfect tense and perfective verbs have perfect tense. However, both rules can be violated in embedded clauses. Aorist is aspect-neutral and can freely appear with both imperfective and perfective verbs.

See Wikipedia for a list of possible aspects.

Documentation of the METU Sabanci treebank classifies causative as voice (page 26). Note that this is a feature of verbs. There are languages that have also the causative case of nouns.

Evidentiality: what is the speaker's source of information?

Is this an abbreviation?

Is this a part of a hyphenated compound?

Is this a reduplicative or echo word? Such words occur in Hindi. In Hyderabad Dependency Treebank they get their own part-of-speech tags RDP and ECH, respectively. We do not want to treat them as separate parts of speech because they could be assigned a POS independent of their RDP or ECH status (same as the word that they echo). We may want to merge this feature in future with hyph into a new feature called compound.

For more details see Rupert Snell and Simon Weightman: Teach Yourself Hindi, Section 16.4 and 16.5, pages 210 – 211.

Indicates bad spelling, grammatical error etc. Does not say how the correct form looks like. Must be separated from the style feature because there can be typo in archaic as well as colloquial form.

Distinguishes between strong and weak forms of adjectives or pronouns. Used e.g. in Romanian UD. See also the variant feature. Some tagsets use variant=long instead of strength=strong, and variant=short instead of strength=weak. However, the strength feature has been tentatively added to Interset because it is slightly more specific and also because we want to be able to seamlessly read the features from the UD corpora that use it.

Allows for distinguishing between word forms that otherwise would share values of all features.

The tagset feature identifies the source tag set driver. Value should be identical to the name of the driver that filled the feature values. (The recommended format is same as in the use statement, without the tagset:: prefix, e.g. “de::stts”. Works together with the “other” feature.

Feature “other”
Any value or reference to array or hash is allowed. Serves to preserve information if the decoding driver happens to be the one who did the encoding. No other driver should expect anything meaningful here.
Only information that cannot be stored in other features should be stored here.
The apparently easiest approach — to store the complete original tag — will not work if the user needs to change feature values between decode() and encode().

Note that the tagset feature may sometimes refer to a related but different driver. For instance, drivers cs::pmk and cs::pmkkr are related. The latter is a reduced version of the former and its implementation uses the code of the driver for cs::pmk. Both drivers share their interpretation of the values of the other feature. They thus also use the same value of the tagset feature, namely cs::pmk. Sharing this identifier helps one driver understand the other values set by the other driver. The derived driver has its own identifier, cs::pmkkr, but this identifier is never mentioned in the feature structures. It would be more precise to say that the tagset feature identifies the language used in the feature structure, rather than the source tagset.