Entities constitute one of the three main types of concepts in UMR, alongside states and processes. They typically correspond to physical objects, but there are also abstract entities such as soul.
A mention of an entity in text can be specific or generic. A specific mention refers to a concrete, unique instance of the entity. Not necessarily by its name (meaning that not every specific entity is a named entity), but there is one specific instance that the speaker has in mind.
In (1a), there are three mentions of the same specific entity, the unique institution in Prague, called “National Museum”. It is first referred to by its name, then by the personal pronoun mu, and finally by the common noun muzeum. In contrast, the common noun muzea (plural of muzeum) in (1b) is a generic entity: It refers to various institutions that all belong to the category of museums.
- (1a) [cs] Národní muzeum v Praze získá nový bezpečnostní systém, který mu dodá firma CESS. Muzeum za něj zaplatí necelé 2 milióny korun. “The National Museum in Prague will get a new security system, which will be supplied by CESS. The museum will pay almost 2 million crowns for it.”
- (1b) [cs] V každé zemi podléhají muzea jiné legislativě. “In each country, museums are subject to different legislation.”
If a name is used, it typically refers to a specific entity, but as we see in (1a), specific entities can be referenced by other means, too. Even if the name were not present in the sentence, the context would tell us that we are talking about one specific museum, which probably has a name, and perhaps the context would be specific enough to allow us to identify the entity and its name in the real world. However, that is not a necessary condition for a specific entity. In (2a), staršího muže “an elderly man” refers to a person whom we do not know and who may not even exist in the real world (the text may be a work of fiction). The man may not be mentioned again and we may not learn anything else about him, yet in this local context he is a specific entity and not a generic one.
- (2) [cs] Když opouštěl budovu, zahlédl staršího muže, jenž nesl v náručí žlutou krabici. “As he was leaving the building, he saw an elderly man carrying a yellow box in his arms.”
A specific entity is a named entity if it is referenced by its name. Name is a word or a sequence of words whose purpose is to label a specific instance, not to describe a category of entities by their properties or relations to other entities. Thus muzeum is not a name because it can be used to refer to any institution that meets certain parameters. The phrase Národní muzeum is a name because it was specifically designed to label one particular museum.
The name does not have to be unique: An important museum in another country may also be called Národní muzeum, just like there are multiple people called John Smith. People may have to add more information if misinterpretation is possible, but the intended purpose of a name is to give the entity a reasonably locally unique identifier, and the purpose is what matters.
Furthermore, it is not necessary that a specific entity has only one name. For example, Spojené státy americké, Spojené státy, USA and Amerika are all names and all refer to the same country. One can even encounter just Státy used as a name and referring to the USA. When used in a Czech sentence in this manner, it cannot refer to, e.g., the United States of Mexico. The same word, státy “states” (not capitalized, unless sentence-initial), can be used as a common noun (hence not a name), referring to a group of entities (states or countries) that may be specific or generic. On the other hand, depending on context, Amerika may refer to a continent rather than to a country (North America, South America), or it may refer to a quarry southwest of Prague.
Names are thus designed to label specific instances, while common nouns are meant to describe broader categories (types). The borderline may be occasionally blurry when a common noun is repurposed as a name (as we have seen with Státy above) but it is much less likely that a name will be used for a generic entity. We can certainly define a category of all people named Václav, as in (3a), but that does not convert the name into a common noun – all these people first got that name with the hope that it will make them identifiable and distinguishable from other people, and only later the speaker artificially grouped them, using their name as the property defining the group. In (3b), the plural of a personal name denotes a group of people who do not even bear this name but rather share similar character features as someone whom the name refers to.
- (3a) [cs] Všichni Václavové by měli znát své slavné jmenovce. “All Václavs should know their famous namesakes.”
- (3b) [cs] všichni ti Havlové “all those Havels” (a critical mention from a political opponent, about the politian Havel and people sharing his point of view)
While the use of Václavové in (3a) is unusual, there are names that denote a type rather than an instance. A primary example is product names, as in (4) (the specification of product category is enclosed in parentheses in the example because it is optional):
- (4) [cs] Používám (prací prostředek) Persil. “I use Persil (detergent).”
Clearly, Persil is a name rather than a common noun, as it was invented specifically to distinguish this detergent from other detergents; it is not a common noun that we expect to find in dictionaries. However, the name denotes a type of product, not one particular instance. There are millions of packages of Persil, and they all share this name. And while the name could be used when referring to a specific package, in (4) it actually refers to a generic entity. We will use the term categorial names / categorial named entities with names that denote types (categories) rather than instances.
Speaking of terminology, common nouns
(NOUN in UD) are
contrasted with proper nouns
(PROPN in UD). These
are part-of-speech categories, used always for a single word. Proper nouns
are single-word names. However, many names of entities are multiword
expressions, and the individual words in those names may belong to other
categories than PROPN.
An entity that is referred to by a common noun is represented by a regular concept (node), typically with the lemma of the noun as the label of the concept (but occasionally the label may be a multi-word string). This is done no matter if the entity is specific or generic.
In UMR, An entity that is referred to by a name is represented by an abstract
concept corresponding to the semantic class of the entity (e.g., “person” or
“organization”; see below for the taxonomy of semantic classes). The name of
the entity is in a separate node, which has the abstract concept “name” and
is attached to the class concept via the relation :name. Individual
orthographic words of the name are listed in the name concept each in its own
attribute, the attributes are named :opX where X is the ordinal number of
the word within the name. The words are not always exact copies from the
sentence, as the name is converted to its canonical form. Note however that
this does not mean that all words in the name are replaced by their lemmas;
some will be lemmatized, others will stay in a frozen inflected form.
An entity that is referred to by a pronoun is represented by an abstract concept corresponding to the semantic class of the entity. Unlike named entities, there is no child node with the “name” concept.
Common noun muzeum “museum”:
(m/ muzeum)
Named entity Národní muzeum “National Museum”:
(o/ organization
:name (n/ name
:op1 "Národní"
:op2 "muzeum"))
- (5) [cs] Získal práci na Ministerstvu školství, mládeže a tělovýchovy. “He got a job at the Ministry of Education, Youth and Sports.”
(o/ organization
:name (n/ name
:op1 "Ministerstvo"
:op2 "školství"
:op3 ","
:op4 "mládeže"
:op5 "a"
:op6 "tělovýchovy"))
Note that the canonical form of the multi-word name of the ministry in (5) is
composed of the canonical form of the head (Ministerstvu was converted to
nominative singular, but its capitalization was retained) and the inflected
forms of the dependent words; the comma is also a separate :opX attribute.
With categorial named entities, the :name relation can occur even with a
generic entity:
(p/ product
:name (n/ name
:op1 "Persil"))
If an entity is mentioned both by its name and by a descriptive common noun, we use a lexical concept of that common noun and add the name relation. If, in contrast, the entity is mentioned only by the name, we use an abstract concept (labeled with English words) from the taxonomy introduced below. Compare:
- (6a) [cs] Estonské volby vyhrál blok Vlast. “The Fatherland bloc won the Estonian elections.”
- (6b) [cs] Vlast má více hlasů než všichni ostatní. “Fatherland has more votes than everyone else.”
In (6a), the name will be attached to the (Czech) concept blok. In (6b),
the name will be attached to the (abstract, English-derived) concept
organization, or here, the more specific political-organization.
(b/ blok
:wiki "Q163347"
:name (n/ name
:op1 "Vlast")
:ref-number Singular)
(p/ political-organization
:wiki "Q163347"
:name (n/ name
:op1 "Vlast")
:ref-number Singular)
UMR defines the (optional) :wiki attribute, which can be used to link a
concept to a corresponding article in Wikipedia. The examples in the UMR
guidelines currently show names of English Wikipedia articles in these
attributes; however, a more robust and thus preferred solution is to use
Wikidata identifiers. They are not bound to a particular language mutation of
Wikipedia (all Wikipedias that have an article about the concept are linked
from the Wikidata page) and they should be more stable (e.g. when one of the
Wikipedias decides that a different title should be used for the article and
the old title should become a redirect). Obtaining Wikidata identifiers is
easy: Let's assume we want to anchor the Czech entity Národní muzeum and we
find its article in the Czech Wikipedia at
https://cs.wikipedia.org/wiki/N%C3%A1rodn%C3%AD_muzeum.
In the menu on the right-hand side we see a link labeled “Položka Wikidat”
and leading to
https://www.wikidata.org/wiki/Q188112.
Although the attribute is optional in UMR, in our data we should strive to provide it for every mention of a specific entity that has a Wikidata entry. (In practice, we could use the coreference annotation in UMR to automatically propagate the anchor from one mention to all other mentions. Note however that it would be a mistake to say that we only fill the attribute manually for named entities. It can happen that a specific entity is never mentioned by its name in a document, yet the context doubtlessly points to a known entity described in Wikipedia.)
(o/ organization
:wiki "Q188112"
:name (n/ name
:op1 "Národní"
:op2 "muzeum"))
If a specific entity has no Wikidata presence, we have to register it in a
local ontology that becomes part of the annotation, and provide a local
identifier instead. Note that the entries in the local ontology are not
always local to just one document. They are still part of the same universe
that is partially described in Wikipedia. Consider, for example, a news
article reporting that A man (80) was killed this morning in a traffic
accident. There could be several other documents reporting on the same
event, and if it can be established that they are indeed talking about the
same accident, then all the mentions of the nameless man should be anchored
to the same entry in the ontology. TODO: Implement a
prototype of the local ontology and specify how it should be linked from the
annotation. We should probably use a different attribute, e.g.,
:lwiki.
Every entity concept should have the attribute :ref-number, with the value
reflecting the grammatical number. UMR defines a number of possible values
for the attribute, based on grammars of various languages. For Modern Czech
data the value will probably (almost?) always be Singular or Plural. We
will not use Dual just because we know that we are speaking about two
people. The dual as a grammatical number has mostly vanished from Czech, and
UMR has other means to annotate quantity (there is not a separate value of
:ref-number for each integer number). A possible exception in Modern Czech
is paired body parts (nohy, ruce, oči, uši) because that is where
grammatical dual still survives.
Abstract entity concepts that correspond to personal pronouns (or their
reflections in verbal morphology) will additionally have the attribute
:ref-person. This attribute is not used with other entity mentions (for
which it would be very unusual to interpret them as anything else than 3rd
person).
We now repeat example (1a) as (6) here and show the annotations of entities from the example, using all the rules specified so far. (We omit the monetary entity from the end because such types of entities have not been discussed yet.)
- (7) [cs] Národní muzeum v Praze získá nový bezpečnostní systém, který mu dodá firma CESS. Muzeum za něj zaplatí necelé 2 milióny korun. “The National Museum in Prague will get a new security system, which will be supplied by CESS. The museum will pay almost 2 million crowns for it.”
(o/ organization
:wiki "Q188112"
:name (n/ name
:op1 "Národní"
:op2 "muzeum")
:ref-number Singular)
(c/ city
:wiki "Q1085"
:name (n2/ name
:op1 "Praha")
:ref-number Singular)
(s/ systém
:lwiki "L1"
:ref-number Singular)
(o2/ organization
:wiki "Q188112"
:ref-number Singular
:ref-person 3)
(o3/ organization
:lwiki "L2"
:name (n3/ name
:op1 "CESS")
:ref-number Singular)
(m/ muzeum
:wiki "Q188112"
:ref-number Singular)
(t/ thing
:lwiki "L1"
:ref-number Singular
:ref-person 3)
The UMR guidelines give a taxonomy of entity classes, types and subtypes in Section 3-1-2. They can be used as abstract concepts for named entities and for entities represented by pronouns. As of now (June 2023), the taxonomy is reportedly under revision by the UMR team. At any rate, the current table has a number of issues. Below we examine some of the entity types and discuss their utility.
Some other resources:
- Live Google spreadsheet where the UMR team (Julia & Co.) update their view of the UMR taxonomy. This is probably more up-to-date than the guidelines on Github. Some types have links to wikidata, which should provide their definition.
- Comparison of AMR and Prague taxonomy of named entities from the JHU-PIRE workshop 2014.
A top-level class without subordinate types and subtypes. Besides humans, the class could serve as a natural fallback for human-like beings that do not have a class of their own: deities, dwarves, hobbits, elves etc. What about robots?
A top-level class. As a named entity, it can be used to represent a pet that was given a name by its owner, or an animal character in a fairy tale, including fantastic beasts like dragons. The class is not suitable to represent species of animals. (They have their own type in the taxonomy but it is problematic, see below.)
Analogous to animals but covering plants.
Not listed in the current UMR table but used in their examples (e.g. sentence (3) in Part 1) and clearly needed at least for pronouns that refer neither to persons nor to animals or plants or other types specifically listed in the taxonomy.
A subset of what other named entity taxonomies often label as “location”. This subset contains only phenomena created by nature, not by mankind. The UMR guidelines currently provide 15 types belonging to this class, probably not exhaustive and to be extended in the future. The annotators should use the types as abstract concepts if they know them, otherwise they can fall back to the whole class.
The types are: ocean, sea, lake, river, gulf, bay, strait,
island, peninsula, mountain, volcano, valley, canyon, desert,
forest.
Like geographic-entity but on a cosmic scale. Currently four types: moon,
planet, star, constellation. Again not exhaustive: What do we do with
objects that are smaller than planets but are not moons?
Note that for Měsíc “Moon”, Země “Earth” and Slunce “Sun”, there is a blurred context-based borderline between a name of a celestial body and a common noun. But at least the Czech grammar puts the burden of decision on the shoulders of the author: if it is a name of a celestial body, it has to be written capitalized. So, unless it is the first word of a sentence, the annotator can take capitalization as the cue.
A subset of what other named entity taxonomies often label as “location”. This subset contains only phenomena created by mankind, not by nature. The UMR guidelines currently provide 7 types belonging to this class, probably not exhaustive and to be extended in the future. The annotators should use the types as abstract concepts if they know them, otherwise they can fall back to the whole class.
The types are: country, state, province, territory, county, city,
city-district.
The current selection is too much focused on North America. For example, the
Czech sub-country administrative unit, kraj “region”, is neither a state,
nor a province, territory or county. Czech okres could probably be
annotated as county. See also the region class below.
A class with three types: world-region, country-region, local-region.
There is no definition and it is difficult to guess what the authors had in
mind. But it is not in the geo-political-entity class, so it is probably
not meant as an official administrative unit and it does not solve the
problem with Czech kraj. Maybe it is meant for less official or formal
regions such as Valašsko, Morava or Evropa. Still, we need criteria to
decide between the three types of regions.
A class of man-made entities that have a fixed location but the name does not pertain just to the location but also to the building (or whatever other facility it is). In some cases, a facility is also an organization (example: museum), which is a different entity class, but they may share a name. Then the annotator will have to decide by context whether the utterance is more about the legal entity (organization), or the place or building (facility).
The guidelines currently list 19 types of facilities: airport, station,
port, tunnel, bridge, road, railway-line, canal, building,
theater, museum, palace, hotel, worship-place, market,
sports-facility, park, zoo, amusement-park.
A large class with 6 types: family, clan, ethnic-group,
regional-group, religious-group, organization. The common
characteristic of all six is that they denote groups of people. In the case
of organization, it also has a common property, set of activities, and
usually also some kind of legal existence; this may be also true to some
extent about families and even clans, but typically not about the other types
of groups. More importantly, an organization typically has a name that refers
to the organization as a single entity (IBM), while an ethnic group is
often denoted by a plural form of a name that labels one member of the group
(Baskové, the plural of Bask “(a) Basque (person)”). It is thus unclear
whether and why these diverse types should be analyzed the same way. An
ethnic group is more like a categorial named entity (see also product), it
denotes people who share a language and/or some cultural and historical
heritage. Similarly, religious groups is just a common label for people who
share beliefs. Do we also want a “named entity” for races, or for
tall/short/slim/fat people etc., or for people who share political views, or
love for rock-and-roll, or anything else?
A large type of the social-group class, see above for my doubts about its
(dis)similarity to other types. There is much less doubt (than for the other
types) that organizations are named entities, but the definition of the
subtypes has to be clarified. The type has currently 11 subtypes:
international-organization, business, company,
government-organization, political-organization, criminal-organization,
armed-organization, academic-organization, association,
sports-organization, religious-organization.
A separate top-level class. I do not understand why the authors did not make
it a type of social-group, just like ethnic-group and regional-group,
to which it is very close. For example, Čech “Czech” can be, depending on
context, any of the three: A member of the ethnic group (sharing the Czech
language and traditions, including people who do not have Czech citizenship,
as their ancestors left the country and settled abroad); a member of the
nationality (having the citizenship of the Czech Republic, even if living
abroad and/or having a mother tongue other than Czech); a member of a
regional group, living or born in Čechy “Bohemia” (as opposed to Moravia
and Silesia, which are the other two parts of the country called Česko
“Czechia”).
The nature and usage of the nationality class has been questioned in issue
10 of the UMR
guidelines. Until the
issue is resolved, we tentatively stick to the following rules:
- Nationality is not distinguished from any other affiliation with a country. Such affiliation is contained in expressions like Američan “American”, americký občan “American citizen”, americký usedlík “American resident”, obyvatel Spojených států “resident of the United States”, americké auto “American car” etc.
- Such affiliation is expressed with the
:modrelation and the concept describing the country as the child node. The:wikiof the child node will always refer to the country, while the:namemay differ depending on what was in the text: "Spojené státy" for Spojených států, and "Amerika" for Američan, americký, americké. Nonationalityconcept is used.
(p/ person
:mod (c/ country
:wiki "Q30"
:name (n/ name :op1 "Amerika")))
Note that an analogous approach can be used to encode affiliation with regions other than countries, e.g., Evropan “European”, Moravan “Moravian”, Pražan “inhabitant of Prague, Praguer”.
(p/ person
:mod (c/ city
:wiki "Q1085"
:name (n/ name :op1 "Praha")))
This class is not listed in the current UMR guidelines, which seems to be a
gap that has to be patched. The current taxonomy actually lists some types
that could be classified as special cases of product, such as
aircraft-type or car-make. But there are names for other products, like
Persil in example (4) above. Products are typical examples of what we call
categorial named entity.
This is a separate class with currently 5 types: ship, aircraft,
aircraft-type, spaceship, car-make. Note that aircraft-type and
car-make are categorial named entities that would be better described as
special types of product. For ship and spaceship it is more typical
that a name denotes a single instance (e.g., Titanic). Perhaps aircraft
is also meant to denote an instance rather than a type. People may
occasionally give a nickname to their car, then the name will also denote an
instance and the entity type car-make will not be suitable for it.
This is a separate class with no types. It could be regarded as a type of
product.
This is a separate class with no types. It seems to be a categorial named
entity like product but it has an unsharp boundary between names and
descriptions of dishes, so it is quite questionable whether, how, and where
exactly the class concept should be used.
This is a separate class with currently 8 types: work-of-art, picture,
music, dance, show, broadcast-program, literature, publication;
the last one has subtypes book, newspaper, magazine, journal. Since
there is no description, it is not clear what is the difference between
literature and publication supposed to be. Also, there does not seem to
be a category suitable for movies.
Some works of art could be seen as a categorial named entity similar to
product: Typically there are many copies of a book, a movie, or a CD. But
even here the prototypical reading is that the name refers to the single
intangible work, not to one of its copies.
This is a top-level class with two types: court-decision, treaty.
Supposedly the class itself should be used for actual laws. There is a need
for other types, such as a named dean's regulation at a university.
One could say that laws are close to publications; but they can hardly be categorized as a cultural artifact.
This is a top-level class without types. Supposedly there is no distinction
between languages and dialects, i.e., names of dialects would also be labeled
as language. Not sure about language groups and families. Note that names
of languages are often (but not always) related to names of ethnic groups,
nationalities, regions and countries.
It is not clear how this label is intended to be used. Should it apply only to the name of the language (noun), e.g., angličtina “(the) English (language)”, or also to adverbs (Mluví anglicky. “He speaks English.”) and adjectives (Procvičuje si anglická slovesa. “She is practicing English verbs.”)
Do we also use this label for constructed languages (esperanto)? I think we
do. Do we also use it for programming languages (Pascal, C, Perl, Python)?
I am not sure. Maybe those fall under the class computer-program.
This is a top-level class with currently three types: writing-script (e.g.
dévanágarí), music-key, musical-note. It denotes an abstract entity.
This is a top-level class without definition and without types. It would be
useful to have an example. See also event below.
This is a top-level class with currently 8 types: incident, war,
natural-disaster, earthquake, conference, game, festival,
ceremony. Besides the usual problem that types are not defined, it is not
clear why earthquake shall be distinguished from other natural disasters.
It is also unclear why cultural-activity is a class separate from event.
Furthermore, note that this concept denotes events as entities, although events are typically processes (refer to the main distinction between entities, states and processes, shown in Table 1 in Section 3-1-1 of the UMR guidelines). Section 3-1-1 even uses the term “event” to refer to all processes in any packaging, plus entities and states when used in predication. Nevertheless, if an event has a name (such as Druhá světová válka “World War II”), it is covered by this taxonomy. It would be helpful to have an annotated example here.
Top-level class with no types. Supposedly, Nobelova cena za fyziku “Nobel Prize for Physics” would be an example.
This is a top-level class with currently 18 types:
molecular-physical-entity, small-molecule, protein, protein-family,
protein-segment, amino-acid, macro-molecular-complex, enzyme,
nucleic-acid, pathway, gene, dna-sequence, cell, cell-line,
species, taxon, disease, medical-condition. They are obviously
inspired by the bulk of work on biomedical processing and we would need more
documentation to understand how the authors intended to use them.
However, at least three types reach into layman's language: species,
taxon, and disease. The closely related species and taxon would be
categorial named entities (like product), where the name denotes a whole
category (type) of entities rather than a single instance. That is, if they
deserve to be treated as named entities in the first place. For example,
kočka “cat” is an animal with a particular set of characteristics, just
like dub “oak” is a particular type (hyponym) of tree, and hrad “castle”
is a particular type of building. But the first two words are biological
genuses, hence taxons, while hrad has no special status in the UMR
taxonomy. (In the Czech grammar, all three are common nouns.) There is no
reason why kočka and dub should be named entities. And by extension,
there is little reason why species should be named entities, for example
kočka domácí “cat (Felis catus)”, or dub letní “pedunculate oak (Quercus
robur)”, or why other taxons should, for example šelmy “beasts of prey,
Carnivora”, savci “mammals”, or živočichové “animals, Animalia”. It is
true that some species have names that are less common than others and were
invented by scholars who discovered and described the species, rather than
being part of the language since ancient times. But it would be neither
tractable nor helpful to attempt to distinguish them. Perhaps the only
exception is the scientific names in Latin, provided that the language of the
annotated text is not Latin.
Similarly, diseases may have scientific names but many common diseases are just common nouns or expressions (angína “tonsillitis”, chřipka “flu”, mor “plague”, neštovice “chickenpox”) and it is not clear why they should be handled differently from other common nouns. Moreover, diseases are states rather than entities.
What is this supposed to mean?