Knowledge graph theory can be considered to be one of the methods to deal with natural language processing. The theory belongs to the theories about semantic network, but has an advantage that it uses a very restricted ontology. Knowledge graphs have been successfully applied to represent almost any characteristic feature in English and Chinese. Indonesian language, on the other hand, has a very different structure as compared with English and Chinese. In this research we investigate the application of knowledge graphs to represent some characteristic features of Indonesian language. The characteristic features to be considered are active and passive form of verbs and the derived nouns. The result shows that knowledge graph can also represent those features effectively. It can be concluded that knowledge graphs can be used to represent various languages with different characteristics effectively.
Knowledge graph theory was developed in 1982 at the University of Twente and
Groningen. The theory can be considered to belong to the theories about semantic
networks. However, the theory is essentially different from other theories, in
particular in the fact that a very restricted ontology is used. For graph theoretical
terminology we refer to the book of Bondy and Murty [
For an introduction to the theory, we refer to the theses of Willems [
The theory of conceptual structures was presented by Sowa [
The basic idea of the theory is that in the mind representation of the world is
present that has a discrete mathematical nature, so can be modeled by a knowledge
graph, that is called mind graph. The vertices of this graph correspond to somethings,
the genus of all concepts. "Something" may be a perception unit, then is represented
by a single token but, more generally, will be a complex structure of tokens that are
linked by links of certain types. So then a subgraph of the mind graph is considered,
the elements of which are "taken together" so, literally, form a concept. The first type
of frame is used to indicate that the subgraph is seen as a unit. That frame may be
called an AND-frame, an idea that goes back to Peirce [
We will now focus on the 8 types of links, 3 edges and 5 arcs. For the choice of
these 8 elements of the 13-elements ontology (1 token, 8 links, 4 frames) we refer to
Hoede [
One slogan of the theory is: "FRAMING AND NAMING". Concepts are seen as contents of a frame that is then named, i.e. to which is then attached a word. Note that this procedure is considered to be the same in any language. A second slogan is: "THE STRUCTURE IS THE MEANING". This is an extremely important pillar of the theory. The semantic aspect is equated to the structure of the mind graph. The meaning of a word is the associated structure in the mind of the interpreter of the word. Most linguistic theories try to keep the speaker or listener, i.e. the human, out of the theory.
Attaching a word is represented in two ways by directed links of type ALI and EQU. Note that these links do not connect two tokens but a word with a token. A simple example is :
ALI
EQU dog
Pluto a knowledge graph that is to be read as “something” of type dog instantiated by “Pluto”.
We have herewith shortly described some background and the formalism. The rest is playing with structures, for which the third slogan holds: "THINKING IS LINKING SOMETHINGS".
In knowledge graph theory every word has a corresponding word graph, expressing
the meaning of the word and therefore called a semantic word graph. Next to that, a
word has a certain type, like noun or verb, and in each language ways of linking to
other words. In English "the cat", a determiner followed by a noun, is possible. "Cat
the" is not a linguistic formation that corresponds to a grammatical rule. The rules of a
generative grammar determine for a word type in which way the word can be linked
to other words. The arising graph is called syntactic word graph; see Zhang [
Combining semantic word graphs of words in a sentence leads to a sentence graph. The graph representing the combination of sentence graphs of sentences in a text is a text graph, expressing the knowledge described by the text.
In most linguistic theories the accent lies on the syntax, the way correct sentences are generated. Such theories are strongly influenced by the theories of formal (computer) language. Putting the accent on the semantics from the beginning allows partial sentences to be represented and to be interpreted. For interpretation, i.e. giving meaning to, grammatical rules are of minor importance as, up to the simplest single word, meaning is identical with the word graph. Languages may differ strongly in syntactical respect, but follow the same semantical procedure in knowledge graph theory.
This claim of universality across languages led to the choice of Chinese as object
of study, in particular the very specific features of that language, see Liu [
• The definitions themselves.
Two lexicons of word graphs may differ in the same way, in the complexity of the word graph and in the structure of the word graph. Most dictionaries try to keep things simple, to grasp the "essential" meaning of nouns and verbs. These two classes are very closely related. Verbs differ from nouns in that a time aspect is explicitly understood to be present. This close relationship will be subject of discussion in a later section. In the representation, the difference comes forward in the CAUrelationships used to link subject and object to the verb. Consider the sentence graph: man
dog
ALI hit "Hit" is a transitive verb. The sentence "John sleeps" would be represented by: man
ALI
CAU EQU
ALI John sleep where now "sleep" is intransitive and only has an incoming CAU-arc and "John" is seen as instantiation of "man" (a dictionary might give "name of a man", when looking up "John"). The prepositions are of a quite different nature. They form the "glue" of a language. This is particularly clear in Japanese. Nouns and verbs are hardly differentiated, but particles play a dominant rule, see Hoede (2005). Their word graphs are very small and can be expressed by the links in knowledge graph ontology.
The second set of word graphs of Hoede and Liu (1998) focused on words that attach to nouns and verbs, like adjectives and adverbs, but in particular on classifier words, a linguistic feature of Chinese, and somewhat less often used in Indonesian Language. Classifiers have to be expressed to indicate a typical aspect of a concept, see Liu (2002) or Hoede and Liu (1998). In Indonesian Language the word "ekor" means "tail" and is used whenever an animal is mentioned. Other classifiers are "orang" for people and "buah" for things, like in "tiga buah pisang", "three banana(s)".
The third set of word graphs of Hoede and Zhang(2001), focused on those words that express logical aspects in language, see also van den Berg (1993).
The reader should have enough information now to follow the discussion of some specific features of Indonesian Language, the use of prefixes and suffixes to express the formation of nouns, respectively active and passive forms of verbs.
The noun is indeed a basic type of word. The token seen as AND-frame with concept C as content can be interpreted as expressing "C being", so also might be called "BEframe". As such, framing a part of the mind graph leads to nouns. Let us make this clear by an example. We consider a CAU-arc in total graph form, which means that also the arc is represented by a vertex linked to the incident vertices with auxiliary unlabelled arcs, like in the following figure:
CAU ≡
We choose this form to make our procedure clearer with respect to the CAU-arc. We consider three frames:
Frame 1 can be named "cause", frame 2 can be named "effect" and frame 3 can be named "causation". So, all three frames get names that are nouns. If a certain process is framed in English, the process is referred to by the ending -ing. So instead of 3 1
bite biting 2
ALI ALI
CAU we may also describe by where the word now is a noun. As we remarked before, noun and verb do not differ very much. This also diminishes the difference between adjectives and adverbs. Consider "nice dog", "nice skating", and "skating nicely" as an example why both types of words were collected as "adwords".
We will now go over to a systematic account of noun formation in Indonesian Language.
Kata jadian and kata benda stand for derived word and noun, respectively. One of the important ways to derive a noun is by the prefix pe-. The basic meaning of pe- is "he/she who ... ", where the dots are to be filled in according to the type of word of which the noun is derived.
We consider an example: pe- ajar, which changes into pe(l)ajar. Ajar is the verb for study(ing), so pe-ajar is "he/she who studies/is studying", i.e. a student. The word graph is as follows: orang ajar The inclusion of orang = man/woman is due to the fact that it is felt as definitely belonging to the concept of student.
This at first seems a somewhat strange situation when we see "he/she who ... " as basic meaning of the prefix pe-. We consider the example pe-layar = sailor, where layar = sail, but as a noun. It is remarkable that in English sail gets a suffix -or with the same function as pe- in Indonesian Language. The explanation, of course, is that "using a sail" on a boat has been shortened to "sailing", from which a verb "to sail" has developed. The filling in of the dots therefore maybe given as "he/she who uses a sail". The "expansion" of the concept sail is not "brought under words".
An even more striking example is pe-laut, laut = sea. The meaning is "seaman". So in English too the combination of two nouns can give a new noun. The derivation in graph theoretical sense would be more complicated than for pe-layar and a complicated word graph for pe-laut would in principle be needed. However, we can now fill in the dots as "he/she who is associated with the sea" and represent the association with a PAR-link. The graph then becomes:
PAR
ALI orang
ALI laut
ALI pelaut and for pe-layar the same graph with "laut" replaced by "layar" can be given.
As third major use of pe- we mention the combination with an adjective. Consider "he/ she who is ..." and the possibility of deriving a noun from an adjective is clear. We give only one example. Takut = afraid, so pe- takut is somebody who is afraid, i.e. a coward. The word graph is orang takut
The counterpart of the prefix pe- in English is the suffix -ard. In Dutch takut = laf and the word for "coward" is "laf-aard".
A rather universal way to derive a noun is by the suffix -an. The meaning can be described by "that what..." With respect to transitive verbs the suffix-an can be said to have the same function with respect to the patient as the prefix pe- has with respect to the agent. Simple examples are the verbs makan = eat and pikir = think. Makan-an = food, whereas pikir-an = "that what is thought", i.e. "thought". The wordgraph for makan-an is
Take, for example, hari = day. Adding a suffix -an to this word we get hari-an = daily. In this case “hari” is attributed to something, so the word graph is
CAU
ALI makan
PAR
ALI
ALI
makanan
In Dutch dag = day and blad = journal = majalah. A daily journal is called a "dagblad", so two nouns are simply joint. In English and in Indonesian Language the nouns day respectively hari are modified when joint with journal respectively majalah; "daily journal" and "majalah hari-an". The word hari-an gets the function of an adjective.
Here the suffix -an creates a noun again. Manis = sweet, whereas manis-an is "that what is sweet". The word graph is
According to the meaning given to pe- and -an as respectively "he/ she who ..." and "that what ...", the combination should express "something, -an, that is associated to somebody, pe-, in relation to a stem word". A nice example is pe-ajar-an. The stem is ajar = study, pe-ajar is, as we have seen before, a student. The something associated to the student, pe-ajar-an, is a lesson.
Combining the graph representations we give as word graph for pe-ajar-an p e l a j a r
ALI
CAU
CAU
ALI ALI
ALI orang ajar p e l a j a r a n
A rather standard example is given by the stem lapor = report, as a verb or an activity "reporting". We already remarked that verb and noun do not differ much. Kerja might also be equated with "working". Pe- lapor is a reporter and pe-lapor-an is the report as that what is produced. Note that in English report both refers to the activity and to the result of the activity. Indonesian Language makes more distinction here. 3.4
Consider words in English ending on -ty, like e.g. ability derived from "able". In Indonesian Language, mahir = able and the noun ke-mahir-an = ability.
The meaning of the combination of prefix ke- and suffix -an can be given as "property of being ...". We therefore use a BE-frame, i.e. an AND-frame without content, and let it contain whatever is filled in for the dots. So the word graph for kemahir-an is given as mahir
ALI
ALI kemahiran
In Dutch often the suffix -heid is used. Bekwaam = able and bekwaam-heid = ability. In German the suffix -keit is often used, like in "ewig-keit" = eterni-ty = ke-abadian, where abadi = eternal. A funny example is given by the stem ada = being. Keada-an then has the meaning "property of being being" i.e, ke-ada-an = situation.
We herewith conclude our account on the formation of nouns by prefixes and/or suffixes.
In most languages the verb is the most complicated type of word. Indonesian Language Indonesia is no exception, but is comparatively simple due to a rather systematic use of prefixes. Like for noun formation various stems can be chosen in verb formation. Next to that, prefixes are used to express active and passive form. We will discuss the prefixes ber-, me- and di-, again without discussing the morphological issues of spelling changes. These will come forward in the word graphs given.
A verb formed with ber- is an active form that does not have a patient and indicates the situation in which the agent is.
The verb ber-V has the extra meaning "being in the process of V-ing". The word graph is therefore given as
An example of the change in meaning induced by ber- is given by angkat = lift. Angkat besi = lift iron, i.e. weight lifting. Ber-angkat is an active form without patient and describing a process. The meaning is "leave". “Bus ber-angkat” says that "the bus is in the process of lifting", lifting itself so to say. The word graph is
V
ALI
ALI
ber- V angkat
ALI
ALI berangkat
Like for the formation of a noun from a noun, recall pe-laut = seaman, the derivation from the basic meaning can be rather complicated. A good example is given by the stem malam = night. The verb ber-malam describes an activity associated with "night". The English description is "stay overnight". As word graph we give
PAR ALI
ALI orang malam
ALI bermalam
There are interesting other formations with ber-. It can be combined with an adjective like in ber-gembira, where gembira = happy, and the verb expresses "being happy". In ber-dua = with two, we see a combination with a number word.
Ber- with a doubling of the stem adds another piece of meaning. Ber-dua-dua = in groups of two, ber-puluh-puluh = in groups of ten. Doubling can also be used for expressing intensity. Tahun = year and ber-tahun-tahun is best translated as "year in year out". We only want to stress here that ber- clearly adds that the situation is considered.
The prefix me- knows many morphological changes that we will not discuss. It gives a verb in active form in which the focus is on the action. Di-, on the other hand, gives a verb used in a passive sentence where the focus is on the patient.
The main difficulty we meet is the representation of the focus. Let us consider verb V and the knowledge graph 1
CAU
CAU 2
V
ALI
The frames 1 and 2 focus on the agent of V respectively the patient of V. However, we already met these frames in the formation of nouns by the prefix pe- and the suffix -an.
We will now give some examples without giving word graphs as the structure of these should be clear.
For this formation the verb may have a patient or not. Beli = buy and me-beli = buy + extra meaning as described before. There usually is a patient, the verb is transitive. Desah = sigh and me-desah = sigh + extra meaning as described before. There is no patient, the verb is intransitive.
There are several basic verbs where me- might not be used as a prefix. Tidur = sleep is one such verb, makan = eat, minum = drink or ingin = want, are others.
Me- plus noun is perhaps clearest in me-telepon = phone (use the telephone). The word graph is m e n e l e p o n
ALI
ALI telepon
PAR
CAU
PAR
F
Here, as an alternative for using total graphs, we introduce an extra element in the ontology of knowledge graphs to express focus. The symbol F, like the types of the links to be seen on the meta-level and not on the word-level, might be attributed to a token. The explicit connotation of focus when using me- and di-, as well as intonation, so far not been dealt with in the theory, seems to force the introduction of a link between the symbol F and a token. As focus is typically attributed by the speaker, the presentation chosen is by a PAR-link, like in the word graph given above.
The link between symbol and token may therefore in our theory now be an ALIarc, for typing, an EQU-arc, for instantiation or a PAR-arc, for focusing, seen as a verbal expression.
For me- plus verb and di- plus verb we would now give m e V d i
V respectively
ALI ALI
CAU
CAU F
PAR
ALI
V CAU
CAU ALI
PAR V
F
This gives a clear simplification of the formalism, at the cost of adding a fourteenth element to the ontology.
Me- plus adjective is exemplified by lemah = weak and me-lemah = weaken, an intransitive verb, with, now, word graph m e l e m a h
ALI
ALI lemah
CAU
CAU
PAR
F
We can be rather short now about the prefix di- that is used to express a passive sentence about a patient on which the focus lies.
Consider the sentence: Dia makan nasi = he eats rice. The passive form reads nasi dimakan dia. The sentence graph is simply orang
ALI
CAU
ALI nasi dia
CAU EQU
ALI makan
PAR
F
ALI dimakan focus on "dia" would lead to the active form and one would expect the prefix me-. However, in combination with "makan" this is not used, as we remarked before.
Note that the focus also has influence on the utterance path. In the passive form the sentence has to start with "nasi" and instead of "makan" "dimakan" is used. Also note that "by", the first CAU-arc from "dia", is not uttered, although sometimes oleh = by is mentioned. The implication by di- is very strong. "dijual" just means "sold", a rather normal way of speech, like "deal!", where the exclamation mark should be noted. Instead of F we might have used the symbol "!". In a similar way, the symbol "?" might be attached by a PAR-arc to a knowledge graph frame. In Chinese this is actually expressed in language by the word "ma".
Different to that of English and Chinese, the formation of Indonesian word often involves the use of prefix or suffix to some basic words. However, knowledge graph can still represent the formation of Indonesian word perfectly. Hence we conclude that knowledge graphs can be used to represent various characteristic features of many different languages effectively.