When an OWL ontology, together with SWRL rules, is defined or verbalized in controlled natural language (CNL) it is important to ensure that the meaning of CNL statements will be unambiguously (predictably) interpreted by both human and machine. CNLs that are based on analytical languages (namely, English) impose a number of syntactic restrictions that enable the deterministic interpretation. Similar restrictions can be adapted to a large extent also for synthetic languages, however, a fundamental issue reveals in analysis of given (topic) and new (focus) information. In highly analytical CNLs, detection of which information is new and which has been already introduced is enabled by systematic use of definite and indefinite articles. In highly synthetic languages, articles are not typically used. In this paper we show that topic-focus articulation in synthetic CNLs can be reflected by systematic changes in word order that are both intuitive for a native speaker and formal for the automatic parsing.
One of the fundamental requirements in definition and
verbalization of ontology structure, restrictions, and implication
rules is the unambiguous interpretation (in terms of the underlying
formalism) of controlled natural language (CNL) statements, so
that the CNL user could easily predict the precise meaning of the
specification he/she is writing or reading; that also includes the
resolving of anaphoric references. To enable deterministic
construction of discourse representation structures (DRS), several
widely accepted restrictions are used in CNLs (e.g., in Attempto
Controlled English [
There are several sophisticated CNLs that provide seemingly
informal means for bidirectional mapping between controlled
English and OWL [
Statements defining class hierarchies consist of subject and subject complement. Subject (topic) is always universally quantified, predicate noun — always existentially quantified: (1) Class: Professor SubClassOf: Teacher Every professor is a teacher.
Katrs profesors ir kāds pasniedzējs.
For the universal quantifier there is a corresponding determiner/pronoun both in English and Latvian (as well as in other analytical and synthetic languages). As to the existential quantifier there is no counterpart for the indefinite determiner in Latvian. We could artificially use an indefinite pronoun instead, but such construction would be more than odd in this case. Besides the fact that the subject complement is always indefinite it also always appears in the focus part of a sentence (here and further — formatted in bold), i.e., it always is new information and, thus, the explicit linguistic marker (indefinite pronoun) can be omitted without introducing any ambiguities. Similarly, class equivalence can be defined by stating subclass axioms in both directions (in two separate statements), and class disjointness — by substituting the determiner ―every‖ with its antonym ―no‖: (2) DisjointClasses: Assistant, Professor No assistant is a professor.
Neviens asistents nav profesors.
In Latvian (and in many other synthetic languages), a negated pronoun is used for the negative universal quantifier, and the statement is negated twice by the copula, but these are minor syntactic differences; the information structure remains the same. This assumption can be directly extended to complex classes that are combined from atomic ones by applying logical constructors: (3) Class: Course SubClassOf: owl:Thing and (MandatoryCourse or OptionalCourse) Every course is something that is a mandatory course or that is an optional course.
Katrs kurss ir kaut kas, kas ir obligātais kurss vai kas ir izvēles kurss.
So far about cases when the verb phrase (VP) is a predicate nominal. Another type of constructors for complex classes are property restrictions — VPs consisting of a transitive verb complemented by a direct object. In the following statement such VP is used to implicitly specify an anonymous superclass. (4) Class: Teacher SubClassOf: teaches some
Course Every teacher teaches a course.
Katrs pasniedzējs pasniedz [kādu] kursu.
Here the role of word order comes in. In Latvian, if the object comes after the verb (the neutral word order) it belongs to the focus part of the sentence (new information), but if it precedes the verb — to the topic part (given information). In the case of the inverse use of a property, the word order is changed for the whole statement (in both languages), moving the agent to the focus: (5) Class: Course SubClassOf: inverse (teaches) some Teacher Every course is taught by a teacher.
Katru kursu pasniedz [kāds] pasniedzējs.
Combinations of the introduced syntactic phrases can be further used to explicitly specify complex superclasses (Statement 6) and general class axioms, where anonymous is either the subclass (Statement 7), or both the super- and the sub-class (Statement 8). (6) Class: Student SubClassOf: Person and
(inverse (enrolls) some AcademicProgram) Every student is a person that is enrolled by an academic program.
Katrs students ir persona, ko uzņem [kāda] akadēmiskā programma. (7) Class: owl:Thing and (teaches some
MandatoryCourse) SubClassOf: Professor Everything that teaches a mandatory course is a professor. Katrs, kas pasniedz [kādu] obligāto kursu, ir profesors. (8) Class: owl:Thing and (inverse (includes) some AcademicProgram) SubClassOf: inverse (teaches) some Teacher Everything that is included by an academic program is taught by a teacher.
Katru, ko ietver [kāda] akadēmiskā programma, pasniedz [kāds] pasniedzējs.
+enrolls <<owlClass>>
Student <<objectProperty>> takes
<<owlClass>>
AcademicProgram +constitutes
+includes +takes <<owlClass>>
Course <<owlClass>> MandatoryCourse +teaches <<owlClass>>
OptionalCourse <<disjointWith>> <<owlClass>>
Person <<owlClass>>
Professor <<objectProperty>> teaches <<disjointWith>> <<owlClass>>
Teacher In English, both active and passive voice sentences are still SVO sentences; the inverse direction of the property is indicated by the passive voice. In Latvian, the voice remains active, but the syntactic functions of both NPs are interchanged in such a case, making it an OVS sentence (semantic roles remain the same in both languages). Thus, it turns out that the object stands to the left from the verb — in the topic part, indicating that it should be given information. However, recall that in TBox statements there is no doubt which determiner has to be assigned with the topic — it is always universally quantified (unless it is an anaphoric pronoun that links a relative clause to its anchor).
Also, it should be mentioned that in the above provided SVO statements (4–8) the indefinite pronoun ―kāds‖ is given in square brackets, which means that in these cases it might be optionally used as a counterpart for the indefinite article. This conforms to the language intuition (see Section 4) and emphasizes the indefiniteness of the NP — the explicit marker improves readability (interpretation) as there is no relative clause associated with the NP, which would then serve as an indicator of indefiniteness.
Another aspect that should be mentioned is that there has not been
made any differentiation between animate and inanimate things —
quantifiers ―everything‖, ―something‖, ―nothing‖, and the relative
pronoun ―that‖ are used in all cases. This makes Statement 7 in
English and Statement 8 in Latvian odd, which has been noticed
by our respondents (see Section 4). However, we ignore this issue
in this paper for the sake of simplicity and for compliance with the
ACE verbalizer [
The already introduced syntactic constructions can also be used to define properties and their restrictions. Thus, to specify the domain and range of a property, in the topic part of the statement one has to refer the universal class, which is then specified in the relative clause, referring the property of interest. The subject is complemented by a predicate nominal in the focus part (see the statements 9–10). Note that usage of definite and indefinite NPs (i.e., references to concrete classes) is not possible in property definitions (except when stating the domain and range), as this would go beyond the expressivity of OWL (see the next section). (9) ObjectProperty: teaches Domain: Teacher Everything that teaches something is a teacher.
Katrs, kas kaut ko pasniedz, ir pasniedzējs. (10) ObjectProperty: teaches Range: Course Everything that is taught by something is a course.
Katrs, ko kaut kas pasniedz, ir kurss.
Although property hierarchies, characteristics, and chains can be defined in the same manner, by additionally exploiting the anaphoric pronoun ―it‖, in many cases usage of if-then constructions together with variables is at least more concise, if not more comprehensible as well (especially in property chaining). Exceptions are functional and inverse functional properties that are defined by cardinality restrictions and can be stated more naturally without anaphoric references, and reflexive and irreflexive properties — their verbalization in CNL is more natural if the reflexive pronoun ―itself‖ is used.
For instance, the definition of a property chain given in Statement 11 could be paraphrased by using pronouns instead of variables — ―Everything that includes something that is taken by something enrolls it.‖ — but such paraphrase would more likely confuse a human interpreter, especially when resolving the anaphoric reference.
(11) ObjectProperty: enrolls SubPropertyChain: includes o inverse (takes) If [something] X includes something that is taken by [something] Y then X enrolls Y.
Ja [kaut kas] X ietver kaut ko, ko ņem [kaut kas] Y, tad X uzņem Y.
Note that the pronoun ―something‖ may be omitted in the apposition phrases — this not only makes the statements more comprehensible, but also allows to reduce or even to hide the issue of discrimination between animate and inanimate things (accordingly, X and Y in Statement 11). For instance, when declaring that two properties are disjoint, we can avoid the use of the indefinite pronoun at all: (12) DisjointProperties: teaches, takes If X teaches Y then X does not take Y.
Ja X pasniedz Y, tad X neņem Y.
The concise form, however, has a drawback in the case of a highly synthetic CNL. In English, the strict word order (and the change of the voice) enables the unambiguous detection of which variable represents the agent in any of the SVO chunks. In Latvian, provided that all the properties involved are represented by transitive verbs (instead of comparative phrases, for instance, as in ―X is smaller than Y‖), the agent/subject can be recognized only due to the different ending if compared with the object; the verb itself does not change. Plain variables, of course, are not inflected. Although for a human interpreter it usually causes no ambiguities (due to the rich background knowledge, and knowledge of lexical semantics), suffixes have to be added to the variables to enable the automatic parsing. Nevertheless, this is still a more user-friendly solution (see Statement 13) than the use of the artificial apposition phrases. Moreover, even if indefinite apposition phrases are used, they are applicable only in the if-clauses; for the then-clauses definite apposition phrases should be introduced, making such statements even more unnatural.
(13) ObjectProperty: includes InverseOf:
constitutes If X includes Y then X is constituted by Y.
Ja X-s ietver Y-u, tad X-u veido Y-s.
Although property axioms can be seen as a special case, variables may be used in statements defining classes as well (e.g., Statement 7 in Section 2.1 can be paraphrased in ACE as ―If X teaches a mandatory course then X is a professor.‖). The formal nature of CNL then becomes explicit more widely, losing the seeming naturalness that, of course, is not a self-purpose; variable constructions should be allowed as an alternative to improve readability in certain cases (e.g., for tracking coreferences in complex rules). Allowing for such alternatives, however, introduces an issue in the verbalization direction — how to decide (encode in the grammar) in which cases variables are preferable over indefinite and definite NPs.
Nevertheless, variable constructions are partially out of the scope of this paper, as there is no need for information structure analysis to cope with utterances of anaphoric pronouns and variables. Note that we have already violated the word order guidelines in some of the previous examples — in Latvian, the indefinite pronoun ―kaut kas‖ typically goes before the verb if it is not specified by a relative clause (see the statements 9–10). Thus, formally it belongs to the topic, although it is always new information. But, again, this causes no ambiguities.
In overall, if we are restricting our synthetic CNL to cover terminological statements (class and property definitions) only, information structure analysis is not necessary at all: since OWL axioms are variable-free, any noun phrase that is not explicitly universally quantified is existentially quantified. However, we have now laid the foundations to extend controlled Latvian for support of data integrity constraints. 3. DATA INTEGRITY CONSTRAINTS In this section we will add some implication rules and data integrity queries to our example ontology, making an actual exploitation of the topic-focus articulation (TFA) — when specifying integrity constraints, one cannot avoid the usage of variables or definite/indefinite NPs.
In SWRL rules [
Katru obligāto kursu, ko ietver [kāda] akadēmiskā programma, ņem katrs students, ko [šī] akadēmiskā programma uzņem.
In the above statement, inverse properties are used in both relative clauses, causing the swapping of the subject and the object. In the first case, the subject (―akadēmiskā programma‖) stands to the right from the verb, indicating that it belongs to the focus part — new information. In the second case, the subject (again, ―akadēmiskā programma‖) stands to the left from the verb — in the topic part of the clause, indicating that this is already given information (a reference to the individual introduced in the first relative clause). Thus, in the latter relative clause, the property (verb) is alone in the focus — the new information is the relationship between the two already given individuals (the student and the academic program).
The English and Latvian verbalizations of the above rule are more clearly aligned in Figure 2 (the optional ―articles‖ are not used).
every mandatory course
that is included by
an academic program is taken by every student
that is enrolled by
the academic program katru obligāto kursu
ko ietver akadēmiskā programma ņem katrs students
ko akadēmiskā programma uzņem To emphasize that an NP is an anaphoric reference, we could optionally use the demonstrative pronoun ―šis‖ (―this‖), which usually sounds natural in Latvian. Theoretically, this would allow us to place the NP also to the right from the verb, but both stylistically and intuitively the preferable position is still to the left, which causes the correct intonation.
Let us consider one more rule (Statement 15) where the indefinite pronoun ―kāds‖ is not anymore offered as an optional attribute. It is omitted due to the cascade of relative clauses that modify all the indefinite NPs.
(15) Rule: Person(?x1), MandatoryCourse(?x2), AcademicProgram(?x3), enrolls(?x3, ?x1), includes(?x3, ?x2), takes(?x1, ?x2) ->
Student(?x1) Every person that takes a mandatory course that is included by an academic program that enrolls the person is a student.
Katra persona, kas ņem obligāto kursu, ko ietver akadēmiskā programma, kas [šo] personu uzņem, ir students.
Due to the cases when the usage of the indefinite and demonstrative pronouns might improve the readability of a rule, an implementation of an additional concrete grammar of surrogate Latvian, where the usage of indefinite and demonstrative pronouns in the role of articles is always mandatory, would be a simple but naive trade-off. Such a coarse-grained grammar could be used to paraphrase (on demand) the purely TFA-based sentences, both user-provided and auto-generated (verbalized). Although reading of such surrogate statements perhaps is easier than writing, this would be confusing to the end-users anyway (see the next section). To provide more fine-grained paraphrases and to protect the users from a confusingly verbose look-ahead editor, the TFA-based grammar can be improved at least by distinguishing two types of NPs: those that are modified by relative clauses, and those that are not. In the latter case, usage of the indefinite or demonstrative pronoun is preferable. However, if a statement is written in synthetic language, the best paraphrase could be its translation into analytical language (e.g., English), or alternatively (for an advanced user) — in a humanreadable formal syntax like the Manchester OWL Syntax.
When executing SWRL rules, potentially new facts are inferred and added to the ontology (ABox) whenever the body of a rule is satisfied. If this is the intention and if the rule (its verbalization) does not include negated atoms or disjunctions then it is the right choice; otherwise we would end up with unwanted entailments or would not be able to translate the statement into SWRL. For example, if we would try to redefine the property chaining defined in Statement 11 (in Section 2.2) as a more specific rule (by referring to the concrete classes) — ―Every student that takes a course that is included by an academic program is enrolled by the academic program.‖ — the effect would be that a student, taking a course that is included by another academic program, is automatically enrolled by that program.
By asking SPARQL queries we can verify integrity constraints
relying on the closed world assumption (negation as failure) [
NOT EXISTS {?x3 includes ?x2}} Is there a student that takes a course that is not included by an academic program that enrolls the student? Vai ir kāds students, kas ņem kursu, ko neietver akadēmiskā programma, kas studentu uzņem? In the case of consistency checking, the ASK form of a query (yes/no question) is entirely appropriate and its verbalization syntax is not much different from that of rules. Note that in such queries the first NP is always indefinite (although appearing in the topic) and the corresponding indefinite pronoun is always explicitly attached to it.
The current SPARQL specification [
To verify whether the proposed assumptions on which the proofof-concept implementation [20] is based are linguistically motivated and universally applicable in highly synthetic CNLs, an initial evaluation was performed. About ten linguists (both Latvians and Lithuanians), specialized in Baltic languages, received nearly twenty examples, covering different types of statements and different levels of complexity. For each example several alternative translations were given in Latvian and Lithuanian in parallel (see Table 1). Among the alternative choices were the ―literal‖ (surrogate) translation, the pure TFAbased translation, and a combination of the previous two, in order to seemingly improve the readability. The respondents were asked to sort the choices (in their native language) by priority from 1 to 3 (1 goes for the best translation of the original statement in ACE), or rejected at all (0). The respondents were introduced with the basic limitations of CNL, but they were also asked to follow their language intuition.
Every student is a person that is enrolled by an academic program.
ALV ALT BLV BLT CLV CLT
Katrs students ir kāda persona, ko uzņem kāda akadēmiskā programma.
Kiekvienas studentas yra koks nors asmuo, kurį priima kokia nors akademinė programa.
Katrs students ir persona, ko uzņem akadēmiskā programma.
Kiekvienas studentas yra asmuo, kurį priima akademinė programa.
Katrs students ir persona, ko uzņem kāda akadēmiskā programma.
Kiekvienas studentas yra asmuo, kurį priima kokia nors akademinė programa. 0 1 2 The respondents were also invited to give an alternative translation for each example, if none of the proposed ones was enough satisfactory. This option was used rather frequently, resulting in some interesting suggestions. Those that can be systematized will be taken into account.
In overall, most of the literal translations, using the artificial ―articles‖, were rejected or assigned with the lowest priority. There was no consensus, however, whether the indefinite and demonstrative pronoun in certain cases should be used to improve readability or not; even the same respondent usually did not act consistently among different examples. However, in most cases, the usage of the pronoun is preferred, if the NP is not modified by a relative clause.
It should be mentioned that almost all the respondents were disappointed with the uniform approach to animate and inanimate things. Although this is not directly related to the topic of this paper, this issue has to be taken into account, which means that one more feature has to be incorporated in the domain-specific lexicons (in the noun and pronoun entries) and exploited in the grammars. However, the issue will still remain, if other tools are used in the workflow (e.g., the ACE verbalizer).
Based on these results, an improved grammar is being developed, which will be evaluated by a wider audience.
We have shown that in controlled Latvian, which is a highly
synthetic CNL, where definite and indefinite articles are not used,
the topic-focus articulation can be reflected by systematic changes
in the neutral word order. This provides a simple and reliable
mechanism (guidelines) for deterministic (predictable) analysis of
the information structure of a sentence, enabling automatic
detection of anaphoric NPs. As the very initial evaluation
confirms, native speakers tend to follow such guidelines rather
intuitively. Moreover, in languages where the semantic and
pragmatic aspects of the sentence are more studied [
At the time of writing, the proof-of-concept implementation of
Latvian-English CNL covers most of the syntactic constructions
that were introduced in the previously given examples. The aim
for the near future is to extend the TFA-based grammar to cover
the full expressivity of terminological statements and rules, while
remaining compliant with ACE. Future work is (a) to introduce
support for assertional statements — the problem is how to
determine, whether the subject noun (in the case of the neutral
word order) represents given or new information, (b) to make a
more detailed investigation on data integrity queries that are
important in practical applications and will make a rather
extensive use of anaphoric references, and (c) to consider pros and
cons for using the GF Resource Library [
This research has been supported by the European Social Fund. The author would like to thank Gunta Nešpore, Baiba Saulīte, Guntis Bārzdiņš and Kārlis Čerāns for the valuable discussions on linguistic and semantic aspects, as well as the reviewers for the detailed comments and suggestions, and the respondents for their help in the initial evaluation of the proposed approach. [19] http://protegewiki.stanford.edu/index.php/OWL2UML [20] http://eksperimenti.ailab.lv/cnl/