Traditionally, Controlled Natural Languages (CNLs) are designed either to avoid ambiguity for human readers, or to facilitate automatic semantic analysis, so that texts can be transcoded to a knowledge representation language. CNLs of the second kind have recently been adapted to the requirements of knowledge formation in OWL for the Semantic Web. We suggest in this paper a variant approach based on automatic generation of texts in CNL (as opposed to automatic analysis), and argue that this provides the best of both worlds, allowing us to pursue human readability in addition to a precise mapping from texts to a formal language.
Several research groups have proposed interfaces for the Semantic Web
based on Controlled Languages [
Our purpose in this paper is to introduce an alternative approach in
the same tradition. We agree with the research groups cited above that
CNL-based interfaces for editing (or viewing) ontologies and other
metadata on the Semantic Web are plausible solutions to an urgent problem.
Our distinctive proposal is to avoid any reliance on automatic
interpretation of human-authored text. The CNL texts representing the content
of a knowledge base are produced not through human authoring but
through automatic Natural Language Generation (NLG). This brings
several immediate advantages: for instance, it eliminates any possibility
of an interpretation error (by the system); it also eliminates any need to
train users to adhere to the CNL grammar. However, the idea is at rst
sight paradoxical: if all texts are generated by the system, how does the
human author convey the desired content?
Over the last decade, our group has pioneered a method by which this
can be done. The key idea is that the author speci es content by direct
manipulation of a generated `feedback text' which expresses both the
current content of the knowledge base and the options for extending it.
By preserving the link between constituents of the feedback text and the
formal representation of its meaning, the interface supports editing at the
level of meaning, not text. The user can select spans linked to individual
objects or events, and perform operations like classi cation, or assigning
values to properties, presented through menus (again expressed in the
CNL) which pop up when a span is selected. This method, which we have
called Wysiwym3 or more generically Conceptual Authoring, has been
employed in several application domains [
Conceptual Authoring was originally developed as a way of de ning the
input for multilingual NLG, in applications where content could be
formally represented by an ABox [
P astArrival(e1) & M edicalDoctor(e2) & hasAgent(e1; e2)
3 What You See Is What You Meant
4 Methods have been proposed for making Wysiwym open-ended by allowing users to
add new terms for atomic classes [
Editing begins by o ering the user a place-holder text indicating a constraint on the individual that embraces the whole sentence (e.g., it must be an event or situation). When this span is clicked, a menu pops up offering a list of options computed from the speci c event/situation classes in the TBox, also expressed in CNL. Typically these options will include further place-holder texts constraining the participants in the event. Here is a typical editing sequence:
User action Feedback text Choose New Event [Something happened] Choose from pop-up menu [Someone] shaved [someone] Choose from pop-up menu A barber shaved [someone] Copy the barber individual A barber shaved [someone]
Paste on to place-holder A barber shaved himself The important point to understand here is that all editing operations are de ned on the ABox graph, not on the text, which is used only as a means of presenting the ABox and the options for updating it. The editing process depends on a model for aligning a text to a graph, in which text spans are mapped to nodes, and span-subspan relations are mapped to arcs.
Can a similar process be devised for editing a knowledge base
embracing ABox and TBox? The simplest approach, in our view, is to adopt an
alignment model in which sentences denote axioms, and major sentential
constituents denote classes; editing can then be performed by selecting
any span denoting a class, and replacing it by a span denoting another
class, chosen as before from a menu of options generated by the
system. ABox assertions can be incorporated into this model by treating
individuals as enumerated classes containing a single element [
Description Syntax OWL
atomic class AN ame AName
universal class > Thing
enumerated class fag oneOf(a)
exists restriction 9R:C someValuesFrom(R,C)
class inclusion C v D subClassOf(C,D)
The author is allowed to add new terms to the knowledge base
(individual, class or property names), and must link them to words (drawn from
a wide-coverage lexicon) of a syntactic category constrained by the CNL.
In a very simple CNL, individuals could be expressed by proper names,
classes by count nouns, and properties by transitive verbs. A xed CNL
grammar for expressing the description language takes care of the rest.
A knowledge base comprises a list of axioms; when the user invokes New
axiom, a trivial axiom > v > is added, and can be edited by
substitution operations on the classes. Here is an example from the People+Pets
domain [
Logical form
Everything is a thing Every pet is a thing
Every pet is an animal > v > P v >
P v A Obviously P and A here are arbitrary labels that would be replaced in a Semantic Web application by URIs meeting the W3C namespace conventions. Here is a more complex sequence using both individuals and properties:
Logical form
> v > Everything is a thing fmg v > Mary is a thing fmg v 9R:> Mary owns one or more things fmg v 9R:P Mary owns one or more pets In these sequences, transitions from one line to the next are made by clicking on a word/phrase representing a class, and choosing from a menu of permissible substitutions. For instance, starting from `Everything is a thing', the user might click on `Everything' and then choose `Every pet' from the following pop-up menu:
Mary Every animal Every dog Every pet Everything that likes one or more things Everything that owns one or more things . . .
This method raises an obvious problem of scale: for any non-trivial
ontology, classes will have to be selected from thousands of alternatives, and
some kind of search mechanism will therefore be needed. One solution
already used in Wysiwym applications [
We have described a new approach to editing hybrid knowledge bases using controlled natural language. The novelty of this approach is that it relies entirely on natural language generation, thereby avoiding possible pitfalls of interpretation-based methods (e.g., interpretation errors, training e ort) as well as bringing potential bene ts (e.g., multilinguality, more exibility in designing the CNL). However, our proposals have not yet been developed in detail or applied and evaluated for large-scale ontologies5. It remains to be seen (a) whether the proposed editing method is intuitive for users, (b) whether users are able (and willing) to de ne the necessary lexical resources during ontology building, (c) whether such an 5 These tasks are being addressed in SWAT (Semantic Web Authoring Tool), a joint project between the Open University and the University of Manchester. editing tool will scale up to large ontologies and expressive DLs like OWL Full, and (d) whether the di erences between interpretation-based and generation-based approaches are important or merely an implementation detail.