=Paper=
{{Paper
|id=Vol-448/paper-4
|storemode=property
|title=How to Evaluate Controlled Natural Languages
|pdfUrl=https://ceur-ws.org/Vol-448/paper4.pdf
|volume=Vol-448
|dblpUrl=https://dblp.org/rec/conf/cnl/Kuhn09a
}}
==How to Evaluate Controlled Natural Languages==
https://ceur-ws.org/Vol-448/paper4.pdf
How to Evaluate Controlled Natural Languages
Tobias Kuhn
Department of Informatics, University of Zurich, Switzerland
tkuhn@ifi.uzh.ch
http://www.ifi.uzh.ch/cl/tkuhn
Abstract. This paper presents a general framework how controlled nat-
ural languages can be evaluated and compared on the basis of user ex-
periments. The subjects are asked to classify given statements (in the
language to be tested) as either true or false with respect to a certain
situation that is shown in a graphical notation called “ontographs”. A
first experiment has been conducted that applies this framework to the
language Attempto Controlled English (ACE).
1 Introduction
Controlled natural languages (CNL) are claimed to be easier to learn and un-
derstand than other formal languages. User studies are the only way to verify
this claim. Many such studies have been conducted on tools that use some kind
of CNL, e.g. the one described in [1]. However, it is hard to determine in these
cases how much the CNL contributes to the understandability and how much
is due to other aspects of the tool. Furthermore, it is hard to compare CNLs
to other formal languages with such studies because different languages often
require different tools. For those reasons, it would be very desirable to be able
to test a CNL in a tool-independent way.
[3] presents one of the rare attempts to test a CNL (concretely the Rabbit lan-
guage that is based on English) independently of a particular tool. The authors
conducted an experiment where the subjects were given one Rabbit statement at
a time and they had to choose from four paraphrases in natural English only one
of which was correct. There are two problems with this approach. First, since
natural language is highly ambiguous, we have to make sure somehow that the
subjects understand the natural language paraphrases in the right way, which
just takes the same problem to the next level. Second, since the formal statement
and the paraphrases look very similar in many cases (both rely on English), it
is hard to determine whether understanding is actually necessary to fulfill the
task. The subjects might do the right thing without understanding the sentences
(e.g. just by following some syntactic patterns), or by misunderstanding both —
statement and paraphrase — in the same way.
2 Approach
The approach that we propose here solves the discussed problems and relies on
a graphical notation which we call “ontographs”. Each ontograph consists of a
legend that introduces types and relations and of a mini world that introduces
individuals, their types, and their relations. Individuals are represented by dif-
ferent symbols depending on the types they belong to. Relations are represented
by arrows that point from one individual to another. Figure 1 shows four such
ontographs. Ontographs require that we have complete information about the
mini world and, as a consequence, do not need an explicit notation for negation.
Everything that is not shown in the ontograph is not true.
We assumed that ontographs are very easy to understand and our results
confirm this assumption. The use of intuitive graphical icons is one of the rea-
sons of the good understandability. More important, however, is the fact that
the ontograph notation has no generalization capabilities and does not support
partial knowledge. This excludes many potential misunderstandings. An onto-
graph explicitly shows every existing individual and depicts every single relation
instance between them. There is not much to misunderstand with these basic el-
ements. While we can make general statements like “every man loves a woman”
about the mini world defined by an ontograph, there is no way to express such
statements in this general way in the ontograph notation itself.
Ontographs are designed to be used in experiments to test the understand-
ability of formal languages. Ontographs could also be used to test the writability
by asking the subjects to describe the given situation. However, only the first
approach has been investigated so far.
In order to test the understandability of a language in an experiment, an
ontograph and several statements (written in the language to be tested) are
shown to the subjects who have to decide which of the statements are true and
which are false with respect to the mini world depicted by the ontograph.
An important property of ontographs is that they use a graphical notation
that is syntactically very different from textual languages like CNLs. This makes
it virtually impossible to accomplish a task like the one described above just
by looking at the syntax. If subjects manage to systematically classify given
statements correctly as true or false with respect to a certain ontograph then we
can conclude that the subjects understood the statements and the ontograph.
Using the presented testing framework, we conducted a small experiment that
tests the language Attempto Controlled English (ACE) [2] which is a mature con-
trolled natural language. The goal was to find out how well the framework works
and at the same time how understandable ACE is. We recruited 15 subjects who
were not experts in knowledge representation.
Four ontographs (see Figure 1) — each accompanied by 10 ACE statements
— were shown to the subjects who had to classify each of the statements as true,
false, or “don’t know”. There was a time limit of five minutes for each ontograph.
There was no explanation how the ACE statements have to be interpreted with
the one exception that it was explained that “something” can stand for persons
and objects.
The ACE sentences are chosen in a way that they cover a broad variety of
semantic structures. All sentences would be expressible in OWL and would cover
most of the axiom types provided by the OWL standard. The first ontograph
Mini World Legend
Ontograph 1
person
1/1 – Mary is a traveler.
man
1/2 + Bill is not a golfer.
Tom
John 1/3 + Mary is an officer or is a woman.
1/4 – Sue is an officer and is a traveler.
woman
1/5 – Every traveler is a man.
Bill
1/6 + No golfer is a woman.
traveler
1/7 + Every woman is an officer and every officer is a
woman.
Sue
officer 1/8 – Every traveler who is not a woman is a golfer.
1/9 + Every man is a golfer or is a traveler.
Lara
golfer 1/10 + Every woman who is a golfer is a traveler.
Mary
Ontograph 2
person
2/1 + John sees Tom.
$ 2/2 + Mary does not see Tom.
man
2/3 – Tom buys a picture.
Mary
2/4 – John sees no woman.
woman
2/5 + Tom sees every woman.
$
2/6 + Tom sees nothing but women.
picture
2/7 + Lara buys nothing but presents.
Tom Lara present 2/8 – No woman sees herself.
2/9 + Every woman buys nothing but pictures.
sees
2/10 + No man who buys a picture is seen by a woman.
$ buys
$ John
$
Ontograph 3
$ person
3/1 + Everything that loves something is a person.
traveler
3/2 – Everything that is loved by something is a person.
Tom
$ 3/3 + Everything that sees something is an officer or is a
traveler.
Lara
officer
3/4 – Everything that is bought by something is an
aquarium or is an officer.
$
aquarium 3/5 – Everything buys at most 1 thing.
3/6 + Everything is bought by at most 1 thing.
present
3/7 – Lara sees at least 2 persons.
Sue
loves 3/8 + Bill is seen by at most 1 traveler.
Bill
3/9 – Every officer is loved by at least 2 persons.
sees
3/10 + Every officer sees exactly 1 traveler.
$ buys
Ontograph 4
person
4/1 + If X asks Y then Y asks X.
loves 4/2 + If X sees Y then Y does not see X.
4/3 – Nothing asks itself.
sees
4/4 – If X loves something that loves Y then X loves Y.
? asks 4/5 + If X admires Y then X sees Y.
4/6 – If X helps Y then Y admires X.
? ?
admires
4/7 – If X admires Y then X does not see Y.
helps 4/8 + If X admires Y then Y does not see X.
4/9 – If X admires Y then X sees Y. If X sees Y then X
admires Y.
4/10 + If X sees something that admires Y then X sees Y.
?
Fig. 1. The four ontographs that have been used for the experiment together with the
respective ACE statements. True statements are marked by +, false ones by –.
100%
percentage of subjects
80%
time exceeded
60% don't know
40% wrong
correct
20%
0%
1/10
2/10
3/10
4/10
1/1
1/2
1/3
1/4
1/5
1/6
1/7
1/8
1/9
2/1
2/2
2/3
2/4
2/5
2/6
2/7
2/8
2/9
3/1
3/2
3/3
3/4
3/5
3/6
3/7
3/8
3/9
4/1
4/2
4/3
4/4
4/5
4/6
4/7
4/8
4/9
statement
Fig. 2. This chart shows for each statement how many subjects classified it correctly,
classified it wrongly, said that they don’t know, or exceeded the time limit.
contains only individuals and types but no relations. The second ontograph in-
troduces relations. The statements of the third ontograph use more complicated
structures like domain, range, and cardinality restrictions. The statements of the
fourth ontograph, finally, use no individuals and no types but talk only about
relations. In this way, we cover a subset of ACE that corresponds to a subset of
first-order logic that is similar to the one used by OWL.
3 Results
Figure 2 shows the result of the experiment. Overall, the decision rate (i.e.
the percentage of subjects who classified a particular sentence as either true or
false) was 93%, and on average each decision took 21 seconds. The decisions
were correct in 85% of the cases (compared to 50% which can be achieved by
mere guessing). The chart shows that the incorrect decisions are not equally
distributed. In the case of five statements the incorrect behavior is predominant.
The statements 1/3, 1/10, 2/7, and 2/9 represent special cases and for this reason
it is not unexpected that they are often classified wrongly.
The statement 1/3 is a very simple statement using “or” that is true but was
often classified as false. One could think that many subjects misinterpreted the
“or” as being exclusive (instead of inclusive as ACE defines it). However, if that
is the case then the subjects should also interpret the “or” of statement 1/9 as
exclusive, but they did not. A more plausible explanation is that the subjects
recognized that the statement 1/3 is imprecise in the sense that using “and”
instead of “or” would be more accurate. This prevented some of the subjects
from realizing that the statement is nevertheless true in a logical sense. Things
are different with statement 1/9 where the replacement of “or” by “and” would
not be more accurate but would make the true statement false. As a result, 1/9
was classified correctly by almost all subjects. It seems that people in such cases
often fail to distinguish accuracy from logical truth.
The statement 1/10 is another special case. It is a conditional statement with
a false precondition. It is not surprising that people with no background in logics
fail to classify this statement in a correct way.
2/7 and 2/9, finally, are common mistakes that were also encountered in OWL:
people tend to think that “nothing but presents” implies “at least 1 present” [4].
We assume that these types of mistakes are not specific to ACE, but can be
encountered in any other language. Disregarding those four special cases, we get
a decision rate of 94% with 92% correct decisions.
Since all three sentences containing “nothing but” were often misunderstood,
we will investigate how this can be improved. Also the statement 3/6 requires
more investigation.
Apart from the five exceptional statements, the correct decision was chosen at
least twice as often than the wrong one. This preference is statistically significant
on a 5% level for all statements except 4/7 and 4/10 (the null hypothesis being
a random 50% decision).
Another interesting point is that we did not tell the subjects explicitly that
the ontograph notation shows the complete information about the mini world
and that everything that is not shown in the ontograph can be considered false.
The result for statement 2/2 shows that the subjects understood this very well
without any explanation. They understood that the fact that the ontograph does
not show a sees-relation from Mary to Tom means that “Mary does not see Tom”
is a true statement.
4 Conclusions
We presented a framework to test controlled natural languages relying on the
graphical notation of ontographs. We applied this framework to test ACE and
come to the conclusion that most ACE sentences are understood very well and
very quickly. These results and individual discussions with the subjects let us
conclude that ontographs are a suitable and powerful concept for understandabil-
ity evaluations of CNLs. A larger and more thorough experiment that compares
ACE to another formal language is currently being set up.
References
1. Abraham Bernstein and Esther Kaufmann. GINO - A Guided Input Natural Lan-
guage Ontology Editor. In International Semantic Web Conference, pages 144–157,
2006.
2. Norbert E. Fuchs, Kaarel Kaljurand, and Tobias Kuhn. Attempto Controlled En-
glish for Knowledge Representation. In Reasoning Web, 4th International Summer
School 2008, Tutorial Lectures, 2008.
3. Glen Hart, Martina Johnson, and Catherine Dolbear. Rabbit: Developing a Con-
trolled Natural Language for Authoring Ontologies. In ESWC 2008, 2008.
4. Alan L. Rector, Nick Drummond, Matthew Horridge, Jeremy Rogers, Holger
Knublauch, Robert Stevens, Hai Wang, and Chris Wroe. OWL Pizzas: Practical
Experience of Teaching OWL-DL: Common Errors & Common Patterns. In EKAW
2004, 2004.