Experimental System for Argument Text Generation:
Dulcinea

ABSTRACT

To establish the component techniques of the text generation for a natural
language interface of FGCS, we have developed and evaluated an experimental
system on the parallel inference machine.

This experimental system generates Japanese text that justifies a given goal
according to the belief contents of the system.

KEY FEATURES

1. Belief content consists of facts, rules and judgments for the state of affairs.
   Exchanging the contents of a belief database enables the system to make
   arguments in various domains.

2. This system can generate not only direct grounds, but also refutations
   for the expected opposing arguments, together with examples that are
   indispensable to persuasive argument.

3. By varying the argument strategies and criteria for sentence structure,
   we perform the text generation experiment and conduct research into the
   objective criteria for "natural" text.

- 109 -

WHAT IS ARGUMENT TEXT GENERATION ?

Argument texts written by humans are appropriately structured to express
judgments and attitudes on topics based on the standpoint of the author.

By computer, however, it is difficult to produce coherent texts. Therefore,
it is necessary to consider coherence and appropriate structure planning if
high quality argument texts are to be generated by computer. The goal of
our research is to investigate the argument strategy for text coherence and
clear standpoint expression.

GENERATION MECHANISM

We set the standpoint of the system by assigning it an argument goal.
Dulcinea puts forward an argument that justifies the given argument goal
according to its beliefs. The argument texts consist of direct grounds for
the argument goal, refutations for the expected opposing arguments, together
with examples. The beliefs consist of three types of belief contents: Fact,
Rule, and Judgment. An example of beliefs is given below.

   Rules:       change[obj2=bus-route,loc=L] => decrease[obj2=passenger, loc=L].
   Facts:       introduce[obj=one-way-system, loc=Midosuji].
   Judgments:   ng[abolish[obj=bus]].

The argument goals we give the system can be of one of the three kinds of
modal expressions listed in the table below, where A stands for some state
of affairs. If a judgment g(A) exists in the system's belief, then the system
believes A to be good.


Dulcinea represents the semantic contents of an argument as an argument
graph. The following figure shows an argument graph that insists on the
argument goal "the two-way lane must be introduced". A two-way lane is a
lane that allows buses to drive the "wrong way" up a one-way street.

Then, the argument strategy on the linguistic text structure is applied to
the argument graph to organize abstract text structure FTS(Functional Text
Structure), which represents not only the semantic contents, but also the text
structure.

- 110 -

Next, Dulcinea produces various surface syntactic text structures by adding
order and connection information to the FTS. It then selects the best text
structure according to the linguistic knowledge criteria.





Finally, Dulcinea generates argument text by omitting redundant words and
selecting words that correspond to each object. The following figure shows
the FTS added order and connection information.


FTS added order and connection information

- 111 -

DOMAINS AND THEMES OF ARGUMENTS

Dulcinea can handle many kinds of argument domain and theme, such as
the following extracts.

   Traffic: introduction of two-way lanes / development of electric cars
   Whaling: moratorium on the the hunting of whales
   Working Conditions: reduction in number of working hours
   Housing: extended families living together
   Sports: prize money for Olympic prizewinners / commercialization of the Olympic Games
   Medicine: regarding brain death as actual death / execution of euthanasia

EXAMPLE OF ARGUMENT

   Argument Domain: Traffic Matters
   Argument Goal: A two-way lane must be introduced
   Generated argument text: (English translation)

In Midousuji street, a one-way system was introduced while a two-way lane for
buses was not introduced. This forced a change in the routes of bus services resulting
in a decrease in passengers of 40%. In this way, if a two-way lane is not introduced with
a one-way system, the routes of bus services changes and the number of passengers
will decrease. Moreover, if the number of passengers decrease, the bus service may
be forced to stop.

On the other hand, introducing a two-way lane seems to put pedestrians in danger.
However the pedestrians were not in danger in London when a two-way lane was
introduced, because the buses turned their headlights on. In this way, if the buses
turn their headlights on, the pedestrians will not be in danger, even though a two-way
lane is introduced.

Therefore, a two-way lane must be introduced.


OUTLINE OF DEMONSTRATION

First of all, we choose traffic matters as the argument domain and make
an argument which asserts "We must introduce the two-way lane". At first,
we limit the argument to its direct grounds and give you a brief explanation
of the function and output for each module. Then, using the same argument
theme, we make a complete argument that includes refutations for the oppos-
ing arguments, together with examples. After that, opposing arguments such
as "must not introduce a two-way lane" will be performed. Other argument
themes in the same area of traffic matters may also be demonstrated. Finally,
we will take another domain, of your choice, and show you some arguments.


- 112 -