=Paper=
{{Paper
|id=Vol-2048/paper05
|storemode=property
|title=Towards an Argumentative Dialogue System
|pdfUrl=https://ceur-ws.org/Vol-2048/paper05.pdf
|volume=Vol-2048
|authors=Niklas Rach,Wolfgang Minker,Stefan Ultes
|dblpUrl=https://dblp.org/rec/conf/icail/RachMU17
}}
==Towards an Argumentative Dialogue System==
https://ceur-ws.org/Vol-2048/paper05.pdf
Towards an Argumentative Dialogue System
Niklas Rach Wolfgang Minker Stefan Ultes
Institute of Communications Institute of Communications Engineering Department,
Engineering, Engineering, University of Cambridge UK
Ulm University Ulm University su259@cam.ac.uk
niklas.rach@uni-ulm.de wolfgang.minker@uni-ulm.de
ABSTRACT aspects. An example is the challenge move discussed in the
In this work we propose a scheme for an Argumentative framework of Prakken [7] that allows to question the validity
Dialogue System that allows a user to discuss a certain topic of an argument. For the considered persuasive scenario we
with a virtual agent using natural language. Starting from an summarize these actions as dodge moves since they do not
agent vs agent case, we address the optimization of the agent present an additional argument component and sometimes
strategy by formulating the problem as a Stochastic Game aim for a distraction from the underlying argumentation
and show that this formalism generally allows the inclusion of structure.
additional strategical moves that are not based on the content We will focus on a debating-like setup in which each player
of the argument. In a second step we propose our approach aims for establishing his own argumentative structure and
for the Natural Language Understanding of arguments by tries to demolish the one of the opponent. To this end, each
combining recent results of argumentation mining with a side is allowed on-turn to either introduce an additional
keyword-based mapping. argument component or to perform a dodge move. Thereby,
each move has to be directly related to a previous (but not
CCS CONCEPTS necessarily the last) one. We assume that both sides have
full knowledge of the underlying argument structure. By first
• Computing methodologies → Natural language process-
employing the formalism of Stochastic Games in an agent
ing; Discourse, dialogue and pragmatics; Machine learning
vs agent scenario, we aim for an optimized agent policy the
approaches; Stochastic games;
user can debate with. The problem of NLU of arguments
is addressed by combining recent results from the field of
KEYWORDS
argumentation mining [11] with a keyword mapping and
Argumentation, Natural Language Understanding, Stochastic established methods from NLU.
games The reminder of this paper is as follows: In Section 2 we
discuss the optimization of the agent policy while Section 3
1 INTRODUCTION covers our approach for the NLU of arguments in our system.
Resolving di↵erent viewpoints of virtual agents can be ad- In Section 4 we conclude by summarizing the presented ideas.
dressed as a dialogical task in which arguments are exchanged.
Thereby it is possible for each side to persuade the opponents
by establishing a convincing argumentation structure. An
2 ARGUMENTATIVE DIALOGUE AS
Argumentative Dialogue System is a system that allows an STOCHASTIC GAME
agent to carry out argumentative tasks such as persuasion or Previous work showed that the problem of optimizing an
negotiation against a human. Despite the strong theoretical agent strategy can be formulated as (partially observable)
grounding [8], implementations of these kind of systems are Markov Decision Process (POMDP) [4, 10] and thus be
rather scarce since they have to overcome di↵erent obstacles addressed by Reinforcement Learning (RL). However, as this
(see [13] for a detailed discussion). approach optimizes only the strategy of one agent, it requires
Following the classification of dialogues presented in [9] either pre-defined rules for the respective opponent or a
we consider a persuasion scenario and discuss approaches to training corpus to learn from. Moreover, the so developed
tackle two important issues, namely the Natural Language strategy directly depends on the rules or the available data,
Understanding (NLU) of arguments and the development respectively.
of a flexible agent strategy. Whereas in recent work the We propose an approach based on Stochastic Games that
first issue is bypassed by using human annotators to map extend the MDP formalism by enabling each involved agent
utterances to argumentative concepts [10] or by presenting to adapt its strategy to previous actions of others. Thereby,
a variety of possible answers to the user [6, 12], the second no prior knowledge about the agents behavior or data is
problem is addressed in various ways. Examples are by use of required as all strategies can be developed by use of RL.
heuristics [12], decision trees [5], a formulation as planning Moreover, the existence of a Nash Equilibrium in Stochastic
problem [2] or machine learning [4, 10]. However, most of the Games can be guaranteed [1]. It should be mentioned that
work done in this direction is focused on the argumentative scenarios with opposite agent goals (which is the case we
structure, i.e. which argument to present when, whereas consider) is a special case of a Stochastic Game and referred
a discussion between humans yields additional strategical to as zero sum game.
18th Workshop on Computational Models of Natural Argument 27
Floris Bex, Floriana Grasso, Nancy Green (eds)
16th July 2017, London, UK
To apply this formalism to our system, an additional exten-
sion of the frameworks proposed in previous work is required
since they did not include dodge moves. The formalism of
MDP (and therefore of Stochastic Games as well) requires the
definition of states in which the actual stand of the debate
is expressed. Argument components and their relations are
usually represented as a tree or (directed) graph as depicted
in Figure 1. If dodge moves are not included, all possible
moves can be tied to argument components and states can Figure 1: Sketch of the graph representation of ar-
thus be expressed as sub-graphs. Since we want to rely on guments and the mapping of a user utterance to a
this representation, we plan to include each dodge move as specific component in it. The subgraph depicted by
a content-free component into the subgraph. We stress that black elements indicate the actual state, the red com-
these additional components are not part of the underlying ponent (node and edge) indicates the argument com-
argumentation structure and only a↵ect the states. ponent the recent user utterance is mapped to and
A second issue is the dimensionality of the associated state gray elements indicate additional components the
space. It is known from previous work [4] that this space system is aware of. Nodes represent a certain con-
rapidly increases with the number of argument components. tent, edges a relation (e.g. support or attack ) which
As we aim for the inclusion of dodge moves as well as a rea- is not specified herein for the sake of readability.
sonable amount of argumentation components, the number of
states in our case will be comparatively large. Therefore, the
to be determined. In some formal frameworks the relation
choice of a proper learning algorithm seems to be crucial for a
itself is given by its type (e.g. [8]), thus allowing to leave
successful optimization of the agent strategies. A selection of
out the first of the above discussed steps. In such case, only
methods that were employed in the field of dialogue systems
a detection of the aim (the component it is related to) is
can be found in [1, 3] and serve as a starting point for our
required and we will focus on suchlike frameworks in the
case.
scope of this work. To determine the aim, the key words
assigned to each component will again be employed.
3 NATURAL LANGUAGE In the case of completely free speech argumentation it is
UNDERSTANDING OF possible that one utterance contains more then one argumen-
ARGUMENTS tative component or that the component the user refers to
is split in more than one turn. To circumvent this problem,
The difficulty in NLU of arguments lies in its dependence
we consider a restricted scenario in which the user is allowed
on the content of the utterance. Given that the complete
to only bring up one argumentative component at a time.
argument (i.e. all argument components) can be represented
Although this approach also restricts the language of the user
as a graph or tree, it requires a mapping of the utterance to
and requires knowledge about the underlying argumentation
a specific node in this representation as depicted in Figure 1.
structure, we consider it an important and reasonable step
Since each node represents an argument component that is
towards natural language argumentation.
associated with a certain content, it is inevitable to include
contextual information into this mapping.
4 CONCLUSIONS
Our approach to this problem consists of two steps. First,
we plan to employ techniques presented in recent work on The purpose of this work is to present and discuss our plan
argumentation mining that allow for a detection of argument for an Argumentative Dialogue System including natural
components and their relations (support or attack ) to others, language. We proposed an approach for optimizing the agent
e.g. in written essays [11]. Although this classification is not policy by employing the formalism of Stochastic Games. In
content dependent, it will decrease the set of candidates in addition we discussed the inclusion of dodge moves into this
question if applied to our scenario. The remaining options will framework and proposed a scheme for the NLU of arguments
be distinguished in a second step where we plan to search the that combines a keyword based approach with recent results
utterance for keywords previously assigned to each argument from argumentation mining.
component. The one with the best match is then picked as
final mapping. REFERENCES
It should be mentioned that the reviewed scenario includes [1] Merwan Barlier, Julien Perolat, Romain Laroche, and Olivier
Pietquin. 2015. Human-machine dialogue as a stochastic game.
an additional task regarding the dodge moves. Whereas the In 16th Annual SIGdial Meeting on Discourse and Dialogue (SIG-
distinction of di↵erent dodge moves and a general argumenta- DIAL 2015).
[2] Elizabeth Black, Amanda Coles, and Sara Bernardini. 2014. Au-
tive move is comparable to the NLU in existing systems and tomated planning of simple persuasion dialogues. In International
will be addressed by established methods, dodge moves may Workshop on Computational Logic and Multi-Agent Systems.
(depending on the type) require further processing in view of Springer International Publishing, 87-104.
[3] Kallirroi Georgila, Claire Nelson, and David R Traum. 2014.
their relations to other components. For example in the case Single-Agent vs. Multi-Agent Techniques for Concurrent Rein-
of a challenge move, the component that is challenged has forcement Learning of Negotiation Dialogue Policies. In ACL (1).
2
28 18th Workshop on Computational Models of Natural Argument
Floris Bex, Floriana Grasso, Nancy Green (eds)
16th July 2017, London, UK
500-510.
[4] Emmanuel Hadoux, Aure?lie Beynier, Nicolas Maudet, Paul Weng,
and Anthony Hunter. 2015. Optimization of probabilistic argu-
mentation with Markov decision models. In International Joint
Conference on Artificial Intelligence.
[5] Emmanuel Hadoux and Anthony Hunter. 2017. Strategic Se-
quences of Arguments for Persuasion Using Decision Trees. In
Proceedings of the AAAI Conference on Artificial Intelligence.
AAAI Press.
[6] Anthony Hunter. 2016. Persuasion Dialogues via Restricted Inter-
faces Using Probabilistic Argumentation. In International Confer-
ence on Scalable Uncertainty Management. Springer International
Publishing, 184-198.
[7] Henry Prakken. 2000. On dialogue systems with speech acts,
arguments, and counterarguments. In European Workshop on
Logics in Artificial Intelligence. Springer, 224-238.
[8] Henry Prakken. 2006. Formal systems for persuasion dialogue.
The knowledge engineering review 21, 02 (2006), 163-188.
[9] Chris Reed and Timothy Norman. 2003. Argumentation machines:
New frontiers in argument and computation. Vol. 9. Springer
Science & Business Media.
[10] Ariel Rosenfeld and Sarit Kraus. 2016. Strategical Argumentative
Agent for Hu- man Persuasion. In ECAI 2016: 22nd European
Conference on Artificial Intelligence, 29 August-2 September 2016,
The Hague, The Netherlands-Including Prestigious Applications
of Artificial Intelligence (PAIS 2016), Vol. 285. IOS Press, 320.
[11] Christian Stab and Iryna Gurevych. 2014. Identifying Argumen-
tative Discourse Structures in Persuasive Essays.. In EMNLP.
46-56.
[12] Tangming Yuan, David Moore, and Alec Grierson. 2007. A human-
computer debating system prototype and its dialogue strategies.
International Journal of Intelligent Systems 22, 1 (2007), 133-156.
[13] Tangming Yuan, David Moore, Chris Reed, Andrew Ravenscroft,
and Nicolas Maudet. 2011. Review: informal logic dialogue games
in human-computer dialogue. The Knowledge Engineering Review
26, 2 (2011), 159-174.
3
18th Workshop on Computational Models of Natural Argument 29
Floris Bex, Floriana Grasso, Nancy Green (eds)
16th July 2017, London, UK