OVERVIEW
Overview of Theoretical Concepts and Results[1]
ASPIC focused on the following four roles of argumentation:
Argumentation Based Inference
Inference is is the act or process of deriving a conclusion from one’s knowledge. Argumentation provides a general framework for reasoning in the presence of conflicting knowledge. Consider a knowledge base encoded as a set of logical ‘if then’ rules and facts. As illustrated in the introduction to argumentation, arguments for conflicting conclusions or claims can be drawn, and the claim of the winning argument is then determined to be a valid inference of the knowledge base.
Consider the example of a doctor formulating a treatment plan for a patient at high risk of developing breast cancer. The doctor can choose between a number of drugs. However the doctor has access to guideline knowledge (obtained from clinical trials) reporting differing levels of efficacy of the treatments that the doctor can choose from. The doctor must therefore first infer the level of efficacy of each available treatment. For example, consider the following two arguments claiming different levels of efficacy for the removal of ovaries:
Removal of the ovaries reduces risk of breast cancer by 70%, since this has been reported by clinical trial 1 = argument A1
Removal of the ovaries reduces risk of breast cancer by 55%, since this has been reported by clinical trial 2 = argument A2
A1 and A2 attack each other. The doctor uses his knowledge about clinical trial designs to construct an argument A3 claiming that clinical trial 2 was based on a statistically flawed design, and so A3 attacks A2, and A1 is evaluated as the winning argument. Hence, the doctor infers that removal of the ovaries reduces risk of breast cancer by 70% (although of course in principle, the argumentation can continue over whether the trial was flawed or not).
The ASPIC project developed a logical formalism of argumentation that constructed arguments based on rules and facts, organised the arguments into a network, and then based on defined interactions between these arguments, evaluated the winning arguments (from hereon referred to as the inferred arguments).
The theoretical formalism represented a significant advance on the current state of the art, in that unlike existing formalisms, it was shown to generate the inferences that one would intuitively expect from a knowledge base.
Specifically, the model was shown to satisfy a number of ‘quality postulates’ that existing argumentation based formalisms of inference fail to satisfy (for further information see: ASPIC-D2.6.pdf and the paper: An Axiomatic Account of Formal Argumentation at: http://www.argumentation.org/list_of_publications.htm)
[1] ASPIC’s theoretical results are reported on in http://www.argumentation.org/Public_Deliverables/ASPIC-D2.6.pdf
Argumentation Based Decision Making
Taking decisions about what to do is referred to in the philosophical literature as practical reasoning. Argumentation based decision making involves generation of arguments for and against decision options, where the decision options represent alternative courses of action.
The ASPIC model integrates argumentation and decision making. Winning arguments for and against decision options are inferred (as described above), and then one of a number of decision making criteria can then be applied to these arguments to determine the preferred decision option.
For example, inferred arguments for and against a decision option (candidate course of action) relate the action to its desired (positive) and undesired (negative) consequences respectively. Furthermore, both sets of consequences can be ranked. Hence, one can represent that one desired consequence is more positive than another. Similarly, one can represent that one undesired consequence is more negative than another. One can then apply decision making criteria that make use of the categorization and ranking of consequences to select the preferred argument and so preferred decision option / course of action.
Returning to the example of the doctor formulating a treatment plan, suppose there are two inferred arguments for two alternative courses of action:
Removal of the ovaries reduces risk of breast cancer by 70%, since this has been reported by clinical trial 1 = argument A1
Administering the drug tamoxifen reduces risk of breast cancer by 60%, since this has been reported by clinical trial 3 = argument B1
Where “reduces risk of breast cancer by 70%”, and “reduces risk of breast cancer by 60%” are both positive consequences, and the former is obviously ranked above the latter. Hence, applying a decision making criterion that exclusively accounts for desired consequences, will select A1 and so removal of ovaries as the preferred course of action.
However, there is also an inferred argument that relates removal of ovaries to the undesired consequence of being sterile. Hence, applying a decision making criterion that additionally accounts for undesired consequences will then recommend tamoxifen rather than removal of ovaries. Note that it may well be that the patient in question may actually not want to have more children, in which case being sterile could be re-categorized as a desired consequence!. We have described relatively simple uses of the categorizations of consequences and their rankings in coming to a decision. ASPIC has defined a range of more sophisticated criteria that account for more involved relationships between the positive and negative consequences of proposed actions, and the rankings on these consequences (for further information see ASPIC-D2.6.pdf and the paper A unified setting for inference and decision: An argumentation-based approach at http://www.argumentation.org/list_of_publications.htm).
Argumentation Based Dialogue
ASPIC developed a framework for conflict resolution dialogues, where the type of reasoning described above under argumentation based inference, is now distributed amongst multiple agents. Clearly, conflict resolution dialogues play a major role in persuasive dialogues. However they also play an important role in dialogues where agents negotiate, and may disagree on statements of fact or opinion pertaining to the negotiation (for example, where a buyer and a seller disagree about the safety assessment of a car being offered). Also, as illustrated by the example below, when agents deliberate to decide a course of action, they may again disagree on statements of fact or opinion pertaining to the deliberation. Conflict resolution dialogues not only involve agents submitting arguments, but also challenging claims and premises of arguments by asking why a premise or claim is believed true, and retracting or conceding claims and premises.
Returning to the treatment planning example, consider now two doctors arguing about the efficacy of removing ovaries for reducing risk of breast cancer, as a precursor to collaboratively deciding a preferred course of action. Below we graphically illustrate how such a dialogue might proceed:
| 1. |
Doctor Kildaire starts the dialogue with argument A1 claiming removal of the ovaries reduces risk of breast cancer by 70% as reported by clinical trial 1 |
 |
|
| 2. |
Doctor Quincy replies with argument A2 claiming removal of the ovaries reduces risk of breast cancer by 55% as reported by clinical trial 2 |
|
|
|
| 3. |
Doctor Kildaire undermines argument A2 with argument A3 claiming that clinical trial 2 was unreliable since it was statistically flawed |
|
|
|
| 4. |
Doctor Quincy questions why the clinical trial was statistically flawed |
At this stage, Quincy is winning the argument since Kildaire is obliged to justify why he claims that trial 2 was statistically flawed. Note that in the dialogue each doctor makes one statement at a time. This might be an agreed rule of the dialogue. On the other hand, the rule might be relaxed so that more than one statement at a time can be made. So, as well as 4, Quincy could also reply to 3 with an argument A4 claiming that trial 2 was published in an esteemed journal and therefore could not be statistically flawed. Note, however, that Quincy is already currently winning the argument with the why question in line 4, and so in this sense submitting argument A4 is superfluous; it is not ‘relevant’ in the sense that the move does not change the outcome of the dialogue in flavour of the participant making the move. On the other hand, A4 is relevant in the sense that it introduces an extra reason for why Quincy is winning.
The dialogue continues:
|
1 |
|
|
|
2 |
|
|
|
3 |
|
|
|
4 |
|
|
| 5. |
Doctor Kildaire replies with the argument A5 that trial 2 was statistically flawed because the international journal of clinical trials reported that there was a major bias in the selection of patients for the trial |
|
|
|
| 6. |
Doctor Quincy now concedes the argument A5 |
So now Kildaire is winning. Note that an agreed rule of the dialogue might be that neither doctor can reply to a previous statement that he has already replied to. This would mean that Quincy cannot now make an alternative reply to argument A1 in line 1. For example Quincy may have an argument A5 claiming that trial 1 was unreliable. Not allowing Quincy to state this argument in reply to A1 would mean that the doctors were not able to explore the full space of argumentation with respect to the issue at hand. This is clearly not desirable, particularly in medical domains where the safety critical nature of the issue at hand means that it is important to be able to do so fully explore all lines of argument. Hence, one could relax the rule to allow Quincy to ‘backtrack’ and reply to A1 with argument A5. Quincy would then be winning.
Note that in general, if a participant in a dialogue replies to an earlier move by its opponent, then it is not necessarily the case that the participant’s reply will change the outcome of the dialogue in favour of the participant. Hence, backtracking to a previous move should be constrained to allow relevant moves in this sense, or at least in the weaker sense that the reply move should add another reason for winning.
The above example illustrates that dialogue rules can specify what kind of statements can be made in a dialogue in reply to other statements (for example it would make no sense to reply to a concede statement with a why statement), whether more than one move can be made at a time, and whether previous moves can be backtracked to. Also, notions of relevance are required to constrain moves to ensure that the move is relevant to determining the outcome of the dialogue in favour of the mover. A chosen set of such rules defines a particular dialogue protocol. ASPIC has defined a general framework for specifying dialogue protocols for conflict resolution, where:
-
The framework makes minimal assumptions about the participants engaging in dialogues, so that, for example, participants might construct arguments in different logics and languages (for example a human agent could engage in a dialogue with a software agent)
-
The framework allows for specification of a variety of different protocols, including protocols that allow backtracking, and when appropriate, protocols that require moves to be relevant (in both the senses described above).
(for further information see ASPIC-D2.6.pdf and the paper Coherence and flexibility in dialogue games for argumentation at http://www.argumentation.org/list_of_publications.htm)
Argumentation and Learning
ASPIC identified the area of Learning as a new area for application of argumentation models. Specifically, two novel integrations of argumentation and learning were developed:
1. Argumentation Based Machine Learning (ABML)
ABML is a
novel approach to machine learning, where classical machine learning is
extended with concepts from the field of argumentation. Usually the problem of
learning from examples is stated as:
- Given examples
- Find a theory that is consistent
with the examples.
<!--[if !supportEmptyParas]--> <!--[endif]-->
To
illustrate the idea of machine learning, consider a simple problem: learning
about credit approval. Each example is a customer's credit application
together with the manager's decision about credit approval. Each customer has
a name and three
attributes: PaysRegularly (with possible values ``yes'' and ``no''), Rich (possible values ``yes'' and ``no'') and HairColor (``black'', ``blond'', ...). The class is CreditApproved (with possible values ``yes'' and ``no''). Let there be three learning examples as shown in the table below.
|
Name |
PaysRegularly |
Rich |
HairColor |
CreditApproved |
|
Mrs. Brown |
no |
yes |
blond |
yes |
|
Mr. Grey |
no |
no |
grey |
no |
|
Miss White |
yes |
no |
blond |
yes |
A typical
rule learning algorithm will induce the following rule from this
data:
<!--[if !supportEmptyParas]--> <!--[endif]-->
IF HairColor = blond THEN
CreditApproved = yes
<!--[if !supportEmptyParas]--> <!--[endif]-->
This rule
looks good because it is short and it correctly covers two of the three given
examples. On the other hand, the rule may not make much sense to a financial
expert. We will now look at how this may change when arguments are introduced.
<!--[if !supportEmptyParas]--> <!--[endif]-->
With
arguments, the learning problem statement changes to:
<!--[if !supportLists]-->· <!--[endif]-->Given examples + supporting arguments for some of the examples
<!--[if !supportLists]-->·
<!--[endif]-->Find a
theory that explains the examples using given arguments
<!--[if !supportEmptyParas]--> <!--[endif]-->
To
illustrate what it means to "explain the examples using given arguments",
consider again the data in Table XXX and assume that an expert gave an
argument for Mrs. Brown:
Mrs. Brown received credit
because she is rich.
<!--[if !supportEmptyParas]--> <!--[endif]-->
Now
consider again the rule above that all blond people receive credit. This rule
correctly
classifies Mrs. Brown, but it does not explain this classification in terms of
the given argument for Mrs. Brown. The rule does not even mention Mrs. Brown's
property in the argument, namely that she is rich. Therefore we say that this
rule is not consistent with arguments given, and an ABML algorithm has to
induce another rule to this effect. Therefore an argument based rule learning
algorithm would induce:
<!--[if !supportEmptyParas]--> <!--[endif]-->
IF Rich=yes THEN CreditApproved=yes,
<!--[if !supportEmptyParas]--> <!--[endif]-->
which is clearly consistent with argument that Mrs. Brown
received credit because she is rich.
<!--[if !supportEmptyParas]--> <!--[endif]-->
A
fundamental problem of machine learning is dealing with large spaces of
possible hypotheses, which is traditionally solved by either biasing learning
towards simpler hypotheses, i.e. applying Occam's razor, or by using experts'
domain knowledge for
constraining search. ABML is special algorithm for using expert's domain knowledge, where experts articulate their knowledge with respect to a chosen learning example. Standard machine learning methods can only accept general prior knowledge (like ILP), however it has been shown that experts face difficulties when providing general knowledge about theory underlying learning examples. Thus, it is much easier to acquire arguments from experts than general rules. Moreover, the ABML algorithm has been shown to improve on the performance of existing machine learning algorithms on several data sets.
2. Machine Learning Based Argumentation (MLBA)
The ASPIC MLBA algorithm takes as input background knowledge facts and a desired claim, and generates an argument that justifies the claim in the context of the background knowledge. For example, consider the example of doctors deliberating over a treatment plan. Thus far we have assumed that rules are available for constructing arguments claiming the efficacy of medical treatments as reported in published clinical trials. However, newly developed drugs are often undergoing trials for considerable periods, and doctors may well make use of such drugs that show promising intermediate results, especially for life threatening conditions. The MLBA algorithm could take input data from ongoing trials about a new drug, and construct an argument for the claim that the new drug reduces the risk of breast cancer, and provide a strength for the arguments (i.e. the degree to which there is a reduction in risk)
(for further information see papers authored by Mozina et.al. at http://www.argumentation.org/list_of_publications.htm)
Overview of ASPIC Technological Results
The ASPIC Inference Component, Knowledge Base Editor and Proof Visualisation Interface
Efficient algorithms were defined for determining whether an argument was a winning argument in the ASPIC formalization of argumentation based inference. These algorithms were implemented in a JAVA programmed inference component, that took as input a knowledge base, and then constructed the arguments, organized them into a network of interacting arguments, and then evaluated the winning arguments whose claims represented the knowledge base’s inferences. A knowledge and query editor was also developed for constructing and maintaining a knowledge bases of facts and rules, and for submitting queries to the inference component. A graphical user interface was implemented to show how the algorithms proved whether a given argument was a winning argument, by showing the network of attacking arguments considered in the proof.
To enlarge the image press the figure 1: Figure 1
The ASPIC inference component represents a significant advance on the current state of the art, in that it is the first to implement a formalization of argumentation based inference that conforms to the quality postulates described earlier, and the first to be implemented using rigorous engineering and design standards, such that it can be used by application developers. For example, the inference component is currently being used by researchers at the Human Genetics Unit, Heriot-Watt University, to resolve inconsistencies arising in gene expression data.
The ASPIC project reviewed a number of research and commercial tools that make use of argumentation as an explanatory paradigm for structuring user interaction with knowledge (see http://www.argumentation.org/Public_Deliverables/ASPIC-D1.1.pdf). These “argument visualisation” tools allow a user, or many users, to construct arguments and counter-arguments, and visualise the arguments and their interactions. None of these existing tools deploy rigorous automated evaluation of the status of arguments. Deployment of the ASPIC inference component will provide such functionality for these tools (as described in the ASPIC roadmap: ASPIC-D1.1.pdf. The utility of the inference component has been demonstrated in ASPIC’s two large scale software demonstrators (see sections 5.1 and 5.2).
The ASPIC Decision Making Component
ASPIC has engineered a decision making component that applies a range of decision making criteria to inferred arguments generated by the linked inference component (see figure 2). The ASPIC project reviewed a number of research and commercial decision making applications that make use of argumentation to decide preferred decision options (see http://www.argumentation.org/Public_Deliverables/ASPIC-D1.1.pdf). A roadmap illustrating deployment of the linked inference and decision making components in such applications was also described (see ASPIC-D1.1.pdf). For example, consider the medical application REACT that supports patient-doctor consultative decision making over a treatment planning for patients at risk of breast cancer (see http://www.acl.icnet.uk/lab/react.html for a description of this application). Our running example illustrates how the deployed components can provide automated support for such decision making tasks. The ASPIC review also described commercial and research tools providing automated support for collaborative decision making amongst multiple stakeholders. The roadmap describes how such tools would benefit from automated support for both argument inference and application of decision criteria to collections of arguments. The utility of the decision making component has been demonstrated in ASPIC’s business large scale software demonstrator (see section 5.2).
To enlarge the image press the figure 2: Figure 2
The ASPIC Dialogue Component
The ASPIC review (see http://www.argumentation.org/Public_Deliverables/ASPIC-D1.1.pdf) surveyed a number of software implementations supporting regulation and mediation of dialogues. The majority of these implementations implemented specialised dialogue protocols implementing automated mediation of conflict resolution in the legal domain.
The ASPIC dialogue component implements a more generic framework that allows for a range of protocols to be specified, and makes minimal assumptions about the languages used for knowledge representation, and the level of automation, of the participating agents. Such a component can be deployed in applications that mediate between agents interested in engaging in conflict resolution. The component can advise on what statements an agent can make in reply to its opponent, and whether the response is relevant (in the sense described in section 4), and who is currently winning the dialogue. An important feature of the component is that it allows for agents to reply to statements made by its opponent earlier in the dialogue, while at the same time ensuring that the replies are relevant. This is important in ensuring that all lines of argument relative to a topic can be explored. The ASPIC roadmap (see ASPIC-D1.1.pdf) described the utility of such a component in legal, medical and business applications where automated support for conflict resolution can be implemented according to provably rational principles, as part of persuasion, deliberation and negotiation dialogues.
Figure 3 illustrates how such a component might be used by human and automated agents each of which can submit arguments in the dialogue that are themselves generated by each agent’s inference component. The utility of the dialogue component has been demonstrated in ASPIC’s medical large scale software demonstrator (see section 5.1).
To enlarge the image press the figure 2: Figure 3
The ASPIC Learning Component
ASPIC engineered an ABML component within machine learning toolkit Orange (www.ailab.si/orange). The name of component is ABCN2, which stands for argument-based CN2 - an extension of the CN2 rule learning algorithm as described in deliverable D3.4. It can accept positive and negative arguments for examples and induces a set of rules that are consistent with learning examples and given arguments. The component was tested for correctness on several testing data sets and some real domains (in medicine and public administration).
The time-consuming core of component is implemented in C++, while user interface tools (like finding problematic examples, giving arguments to examples, loading/reading data, visualization of rules, etc.) are implemented in Orange canvas (see Figure yyy). Furthermore, we also implemented a CORBA interface in Java, which enables the component to communicate with other ASPIC components (especially Inference component).
The Argument Interchange Format
Envisaged deployment of the ASPIC components in stand-alone and agent applications illustrated requirements for representation of and communication of argumentation knowledge between external knowledge repositories, human users, ASPIC components and software
tools. Establishing a shared ontology for representation and communication of argumentation knowledge was thus consdered an important objective. It was deemed vital that development and acceptance of the ontology was shared by the wider community. To this end ASPIC initiated a sn open workshop attended by argumentation, agents, and software researchers from outside of the ASPIC project. The worksop resulted in a first draft of the Argument Interchange Format (AIF) that was subsequently developed and published<!--[if !supportFootnotes]-->[1]<!--[endif]-->, and has since undergone further development, and exposure through publication. Currently, the inference component exports argumentation knowledge in the AIF.
<!--[if !supportFootnotes]-->
<!--[endif]-->
<!--[if !supportFootnotes]-->[1]<!--[endif]--> Carlos Chesnevar, Jarred McGinnis, Sanjay Modgil, Iyad Rahwan, Chris Reed, Guillermo Simari, Matthew South, Gerard Vreeswijk, Steven Willmott. Towards an Argument Interchange Format for Multi-Agent Systems. In: Proc. Third International Workshop on Argumentation in Multi-Agent Systems (ArgMAS 2006 at AAMAS 2006), Hakodate, Japan, May 2006.
The ASPIC large scale demonstrators
ASPIC built two large scale software demonstrators of the components. These illustrated deployment of the components in the medical and business domains.
The CARREL Medical Large Scale Demonstrator
The ZEUS Large Scale Demonstrator
