In this paper, we discuss how Value-based Argumentation can be used as a tool in human and computer story understanding, especially where understanding the story requires understanding of the motives of its characters. It is shown how arguments about motives can be extracted from stories, and how dialogues about these arguments can aid in story understanding.
In this paper, a short version of which was
published as
A central concept in the research on story
understanding is that of scripts
In our opinion, purely script-based approaches
to story interpretation are not suited to
understanding persuasive stories concerning values, such as
parables. Scripts represent the way in which we
expect typical situations to play out: the more a
story adheres to a familiar script, the more
plausible a story is considered to be. However, many
memorable stories such as parables depend on a
twist in the story, something which is out of the
ordinary and which challenges conventional
attitudes
For a computational model of story understanding, we need to add a more fine-grained psychological dimension to the causal narrative, in which conflicts between characters’ attitudes and challenges to common attitudes can be modelled. This gives us an internal perspective that allows us to represent the deliberations of the characters involved, which allows for a much more subtle analysis of character motive and attitude than we can perform with the external causal perspective. This in turn allows us to show how the relevant stories can influence the audience’s attitudes or, in other words, how these stories can persuade an audience to adopt a different attitude.
Recently, we have proposed a model for story
understanding
Stories can be a powerful vehicle of persuasion. A story does not persuade by imparting explicit rules, but by exposing a coherent narrative aimed at changing or reinforcing attitudes, so that the stories exemplify various group cultural norms. Many folktales are of this type, as are parables, both secular and biblical. As an example of a well-known parable, we will consider The Good Samaritan. Since we will be discussing this parable throughout the paper, we will quote it in full. The context is established in Luke 10:25-27:
your God with all your heart, with all your soul, with all your strength, with all your mind, [Deuteronomy 6:5]; and your neighbour as yourself [Leviticus 19:18].”
Jesus, “Who is my neighbour?”
Thus the lawyer asks two questions. The first, “what shall I do to inherit eternal life?”, receives an answer justified by scriptural authority. But the second, “Who is my neighbour?”, is met simply by a story.
ing down from Jerusalem to Jericho, and he fell among robbers, who both stripped him and beat him, and departed, leaving him half dead. By chance a certain priest was going down that way. When he saw him, he passed by on the other side. In the same way a Levite also, when he came to the place, and saw him, passed by on the other side. But a certain Samaritan, as he travelled, came where he was. When he saw him, he was moved with compassion, came to him, and bound up his wounds, pouring on oil and wine. He set him on his own animal, and brought him to an inn, and took care of him. On the next day, when he departed, he took out two denarii, and gave them to the host, and said to him, ’Take care of him. Whatever you spend beyond that, I will repay you when I return.’ Now which of these three do you think seemed to be a neighbour to him who fell among the robbers?”
wise.”
This provides a very clear example of a story being used as an argument to justify a particular answer to a question, “Who is my neighbour?”. However, it is not meant as a theoretical argument: the aim is not that the lawyer should believe that the Samaritan is his neighbour (nor, since the one in the story is a fictional character, that all Samaritans are his neighbour). Nor is the lawyer intended to set out to assist wounded travellers on the road from Jerusalem to Jericho. Unlike practical reasoning proper, there is no specific situation, with a specific choice of actions to resolve. Rather the argument is intended to convince the lawyer (and ultimately of course the reader) to become a different person, the sort of person who will enjoy eternal life.
So how exactly does the story convince its
audience to change their ways? Govier and Ayers
So 3. What matters about being a neighbour is not one’s status or one’s prior knowledge of a person. 4. What matters about being a neighbour is treating another with mercy and kindness when that person is needy and one encounters him. 5. It is good to treat a needy stranger as a neighbour if one encounters him.
6. One should treat other people, when they are in need and one encounters them, as one’s neighbours with mercy and kindness.
Statements 1 and 2, which both can be said to
follow from the story in some way2, lead to
conclusions 3 and 4. These two conclusions together
with the value judgement contained in 5 then lead
to the final conclusion 6. The addition of 5 and 6
is, in our opinion, somewhat contentious because
it transforms the argument into an argument with
a normative conclusion, advocating particular
behaviour. This is perhaps justified by the comment
‘Go and do likewise’ made by Jesus, since this
2It is unclear why
The computational model for story
understanding we propose is based on
Structuralist accounts of narrative argue that actions that represent transitions between states are the basic building blocks of stories. It is for this reason that we choose the mechanism of Actionbased Alternating Transition Systems with Values (AATS+V) as our basic formalization method for stories. An AATS consists of a set of states and transitions between them, with the transitions labelled with joint actions, that is actions comprising an action of each of the agents concerned. In an AATS+V, the transitions are labelled with the values that motivate the characters in the story. A basic version of the parable of the Good Samaritan can be rendered as the AATS+V in Figure 1.
At the beginning of the story q0, the condition of the traveller is wounded. In q4, the traveller’s wounds have been bandaged and he is in a stable condition. In addition to the actions taken by the characters in the story (j1, j3, j6), we have also included the hypothetical actions the characters could have performed: for example, the Priest could also have helped the traveller (j2). Action choice in parables is often more or less binary (help or :help, accept or :accept in the Prodigal Son), so modelling these extra actions does not require much extra information besides the original story text. The values that are promoted by each action are included in the AATS+V: Religious Duty (+RD), Religious Law (+RL), National Solidarity (+NS), Racial Solidarity (+RS), Compassion (+C), Prudence (+P), Convenience (+Cv) and Revenge (+R). Adding the values requires more background knowledge. For example, we need to know that the traveller and the Levite were of the same race, and that Samaritans were a common enemy for the Jewish people. Nowadays, this background information can be gained from Biblical texts, or from the many varied accounts on how parables should be interpreted, but it would have been well-known to the original audience. The values in figure 1 are a selection that the authors have heard from a variety of sources over the years. 3.2
The idea of arguments based on stories is that we look for arguments that instantiate the Practical Reasoning Argumentation Scheme (PRAS). Such arguments are of the following form.
3. Which will result in new circumstances S
Now, given an AATS+V, we can construct these
arguments for the different characters. The
basic idea expressed in
1. In the initial state q0 = qx 2 Q 2. Agent i 2 Ag should participate in joint action jn 2 JAg, where jni = i 3. Such that (qxjn) is qy 4. Such that for some vu 2 Avi, (qx; qy; vu) is +
Here, Q is a finite, non-empty set of states, Ag = f1; : : : ; ng is a finite, non-empty set of agents, i defines the set of states from which action may be executed by agent i, is a partial system transition function, which defines the state (q; j) that would result from the performance of action j in state q, Avi is a finite, non-empty set of values for agent i and is a valuation function which defines the status (promoted (+), demoted (–) or neutral (=)) of a value ascribed to the transition between two states.
Given this mapping of PRAS on an AATS+V, we can generate the arguments from the AATS+V, noting that arguments for different actions attack each other because the actions are mutually exclusive, i.e., one cannot help and not help someone at the same time. First, there are the two arguments that might apply to the priest.
A1: I should help the man because I have a religious duty to do so. This will promote Religious Duty (+RD) A2: I should not help the man because I risk uncleanliness through contact with his blood.
This will promote Religious Law (+RL).
A3: I should help the man because he is a fellow countryman. This will promote National Solidarity (+NS).
A4: I should help the man because he is of my race. This will promote Racial Solidarity (+RS).
None of the above arguments apply to the Samaritan. The following arguments apply to all three characters.
q0 T wounded
P arrives j1: not help +RL +P +Cv
q1 T wounded
L arrives j2: help +RD +C j3: not help +P +Cv
j4: help +C +NS +RS
q3 T wounded
S arrives j6: help +C
q4
T stable j5: not help +P +Cv +R
q5 T wounded A5: I should help the man because he is a fellow human being. This will promote Compassion (+C).
A6: I should not help the man because it may be trap and I may be robbed. This will promote Prudence (+P).
A7: I should not help the man because it will interrupt my journey. This will promote Convenience (+Cv).
Finally there is an argument that applies only to the Samaritan:
A8: I should not help this man, because his people have quarrelled with mine. This will promote Revenge (+R).
All of the arguments A1-A4 relate to duties of one sort or another, arising from religious law or duty, or one form or another of social relationship (nation, race). A5-A8 all arise from natural human instincts, unconnected with any social institution. 3.3
From these arguments, we can construct a
Valuebased Argumentation Framework (VAF). A VAF
is based on
The purpose of building a VAF is to find a subset of the arguments which is at once conflict free (i.e. no two arguments in the subset attack one another), and collectively able to defend itself (i.e. any attacker of an argument in the subset is itself attacked by an argument in the subset). The maximal such subset is called a preferred extension, and represents a maximal consistent position given the arguments presented. The key feature of VAFs is that they allow a distinction to be made between successful attacks (defeats) and unsuccessful attacks, on the basis of the values associated with the arguments: attacks succeed only if the value associated with the attacking argument is ranked by the audience as equal to, or higher than, the argument it attacks. The VAF thus accounts for elements of subjectivity in that the arguments that are acceptable are dependent upon the audience’s ranking of the values involved in the scenario.
We now attempt to explain the actions of the three characters by considering different value orderings, different audiences. Suppose that the Priest puts religion before all else (i.e., Religious Duty and Religious Law are preferred to Convenience, Compassion and Prudence). He then has a conflict between A1, which argues he should help to promote RD, and A2, which argues he should not help to promote RL. In the story, he chooses to observe of the law, which applies specifically to himself because of his special role, over the vaguer practical obligation to serve others. This ranking of strict observance of the law over more human concerns is criticised elsewhere in the Gospels, e.g. Mark 2:27 (Then Jesus said to them, “The Sabbath was made to meet the needs of people, and not people to meet the requirements of the Sabbath.”).
The Levite must be supposed to act on either A6 or A7, overriding the specific duties of A3 and A4 as well as A5. But because we can assume to have a type of a morally respectable man, it must be assumed that we are being invited to conclude that these preferences are acceptable in the eyes of the current moral climate: that it is morally acceptable for prudence and/or personal convenience to override obligations arising from country or race, let alone from natural feelings of compassion.
The Samaritan, in contrast has no duties prompting him help the man, and must balance his compassion against the other natural human instincts. That he helps the man (A5), can only be explained in terms of him putting compassion before all other values, individually and in combination, and this is what we are invited to conclude is what being a neighbour really is. The context supplied in the coda quoted above invites the hearer to adopt these value preferences, to become a person who places compassion above creed, country and convenience and to act in accordance with these priorities in future. 4
In the previous section, we discussed how stories, the characters in them and these characters’ motivations can be understood using VAFs. We can now use the story as an argument, using exactly this interpretation of the story. The conclusion the audience is invited to draw from the story depends on the context in which the story is told. In the case of Good Samaritan, this context is provided by the exchange between Jesus and the lawyer, and specifically the lawyer’s question “Who is my neighbour?”. As we have argued in section 2, the actual question is something like “what does it mean to love your neighbour like yourself?”, and the answer is not the literal “The Samaritan is my neigbour” but rather an understanding of why the Samaritan acts as he does, which encourages one to adopt similar attitudes to the Samaritan. 4.1
In our model, the audience of the story should
identify the value-based arguments in the story
and then reason about which values will explain
the behaviour of the Samaritan. In
In the EAF for the Samaritan there potentially two value-preference arguments for each pair of values, for example:
AV1 Prudence is preferred to Compassion (P > C).
AV2 Compassion is preferred to Prudence (C > P ).
These pairs will mutually attack, but more importantly they will attack the attack from the argument motivated by the less preferred value on arguments motivated by the other value. The complete EAF for the parable will now contain all the base arguments A1-A8 and a value preference argument for each attack between these original arguments. Furthermore, we introduce arguments for the various characters: AC1 (Character is a priest), AC2 (Character is a Levite) and AC3 (Character is a Samaritan). This will enable us to eliminate arguments which do not apply to particular characters from consideration: thus AC1 will attack A3, A4 and A8, AC2 will attack A1, A2 and A8, and AC3 will attack A1, A2, A3 and A4. Adding AC3 to the AF that contains all characters’ arguments A1 - A8 then produces the EAF applicable to just the Samaritan, as shown in Figure 2. Similarly, we can introduce AC2 to get the EAF applicable to the Levite and AC1 to get the EAF applicable to the priest.
A7 not help +Cv
AV3 C > Cv
AV4 Cv > C
AV6 R > C
A8 not help +R A5 help +C
AV5
C > R
Now that we have established appropriate EAFs
for the various characters, we need to evaluate
them to explain the choices they make in the story.
Thus, in the case of the Samaritan, we need to
construct an admissible set containing an argument to
justify helping the traveller, and then to consider
what value preferences it contains. One method of
constructing admissible sets from Dung style AFs
is to use a dialogue game, such as the TPI (Two
Party Immediate Response) Game of
A TPI game proceeds by the proposer playing an argument, the opponent playing an attacker, the proposer playing an attacker of that argument and so on, until one player cannot move. At this point a player can back up to a choice point and play a different attacker. This continues until no moves are possible (note that arguments under attack cannot be played). At this point we will have an admissible set containing the arguments played by the last player to move. If this was the proposer is will contain the original argument and this will have been shown to be acceptable. Because it is the Samaritan’s preference we are trying to determine, we use the EAF in figure 2, rewritten as a regular AF. The dialogue then proceeds as follows: Samaritan: A5 holds. This is an argument justifying what the Samaritan did in the story: current position is {A5 holds}.
Opponent: A6 defeats A5. chooses a way to attack A5.
Opponent Samaritan: AV2 C > P. The preference argument is played: the alternative would eventually require A5 holds to be played, but this is under attack. Current position is {A5 holds, C > P}.
Opponent: A7 defeats A5. Opponent cannot play P > C, because it is under attack, and so backs up and chooses another line of attack. Samaritan: AV3 C > Cv. Current position is {A5 holds, C > P, C > Cv}.
Opponent: A8 defeats A5. Again the opponent must back up since Cv > C is under attack.
A5 holds +C
A6 defeats
A5
A5 does not hold
A6 does not hold
A5 defeats
A6
A6 holds +P
At this point the opponent must stop, since there are no further lines of attack. The Samaritan’s position, fAC3; A5; AV2; AV3; AV5g, comprises an argument justifying his action A5, and the three value preferences required to defend that argument AV2, AV3 and AV5. It is exactly this position that the audience is being urged to adopt, since it provides the answer to the lawyer’s question “what does it mean to love your neighbour like yourself?”.
In our opinion, the argument that the story of the
Good Samaritan presents is accurately captured
by the above dialogue. In contrast to
Generating arguments from stories and presenting the different possible extensions based on the value orderings allows one to gain insight into the point of the story: why did the characters act as they did, and which attitudes are advocated in the story? Whilst this is interesting as a theoretical exercise, one additional aim is to implement a system that allows people to explore the stories and character motives in an interactive and intuitive way. One option is to allow humans to engage in a dialogue akin to the ones in section 4.2, thus allowing users to for example, interrogate an agent representing the Samaritan about his motives, and thus gain a better understanding of the story. This can then be used for educational purposes, for example, schoolchildren learning about values through stories.
For such a system, the following separate elements need to be implemented.
1. Construct initial AATS+V on the basis of a story. 2. Include additional hypothetical transitions: ‘what could the characters have done and why?’. 3. Generate a VAF of arguments and critiques based on AATS+V. 4. Execute a dialogue based on the VAF.
Elements 1 and 2 have been done manually for a
few stories: the fable of the Ant and the
Grasshopper and the Parables of the Prodigal Son and the
Good Samaritan. Ideally part of this process is
automated if we want to build a more substantial
corpus. For element 1, we can first automatically
extract the characters and events from stories,
especially from fairly short and simple stories such
as fables. This is certainly not trivial but very
well possible (see e.g.
For element 3, currently, Prolog and PHP
implementations3 exist
Once the arguments are available, it becomes
possible to reason with them in a dialogue.
Recently a dialogue game for arguing about the
motives found in fables and parables was proposed
In this paper we have shown two important connections between computational models of narrative and computational models of argumentation: how argumentation can be used to understand stories, in terms of the motives and attitudes of the characters, and how stories can themselves be used to present arguments, especially arguments designed to persuade the audience to adopt particular attitudes. We have argued that parables can be interpreted as arguments of this sort, and illustrated our views with the famous parable of the Good Samaritan. We have identified several advantages of using stories in this way.
Using stories enables the consideration of hypothetical choices, so that the choice can be made clear and memorable, allowing us to benefit from the vividness of the concrete, without needing to have had any particular experience. Moreover using stories excludes irrelevant considerations: we need not consider facts and actions not mentioned in the story; this simplifies the construction of the AATS, and disbars irrelevant counter arguments, allowing for focus to be kept on the main point at issue. Stories are intended to reinforce or change attitudes: this is preferred to presenting a specific set of norms, since attitudes tend to produce an instinctive, and hence more immediate, response and can be applied to numerous, as yet unforeseen, situations. Moreover, they go deeper and so are more to be relied on. This is why soldiers are taught the history of their regiments: the tales of heroism and derring-do can inspire the loyalty and camaraderie required to bind them into an effective unit in a way in which standing orders cannot hope to do. Often there is no objective argument for an attitude or a norm, and so we need to rely on an emotional reaction, which is more easily produced by a story, especially one which allows the hearers to draw the conclusion for themselves (as does the good Samaritan parable, where the conclusion is stated by the addressee, not in the parable itself).
Engaging in a dialogue about a story further draws out the message of the story, and thus dialogue can act as an aid for story understanding. Our model, when combined with an application for argumentative dialogue, makes these dialogues about stories possible. Users can engage in meaningful discussions about a story not just with each other but also with the characters in a story which, when asked, will explain their motives and thus clarify the point of the story. In this way, our model comprises not just a theoretical discussion of understanding and arguing with stories, but also provides a first step towards a promising applications that can be used in, for example, educational settings.