Games currently feature simple models of morality and moral reasoning. These typically take the form of a simple binary opposition between good and evil, in which game actions are sorted into these categories, and reputation systems that track the reputation of players (and sometimes NPCs) with ingame factions based on the actions taken. This paper presents a more sophisticated model of moral reasoning based on Lakoff's metaphorical family models: the strict father and nurturant parent morality. This model utilizes more rapid, surface level categorization of different situations in moral categories, and deeper reasoning that characterizes situations in terms of their relationship to moral virtues as determined by the metaphor value system. This model is used in the experimental prototype Argument Box (AB), a social argument simulator where the player argues with clients visiting the shop. Arguments in AB center around the moral virtues and vices of simulated characters in the social simulation Talk Of the Town. This paper presents the current architecture, discussing its technical details.
In many faction and reputation-based games, morality is often portrayed as a simple binary value of good and evil. The reputation of the player character (and sometimes the Non-Player Character (NPC)) is represented as a singledimensional scale where players or NPCs land on a spectrum based on their deeds and the morality points they accumulate.
Unfortunately, NPCs in these types of games tend to
employ surface-level judgments about the characters they
encounter. For instance, games such as World of Warcraft and
Neverwinter Nights
When referring to belief modeling in games, we usually
associate it with facts and theories of mind. For example, a
given character might believe that an NPC has brown eyes
or another NPC likes them. Research games employ this
factual sense of belief modeling, such as characters believing in
false information in (Ryan et al. 2015) and
We believe there is a need to create a more sophisticated moral reasoning system. This paper presents our current progress in Argument Box: a game prototype that employs moral reasoning for NPCs in terms of both surface and deep values.
Our prototype adds a layer of moral reasoning by
incorporating Lakoff’s metaphorical family models: the strict father
and nurturant parent morality into our agent’s belief systems.
In the current prototype, we focus on the Strict Father Model
(SFM), which, as the name suggests,
Our model classifies different situations according to each character’s held moral beliefs at a surface level. Surfacelevel reasoning is used to initially describe an NPC’s stance on an issue. We also incorporated deeper reasoning capabilities that are referenced when an NPC strongly cares for a given topic, granting an NPC the ability to fight for a passionately held belief when the player disagrees with them. Our deeper model allows for deeper reasoning that defines NPC situations related to moral values as determined by the father model metaphor.
In this section, we review related work in terms of morality, reasoning, and social simulation as it relates to various elements of our project.
Several research games incorporate social simulation
and reasoning in their system’s designs. Work such as
PromWeek
Unsurprisingly, social simulation and belief modeling are
closely related to one another. Cif-CK
In Talk of the Town (TotT) (Ryan et al. 2015), a historical town simulator, incorporates characters that can forget information, misremember and lie. Beliefs in TotT are affected by a character’s social network, what characters remember, and the strength of the information supporting a given belief. One interesting application of TotT is Bad News (Samuel et al. 2016), where players converse with real actors portraying different NPCs. The actor acts on instructions from a developer behind the scenes. We incorporate TotT’s generated characters as topics of conversation in Argument Box.
Other work, such as MKULTRA
While social networks do not directly influence this project, we use characters that incorporate social relationships and elements in their design. We essentially used Talk of the Town’s (Ryan et al. 2015) exported characters as conversation topics between the player and the NPC.
Morality in Games There are many games that involve
morality and moral decision-making. In commercial games,
moral decision-making is usually tied to the player’s
actions. Games such as The Witcher 3, Undertale, and those by
Telltale Games
In academia, morality systems in games often target moral
theories or thought experiments. Works such as Togelius
Other authors have built systems based on political
divides as well as belief modeling; Azad and Martens’s
system, Lyra
Lakoff’s Moral Politics Lakoff is well known for
his work on metaphor theory, where he identifies deep
metaphors, such as Life is a Journey or Love is War, that
structure human cognition and the use of language
Lakoff’s
On the other hand, the Nurturant Parent believes that members can grow as a result of nurturance and care. It encourages self-reliance by caring for others. Unlike the SFM, the figurative parents in the NPM value their member’s opinions; the model employs respect between its members rather than dreading punishments or expecting rewards. Like the SFM, the NPM has metaphors and values it believes in, such as Morality as Fair Distribution, Morality as Social Nurturance (e.g., strengthening social relationships with others and mending those relationships), and Morality as Nurturance (i.e., being regularly empathetic).
In our current prototype, we focus on the values employed by the Strict Father Model (SFM). The deep values held by this model include Moral Strength, Moral Boundaries, and Moral Wholeness. We will further describe how we incorporated these metaphors into our system in a later section.
In a classic Monty Python skit called ’Argument Clinic’ (Monty Python ), a man approaches a clinician asking to buy an argument; he then proceeds to argue with the clinician about arguments, debating if their argument is an argument! Likewise, our prototype features an argument simulator game, where NPCs walk into a shop called Argument Box (AB) to procure arguments with the player character, focusing on arguments about moral behavior.
The prototype described in this paper starts when an NPC initiates conversations about characters living in their town. They give the player a piece of gossip about a particular character they heard about, including their opinion and evaluation of that character’s actions. The player can agree, disagree or vouch for the talked-about character by arguing for particular stances. Depending on the player’s response and how passionately the NPC feels about the current argument, the NPC can reference their deep-seated beliefs as moral arguments, which for the presented prototype are based on Lakoff’s strict father model. In the following sections, we will discuss the system in greater detail. First, we examine the components of our system at a high level; we then look at our NPC’s architecture, covering its creation, design, and belief modeling. Lastly, we go through two examples covering a typical argument in our simulation. We note that the characters in Argument Box are modeled as polyhedra. Thus, in the example dialog below, there are references to “shapes” which in our world is the same thing as saying “people.”
Our current prototype features a character coming into a clinic-like shop called Arg Box to argue with the player character about the latest gossip in town. These characters are called conversational NPCs (CNPC). The conversational NPCs (CNPC) are the main NPCs the player converses within the Argument Box.
A single game loop involves the CNPC conversing about one of the town’s characters, stating what they heard, and saying how they feel about it. The conversed-about characters are what we refer to in this paper as the background NPCs (BNPC). The BNPCs include a separate list of NPCs that the player character never interacts with; they are used to seed conversational topics by our CNPCs. These BNPCs are generated from TotT. TotT, as a reminder, is a historical town simulator that generates characters, including various elements such as character relationships, locations, and jobs.
We import the data from TotT in JSON format. We
then search for patterns, such as combinations of
attributes or social connections on BNPCs or temporal
sequences undergone by BNPCs, in a process similar to story
sifting
Once we have our list of BNPCs, the CNPC brings up appropriate dialogue based on the tags present on the BNPCs. The starting dialogue is unbiased and simply expands the tag into a textual utterance. For example, the tag familyPerson gets translated into “Have you heard that X has a large family?”
Each pattern in the system is further mapped to the CNPC’s surface values. These surface values denote how a CNPC generally feels about a set of tags; there is a many-tomany mapping of tags to surface values which is described in more detail below. The CNPC holds each surface value with a strength of high, medium or low, indicating how passionately the CNPC believes in that value.
As long as the player agrees with the judgements of the CNPC, the conversation will stay at the level of surface values. However, when the player disagrees with the CNPC on a surface value they hold strongly, the CNPC will perform deeper reasoning using the strict father model to back up their claim. The conversational options provided to the player allow them to agree, disagree, bring up a specific discussion, or change the conversation topic entirely if they desire. Figure 1 illustrates the general components of the system. The following sections will describe how each component works in greater detail.
BNPCs are generated using the TotT simulation. Our sifting
patterns focus on topics we can have moral debates about,
and so generally exclude details such as locations of homes
and businesses, street names and physical character
descriptions. We currently use 58 patterns to assign tags to BNPCs
as follows:
• Patterns that directly map a single TotT raw attribute.
Examples include isWealthy, departed, and familyPerson,
indicating if a character has wealth, left town, or has a
family, respectively.
• Patterns that are created by combining different TotT raw
attributes. For example, a retired character that is 45 years
old has the pattern retiredYoung assigned.
• Patterns that are based on TotT character jobs; these
patterns are used to make controversial assumptions about
characters, based on the character’s career and its
potential effect on others. For example, any TotT character with
the job miner or cooper is assigned the tag polluterRole.
This is later used for conversations related to the
environment.
• Patterns that focus on relationships with other generated
characters. such as love triangles and backstabbing.
Examples include friendWithBestFriendsEnemy and
InLoveWithSpouseOfFriend.
• Patterns that combine lower-level tags into higher-level
ones. For example, if a BNPC has the tags
adultButNotworking and IsWealthy (from lower-level patterns) this
results in the higher-level tag notWorkingAndRich.
• The last category of patterns combines tags found by
other patterns with the Five Favor Personality traits
As mentioned above, once all pattern matching is completed and the tags have been assigned, BNPCs are placed in a priority queue based on the number and quality of tags. By starting the conversation about characters at the front of the queue, this ensures that the CNPC will have a good number of debatable topics to argue about with the player.
The CNPC starts the conversation by picking the BNPC at the front of the queue to talk about, doing this until it has exhausted all the BNPCs tags or the player chooses another character to talk about.
Once a tag has been chosen, the CNPC starts the conversation by commenting on the tag in a value neutral manner. For example, if our BNPC named Mike had the tag familyPerson selected by the system, the CNPC states, “Oh, have you heard that Mike has a big family?” The CNPC then states how they feel about this tag by relating it to their surface values.
There are currently 28 surface values defined in our system, of which a subset will be held (with varying strengths of low, medium or high) by a CNPC. Examples include LoveIsForFools, LoveAboveAllElse, FamilyPerson, and ShapesArenothingIfNotSocial. BNPC tags map in a many-to-many way with surface values. The tags that map to a given surface value are called core tags of that surface value. The core tags have a many-to-many relationship with the surface values. For example, the tag willActOnLove is a core tag of both the surface values BeTrueToYourHeart and LoveIsForFools. In the case that a core tag maps to two or more mutually held surface values, this provides some non-determinism on how the CNPC will comment on the presence of this tag, depending on which surface value is taken as being activated.
When a CNPC is instantiated, the system randomly assigns the surface values and their accompanying strength. Some surface values are mutually exclusive, so can not be simultaneously held with high strength. For example, if our CNPC holds LoveIsForFools with a high rating, it cannot hold BeTrueToYourHeart with any strength other than low.
Additionally, the mapping from core tags to surface values is used for the conversational options presented to the player. This allows the player to bring up BNPC characteristics during the conversation that explicitly agree or disagree with the CNPC at the level of surface values.
The surface values are used for immediate value-laden reactions during the conversation. As long as the player and CNPC agree with each other, the conversation can stay at the surface value level. However, when the player’s judgements disagree with the CNPC, the system switches to reasoning about deep values as determined by the Strict Father Model (SFM). Implementing a version of the Nurturant Parent Model (NPM) is ongoing work. Switching to this deeper model allows the CNPC to marshal arguments by bringing up characteristics that relate to more deeply held values. This prevents the conversation from immediately degenerating into repeated assertions (e.g. “Yes it is! Not it isn’t! Yes it is! Not it isn’t”) at the surface level.
We define six deep values drawn from Lakoff’s book How
Liberals and Conservatives think
As we mentioned earlier, each surface value includes a set of core tags. Our SFM deep values provide a mapping for each core tag to high morality or low morality. High indicates that the core tag exemplifies the deep value; low indicates it violates the deep value. For example, the surface value BeTrueToYourHeart and one of its core tags, inLoveWithSpouseOfFriend is evaluated as low by the Moral Boundaries deep value. If our CNPC supports the surface value BeTrueToYourHeart and is involved in a argument with the player over this value, it would not make use of the BNPC tag inLoveWithSpouseOfFriend as an argument for their claim, as this violates a deep value. Our upcoming examples will further clarify these situations.
Furthermore, the SFM can check if a CNPC is searching for pro or contra argument to back up a claim about a given surface value. Generally, if a CNPC is supporting an argument for a given surface value, the SFM is looking for core tags that evaluate as high against the deep values, possibly switching the argument to another surface value if no such tags can be found for the current surface value (this is an example of deflecting to another topic to continue to argue for the moral virtue or vice of the BNPC if the CNPC can no longer argue for the current surface value). Alternatively, the SFM can look for core tags that evaluate as low against a deep value to provide a cautionary tale in the argument. For example, if the CNPC holds the surface value LoveIsForFools, the SFM can look for a core tag, such as inLoveWithSpouceOfFriend, that evaluate as low against a deep value (e.g. Moral Boundaries), to make a negative statement about the CNPCs behavior, and thus support their LoveIsForFools stance.
Our first scenario will illustrate a typical conversational loop showing what happens when the player agrees with our CNPC on a medium-rated surface value.
At the start of the loop, the CNPC picks a BNPC to converse about, in this case Caroline Milliem. Caroline has three tags available, inLoveWithSpouseOfFriend, familyPerson and willActOnLove.
The CNPC selects Caroline’s first pattern, inLoveWithSpouseOfFriend, and opens the conversation by gossiping: “Oh, Have you heard! Caroline Milliem is In love with their friend’s spouse!” Under the hood, the pattern was mapped to the surface value LoveIsForFools as ”inLoveWithSpouseOfFriend is one of its core tags. We note, this assignment happens at random; it could have mapped to other surface values as long as the pattern belonged to that core set of the selected surface value.
It so happened that our CNPC is relatively indifferent to this value, holding it with medium strength. They care about the value but not so much that they would raise a fuss if the player disagreed with their opinion (represented as a “disagree” button). The CNPC then presents its opinion based on this surface value, stating “Well, I think that shapes should never let their emotions cloud their judgment.”
The player then has the option to agree, disagree or bring up another topic or BNPC as clickable options. In our scenario, the player agrees: “I couldn’t agree with you more.” The CNPC, satisfied with the result, says, “Yeah, it’s good that you see it my way!”. We note that this scenario presents a medium stance on a surface value; if the player disagrees, the CNPC simply notes its distaste for the player’s choice.
The system then checks if the other tags of the current BNPC have been explored; if they haven’t been explored, the system proceeds with the same loop but with the newly appointed tag familyPerson. Otherwise, the system moves on to the next BNPC in the queue.
Let us take the previous example but assume our CNPC actually cares deeply about the surface value LoveIsForFools. The CNPC conveys this as, “Well, I think that shapes should never let their emotions cloud their judgment AND doing what’s right is always better than doing what feels right.” As a note on authoring, we use AND and BUT to emphasize the CNPC’s high or low stances by extending the same sentence with a modifier that enhances how strongly or weakly they feel about a given topic. This helps minimize the combinatorial amount of dialog we have to write and signals the underlying model more strongly to the player.
In this example the player disagrees with the CNPC, selecting the dialog option: “Are you kidding?? Love is the best thing ever.” Again, the tone mimics that of the CNPC. The CNPC then responds with, “You’re joking! Love is for fools.” It consults the underlying model, in this case the SFM, to back up why love is indeed for fools.
The currently selected tag inLoveWithSpouseOfFriend is validated as a core tag of LoveIsForFools, and determined to score low against the deep value Moral Boundaries. Thus the SFM validates the surface value stance, and allows the CNPC to make an argument based on deep values: “Love can be immoral. Honestly, they have no boundaries. This shape went after their friend’s partner; that’s just wrong.” We note that the text is written in a way that references the specific tag, illustrates the deep value, and provides reasoning as to why the surface value LoveIsForFools should be held. The player can then agree or disagree with the pattern, choose another BNPC pattern to bring up, passing it as an argument for the current surface value.
If the CNPC is presented with a tag that the SFM maps in a way contradictory to the surface value argument being made, it then searches through the BNPC’s available tags for an alternative argument. If any of the remaining tags are core tags of the surface value and the SFM mapping supports the argument, it presents it as a backup argument. Otherwise, the CNPC mimics a person backed into a corner, randomly firing off defenses based on the tags found. This results in more generic arguments, for example, stating “But that BNPC has a family!” in response to the situation where the BNPC is in love with the spouse of a friend.
This paper presents our current work on Argument Box, a game prototype incorporating a moral reasoning system that operates on both surface and deep values (beliefs). The model is inspired by Lakoff’s work on family-based metaphors in moral reasoning. As future work we are implementing the Nurturant Parent Model, refining the conversation loop so that under appropriate (but challenging) circumstances the player can change the CNPCs mind, and playtesting the game.
Monty Python, u. c. Argument - monty python.
Morais, L.; Dias, J.; and Santos, P. A. 2019. From caveman to gentleman: a cif-based social interaction model applied to conan exiles. In Proceedings of the 14th International Conference on the Foundations of Digital Games, 1–11. Nelson, M. 2012. Prototyping kant-inspired reflexive game mechanics. In Proceedings of the 2012 Workshop on Research Prototyping in Games.
ObsidianEntertainment, B. 2002. Neverwinter nights. [CDROM].
Ryan, J.; Summerville, A.; Mateas, M.; and Wardrip-Fruin, N. 2015. Toward characters who observe, tell, misremember, and lie. In Proceedings of the AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment, volume 11.
Samuel, B.; Ryan, J.; Summerville, A. J.; Mateas, M.; and Wardrip-Fruin, N. 2016. Bad news: An experiment in computationally assisted performance. In International Conference on Interactive Digital Storytelling, 108–120. Springer.