A number of recent models for empirically-grounded social simulation have emerged recently from games and interactive narrative research, generally exploring models of trust, emotion, and social graph changes that occur in the process of inter-character interactions. However, these models so far failed to provide realistic models of opinion change and predisposition to new knowledge. Equipped with such a notion, these emergent social simulations can express both real and fictionalized depictions of modern phenomena like adverse media influence, the spread of “fake news,” and the polarization of ideological sects. We present a preliminary computational investigation into modeling opinion change in virtual characters with this goal in mind.
Humans are rational and emotional beings. Their social systems are complex and contextual. Understanding and simulating humans with virtual characters requires reasoning not just about observable social network graphs or social interactions, but also about geography, economics, and increasingly, online participation and discourse. However, these simulated models typically do not account for some of the most important features of social networks, namely that of the social dynamics of opinion change.
In this paper, we describe a simulation we have designed for a society of virtual characters that can discuss and exchange their views amongst one another. We define a character’s view on the topic as a combination of their internal, personal attitude on a topic of discussion and their externally expressed opinion. We allow for conversationalists to influence each others’ opinions based on existing literature from the social science regarding group conformity and acceptance, as well as by the strength of the overall public opinion. We describe a conversational model that allows virtual characters to subscribe to information sources based on the source biases and opinions, share new information with one another, and form and exchange their opinions on the various issues at hand.
For virtual characters to be socially adept and add to the experience of the player, they must have a sizable expressive range of conversational repertoire. We advance an abstract knowledge base for the characters that groups various objects of discussion under overarching topics, tracks the sources from which they originate with their inherent bias or ratings, and allow the non-player characters (NPCs) to form opinions based on individual preferences or cultural norms. Our system can track the spread of influence (adverse or otherwise) and change in the views of the participant NPCs.
Finally, we demonstrate our system with a case study that showcases a series of conversations where virtual characters discuss current political news from the U.S., exchange their views on individual news articles or issues of interest, and reevaluate their political ideologies and affiliations over time. For instance, an NPC growing up in a more liberal society may eschew conservative ideals, and have a low opinion of the same. Our simulated NPCs are aware of the difference in their internal attitude on a topic of discourse as well as the public opinion shared by other NPCs during their interactions. These differences can lead to the NPCs changing their attitudes over time or expressing opinions different from their attitudes to conform to the society they reside in over time.
We posit that the holding of these opinions on the various topics could lead to the virtual character having access to choices and interactions that would otherwise need to be authored in the character’s preferences or goals. For instance, an NPC growing up in a country with strict gun control, or that holds an unfavorable opinion of gun ownership may choose never to buy a gun.
We imagine that in the future our opinion model could
be used to evaluate how a virtual society would integrate
and accept new additions with new members learning of
the views and opinions of the society while bringing with
them new ideas and concepts from their own culture.
Similarly, opinion modeling for virtual characters could be used
to study the spread of debatable ethical or moral influence
and media bias. Characters could choose to accede to peer
pressure (from the media or society) and change their
behaviors in order to feel a mix of both private acceptance
(that they are acting based on their views) and public
conformity (to gain acceptance by the group). We believe the
behaviors resulting from virtual characters modeled by this
system would be more believable and improve a player’s
interactive experience.
Prior work discusses how designing for richer social
behaviors and interactions amongst virtual characters improve
the believability of the character and the player’s
interactive experience with the system
Research has been conducted on social rules and
practices that virtual agents in the system must conform to as
designed by authors and designers
Other research explores the simulation of conversations
and influence amongst NPCs. They view how societal norms
and popular opinions could affect the behavior of
generated NPCs in a multi-agent system. In Social Role
Awareness
Group formation has been studied in depth by social
scientists, historians, and psychologists to understand how
humans respond to group (or societal) archetypes and opinions.
When modeling group conversations, the physical or virtual
space where conversationalists congregate can be used to
contextualize the interaction, allowing us to incorporate the
history, physical affordances, or cultural significance of the
geographic location or the topic in question. Merely
reading the news enables one to gain a perspective of humans
forming groups to support various issues. These could be
geographic groups, with articles describing how the
Scottish voted to “overwhelmingly remain” in the Brexit vote;
or political ideology groups, with reports on Democrats
discussing immigration resolutions; groups based on shared
interests, with news on Whovians that approve or condone
representation of women in Doctor Who
We review the problem of simulating agents capable of
conversing and sharing opinions with one another. We model
the change in the opinions of the agents based on the model
proposed by Wang, Huang, and Sun in their 2014 paper
Agents are modeled as individual nodes in a social
network graph. Agents may exchange opinions with other
agents if an edge links the nodes in the graph. Wang, Huang,
and Sun defines how agents every agent’s feelings on a topic
is informed by an inner “attitude” towards the topic that
cannot be perceived by other agents, an outward expressed
“opinion” and the level of “uncertainty” they feel about
their opinion. Agents may adjust their internal attitudes or
express modified opinions from their attitudes, on hearing
the opinion of other agents
Due to space restrictions, we refer readers to the Wang,
Huang, and Sun (2014) paper, and the
Source Online or Print Media Articles, Interviews, Candidate Rally Journals, Conference Proceedings Movie Studios Rating Political Bias or Affiliation Approval Rating Journal or Conference Rankings Rotten Tomatoes ratings their proposed agent model combines normative social influences with a continuous dynamics model in a novel approach. Our objective is to extend these current theories of dynamic opinion modeling research to the narrative intelligence community with the goal to simulate virtual societies capable of exploring complex issues of politics, religion, or even simply movie ratings.
Towards this goal, our contribution builds on that of Wang, Huang, and Sun’s in the following ways:
Prior work fails to model the complex and ever-changing social relationships between conversationalists. The authors assume a grid-based society where the same neighboring agents surround an individual throughout their simulation. Our method proposes a more utilitarian definition of social relationships where NPCs with differing or similar opinions could change relationships over time, allowing their old social connections to dissolve over time. Instead of a single object of discussion, we allow characters to discuss a variety of information clustered by topics. This allows for relationships where characters that agree over a few views but disagree over others to change their affinity for one another over time.
We allow for the simulation to add new concepts and topics over time. We believe this could lead to virtual characters to extend their knowledge base while retaining their individual views on existing knowledge.
We list our goals for the project as follows:
Account for bias in characters where agents may have a predisposition to adopt a specific view from prior experience.
Account for bias in the information. Information and sources producing information may have an inherently biased perspective.
Ability for characters with similar opinions to form relationships, and allow ad-hoc groups developing during social interactions to discuss their opinions on various topics.
Be able to use the same discussion model for a variety of different data sources to simulate opinion modeling on discussions.
The model we use to define in-game knowledge is described as follows. For a single discussion, the participants in the discussion choose an Object of Discussion to converse on, obtained from a Source. The Source and the Object of Discussion are associated with a Rating. Multiple objects of discussion can be clustered to form a Topic.
A major contribution of our paper is that this model of the knowledge base can be used for a large variety of datasets while affording the same discussion and opinion modeling. For instance, simulating debates among NPCs about current news articles clustered by political issues and ranked by their bias. Similarly, we could use our model to discuss the merits of various journal articles clustered together by research topics and ranked by journal rankings or have audience members discuss their movie preferences clustered by movie genres and ranked by their Rotten Tomatoes rankings. Some datasets considered during the design phase have been highlighted in Table 1.
We define the rating as the value of the information learned by the NPC in the system. This rating could represent either (1) the personal judgment or favor associated with the presentation of the information, or (2) a measure of the impartiality of the unit of information. The rating is the merit or value of the information that is debated by the NPCs in our system. For instance, this could represent the ratings for a movie, reviews for a paper, or a bias rating for a media source.
We describe topics as a clustering of information regarding a specific subject, or field of information. A specific information unit can be a part of multiple topics at the same time. For instance, a discussion of procedural content generation could belong to the topics of both artificial intelligence or game design. A virtual character may periodically reevaluate his rating of a topic by considering the rating of all the objects of discussion within a topic.
This single unit of information forms the basis of our discussion model. While interacting with one another, virtual characters search through their knowledge base and conversational repertoire, choosing a single object of discussion to debate. An NPC that adds a new object of discussion to his knowledge base will note the original authorial rating intended to be affiliated with the information, and associate with it their own opinions on the topic. These views could be based on prior discussions of the information with conversationalists that introduce the character to the information, as well as on the character’s current view of the topic to which the information belongs. may create information covering a wide variety of objects of discussions and topics. Sources may also have associated with them a rating, representing the expected rating of the information they produce. NPCs may use this rating to choose to subscribe or unsubscribe to these over time based on their current inclinations. For instance, an NPC studying in the computer science domain may subscribe to AAAI for periodic information on the research in their field.
Every participant in the discussion has their own Bias and View on the information and can express their opinions on the object of discussion at hand. These elements and our dataset have been described in further detail below. The attributes of an agent’s view are modeled based on those by Wang, Huang, and Sun.
Bias: We define an agent’s bias to be the agent’s predisposition to adopt a particular view on a topic in a discussion. This bias is informed by either (1) the agent’s views inherited from their parents or (2) a mean of their views on all objects of discussion under the said topic or (3) the initial bias they learn from the conversationalists when the topic was added to their knowledge base during a discussion. Attitude (att): the agent’s private views on a specific issue. Attitude is a real number in the range [ 1; 1], and represents an evaluation of the object of discussion. Opinion (op): an agent’s outwardly expressed or shared views on a specific issue. Like attitude, opinion is a real number in the range [ 1; 1] and reveals the agent’s opinion on the object of discussion to the other dialogists. There may be a discrepancy in the attitudes and opinions of the character since a character may not represent their attitudes accurately to participants. A human example of the situation where this is apparent can be seen in examples of an employee in conversation with his managers who choose not to express his disagreement to avoid being punished.
Uncertainty (unc): a measure of an agent’s confidence in their view. The higher the uncertainty, the more likely the agent is to change his mind or accept other perspectives. As an example, an NPC may express opinions about the legality of abortion in their town. However, the agent may have lower confidence in their attitude if (1) information in their existing knowledge base inadequately back them, (2) if contradictory opinions are presented to the agent with high certainty, or (3) if the agent is surrounded by a society a majority of whom disagrees with him. unc is a real number in the range [0; 1].
strength of the public opinion exceeds this value, the agent will choose to comply with the public opinion to feel accepted within the community. pub thr has a default value of 0:6.
strength of the public opinion is below this value, the agent will choose to stand by their views. The pri thr is a real number in the range [0; 1]. Professors or experts on a particular topic in our simulation would have higher values to indicate their expertise.
During initial generation, the NPC population is assigned random cultural biases on topics in the knowledge base. From this stage on, children inherit as bias the mean of their parent’s biases (i.e., representing “nature”) while tending to agree more with one parent or the other about individual issues. However, the children may change these opinions over time (i.e., representing “nurture”) over the course of several social interactions between the agents.
We begin by clustering the expressed opinions of all
participants of the conversation using the Jenks optimization
method
Public Opinion Formed If the agent has high uncertainty (i.e., agent:unc > 0:8), they are more likely to accept the views of their fellow dialogists. We assign these agents the attitude and opinion equal to the mean of the largest group in the consensus.
If the agent has low uncertainty (i.e., agent:unc 0:8), we find the largest clustered opinion group with views closest to that of the agent. We then calculate the public opinion strength for the selected group and decide if an agent’s attitudes or opinions are affected. The public opinion strength (op str) is calculated as described by Wang, Huang, and Sun by normalizing and finding the mean of the sum of the following factors:
The size (fa) of the group. The larger the group, the stronger the public opinion.
fa = xa=10,if 1 < xa
10 8 < : 0,if xa
1 1,if xa > 10 The homogeneity (fb) in the opinion of the group defining if the group come to a consensus fb = 1=(1 + e24xb 6) fc = 1=(1 + e 12xc+6) The discrepancies (fc) in the agent’s opinion and attitude.
Next, the agent measures their own uncertainty with the strength of the public opinion by calculating two threshold values, th1 = 1 agent:unc and th2 = max(0:6; th1).
Low Opinion Strength (op str < th1): If the opinion strength is too weak, the conversationalist does not change their mind, recognizing the discrepancy between their internal attitudes and ideas and those of the group. Moderate Opinion Strength (th1 op str < th2): – Members with a low uncertainty find the opinion strength of their group strong enough to modify their opinions to the mean of the group. Agents then find their internal attitudes, and their expressed behaviors are inconsistent, and so change their attitudes to match. In this case, agents believe that the change in their views are a natural and expected evolution, and do not realize they are bending to public opinion. – Agents with large uncertainty realize that they are conceding the discussion, and bending to public opinion. They change their external opinions and internal attitudes to match.
High Opinion Strength (op str th2): The agent realizes the strength of the opinion. In this case, the agent may choose to conform to the public opinion with their outwardly expressed views and change their opinion to the mean of the group. However, they do not change their inner attitudes, and in the absence of external pressure will revert to their attitudes.
formed yet, then after clustering the agent finds the cluster of opinions with the opinions most similar to that of the NPC. The NPC modifies their opinion to the mean of the cluster and their internal attitudes on the information being discussed.
In this divisive age, it is difficult (yet unavoidable) to
discuss current political events with family or friends. APIs for
major media sources are available with access to news
articles on various topics. As a case study, our simulation uses
a corpus of news articles
New York Daily News New York Times Associated Press Boston Herald Fox News Editorial
Left Lean Left Center Lean Right
Right
Rating: We use media bias as our rating and associate
with each bias a value as follows: Left( 1:0), Lean
Left( 0:5), Center(0:0), Lean Right(0:5), Right(1:0).
The bias ratings in our dataset are obtained from
AllSides using a combination of blind bias surveys, editorial
reviews, third-party research, independent research, and
community votes to calculate media bias of the
information
Objects of Discussions: Individual news articles are our objects of information. A character will note the original authorial bias of the information and associate with it their views based on their current attitude towards to the topic, their overall political affiliations, and their discussions on the article with other conversationalists.
Sources: Sources are media sources that publish articles on a wide variety of issues. NPCs may subscribe or unsubscribe to these over time based on their current political inclinations. Overall Political Affiliation: is a weighted average of the agent’s attitudes of all topics in the agent’s knowledge base (ranked by an agent’s priorities). For instance, a simple measure how Liberal or Conservative a person is could be expressed as a weighted average of their attitudes on the topics of gun control, abortion, homosexuality, tax reform, and so on.
We simulate a town where characters can interact with one another. Our preliminary experiment allows for two types of organizations, Schools, and Businesses, to facilitate group discussion. Schools Schools choose a subset of topics from the world to teach their students. Professors are modeled to have a low uncertainty value regarding their views. This in combination with the fact that they are regarded as authority figures in the simulation implies that a student is more likely to adopt their views. In Fig. one can see the knowledge base of a recent graduate after he reevaluates his views on Immigration. Businesses NPCs may apply to work at open positions in various local businesses. The application to these positions is based on the knowledge as well as the opinions an NPC acquires over time. For instance, an NPC may be required to have specific views on the topic of abortion as a qualification to work at a local hospital that matches those of their colleagues.
ical outcome for a discussion from our simulation as shown in Fig. 2. NPCs discuss an article titled “Room for Debate: Should ‘Birthright Citizenship’ Be Abolished” at work with colleagues. The article falls under the topic of Immigration and is published by the source NY Times with an original authorial bias calculated by AllSides as Leaning Left. The duration of the discussion is 11 minutes, representing the number of times the algorithm is run, and the views of the participants are updated.
Ruth and Suzanne learn about the article for the first time. They choose to accept the outcome of the discussion as their opinion after applying any pre-existing bias on the topic of Immigration. Richard, whose political views Leaned Left (att = 0:5) before the discussion, is more convinced about his views after reading the article. As such his uncertainty on the subject reduces, but his views stay the same. Vickie, whose political views were aligned Right (att = 1:0) before the discussion changes her views slightly over the course of discussion (att = 0:948) and finds herself a little more uncertain about her view on the article.
However, since the internal attitudes of all four participants on the article and the topic of Immigration (not shown in Fig. 2) remain the same, their overall Political Affiliations do not change. . . yet.
In the future, we hope to be able to simulate cultural or regional opinion preferences by associating opinions with locations at the beginning of the population generation. For instance, NPCs originating in Japan may have a bias for highly restrictive gun control laws. Additionally, some articles or knowledge may be regional, prompting stronger opinions among members directly associated with an issue or allowing for the modeling of an exchange of cultural knowledge. We believe that NPCs with an ability to share and exchange opinions could lead to the generation of a virtual society that has more diversity in beliefs and preferences. A greater awareness regarding the variety of opinions that exist could be extrapolated in the future causing them to revisit the certainty of their opinions on other topics of discourse; thus allowing for more realistic machine enculturation.
We aim to enable the creation of virtual communities sharing opinions that form to discuss their views. These groups could then inform classes of actions available to their members. For instance, a group of students could petition to reduce the school’s carbon footprint. We hypothesize the spread of opinions and influences will enable us to study how more believable information dissemination could occur in simulated populations and narratives.
We thank Chung-Che Hsiao from NC State University for his insight towards the research of this paper. We would also like to express our gratitude to Scott J. McDonald and the AllSides team for providing us with API endpoints and bias rated media articles to use in our Political News Case Study.
Riedl, M. O., and Stern, A. 2006. Believable agents and intelligent scenario direction for social and cultural leadership training. In Proceedings of the 15th Conference on Behavior Representation in Modeling and Simulation.
Ryan, J.; Mateas, M.; and Wardrip-Fruin, N. 2016. A simple method for evolving large character social networks. Proc. Social Believability in Games.
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. a survey. In Proceedings of the 10th international conference on Multimodal interfaces, 61–68. ACM.