Data-Driven Alibi Story Telling for Social Believability
Boyang Li, Mohini Thakkar, Yijie Wang, and Mark O. Riedl
School of Interactive Computing
Georgia Institute of Technology
{boyangli, mthakkar, yijiewang, riedl}
As computer games adopt larger, more life-like virtual worlds,
socially believable characters become progressively more
important. Socially believable non-player characters (NPCs) must
be able to act in social situations and communicate with human
players. In this paper, we address one aspect of social
believability: the construction and telling of alibi stories, or an
artificial background that explains what a character has been
doing while not in the presence of the human player. We describe
a technique for generating alibi stories and communicating the
alibi stories via natural language. Our approach uses machine
learning to overcome knowledge engineering bottlenecks
necessary to instill intelligent characters with social behavioral
knowledge. Alibi stories are subsequently generated from learned
social behavioral knowledge. By leveraging the Google N-Gram
Corpus and Project Gutenberg books, natural language is
generated with a discourse planner and text generation that
incorporate different expressivity and sentiment, which can be
employed to create NPCs with a variety of personal traits.
Categories and Subject Descriptors
I.2.1 [Artificial Intelligence]: Applications and Expert Systems—
Games; I.2.7 [Artificial Intelligence]: Natural Language
Processing—Language generation
General Terms
Algorithms, Design, Human Factors.
socially believable characters, alibi generation, natural language
Many computer games invite players to temporarily suspend their
disbelief and enter a rich fictional world populated with non-
player characters (NPCs). To create the illusion that these NPCs
lead their own lives in the virtual world and are not just part of a
show that disappears when the player looks away, each NPC
needs to have a background story or an alibi, which can explain
where they have been and what they have done while they are not
with the player [24]. NPCs should be able to tell these background
stories and recall details to substantiate their stories when asked
to. Otherwise, the suspension of disbelief may quickly fade.
Social believability is thus partially a problem of automated story
The benefits of NPCs with background stories are not limited to
only computer games. For example, Bickmore and Schulman [4]
found that the ability to tell autobiographical stories increases the
likelihood that human users will interact with virtual characters
over an extended period. This can be advantageous when we need
to encourage users to keep interacting with some programs or
electronic devices, such as educational software or medical
devices for self-monitoring.
In this paper, we tackle two challenges of supporting social
believability in NPCs: generating socially believable behaviors
and communicating alibi stories according to personal traits. To
acquire social believability, behaviors in the alibi stories must
adhere to socio-cultural norms in the virtual world. For example,
in modern American restaurants, drinks are typically ordered
before food. Further, NPCs with different personal traits may tell
the stories with different language and levels of detail. Typically,
knowledge about how social situations unfold is encoded as a set
of scripts—knowledge structures that explain what to do in certain
situations and when to do it. However, manually authoring
sufficient script knowledge for NPCs with different personal traits
is an expensive process. This authoring bottleneck limits the
amount of domain knowledge and personal variation available to
human-agent interaction and threatens suspension of disbelief.
We create socially believable NPCs by incorporating data about
the real world into the game world and character decision-making.
These data provide an intelligent agent with observations about
how the real world works and the language used by real humans,
from which believable behaviors can be derived. Our previous
work [11, 12] describes an approach to story generation that
learns about a priori unknown social situations from exemplar
stories acquired using crowdsourcing. This paper applies the
technique to alibi generation and tackles the problem of
communicating alibi stories via natural language. Making use of
the Google N-Gram corpus [16] and books from Project
Gutenburg (, we offer methods to tune
the alibi stories with different levels of detail, and to tell the
stories with different styles and sentiments. The combination of
these tunable parameters allows NPCs to speak with differing
personal traits when telling their alibi stories of the same situation.
2.1 Social Non-Player Characters
Humans respond to virtual agents much the same way they
respond to each other [7, 22]. There are numerous attempts to
create intelligent virtual agents that are capable of interacting with
humans in social contexts, including entertainment [2, 5], learning
[27], and healthcare [3]. There are many aspects to social
believability that must be addressed, including theory of mind
[25], emotion [6], and storytelling [4].