Believable Suspect Agents: Response and Interpersonal Style Selection for an Artificial Suspect

Abstract

The social skills necessary to properly and successfully conduct a police interrogation can and need to be trained. In the thesis I will describe the steps I took towards a virtual character that can play the role of a suspect in a police interrogation training. Students of the police academy will be able to use this ‘virtual suspect’ to practise their social skills. The virtual suspect needs to behave as a human suspect would. An important first step towards this goal is an analysis of the behaviour of human suspects in a police interrogation. We investigated whether observers could agree on what interpersonal stance was taken in the DPIT corpus and whether the observers agreed on the way they perceived the various aspects of stance taking. It turned out that agreement between observers was very low on the level of individual turns of speech. However, we showed that a ‘majority vote’ of multiple observers can indeed reveal the dynamics of stance taking in the entire interview. Additionally, we found that for some of the stance types observers agreed more than for others. Next, we explored the relation between the stance taken by the suspect and the turn-taking behaviour. Stances and roles seemed to be mediating factors for the meaning of overlaps and silences in suspect interviews. We analyse the behaviour of actors and not of real suspects to investigate how a virtual suspect should behave in order to behave as a human suspect. Therefore it is important to investigate the effect of using actors. We found that some actors are better at portraying an interpersonal stance than others. Also, validity (recognizing which stance is acted) and agreement between observers did not always go hand in hand. The virtual suspect should behave as a human suspect would behave and thus it should respond to the human interrogator as a human suspect would respond. For this we need to understand why a suspect responds in the manner he or she responds. Therefore, the next step is investigating which social and psychological theories can give an explanation for the behaviour of a (human) suspect in a police interrogation. These theories were then used to create a model that can determine appropriate behaviour of the virtual suspect in response to the behaviour of the interviewer: a Response Model. The credibility of virtual humans, such as the virtual suspect, is crucial for a ‘serious game’ with which users can train their social skills. Users need to be willing to join in the role-play and the probability that they will do this is greater when the game has a compelling story and realistic virtual characters. This requires consistency within the possible behaviours of the virtual human: the virtual human should behave in a manner that is in agreement with his or her nature. The Response Model contains a number of personality settings which can describe the personality of the virtual suspect. This allows the system to play different suspects depending on what the persona of a suspect is according to the scenario of the game. The evaluation had the form of a ‘Guess who you are talking to?’ task. Participants interacted with the virtual suspect and were unaware of its personality setting. Afterwards the participants were asked to choose which of a number of personas was most similar to the personality of the suspect they had just interacted with. Participants were able to ‘Guess who they were talking to’ better than chance. Additionally, we found that personas that differed more were less likely to be confused. This means the response model was indeed able to select different behaviour for different personas and that the behaviour differed more when the personas were more different.
Original languageUndefined
Awarding Institution
  • University of Twente
Supervisors/Advisors
  • Heylen, Dirk K.J., Supervisor
  • op den Akker, Hendrikus J.A., Advisor
Sponsors
Date of Award7 Oct 2016
Place of PublicationEnschede
Print ISBNs978-90-365-4203-6
DOIs
StatePublished - 7 Oct 2016

Fingerprint

suspect
human being
behavior
police
social competence
personality
playing
interview
participant
model
psychological theory
user

Keywords

  • Social Skills Training
  • Response Style
  • IR-101371
  • METIS-317976
  • HMI-IA: Intelligent Agents
  • Interrogation Training
  • Interpersonal stance
  • Virtual Human
  • Virtual Agent
  • Police Interview
  • EWI-27787

Cite this

@misc{e32c44ab075043108c4a0e1e8745dcac,
title = "Believable Suspect Agents: Response and Interpersonal Style Selection for an Artificial Suspect",
abstract = "The social skills necessary to properly and successfully conduct a police interrogation can and need to be trained. In the thesis I will describe the steps I took towards a virtual character that can play the role of a suspect in a police interrogation training. Students of the police academy will be able to use this ‘virtual suspect’ to practise their social skills. The virtual suspect needs to behave as a human suspect would. An important first step towards this goal is an analysis of the behaviour of human suspects in a police interrogation. We investigated whether observers could agree on what interpersonal stance was taken in the DPIT corpus and whether the observers agreed on the way they perceived the various aspects of stance taking. It turned out that agreement between observers was very low on the level of individual turns of speech. However, we showed that a ‘majority vote’ of multiple observers can indeed reveal the dynamics of stance taking in the entire interview. Additionally, we found that for some of the stance types observers agreed more than for others. Next, we explored the relation between the stance taken by the suspect and the turn-taking behaviour. Stances and roles seemed to be mediating factors for the meaning of overlaps and silences in suspect interviews. We analyse the behaviour of actors and not of real suspects to investigate how a virtual suspect should behave in order to behave as a human suspect. Therefore it is important to investigate the effect of using actors. We found that some actors are better at portraying an interpersonal stance than others. Also, validity (recognizing which stance is acted) and agreement between observers did not always go hand in hand. The virtual suspect should behave as a human suspect would behave and thus it should respond to the human interrogator as a human suspect would respond. For this we need to understand why a suspect responds in the manner he or she responds. Therefore, the next step is investigating which social and psychological theories can give an explanation for the behaviour of a (human) suspect in a police interrogation. These theories were then used to create a model that can determine appropriate behaviour of the virtual suspect in response to the behaviour of the interviewer: a Response Model. The credibility of virtual humans, such as the virtual suspect, is crucial for a ‘serious game’ with which users can train their social skills. Users need to be willing to join in the role-play and the probability that they will do this is greater when the game has a compelling story and realistic virtual characters. This requires consistency within the possible behaviours of the virtual human: the virtual human should behave in a manner that is in agreement with his or her nature. The Response Model contains a number of personality settings which can describe the personality of the virtual suspect. This allows the system to play different suspects depending on what the persona of a suspect is according to the scenario of the game. The evaluation had the form of a ‘Guess who you are talking to?’ task. Participants interacted with the virtual suspect and were unaware of its personality setting. Afterwards the participants were asked to choose which of a number of personas was most similar to the personality of the suspect they had just interacted with. Participants were able to ‘Guess who they were talking to’ better than chance. Additionally, we found that personas that differed more were less likely to be confused. This means the response model was indeed able to select different behaviour for different personas and that the behaviour differed more when the personas were more different.",
keywords = "Social Skills Training, Response Style, IR-101371, METIS-317976, HMI-IA: Intelligent Agents, Interrogation Training, Interpersonal stance, Virtual Human, Virtual Agent, Police Interview, EWI-27787",
author = "Merijn Bruijnes",
note = "SIKS dissertation series no. 2016-39",
year = "2016",
month = "10",
doi = "10.3990/1.9789036542036",
isbn = "978-90-365-4203-6",
school = "University of Twente",

}

Believable Suspect Agents: Response and Interpersonal Style Selection for an Artificial Suspect. / Bruijnes, Merijn.

Enschede, 2016. 186 p.

Research output: ScientificPhD Thesis - Research UT, graduation UT

TY - THES

T1 - Believable Suspect Agents: Response and Interpersonal Style Selection for an Artificial Suspect

AU - Bruijnes,Merijn

N1 - SIKS dissertation series no. 2016-39

PY - 2016/10/7

Y1 - 2016/10/7

N2 - The social skills necessary to properly and successfully conduct a police interrogation can and need to be trained. In the thesis I will describe the steps I took towards a virtual character that can play the role of a suspect in a police interrogation training. Students of the police academy will be able to use this ‘virtual suspect’ to practise their social skills. The virtual suspect needs to behave as a human suspect would. An important first step towards this goal is an analysis of the behaviour of human suspects in a police interrogation. We investigated whether observers could agree on what interpersonal stance was taken in the DPIT corpus and whether the observers agreed on the way they perceived the various aspects of stance taking. It turned out that agreement between observers was very low on the level of individual turns of speech. However, we showed that a ‘majority vote’ of multiple observers can indeed reveal the dynamics of stance taking in the entire interview. Additionally, we found that for some of the stance types observers agreed more than for others. Next, we explored the relation between the stance taken by the suspect and the turn-taking behaviour. Stances and roles seemed to be mediating factors for the meaning of overlaps and silences in suspect interviews. We analyse the behaviour of actors and not of real suspects to investigate how a virtual suspect should behave in order to behave as a human suspect. Therefore it is important to investigate the effect of using actors. We found that some actors are better at portraying an interpersonal stance than others. Also, validity (recognizing which stance is acted) and agreement between observers did not always go hand in hand. The virtual suspect should behave as a human suspect would behave and thus it should respond to the human interrogator as a human suspect would respond. For this we need to understand why a suspect responds in the manner he or she responds. Therefore, the next step is investigating which social and psychological theories can give an explanation for the behaviour of a (human) suspect in a police interrogation. These theories were then used to create a model that can determine appropriate behaviour of the virtual suspect in response to the behaviour of the interviewer: a Response Model. The credibility of virtual humans, such as the virtual suspect, is crucial for a ‘serious game’ with which users can train their social skills. Users need to be willing to join in the role-play and the probability that they will do this is greater when the game has a compelling story and realistic virtual characters. This requires consistency within the possible behaviours of the virtual human: the virtual human should behave in a manner that is in agreement with his or her nature. The Response Model contains a number of personality settings which can describe the personality of the virtual suspect. This allows the system to play different suspects depending on what the persona of a suspect is according to the scenario of the game. The evaluation had the form of a ‘Guess who you are talking to?’ task. Participants interacted with the virtual suspect and were unaware of its personality setting. Afterwards the participants were asked to choose which of a number of personas was most similar to the personality of the suspect they had just interacted with. Participants were able to ‘Guess who they were talking to’ better than chance. Additionally, we found that personas that differed more were less likely to be confused. This means the response model was indeed able to select different behaviour for different personas and that the behaviour differed more when the personas were more different.

AB - The social skills necessary to properly and successfully conduct a police interrogation can and need to be trained. In the thesis I will describe the steps I took towards a virtual character that can play the role of a suspect in a police interrogation training. Students of the police academy will be able to use this ‘virtual suspect’ to practise their social skills. The virtual suspect needs to behave as a human suspect would. An important first step towards this goal is an analysis of the behaviour of human suspects in a police interrogation. We investigated whether observers could agree on what interpersonal stance was taken in the DPIT corpus and whether the observers agreed on the way they perceived the various aspects of stance taking. It turned out that agreement between observers was very low on the level of individual turns of speech. However, we showed that a ‘majority vote’ of multiple observers can indeed reveal the dynamics of stance taking in the entire interview. Additionally, we found that for some of the stance types observers agreed more than for others. Next, we explored the relation between the stance taken by the suspect and the turn-taking behaviour. Stances and roles seemed to be mediating factors for the meaning of overlaps and silences in suspect interviews. We analyse the behaviour of actors and not of real suspects to investigate how a virtual suspect should behave in order to behave as a human suspect. Therefore it is important to investigate the effect of using actors. We found that some actors are better at portraying an interpersonal stance than others. Also, validity (recognizing which stance is acted) and agreement between observers did not always go hand in hand. The virtual suspect should behave as a human suspect would behave and thus it should respond to the human interrogator as a human suspect would respond. For this we need to understand why a suspect responds in the manner he or she responds. Therefore, the next step is investigating which social and psychological theories can give an explanation for the behaviour of a (human) suspect in a police interrogation. These theories were then used to create a model that can determine appropriate behaviour of the virtual suspect in response to the behaviour of the interviewer: a Response Model. The credibility of virtual humans, such as the virtual suspect, is crucial for a ‘serious game’ with which users can train their social skills. Users need to be willing to join in the role-play and the probability that they will do this is greater when the game has a compelling story and realistic virtual characters. This requires consistency within the possible behaviours of the virtual human: the virtual human should behave in a manner that is in agreement with his or her nature. The Response Model contains a number of personality settings which can describe the personality of the virtual suspect. This allows the system to play different suspects depending on what the persona of a suspect is according to the scenario of the game. The evaluation had the form of a ‘Guess who you are talking to?’ task. Participants interacted with the virtual suspect and were unaware of its personality setting. Afterwards the participants were asked to choose which of a number of personas was most similar to the personality of the suspect they had just interacted with. Participants were able to ‘Guess who they were talking to’ better than chance. Additionally, we found that personas that differed more were less likely to be confused. This means the response model was indeed able to select different behaviour for different personas and that the behaviour differed more when the personas were more different.

KW - Social Skills Training

KW - Response Style

KW - IR-101371

KW - METIS-317976

KW - HMI-IA: Intelligent Agents

KW - Interrogation Training

KW - Interpersonal stance

KW - Virtual Human

KW - Virtual Agent

KW - Police Interview

KW - EWI-27787

U2 - 10.3990/1.9789036542036

DO - 10.3990/1.9789036542036

M3 - PhD Thesis - Research UT, graduation UT

SN - 978-90-365-4203-6

ER -