Using the Crowd to Generate Content for Scenario-Based Serious-Games
In the last decade, scenario-based serious-games have become a main tool for learning new skills and capabilities. An important factor in the development of such systems is the overhead in time, cost and human resources to manually create the content for these scenarios. We focus on how to create content for scenarios in medical, military, commerce and gaming applications where maintaining the integrity and coherence of the content is integral for the system’s success. To do so, we present an automatic method for generating content about everyday activities through combining computer science techniques with the crowd. We use the crowd in three basic ways: to capture a database of scenarios of everyday activities, to generate a database of likely replacements for specific events within that scenario, and to evaluate the resulting scenarios. We found that the generated scenarios were rated as reliable and consistent by the crowd when compared to the scenarios that were originally captured. We also compared the generated scenarios to those created by traditional planning techniques. We found that both methods were equally effective in generated reliable and consistent scenarios, yet the main advantages of our approach is that the content we generate is more varied and much easier to create. We have begun integrating this approach within a scenario-based training application for novice investigators within the law enforcement departments to improve their questioning skills.
Using the Crowd to Generate Content for Scenario-Based Serious-Games
|Bar-Ilan University, Israel|
|Bar-Ilan University, Israel|
|Jerusalem College of Technology, Israel|
Simulations and scenarios-based games, which constitute an important subset of serious-games, are an important tool for learning new skills and capabilities. Such systems are currently being used in a broad range of applications such as military, government, educational and health care [?, ?, ?, ?]. One main factor in the development of such systems is the overhead in time, cost and human resources to manually create the textual content for these scenarios. Specifically, while the discipline of Procedural Content Generation (PCG) for Games has focused on automatic generation of artificial assets such as textures, models, terrain and game rules, the automatic tools for scenario and textual content generation or adaptation for games and education are much less common [?]. In this paper, we provide an automatic method for generating textual content about everyday activities through combining computer science techniques with the crowd– an approach that can be generally applied to education, government or health care scenarios.
This paper focuses on how one can easily generate textual content within a scenario that can provide narratives within a variety of domains. Our approach presents a novel solution which takes manually written scenarios of everyday activities and uses it to automatically create a new revised, reliable and consistent scenario. At the core of our approach is a database of daily, social scenarios and narratives. We demonstrate how this database is created from Amazon Mechanical Turk (AMT) workers, an established method for data collection from crowd [?]. We are able to ensure the consistency of these scenarios by providing the AMT works with a semi-structured form. This facilitates the creation of a varied database of daily activities’ scenarios, their descriptions (narratives) written in natural language along with key attributes and statistical information regarding possible content replacements to the everyday activities scenarios. Once this database is large enough, we can match the best possibility for a given scenario using the k-nearest neighbor algorithm given the constraint that the replacement must preserve the integrity of the modified data and then generate a personalized description for it using a ‘fill and adjust” approach. Using crowdsourcing offers several advantages: first, it enables us to construct the activities’ scenarios and narratives database rapidly and at a low cost. Second, it gives us a large and diversified content of daily activities from the workers. Last, we also use crowdsourcing techniques to quickly evaluate and demonstrate the efficacy of our approach in an extensive user study with another pool of workers.
While the approach that we present is general and can be used in many scenario-based games, we specifically focused on training scenarios. Specifically, our motivation for generating everyday content comes from a joint project VirtualSuspect with the law enforcement training department. The project’s purpose is to train new law enforcement detectives of property felonies to efficiently extract relevant information from an interviewee. As our approach generates content with relatively low cost and maintenance, we can easily add new training cases, allowing investigators to have repeated practice sessions using different types of investigation techniques for different cases of property felonies. In the VirtualSuspect project, the generated content can also be applicable as an alibi for a specific case, for example, consider a case where a robber broke into a private house on Sunday night and stole a laptop, jewelry and some cash money. The law enforcement investigator (the human trainee) can question a suspect, focusing on what he did on Sunday night. The interviewee (our virtual agent), which can be innocent or not, needs a coherent scenario of his Sunday activities that is consistent with the facts that are known to the investigator and/or are common knowledge. While this paper focuses only on describing the interviewee’s scenario generation portion of this project, it is important to stress that our proposed approach can also be useful for other applications with everyday content, especially in training scenarios such as training doctors to ask a patient the right questions or to help train candidates in a job interview application.
We found that the scenarios and their activities’ details (narratives) we generated were rated as being as believable and consistent as the original scenarios and the original activities’ details, and also compared them to activities’ details that were generated with the more costly planning technique. In addition to the low cost of this technique, another main advantage of our approach is that the generated activities’ details of the revised scenarios and narratives are much more diversified than those with a traditional planning technique. This advantage is an important key when implemented on large amounts of data, as one needs as varied a set of replacements as possible in order to keep the modified scenarios believable.
The development of serious games for training is a complex and time-consuming process [?]. While early works considered a whole story creation without getting any real-time input from the user [?, ?], the later works focus on interactive narratives where the user is part of the story and can affect the plotline [?, ?, ?, ?]. Some works use also hybrid methods whereby a predefined plot is created and an autonomic agent can later add to or adapt the plot in real-time [?]. However, these studies focus on the general plot of the story but not on the story’s details, which were almost exclusively manually created by content experts. A second important direction is the Procedural Content Generation (PCG) for Games discipline. However, while there has been significant in PCG generation for serious games in the past decade, to date, tools for textual content generation are still lacking [?]. To address this need, we created a framework for generating coherent scenarios, which include a sequence of activity events (activities) and also the activities’ descriptions (narratives).
The term scenario have several definitions depending on the context it is used. According to Hendrikx et al. [?] “Game scenarios describe, often transparently to the user, the way and order in which game events unfold. Two types of game scenarios can be distinguished, abstract and concrete. The abstract game scenarios describe how other objects inter-relate. The concrete game scenarios are explicitly presented in the game, for example as part of the game narrative.”. In this paper, we refer to concrete scenario types. As such, we use the term scenario to describe a sequence of activity events (activities) and their descriptions (narratives).
As is the case with several recent works [?, ?, ?], we implement a hybrid method which constitutes a “fill and adjust” semi-automatic narrative generation method. However, unique to our work, we leverage the crowd to create the textual content. The first step of our task is to identify the activities that should be modified. This problem has been studied extensively in database systems [?, ?], and we built upon previous work [?] that uses a maximal satisfiability solver (MaxSat) in order to identify these activities. Specifically, we used the off-the-shelf “akmaxsat” solver [?], which was one of the winners at “MaxSat 2010” for the “Weighted Partial MaxSat” category. Once the portions of the data that must be modified were identified, we found that using crowdsourcing allowed us to create a database of alternatives quickly and at a low cost. In crowdsourcing platforms, such as AMT or MicroTask, requesters typically break up a complex task into a number of simpler sub-tasks to make them easily applicable for layman workers. Previous work found that although micro-task markets have great potential for rapidly collecting user measurements at low costs, special care is needed to formulate tasks in order to properly harness the capabilities of the crowd [?]. In order to facilitate the successful creation of the database needed for our application, we built questionnaires with precise instructions for each task and with proper compensation for the workers.
One traditional approach for activity-details generation is planning-based systems [?, ?, ?, ?]. The planning-based approach uses a causality-driven search to link a series of primitive actions in order to achieve a goal or to perform a task. For the domain of descriptive, every day activities, hierarchical scripts can capture common ways to perform the activity. Therefore, we implemented the planning-based generator using a Hierarchical Task Network (HTN), which is one of the best-known approaches for modeling expressive planning knowledge for complex environments. We used the state-of-the-art SHOP2 planner [?], a well known HTN planner, which has been evaluated and integrated in many real world planning applications and domains including: evacuation planning, evaluation of enemy threats and manufacturing processes. In order to validate the significance of modified scenarios with the activity-details we created, we also implemented a planning-based activity-details generator. As we later report, this approach is more costly to implement and generates less varied scenarios than the crowdsourced approach we now describe.
We propose and build a system (presented in Figure Using the Crowd to Generate Content for Scenario-Based Serious-Games) which ensures the scenario’s integrity while replacing a given set of information. In the paper we will use the following running example to explain different stages within the system. For clarity reasons, we are using a simple example.
John is a 21-year-old male who is single and has no children. He broke into a private house on Sunday night and stole a laptop, jewelry and cash money. He is now being questioning and need an alibi story for Sunday night.
In this example, our system needs to preserve John’s activity by concealing the information that John broke into a house. This is done by replacing the activity with a common activity, such as . After this activity switch has been made, a modified scenario with a new activity-details will be generated. Our system does this by basing itself on details from a collection of reported activities, which it modifies to better match the scenario main character’s profile. Referring back to example 1, we base the revised scenario on a reported activity of a 26-year-old male with no children who goes out to lunch (Example 2). Note that in this case we need to change the details about the time and location to match the required alibi (Example 3).
“I went and got lunch and a beer at a local bar “The Liffey”. It was during March Madness, so I was watching some basketball. I sat at the bar and got chicken wings. I watched a few basketball games and ate. I read the newspaper a bit too. The food at “The Liffey” is always good. The team I picked won so that was also good. ”
: “On Sunday night I went out for dinner. I did not really want to spend too much so I went to “54th Street”. I sat at the bar and got chicken wings. I watched a few basketball games and ate. I read the newspaper a bit too. The food at “54th Street” is always good. The team I picked won so that was also good.”
We now focus on how the system identifies what needs to be modified in the scenario, how it finds a similar record on which to base the modified scenario, and how it generates the final scenario with the appropriate activity-details. As we analyzed the coherent problem in scenarios, we identified three main questions the system needs to address. These questions are: (1) What should be removed from the scenario? (2) With what should we replace it? (3) How should we replace it? Accordingly, our system consists of three main modules, one to address each of these questions. The modules are: (1) A Maximal Satisfiability Solver (MaxSat) which ensures the scenario’s integrity and identifies the places where modifications are required. Referring back to our example, this module will identify the need to replace the activity; (2) Activities Generator which uses an algorithm based on the k-nearest neighbor algorithm in order to choose the most appropriate activity replacement for the required modification. In our example, it suggests replacing the activity with the activity; and (3) Activity-Details Generator which expands the required activity replacement into a realistic, reliable, descriptive activity, as can be seen in examples 2 and 3.
In this paper we do not focus on the MaxSat solver as this approach has been previously studied [?]. We use an off-the-shelf MaxSat solver to generate the optimal solution with the minimum number of modifications to the scenario. We expand the traditional usage of this approach by using placeholders to protect facts which will later be replaced. Once the constraints have been identified, the system uses the activities generator and activity-details generator modules to fill in the missing details in the scenario for these placeholders. After the scenario has been finished but before the modified scenario is returned, the MaxSat solver is run again in order to confirm that no constraints have been violated in the final modified scenario.
This paper focuses on a novel approach using crowdsourcing techniques to generate replacement activities and their details. This is done through creating an activities dataset that contains a collection of daily schedule records. These records are composed of a list of activities and a collection of possible activities replacements, which can form the skeleton of a modified scenario. The details in these activities are selected from a dataset containing a collection of descriptive reported activities written in natural language. The selected replacement activity is associated with revised details that are consistent with a specific user profile. We now detail how these modules are defined and applied.
Before we present the system implementation, we define several concepts to be used in its description.
User Profile (P) - describes the user (i.e. the scenario’s writer or the scenario’s subject) properties and consists of gender, age, personal status and number of children.
Scenario - a sequence of activities and their descriptions represented as a list of pairs <AI, ADR> where each activity instance AI is accompanied with an activity-details record ADR. The description of these two fields follows.
Activity Instance (AI) - is a specific occurrence of activity which is part of the scenario and is composed of the activity name, a day, start and end time, location and participants. In our example: .
Activity-Details Record (ADR) - is a tuple <P,ADA,ADP> where: P is a user profile, ADA is the activity-details attributes vector and ADP is the activity natural language presentation. A detailed description of the latter two fields follows immediately.
Activity-Details Attributes (ADA) - contains a vector of attributes which accompanies the activity-details. This vector is a superset of the activity instance AI, which contains the general attributes such as participants, a day and location, but it also contains information specific to the activity-details domain, such as restaurant name and type. It can contain optional values, and thus can be full or partial, for example in the eat-at-a-restaurant activity a person can eat at a restaurant alone, but can also go with a spouse. Within example 2 above, we represent this vector as: day (Thursday), part-of-day (noon), name (The Liffey), type (Bar and Grill), location (downtown) and participants (alone).
Activity Presentation (ADP) - is the activity’s detailed description written in natural language. It is composed of three parts: (1) The activity Introduction describes the main facts of the activity, such as who went, when, what are the main objects’ names (which movie/restaurant), where and why; (2) The activity Body describes the activity in detail, what was the course of events and what happened during the activity; and (3) The activity Perception describes how good or bad the experience was from the user’s perspective. Note that we intentionally split the activity presentation into these three parts. This semi-structured free text writing is very applicable when describing social, everyday situations. It also centralizes most of the activity specific details in the introduction part, which facilitates adjusting the activity to a new user profile and attributes vector. Accordingly, the presentation of example 2 is: (1) Introduction: “I went and got lunch and a beer at a local bar” (2) Body: “I sat at the bar and got chicken wings” and (3) Perception: “The food at “The Liffey” is always good”
We developed an innovative methodology to build the datasets using crowdsourcing. In all tasks, the AMT workers were first asked to provide their profiles P (gender, age, personal status and number of children). Then, they were presented with a semi-structured questionnaire containing a list of questions and were asked to fill it out. Examples of these forms can be found at http://aimamt.azurewebsites.net/.
One main challenge is how to best select the most appropriate record from within the entire dataset. To accomplish this task, we define a compatibility-relevant measure (as we describe in the algorithm flow) which is based on the similarity measure between attributes in order to predict which record is the best replacement. The basic component of the similarity comparison is the decision whether two values are similar. To make this comparison we associate each one of the attributes with a specific comparison function which gets as input two values and returns one of the three values: same, similar and other. Note that in the case one of the values is missing, it returns the similar value, as we assume the generators will fill this attribute with a similar value. For example, we consider the number-of-children attribute to be the same if the difference between the two values is 1 or less, similar if it is less than or equal to 3 and other if one person has children and the other does not or when the difference is greater than 3. For the day attribute, the comparison function returns same if both values equal, similar if both values are weekdays or weekends and other otherwise.
The activities generator’s goal is to find the most appropriate activity replacement, such as , for the scenario’s placeholder provided by the solver, such as . To accomplish this goal we built the KAR (K-nearest neighbor Activity Replacement) generator whose input is a user’s profile, the scenario’s activities list and an indication which activity instance to needs be replaced. KAR returns a revised scenario with a replaced activity which will later be associated with a natural language description and details.
The Dataset The activities dataset, denoted , contains a collection of daily schedule records, SR, and activity records, AR. The SR is a tuple <P,Sch> where P is a user profile and Sch is a daily schedule represented as a list of activity instances AI. The AR is a tuple <P,Act> where P is a user profile and Act is the activity properties and consists of the activity name (such as see-a-movie or have-a-meeting) and six attributes: a day (a weekday or weekend), part of day, duration, location, participants and frequency. We use two types of questionnaires in order to acquire the two record types. The first form is used to define the set of possible activities and the second form is used to collect additional data on each of the activities from a variety of profiles for the activities generator (described below). In the first questionnaire, we acquire weekday and weekend schedules (in an hour resolution), where we asked the workers to describe the activities as specifically as possible and limited each activity to up to a 3 hour duration. As we later describe, this data was also used to evaluate our system. For each activity in the schedule, the worker is asked to fill in the activity name (written in free text), the participants and the location of the activity. Defining the set of possible activities requires only a few schedules, and thus we collected 16 schedules from 8 subjects (4 male, 4 female, ages 23-53) which were paid 25-40 cents each for writing two schedules. We then map all of the activities in these schedules into an enumerated list and store the converted schedules Sch with their profile P at as the SR records. The second questionnaire used to acquire additional properties, Act, for each of the activities in the collected schedules. The workers were presented with a form which included a list of activity record fields (Figure Using the Crowd to Generate Content for Scenario-Based Serious-Games), and were asked to fill it out using predefined selection lists. The AR records were collected from from 60 subjects (23 male, 37 female, ages 21-71) which were paid 35 cents each and were also stored at .
KAR Our activities generator, KAR, is implemented using the k-nearest neighbor algorithm and it uses a compatibility measure to predict which activity record is the best replacement for the placeholder. To calculate this measure, we select 10 attributes: the 4 attributes of the profile P and the 6 attributes of the activity Act. KAR first calculates the similarity measure of each of these attributes for all of the activity records AR within the dataset compared to the given user’s profile P and the activity placeholder AI, which it needs to replace. For example 1 with a profile (Male, 21, single, no children) and a placeholder compared to the following activity record (Female, 31, married, 2 children), (EatAtRest, weekend, night, one hour, downtown, spouse, once a month), the similarity measures are: gender (other), age (other), number-of-children (other), personal-status (other), day (same), part-of-day (same), duration (similar), location (same), participants (other), frequency (similar). The importance for any two values to be the same or at least similar depends on the specific attribute. For example, having the same gender value in the generated scenario is much more important than having the same age. To associate different importance levels for each attribute similarity measure, we developed a scoring function that gets an attribute and a similarity measure and returns a score within the range [-15,15]. We refine this score function using several preliminary trial and error iterations. The KAR generator then calculates the compatibility measure for each of the records, AR, as a summation of the scores of each of these attributes and its calculated similarity measure. Last, KAR sorts the activities records according to this measure and uses the k-nearest neighbor algorithm in order to choose the best candidate. Specifically, we implemented two variations of this algorithm, one with K=1 and the other with K=11. While the K=1 variation returns the activity with the highest measure, the K=11 variation also takes into account the number of similar records and thus returns the activity with the highest probability from the top 11 measures.
The activity-details generator module is responsible for turning a given activity instance in the revised scenario into a realistic, reliable, descriptive activity. It gets as input the user’s profile P and a partial activity details attributes vector ADA (which is made up of the values given in the activity instance AI). It returns as output a new activity-details record ADR which contains a reasonable, consistent and realistic activity presentation ADP written in natural language, which substitutes activity in the revised scenario. We implemented two types of generators: our approach, SNACS (Social Narrative Adaptation using CrowdSourcing), which uses the activity-details records we collected from the crowd, and for comparison a traditional planning-based generator, which is a common technique for content generation in many real world domains.
The Dataset We again use the crowd as the source of the activity-details dataset, denoted as , and build a collection of human activity-details written in natural language for a specific activity, such as see-a-movie. We used a dedicated, semi-structured questionnaire on AMT to collect the activity-details record ADR which includes: the profile P, the activity attributes vector ADA and the activity presentation in natural language ADP. Here, the workers were asked to describe daily, social activities in natural language in as much detail as possible according to the three activity presentation parts - introduction, body and perception. Then, the workers were presented with a list of specific questions used to collect the activity-details attributes vector, such as “What was the name of the movie/restaurant?”, “With whom did you go?” and “on what day?”. The completed records ADR were then stored at . We intentionally split the activity’s detailed description into three parts. On one hand, this semi-structured free text writing is very applicable when describing social, everyday situations, and it helps us to elicit a detailed description of the activity from the workers. On the other hand, it centralizes most of the activity-specific details in the introduction, which allows us to adjust the activity-details to a new user’s profile and attributes vector during the activity details generation without the need for intensive usage of NLP tools. Specifically, we collect and store 10 activity-details for 4 activities: two are entertainment activities (see-a-movie and eat-at-a-restaurant) and two are errand activities (buying-groceries and dry-cleaning). These records were collected from 20 subjects (6 male, 14 female, ages 19-55) which were paid 50 cents each for writing two activity-details.
SNACS Our activity-details algorithm, SNACS, first selects a candidate record from the activity-details dataset . We present 3 variations of this selection process below. Then, the algorithm completes the missing activity-details attributes. It generates attributes which are similar to the selected, original record’s attributes and matches them to the new user’s profile. It starts with the participant: who went and how many people participated in the activity. It then generates the objects’ names (movie, restaurant, location) and time frame attributes. For example, if in the original activity someone went to see a children’s movie with his son and the new user has no children, SNACS can choose to include his niece/nephew among the participants. Next, the algorithm generates the activity’s natural language presentation. First, it replaces the original activity’s introduction, i.e. its first part (who went, when, where, why), with a newly generated introduction according to the new profile and the new vector of attributes. This is done by using SimpleNLP [?], a Natural Language Generation (NLG), template-based surface realization, which creates an actual text in natural language from a syntactic representation. We created several NLG templates for each activity type, which were randomly chosen during the introduction generation. For example, one of the NLG templates used in order to build an activity introduction for the see-a-movie activity was: “Last I went to a movie with my . We went to see the movie at ”. Each such template can generate a few variations according to the chosen attributes. For example, the first part of the above template, where the participants are a wife and son and the time is Sunday afternoon, can generate (a) Last weekend I went to a movie with my family or (b) Last Sunday afternoon I went to a movie with my wife and my son. Finally, SNACS applies some adjustments to the body and perception parts of the chosen activity’s presentation (the second and third parts). This is done by replacing the references of the original attributes’ vector with the new corresponding activity attributes’ vector. In example 3, we replaced the restaurant’s name.
We implemented 3 variations of the SNACS algorithm which differ in how the original candidate activity-details record is chosen: SNACS-Any, SNACS-Bst and SNACS-Tag. The SNACS-Any variation is a baseline measure that randomly chooses one activity-details record from . No further logic is performed to check how appropriate that choice is. In contrast, both the SNACS-Bst and SNACS-Tag variations use a compatibility measure to select which candidate from among all records in will serve as the base for the generated activity description. The compatibility measure is based on 7 attributes: the 4 attributes of the profile P and only 3 attributes from the activity-details attributes vector ADA (participants, type and part-of-day). However, when assessing the compatibility of the activity-details record ADR, we also have to account for the activity presentation as written in natural language. Thus, we define an importance level vector ILV, which corresponds to these 7 attributes, for each activity-details record ADR within the dataset . Each value in ILV is a value SM and is used to represent the importance of the compatibility of a given attribute within the activity body and perception parts of the activity presentation. These values control how much importance should be given to having similarity between the original and generated activities’ attributes. Accordingly, if a given attribute within ADR can be modified without violating any common sense implications, then the value is other. At the other extreme, if that attribute is critical and even small variations can make the activity implausible, then the value is same. SNACS considers two approaches in which the vector ILV can be built for every record. The first approach, denoted as SNACS-Bst, uses a fixed (automatic) ILV across all records within . Specifically, it contains the same value for the gender attribute and a similar value for all of the other attributes. The second approach, denoted as SNACS-Tag, utilizes a content expert to manually tag every record within . For example, the manual ILV for example 2 is gender (same), age (similar), number-of-children (other), personal-status (other), participants (other), type (similar), part-of-day (similar).
During runtime, SNACS first calculates the similarity measure of each of these attributes for all of the records ADR within the dataset compared to the given user’s profile P and the (partial) activity instance AI it needs to replace in order to select the best candidate activity-details record. Recall that we used a similar value in case of missing values. In example 1, we evaluate John’s profile and the activity instance compared to the record ADR from example 2. Thus, the similarity measures are: gender (same), age (similar), number-of-children (same), personal-status (same), participants (same), type (similar), part-of-day (other). We again build a score function (which was refined using trial and error iterations), but this time it gets as input an attribute, a similarity measure and an importance level and returns a score within the range [-15,15]. SNACS then calculates the similarity compatibility measure for each of the records, ADR, as a summation of the scores of each of these attributes, its calculated similarity measure and its given (fixed or manual) importance level. Finally, the record with the highest measure value is chosen as the best activity-details candidate.
Planning-Based Generator In order to validate the significance of SNACS, we also implemented a HTN planning-based generator, denoted as Planner, using a plot graph that we built manually. A plot graph [?] is a script-like structure, a partial ordering of basic actions that defines a space of possible action sequences that can unfold during a given situation. As we wanted to get richer activity-details which include a detailed description of the activity written in natural language, we gave the planner an option to tailor natural language descriptions in the basic actions portion of the activity. We defined a set number, 10-15, of different descriptions that were tailored to each one of the selected actions, which assured the implemented planner had a variety of descriptions with which to build activity-details. These descriptions were manually handwritten by two experts, which needed approximately one hour to write the set of descriptions for each basic action option. Part of these descriptions were also manually tagged with specific tags, such as movie or restaurant types. The tagging gave the generator an option to choose between a generic description which can be associated with any movie/restaurant type or a specific description which can be associated with the current selected type, such as action movie or Italian restaurant. Note that in SNACS, this step is not necessary as it automatically gets the activity’s detailed descriptions from the original activity-details record. We also implemented dedicated actions’ realizators (SimpleNLP based) that took the planner output, a semi-structured plan, and translated it into a natural language activity presentation in addition to the introduction’s realizator we also used in SNACS.
Overall, the HTN-based generator has an inherently higher cost associated with it as compared to SNACS for the following reasons: Both SNACS and the planner have the steps of building the activity introduction templates and the implementation of the logical constraints. However, the planning-based algorithm implementation also required the following additional manual steps: the manual building of the plot graph; the writing, associating and tagging of several detailed descriptions for each basic action; and writing a specific realizator for each basic action. Each one of these steps requires both time and resources from a content expert or a system’s designer. In fact, because of this cost overhead, we only used the HTN-planner in order to define the two entertainment activities (see-a-movie or eat-at-a-restaurant) and intentionally did not implement the HTN-based generator for the errand activities. Nonetheless, the activities’ detailed descriptions produced by SNACS were as good as those developed by this costly process, as our results detail in the next section.
Random Baselines We also implemented two random methods as baselines to ensure the validity of the experiment. The first random method, denoted as Rnd-SNACS, uses the SNCAS infrastructure. It randomly chooses one of the activity-details records in the dataset and then randomly fills in the rest of the activity attributes. The second random method, denoted as Rnd-Planner, uses the planning-based generator. We removed the plan’s logical built-in constraints and used random selections instead.
In this section we present the evaluation of the generated, revised scenarios and their associated descriptive activities. Note that we don’t include the evaluation of the MaxSat solver, as we use an off-the-shelf, previously studied solver. We evaluated separately the activities generator (KAR) and activity-details generator (SNACS) as described in the following sections. We use crowdsourcing to evaluate the efficacy of our approach in an extensive user study (200 subjects), again in AMT but with another pool of workers. We ensure that subjects answer truthfully by including open test questions and reviewing it manually before accepting the grades. We also estimate that completing a survey should take 8-15 minutes, so we filtered out forms which were filled out within less than 4 minutes. Examples of these questionnaires can be found at http://aimamt.azurewebsites.net/.
The purpose of this experiment is to check the integrity and coherence of the scenario’s activities list after the replacement of one of its activities.
Setup We chose to use the daily schedules we already collected from the crowd for , as the original scenario (without the activity-details that will be evaluated next).We randomly cut a section of 7-8 hours from each of the original activities list, to which we refer as Original. We then randomly chose one activity for the list to be replaced. We generated three revised lists by replacing the chosen activity. Two of the lists were generated using KAR algorithm, one with K=1 and the other with K=11, to which we refer as KAR K=1 and KAR K=11 respectively. The third list Random was generated using a random replacement and was implemented as a baseline to ensure the validity the experiment. We then evaluated these 4 variations of each activities list using AMT questionnaires. Each activities list was associated with a user profile and a day and the workers were asked to rate it with reference to the profile attached. The grades were valued from 1 (Least) to 6 (Most). An even number of choices was given as we didn’t want the users to choose a middle value. The users were asked to grade three aspects: reasonable, matching to profile and coherent. The users were also asked to try to recognize which, if any, activity was replaced, and explain their answers in free text. Each activities list was rated by 15-20 subjects, to ensure we had enough independent grades for each generation method. A total of 80 subjects (38 male, 42 female, ages 19-61) participated in the evaluation and were paid 25-40 cents each.
Results Table Using the Crowd to Generate Content for Scenario-Based Serious-Games presents the average grades for each aspect and also the total average grade, which was calculated as the average of these three grades. The results show that our generated method KAR K=11 got the highest results, higher than the Original, however, there is no significant difference between the results of Original, KAR K=11 and KAR K=1. As expected the Random variation got lower grades, which are significantly lower than the others (specifically, the ANOVA test of Random compared to Original, KAR K=11 and KAR K=1 had a much smaller than 0.05 threshold level with p=1.72E-9, 2.8E-11 and 1.13E-8 respectively). It also can be seen from the replacement identification percentage (the last row in Table Using the Crowd to Generate Content for Scenario-Based Serious-Games), that only 7% of the users identify the generated activity in the KAR K=11 and KAR K=1 methods compared to 61% in the Random replacement. These percentages are significantly different from an uniform random selection, which has a probability of 20% for being chosen.
The purpose of this experiment is to check the authenticity, integrity and coherence of the generated descriptive activities which accompany the revised scenario.
Setup We chose to evaluate four types of activities: see-a-movie, eat-at-a-restaurant, buying-groceries and dry-cleaning. For each activity type, we generated 4 user profiles. Then for each profile we generated activity-details for all of the generation methods111We implemented the planning-based methods only for the see-a-movie and eat-at-a-restaurant activities because of the cost overhead.. We also randomly selected 4 additional activity-details out of the original activity-details dataset for each activity type. As before, we ran the experiment using AMT questionnaires for comparison. Each activity-details was associated with its user profile and the AMT workers were asked to rate (with the same 6-value scale as before) six aspects of the activity-details: authenticity, reasonable, matching to profile, coherent, fluency and grammar. As before, subjects were also asked to explain their choice in free text. Each activity-details was rated by 8-10 workers to ensure we had enough independent grades for each activity type and generation method. A total of 120 subjects (59 male, 61 female, ages 18-69) participated in the evaluation and were paid 40-50 cents each.
Results Table Using the Crowd to Generate Content for Scenario-Based Serious-Games presents the average grade for each activity type and generation method, which was calculated as the average of the six aspects’ grades. The results show that our approach produced revised activity-details which were rated as being as reliable and consistent by workers when compared to the original activity-details and the planning-based technique. The main advantage of our approach is that the descriptive activities are much easier to create. Both of the random variations Rnd-SNACS and Rnd-Planner got significantly lower grades than all the others (specifically, the ANOVA test for the movie activity-details of Rnd-Planner compared to Original, Planner and SNACS-Bst had a much smaller than 0.05 threshold level with p=1.95E-4, 4.99E-3 and 9.30E-3 respectively). The results also show that all SNACS-based algorithms got very similar grades where SNACS-Bst and SNACS-Tag are slightly better than SNACS-Any. Overall for the entertainment activities details, although the Original grades are slightly higher than all of the other methods, there is no significant difference between all of the SNACS-based generator methods or the planning-based generator or the original activity-details. For the errands activity-details, there is also no significant difference between the grades of SNACS-Tag, SNACS-Bst and the Original activity-details, although the SNACS-Any variation got lower grades for the buy groceries activity-details. We found that the results for each of the six aspect grades were very similar to the average grade222We omit this table due to lack of space..
The paper makes the following key two contributions: (i) It is the first work to address the problem of modifying scenarios to generate personal information but yet maintains consistency even when varied scenarios are generated. (ii) It provides a methodology to use crowdsourcing in a principled way for this task. Instead of manually modifying scenarios, which makes the development process costly in both time and resources, we used the crowd as the source of our dataset, thus reducing the time and effort needed by to maintain coherence in these scenarios. To accomplish this task, we use the MaxSat logical engine in combination with a novel approach for the generation of everyday activities scenarios using the crowd. Our evaluation showed that our revised scenarios and their activities’ details were rated as being as reliable and consistent as the original scenarios and the original activities’ details, and also compared them to activities’ details that were generated with the more costly planning technique.
This work was supported in part by ERC grant # 267523.
- 1 Bertossi, L. Consistent query answering in databases. SIGMOD Record 35, 2 (2006), 68–76.
- 2 Biskup, J., and Bonatti, P. Controlled query evaluation for enforcing confidentiality in complete information systems. International Journal of Information Security 3, 1 (2004), 14–27.
- 3 Biskup, J., and Wiese, L. Preprocessing for controlled query evaluation with availability policy. Journal of Computer Security 16, 4 (2008), 477–494.
- 4 Cavazza, M., Charles, F., and Mead, S. Character-based interactive storytelling. IEEE Intelligent Systems 17, 4 (2002), 17–24.
- 5 Gandhe, S., Whitman, N., Traum, D., and Artstein, R. An integrated authoring tool for tactical questioning dialogue systems. In 6th Workshop on Knowledge and Reasoning in Practical Dialogue Systems (2009).
- 6 Gatt, A., and Reiter, E. SimpleNLG: a realisation engine for practical applications. In 12th European Workshop on Natural Language Generation (2009), 90–93.
- 7 Graafland, M., Schraagen, J., and Schijven, M. Systematic review of serious games for medical education and surgical skills training. British Journal of Surgery 99, 10 (2012), 1322–1330.
- 8 Hajarnis, S., Leber, C., Ai, H., Riedl, M., and Ram, A. A case base planning approach for dialogue generation in digital movie design. In Case-Based Reasoning Research and Development (2011), 452–466.
- 9 Hendrikx, M., Meijer, S., Van Der Velden, J., and Iosup, A. Procedural content generation for games: a survey. ACM Trans on Multimedia Computing, Communications, and Applications (2013).
- 10 Kittur, A., Chi, E. H., and Suh, B. Crowdsourcing user studies with mechanical turk. In SIGCHI Conference on Human Factors in Computing Systems, ACM (2008).
- 11 Kuegel, A. Improved exact solver for the weighted max-sat problem. Workshop Pragmatics of SAT (2010).
- 12 Lane, H. C., Schneider, M., Michael, S. W., Albrechtsen, J. S., and Meissner, C. A. Virtual humans with secrets: Learning to detect verbal cues to deception. In Intelligent Tutoring Systems, Springer (2010), 144–154.
- 13 Meehan, J. Tale-spin, an interactive program that writes stories. In Fifth International Joint Conference on Artificial Intelligence. (1977).
- 14 Nadolski, R. J., Hummel, H. G., Van Den Brink, H. J., Hoefakker, R. E., Slootmaker, A., Kurvers, H. J., and Storm, J. Emergo: A methodology and toolkit for developing serious games in higher education. Simulation & Gaming 39, 3 (2008), 338–352.
- 15 Nau, D., Au, T., Ilghami, O., Kuter, U., Murdock, J., Wu, D., and Yaman, F. SHOP2: An HTN Planning System. Journal of Artificial Intelligence Research (JAIR) 20 (2003), 379–404.
- 16 Niehaus, J., Li, B., and Riedl, M. Automated scenario adaptation in support of intelligent tutoring systems. In Twenty-Fourth International Florida Artificial Intelligence Research Society Conference (2010).
- 17 Paolacci, G., Chandler, J., and Ipeirotis, P. Running experiments on amazon mechanical turk. Judgment and Decision Making (2010).
- 18 Riedl, M., and Stern, A. Believable agents and intelligent story adaptation for interactive storytelling. In Technologies for Interactive Digital Storytelling and Entertainment, Springer (2006), 1–12.
- 19 Riedl, M., and Young, R. Narrative planning: Balancing plot and character. Journal of Artificial Intelligence Research (JAIR) 39 (2010), 217–268.
- 20 Susi, T., Johannesson, M., and Backlund, P. Serious games: An overview.
- 21 Turner, S. MINSTREL: a computer model of creativity and storytelling.
- 22 Ware, S., and Young, R. CPOCL: A Narrative Planner Supporting Conflict. In Seventh AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment, AIIDE (2011).
- 23 Weyhrauch, P., and Bates, J. Guiding interactive drama. Carnegie Mellon University, 1997.
- 24 Zook, A., Lee-Urban, S., Riedl, M. O., Holden, H. K., Sottilare, R. A., and Brawner, K. W. Automated scenario generation: Toward tailored and optimized military training in virtual environments. In International Conf on the Foundations of Digital Games, ACM (2012).