Within the field of Computational Creativity, significant effort has been devoted towards identifying variegating aspects of the creative process and constructing appropriate metrics for determining the degree that an artefact exhibits creativity with respect to these aspects. However, the formalization of a person's creativity (i.e. a creativity user profile) as a derivative of such creations is not straightforward, as it requires a transition to a space reflecting the core principles of creativity as perceived by humans. This becomes a necessity in domains where personalization goes beyond timely and personalized knowledge provision, targeting the encouragement and fostering of creative thinking. Thus, it becomes essential to develop methodologies for modelling creativity to support personalization based on creativity aspects / characteristics of users. The paper proposes a user modelling framework for formulating creativity user profiles based on an individual's creations, by transitioning from traditional computational creativity metrics to a space that adheres to the principal components of human creativity. Furthermore, the paper presents the Creativity Profiling Server (CPS), a system implementing the aforementioned user modelling framework for computing and maintaining creativity profiles and showcases the results of experiments over storytelling educational activities.
Human creativity is a multifaceted, vague concept, combining undisclosed or
paradoxical characteristics. As a general notion, creativity adheres to the ability to move
beyond traditional and established patterns and associations, by transforming them to
new ideas and concepts or using them in innovative, unprecedented contexts and
settings [
creativity expressed by given creativity artifacts; (b) associating these measurements with dimensions of human creativity corresponding to the given dimension.
More specifically, Novelty reflects the deviation from existing knowledge/
experience and can be measured as a difference metric between what is already known and
the given piece of content. Novelty is a generally accepted dimension of creativity
within the area of computational creativity and an essential candidate for measuring
elements of creativity within the human-created content when interacting with the
machine. It has been used as a heuristic for driving the generation of novel artefacts in
exploratory creativity [
In our previous work [
proposed metrics with the human perception of creativity. Afterwards, we build on these observations to propose a transition model from computational metrics to a twodimensional orthogonal space which aims to closely reflect the way human beings perceive creativity. We present the experiments for assessing the effectiveness of the proposed model towards this goal, describe the architecture and functionality of the
2
Correlation of Computational Creativity Metrics With the Human Perception of Creativity In order to assess the adherence of the proposed metric formulization with the human perception for creativity, we organized and conducted an experimental session based on storytelling activities. For the execution of the experiment, we employed forty (40) human participants, split in ten (10) teams of four (4) members each. All teams were asked to construct a story, on a specified premise, the survival of a village’s habitants under a ravaging snow storm. The stories were created incrementally, with twenty (20) fragments produced for each story.
Following the completion of the stories, the teams were organized in two groups, each consisting of five teams. Without any interaction between the groups, each team was called to rate the stories of the remaining four teams belonging to their group, using a rank-based 4-star scale (i.e. the best story received 4 stars, the second-best story received 3 stars etc.). In this way, we obtained a ranked list of the five stories in each group. The goal of our experiment was to determine if, using the ranked lists of one of the test groups and a formalized representation of the computational creativity metrics, we can identify their correlation and examine if the distribution of values for the metrics follow the pattern of human judgment. To this end, we define a constrained optimization problem over functions of the aforementioned metrics, which is described below. 2.1
Extracting a Model for the Human Perception of Creativity
Each artefact (story) is characterized (via the application of the computational
creativity metrics presented in the previous section) [
∗ 1− ∗ +
2
, where ∈ [
− where P is a strict preference relation and I is an indifference relation, as perceived by humans when evaluating the creativity of these artefacts.
differences = 1, 2, … , −1 , where q is the number of artefacts in the reference +1 ≥ 0 is the creativity difference between two subsequent
= − +1 + ≥ 0 We can then solve the following constrained optimization problem: −1 parameters for each property .Based on these values and the human assessment of the story rankings, the results of the constrained optimization problem defined in the previous section resolves in the calculation of the partial creativity parameters (a, b, c, d and w). Regarding the impact of the various metrics in the computation of the overall creativity, we observed that Novelty is generally considered a particularly positive attribute creativity-wise for the stories, its partial creativity (PC) increasing as its value increases (see Figure 1). In contrast, the remaining metrics reached their maximum partial creativity at a certain value, after which their partial creativity started to decrease, indicating that e.g. recreational effort greater than a certain point is not perceived as a direct indication of creativity (see Figure 1).
metrics are correlated with the perception of humans for creativity, this correlation is not direct for all metrics. The following section discusses on the implications of these observations and details our approach for using the proposed metrics towards building a dimensional plane that more accurately reflects the human perspective for creativity. 3
Transferring Computational Creativity Metrics to the
Human Perspective As stated, each textual artefact can be described by 4 computational creativity metrics, namely, Novelty, Surprise, Rarity and Recreational Effort. Following the formulation of the creativity metrics, therefore, the next hypothesis that was examined was the reduction of the dimensional space for representing creativity as expressed through creative artefacts, in an orthogonal space. In order to effectively conceptualize human creativity, orthogonality is a particularly desirable attribute of the conceptualization space to be used, since it allows the examination of independent variables when trying to analyse and influence / encourage certain creativity aspects. Hence, the first step towards identifying the adherence of the computational creativity metrics with the human perspective is to examine the orthogonality of the proposed metrics formulation. To this end, we ran an experiment for calculating the four basic computational creativity metrics on two datasets derived from distinct and distant domains, and determined whether the four metrics are orthogonal.
The first dataset comprised transcriptions of European Parliament Proceedings
[
In total, we examined 50 distinct parliament sessions from the Europarl dataset and 40 chapters from the storybook. Based on the obtained results, we calculated the correlation between the four computational creativity metrics. Tables 1 and 2 provide the correlation values between the four metrics. It is evident that the computational creativity metrics by themselves are not orthogonal. In order to better approximate the human perception for creativity, we propose the following abstraction for modelling the examined aspects of creativity to a space more closely resembling human thinking:
aspects of creative work: whether the said work is visionary, i.e. it provides a groundbreaking approach on a given field; and whether it is constructive, i.e. it uses in a novel way established techniques and ideas in order to produce a high-quality artefact.
orthogonality, we consider Novelty as the strictly defined dimension of our space and seek for the formulation of Atypicality that results to a dimension orthogonal to Nov
Surprise, Rarity, and Recreational Effort: ( ) =
, , ∈ [
( )+
+ +
( ) , with
∑ =1(
( ,
) )2, s.t. , , ∈ [
Although the search space of the optimization problem above is highly non-linear solving this problem is straightforward. The resulting model defines two orthogonal axes, Novelty and Atypicality, which define the space for measuring and characterizing the observed creativity, as an Euclidean vector, the length of which indicates the quantitative aspect of the creativity exhibited by the artefact, while its direction indicates the tendency for either Novelty (visionary creativity) or Atypicality (constructive creativity). The following tables present the novelty and atypicality in the two datasets, as well as, the correlation between these two dimensions for the found optimum weight values.
creativity profiles of users using creativity exhibits that are produced from applications of the outside world. CPS provides a simple and straightforward API in order to expose its functionalities and to facilitate the communication with the outside world.
Computational Creativity Metrics
Calculator Creativity Exhibits
Analysis User Profile Update
Submit CreativityExhibit Computational Creativity Measurements User Validation Creativity Exhibit Model Controller Creativity Evidence &
Measurements Creativity Exhibits User Manager Application Validation
Client Application
Validator Machine Learning
Components
send the request to the User Management. Afterward User Management verifies through the Application Validation module that the application is registered to CPS. Since the application is validated User Management creates a unique user profile id and sends it to the application. Since a user profile is created then the application can submit creativity exhibits of this user. More specifically the application submits the creativity exhibit to the CPS API along with type of the creativity exhibit and the timestamp the creativity exhibit was created. After submission the API sends the creativity exhibit and its type to the Creativity Exhibit Model Controller module. After validating the user and the application through the User Management and the Application Validator respectively, the module sends the creativity exhibit to the Computational Creativity Metrics Calculator module. The Computational Creativity Metrics
the measurements to the CPS database. Finally, the Creativity Exhibit Model Controller invokes the Machine Learning Components. Machine Learning Components take as input the creativity exhibit and calculate the values of the profile properties of the user. Afterwards the newly calculated values are send to the Creativity User Modelling Controller module, which stores the values to the CPS database.
the CPS API the User Profile Properties and also the Model which describes the profile. After sending the request to the API, the request is redirected to the Creativity
5
and Recreational Effort metrics in order to form another metric, which we call Atypicality and is orthogonal to Novelty. Atypicality is calculated as a weighted average of Surprise, Rarity and Recreational Effort, as follows: = ∗ + ∗ + ∗ , where: + + i refers to an artifact, Si , Ri ,Ei and ATi to Surprise, Rarity, Recreational Effort and
tive weights assigned to Surprise, Rarity and Effort respectively, in order to calculate the Atypicality metric in a way as much uncorrelated (and thus, orthogonal) with
vector expressing two types of user’s creativity. The Visionary Creativity, which is measured by the Novelty metric, and the Constructive Creativity, which is measured by the Atypicality metric. CPS gathers all Creativity Exhibits (artefacts) that are produced by its users within external applications. In discrete time intervals, which we call Time Window, CPS calculates and/or updates the Creativity Profile of each user.
Calculation of optimum Computational Creativity Metric Weights for the Application Domain
. ( ,
)2, . , , ≥ 0 , + + ≠ 0
Each time where a new CPS Time Window starts, we solve the above minimization problem for all the artefacts of the application domain (all the creativity exhibits collected for all CPS users and for all CPS Time Windows so far). It is evident that in each execution of this process there is a strong probability of discovering a new vector [wS,wR,wE] that makes Atypicality (AT) more orthogonal to Novelty (N). In order to reduce the sensitivity of the system to this continuous change, we update the vector [wS,wR,wE] to be used in Phase B with the new vector retrieved by solving the optimization problem defined in Eq. 1 only when the improvement (minimization) in Correl(N, AT)2 exceeds 5%.
Phase B: Construct/update of Users’ Creativity Profiles A user’s creativity profile is determined by the creativity exhibits produced by the user alone or as a member of a group. Groups are treated by CPS as a user, meaning that CPS will construct a creativity profile also for each group. In this case, the creativity profile is constructed/ updated based on the creativity exhibits of the group during the last (just finished) time window. In the case of simple users (not groups) their creativity profile is constructed/ updated based on all the creativity exhibits they constructed (either alone or as a group member). The first step for computing the creativity profiles is to transform the space (N,S,E,R) to the space (N, AT) and compute the average of N and AT measures for the creativity exhibits for a given user and for the time window that just finished, as follows: B1. Calculate Average Novelty and Atypicality of Creativity Exhibits
≡ [
Such , a user cannot be part of other groups.
Let ] of a user U, calculated for the creativity exhibits in the time window T, after the transformation of the space (N,S,E,R) to the space (N, AT) using the optimal weight vector [wS,wR,wE] (calculated in Phase A). Let also ≡ [ ,
] of a user U, calculated for the creativity exhibits of UG in the time window T, after the transformation of the space (N,S,E,R) to the space (N,
U is calculated as a fusion of ET and GT, relying on the analogy of the user’s and the groups’ achievements. If the user’s creativity (ET) surpasses the creativity exhibited within his/her participation in groups (GT), then only ET is considered. Otherwise, a part of the difference between groups’ creativity and user’s creativity is also considered, as follows:
= + ∗ ( − ) < ≥ , with = 1
Though all exhibits must be taken into account, the recent ones are considered more important, as they depict the exact current status of the user’s creativity whereas past exhibits play a less vital role. To give our model an essence of decay through +
−1 time, we use this formula: =
∗ −1, where: is the vector describing the Creativity of the user (or group) at the time window T, and −1 at the time window T-1 respectively
≡ [ D, a proportional constant of decaying analogous to the timespan.
,
proach, we conducted a two-phase experiment in order to determine (a) the degree to which the selected computational creativity metrics conform to the opinion of experts regarding the creativity exhibited in a textual artefact and (b) the degree to which the proposed model for human creativity reflects the opinion of such experts.
to produce five stories each under pre-defined topics. For the first stage of the experiment, we sampled the stories produced during the aforementioned story writing session, randomly selecting two stories by each student, and asked five experts to rank them with respect to their creativity, as the latter is perceived by each of these experts.
duced by the CPS. For the second stage of the experiment, we picked the complete set of stories (i.e. five stories) for five of the users and asked from the same five experts to rank these users with respect to their creativity, using as evidence the produced stories. We then compared the expert ranking to the one produced by the CPS.
In order to evaluate the similarity between the rankings of the experts and the rankings of the CPS, for the textual exhibits’ and the users’ ranks, we employed a metric based on Kendall’s Tau, defined by the following equation: = 1 2 + −
( −1) exhibits or users,
stands for the concordant pairs of ranked stands for the discordant pairs when comparing the ordering of the experts and the CPS and is the number of the examined exhibits or the users. We calculated this metric for the series of textual exhibits rankings and the series of participating users rankings. The following table presents the summary statistics of the two Success metric series we had as an outcome.
from computational creativity metrics associating specific attributes of text artefacts with creativity aspects to a creativity calculation model that better reflects the human perception of creativity. Furthermore, the present manuscript provides a summary of the architectural design and functionality of the Creativity Profiling Server (CPS).
Towards the continuation of our research, we aim to examine the effectiveness of the model in more complex experiments, examining textual exhibits from different domains and modalities (prose, poetry, speech) in order to obtain a more general reflection of the human perception of creativity. Observation over more open-ended experiments will likely lead to further refinements and extensions of the proposed human creativity model. 8 14. Lehrer, A.:Brendalicious. Lexical creativity, texts and contexts. Amsterdam, John
Benjamin, , pp. 115-136, (2007). 15. Chiru, C. G.:Creativity Detection in Texts. In ICIW 2013, The Eighth International