When Reflective Feedback Triggers Goal Revision: a Computational Model for Literary Creativity∗ Pablo Gervás Carlos León Instituto de Tecnologı́a del Conocimiento Facultad de Informática Universidad Complutense de Madrid Universidad Complutense de Madrid 28040 Madrid, Spain 28040 Madrid, Spain pgervas@sip.ucm.es cleon@fdi.ucm.es Abstract The present paper focuses on literary creativity, and pro- poses a computational model for the creative process in this Existing models of the writing task from a cog- domain based on a number of cognitive model of the task of nitive viewpoint agree on the importance of draft writing as carried out by humans. revision in the overall process. This is generally assumed to focus on reviewing intermediate drafts in search for feedback on how to modify them to 2 Previous Work match the driving constraints. However, in literary The present paper puts forward a proposal that captures in creativity it is often the case that the feedback leads computational terms the operations described in two existing not to a revision of the current draft but to a redef- cognitive models of the writing tasks. This section reviews inition of the constraints that are driving the pro- these two models and two competing computational models cess. This phenomenon is explicitly described in also based them. Sharples’ model of writing as a creative task. Yet existing computational models of literary creativity 2.1 Cognitive Models of the Writing Task do not contemplate it. The present paper describes Flower and Hayes [Flower and Hayes, 1981] define a cog- a computational model of the creative processes in nitive model of writing in terms of three basic process: plan- literary creativity that contemplates the explicit rep- ning, translating these ideas into text, and reviewing the result resentation of the constraints driving the process, with a view to improving it. These three processes are said to and allows for the feedback from the validation to operate interactively, guided by a monitor that activates one modify not just the ongoing draft but also the con- or the other as needed. The planning process involves gener- straints that it is expected to satisfy. This allows the ating ideas, but also setting goals that can later be taken into model to represent cases of serendipitous discovery account by all the other processes. The translating process of interesting features. involves putting ideas into words, and implies dealing with the restrictions and resources presented by the language to be employed. The reviewing process involves evaluating the text 1 Introduction produced so far and revising it in accordance to the result of Creative processes as carried out by humans are known to the evaluation. Flower and Hayes’ model is oriented towards involve a significant amount of trial and error. Writers, musi- models of communicative composition (such as writing es- cians, painters, poets... rely on a succession of drafts that get says or functional texts), and it has little to say about literary polished over many iterations, each one involving feedback creativity in particular. Nevertheless, a computational model from the previous version, and resulting from a process of of literary creativity would be better if it can be understood revision or regeneration of it. Yet computational models de- in terms compatible with this cognitive model. An important veloped in AI over the years to emulate these same processes feature to be considered is that the complete model is framed very rarely capture this type of dynamic operation. Some- by what Flower and Hayes consider “the rhetorical problem”, times they do, in a limited fashion, when an AI program in- constituted by the rhetorical situation, the audience and the cludes an evaluation function that defines the desired form writer’s goals. for outcomes, and this is run over the results of a generative Sharples [Sharples, 1996] presents a description of writing process that produces candidate artefacts of the desired type. understood as a problem-solving process where the writer is However, the dynamics of the creative process in humans are both a creative thinker and a designer of text. For Sharples, known to be more complex than this, possibly differing sig- the universe of concepts to be explored in the domain of writ- nificantly across different domains. ing could be established in a generative way by exhaustively ∗ This paper has been partially supported by the project WHIM applying the rules of grammar that define the set of well- 611560 funded by the European Commission, Framework Program formed sentences. The conceptual space on which a writer 7, the ICT theme, and the Future Emerging Technologies FET pro- operates is a subset of this universe identified by a set of con- gram. straints which define what is appropriate to the task at hand. 32 Sharples explains that the use of a conceptual space “eases a mental framework that underlies and supports their writing the burden of writing by limiting the scope of search through efforts. For beginners, the problem must be addressed with long term memory to those concepts and schemas that are ap- the aid of general knowledge about how to design artefacts, propriate to the task” [Sharples, 1996, p. 3]. To Sharples, how to transform mental structures and how to solve prob- the imposition of these constraints enables creativity in the lems. Because this is difficult to do in the head, some writers sense that he identifies creativity in writing (in contrast with resort to capturing the ideas involved in paper, as sketches, simple novelty) with the application of processes that ma- lists, plans, notes etc. These external representations stand nipulate these constraints, thereby exploring and transform- for mental structures, and they are easier to manipulate. The ing the conceptual space that they define. Sharples provides writer can then explore different ways of structuring the con- a specification of what he envisages these constraints to be. tent, apply systematic transformations, establish priorities, Constraints on the writing task are described as “a combi- and reorder or cluster items. The task of writing addressed nation of the given task, external resources, and the writer’s in these terms is much closer to recognised design tasks. knowledge and experience” [Sharples, 1996, p. 1]. He also The arguments outlined above with respect to how Sharples mentions they can be external (essay topic, previously writ- models the differences between beginners and experts sug- ten material, a set of publishers guidelines. . . ) or internal gests further consideration of the role of the evolution of rep- (schemas, inter-related concepts, genres, and knowledge of resentation in the progressive acquisition of expertise. In this language that form the writer’s conceptual spaces). respect, Karmiloff-Smith [Karmiloff-Smith, 1995] proposes Sharples also provides a description of how the typical a model of evolving representation called Representational writer alternates between the simple task of exploring the Redescription model. conceptual space defined by a given set of constraints and the This model analyses the development of behavioural mas- more complex task of modifying such constraints to trans- tery in a given domain – meaning consistenly successful per- form the conceptual space. Sharples proposes a cyclic pro- formance in the domain – in terms of how knowledge about cess moving through two different phases: engagement and the domain is represented internally by the individual. The reflection. During the engagement phase the constraints are model considers three phases of learning. During the first taken as given and the conceptual space defined by them is phase the individual focuses on his interaction with the envi- simply explored, progressively generating new material. Dur- ronment, and represents these in the form of raw data received ing the reflection phase, the generated material is revised and from outside. This may lead to an initial achievement of be- constraints may be transformed as a result of this revision. havioural mastery. Over the second phase, internal represen- Sharples also provides a model of how the reflection phase tations are abstracted from the raw data, and processing may may be analysed in terms of specific operations on the var- start to focus on them. As a result of this introspection, fea- ious elements. A three step process of reviewing, contem- tures of the environment may temporarily be disregarded and, plating and planning the result is suggested as a description as a result, observed behaviour may deteriorate. However, of the reflection phase. During reviewing the result is read, this leads to a recuperation of a more flexible achievement of minor edits may be carried out, but most important it is in- behavioural mastery, by then based on having achieved rec- terpreted to represent “the procedures enacted during com- onciliation between internal representation and external data. position as explicit knowledge which can then be integrated This model describes four different levels of cognitive rep- with an existing conceptual space”. Contemplation involves resentation: implicit, focused on the process itself; explicit the process of operating on the results of this interpretation, level one in which basic aggregation of raw data present in which are likely to be explicit representations of specific con- the implicit level is performed in terms of data storage but straints. Planning uses the results of contemplation to create may not yet be accessible to the cognitive system for manipu- plans or intentions to guide the next phase of engagement. lation operations; explicit level two, in which structures from Sharples also provides an account of how the explicit rep- the first explicit level are converted into schemas and thereby resentation of constraints as elements susceptible of modifi- become available; and explicit level three, a final and “cross- cation is fundamental to achieve this type of cyclic opera- system” representation of concepts that can be verbalized and tion. People produce grammatically correct linguistic utter- are fully integrated in a more general cognitive system. ances without being aware of the rules of grammar, but to explore and transform conceptual spaces one needs to call 2.2 Existing Implementations of Sharples’ Model up constraints and schemas as explicit entities, and work on MEXICA [Pérez y Pérez, 1999] was a computer model de- them in a deliberate fashion. For the mind to be able to ma- signed to study the creative process in writing in terms of the nipulate the constraints, they have to be subjected to a process cycle of engagement and reflection [Sharples, 1999]. It was of “representational redescription” [Karmiloff-Smith, 1995], designed to generate short stories about the MEXICAS (also re-representing knowledge that was previously embedded in wrongly known as Aztecs). MEXICA is a flexible tool where effective procedures as elements susceptible of manipulation. the user can set the value of different parameters to constrain The problem is that beginners addressing such a cognitive the writing process and explore different aspects of story gen- task do not have a vocabulary to describe mental processes eration. It takes into account emotional links and tensions to themselves. To learn, they must develop “a coherent men- between the characters as means for driving and evaluating tal framework of plans, operators, genres and text types that ongoing stories. can guide the process of knowledge integration and transfor- MEXICA relies on certain structures to represent its mation” [Sharples, 1996, p. 5]. Experts tend to have such knowledge: a set of story actions (defined in terms of pre- 33 conditions and post-conditions) and a set of previous stories initial attempt to construct a theoretical model based on an (stated in terms of story actions). MEXICA stands out from abstract analysis of the task of story construction in the con- other systems in that it actually builds its own set of schemas text of a basic communication situation. Figure 1 shows a from the set of previous stories. A single type of knowledge graphical representation of the ICTIVS model. The commu- structure, known as a Story-World Context (SWC), is used to nication takes place as an exchange of a linear sequence of represent these schemas. Story-World Contexts represent in- text that encodes a complex set of data that correspond to a stances of contexts (described in terms of emotional links and set of events that take place over a volume of space time, pos- tensions between existing characters) in which an action has sibly in simultaneous manner at more than one location. To appeared in a previous story, and they act like rules during convey this complexity as a linear sequence and recover it the engagement phase: an action is added to the plot if a again at the other end of the communication process requires Story-World Context for that action can be found that matches a process of condensing it first into a message and then ex- the plot so far.1 The reflection phase revises the plot so far, panding it again into a representation as close as possible to mainly checking it for coherence, novelty and interest. The the original. There is a composer, in charge of composing a checks for novelty and interest involve comparing the plot so linear discourse from a conceptual source that may also have far with that of previous stories. If the story is too similar been produced by himself, and an interpreter, faced with the to some previous one, or if its measure of interest compares task of reconstructing a selected subset of the material in the badly to previous stories, the system takes action by setting a conceptual source as an interpretation of the received narra- guideline to be obeyed during engagement. These guidelines tive discourse. In real life, the role of the composer is usually can be considered as a basic implementation of Sharples’ con- played by a writer and the role of interpreter by a reader, but straints, driving which types of action can be chosen from the in the present case a more generic formulation has been pre- set of possible candidates. ferred for generality. In MEXICA, the system is actually aware of the emotions This overarching act of communication is fundamental be- of all the characters (and the emotional tensions between cause it allows the definition of the purpose of the task in them) and uses these to drive and structure the story. But terms of the expected impact of the constructed story on the these emotions and tensions are often not mentioned in the interpreter. Whatever is produced by the composer will have final text of the story. to be processed by the interpreter, and the impact on the in- terpreter cannot in truth be considered without taking into 2.3 The ICTIVS Model account what this process of interpretation involves. With this premise in mind, the ICTIVS model [Gervás and León, 2014] started from a linear description of the complete act of communication from an original purpose in the mind of the composer to a final impression in the mind of the interpreter. This act of communication involves processes of invention of a message and composition of an appropriate form meant to be carried out by the composer. It also involves processes of interpretation and validation carried out by the interpreter. From the point of view of the communicative act, the mea- sure of success of such an act of needs to be established in terms of whether the interpretation by the interpreter matches the message constructed by the composer – success in terms of information transfer – and whether the impression in the mind of the interpreter matches the original purpose of the composer – success in terms of expected impact. As a first approximation, the impression in the mind of the interpreter could be correlated to the results of the validation applied to the message. In order to capture this intuition, the ICTIVS model defines the task of story composition as an iterative cycle of revisions in which the composer progressively gen- erates drafts of his message, and then applies to them an inter- nal process of interpretation and validation intended to match the one that the interpreter will be applying. At each iteration, Figure 1: The original ICTIVS model. This version of the the results of this estimated interpretation/validation are com- model does not take feedback into account. pared with the original purpose. If mismatches are detected, another cycle is started, and only when a successful match has been found does the resulting version of the message get The ICTIVS model [Gervás and León, 2014] arose as an communicated to the intended audience. 1 It is important to note that Story-World Contexts (and not the Five specific stages are included in the model: INVEN- definitions of action in terms of their pre-conditions) are used to find TION – coming up with content for the narrative, possibly the next action to extend the plot. starting from scratch but often from some specification of 34 purpose; a composer task –, COMPOSITION – establishing a successful in terms of how it matches the original purpose, a form to express the desired content; a composer task –, IN- truly creative process may consider not only revision of the TERPRETATION – given a story, fill in the gaps, connect draft but also revision of the purpose. This may arise when- the dots, make assumptions on posible background implied, ever the estimated impact of a given draft on the interpreter is and extend it into a full picture of what the author wants you considered valuable by the composer beyond his original pur- to “see in your mind”; an interpreter task that the composer pose. By means of this extension, the model can capture the needs to model to generate informative feedback for the con- role of serendipity in the creative process [Pease et al., 2013; struction process –, VALIDATION – identify the impact that Corneli et al., 2014]. the story, and/or the material interpreted from it, has on the Third, although computational models of the creative task interpreter; as above, an interpreter task but one that the com- are traditionally formulated as a cycle, in an ideal creative poser needs to model to provide feedback –, and TRANSMIS- process cross-fertilization across the type of stages defined SION – passing over the result of the other processes to an au- would be very positive. This is evident in Flower and Hayes dience; this stage establishes the link between the composer description of the process as a set of transitions between three and the interpreter. Of these five stages, the first four may processes governed by an overall monitoring process that al- take place in an iterative cycle, and the final stage occurs only locates effort to each one of them, and in Sharples’ phrasing once after the iterations have lead to a successful draft with of his model as a dual cycle between two stages that oper- potential for achieving the expected impact on the interpreter ate on different data – the text and the constraints. A similar according to the composer’s purpose. abstraction will need to be considered in our model. The present section analyses these important concepts in 3 Reflective Feedback and Goal Revision in more detail and attempts an initial formulation of such a re- finement to take them into account in a manner that better Computational Models of Literary reflects the intuitions arising from the cognitive models. Creativity The computational models reviewed in section 2.2 and 2.3 3.1 Analysing the Tasks Involved in Creative capture the essence of the cognitive models described in sec- Production tion 2.1, but they both fail to capture the particular features In order to identify the core features that the desired model that concern feedback and goal revision. The model of en- needs, the tasks involved in generation must be examined ex- gagement and reflection in MEXICA has very limited ex- plicitly and compared to what the models can currently repre- plicit representation of the constraints driving the process, sent. Following this, we proposed a categorization of gener- in the form of guidelines set during reflection. The ICTIVS ative system according to their capabilities in terms of feed- model as originally described was formulated at a more ab- back. stract level, but focuses more on the constructive approach to Regarding their internal process, four types of systems can the creative process, with no explicit modelling of the task be identified: of revising an already existing draft. It did include the repre- • those that take no input and generate outputs determined sentation of a seed idea or meaning that the composer wants exclusively by decisions taken during the construction to convey, but no representation of the possibility of this idea of the system (mere generation) being modified as part of the creative process. A refinement • those that take as input some kind of specification that on these models is required that can integrate a specification determines in some way the type of output that is to be of the purpose for the generation task as an input, that can obtained (specification) allow for revision of this specification as part of the process, and that at the same time can take advantage of the existing • those that include a module that quantifies in some way body of work on narrative generation. the degree to which the outputs obtained satisfy the re- Three relevant insights arise from the consideration of the quirements specified as input (diagnostic) original ICTIVS model in this enriched context of purpose- • those that can benefit from the results of a diagnostic driven communication. module to modify the specification and self correct their First, there will probably be a significant difference in com- output (reflective) putational terms between the initial iteration, where at each Taking into account the kind of input that the systems ac- stage new material is generated from the corresponding in- cept, a parallel axis of classification may be whether the sys- put, and subsequent iterations, where two different processes tem can generate outputs only by constructing them from may need to be employed: further generation of new mate- scratch (construction) or by applying transformations to an rial from the specification, and revision of the material gen- initial version of the desired artefacts (revision). erated in previous iterations - where the revision needs to be When these two axes are combined with the issues de- informed by the initial specification, the earlier drafts, and the scribed previously, the following set of possible modes of op- identified mismatches. This is important because the compu- eration arise: tational mechanisms involved in each case may be different, and also because outputs from these two different processes • mere construction: the system generates outputs of a may need to be combined into an integrated output for the given form as determined by its construction corresponding stage. • construction to a specification: the system generates out- Second, at the point of deciding whether a given draft is puts conforming to a given specification 35 • construction with diagnostic: the system generates out- 3.3 Integrating the Reviewed Tasks and Features puts and can provide some quantification of their quality into a Computational Model • reflective construction: the system generates outputs After having analysed both the tasks and the features involved conforming to a given specification and can provide in the creative process from a computational perspective, we some quantification of their degree of satisfaction, and propose the following three extensions for the refined model modify it accordingly. of the computations involved in literary creativity: • mere revision: the system receives an instance of the • to consider the explicit representation of constraints as desired artefact and revises it towards a given goal de- part of the draft itself, so that they can be subject to the termined by its construction same operations as the rest of the draft • revision to a specification: the system receives an in- • to consider a range of operations that includes both con- stance of the desired artefact and a given specification struction and revision and revises the instance of the desired artefact towards • to consider the possibility of focusing system operation the given specification on particular subsets of the draft • revision with diagnostic: the system receives an instance The representation on which the creative process operates of the desired artefact, revises it towards a given goal would therefore need to include at least two different parts: determined by its construction, and can provide some • the set of constraints to be used to drive the construction quantification of the quality of the revised artefact process and/or to validate any resulting drafts, known as • reflective revision: the system receives an instance of the brief the desired artefact and a given specification, revises the • the actual draft at each point of the creative process instance of the desired artefact towards the given spec- ification, and can provide some quantification of their Both the brief and the draft should be represented in such degree of satisfaction of the specification, and modify it a way that different parts of them may be operated upon in accordingly. isolation of the rest. This representation that includes both a brief and a draft 3.2 Summarising the Features of a Creative will be referred to henceforth as the work in progress. Any Process from a Computational Point of View references to operations upon the work in progress can refer to both operations on the draft or on its specification. The cognitive models reviewed in section 2.1 show a number The set of operations to consider would be: of distinctive features that are relevant for the purpose of the present paper: • reject: eliminate from the work in progress a particular item for the next cycle 1. the creative process is iterative in nature • generate: generate anew a particular item of the work in 2. the creative process is driven by a set of constraints progress during the next cycle that restrict the desired outputs in some way; these con- • revise: modify a particular item of the work in progress straints may be considered an input to the process during the next cycle 3. a cycle may involve processes of construction and/or • keep: leave a particular item of the work in progress as processes of revision of prior results it resulted from the previous cycle 4. at the end of each cycle a diagnostic procedure is applied Based on this terminology, a computational model for the to the result obtained so far tasks involved in literary creativity can now be rephrased at 5. part of the diagnostic may involve quantifying degree of a lower level of detail. The same set of general steps can be satisfaction of the given constraints seen, but each one of them now operates over a representation of the work in progress that includes both a brief and a draft, 6. subsequent cycles take into account the diagnostic to at- and at each stage the four types of operation (reject, generate, tempt to improve the results of subsequent cycles revise or keep) may be applied to any subset of the work in 7. consideration of the diagnostic may take the form of progress. planning further operations either on the artefact so far The computational model that we propose may now start or on the set of constraints from a hybrid representation of work in progress. Input may be provided to a creative system either in terms of a brief - a 8. the process as a whole includes a stage of meta-level rea- set of constraints that the output should satisfy – or a partial soning which decides among the various available oper- sample of the desired artefact, or as a combination of both ational options applicable to the task at hand, such as, modes. for instance, whether to iterate further or to stop, or, This initial representation of the work in progress would for a given iteration, whether to construct or to revise, undergo a process of reflection. In this initial reflection pro- whether to act upon the artefact itself or upon the set of cess, each of the sections of the work in progress is consid- constraints, or whether to apply the chosen operation to ered. If only a brief is available, the brief is marked as to be the whole element or to specific parts of it retained for the following construction cycle, and the empty 36 draft is tagged to be generated. If a brief and a partial draft are as compared to the classic version that does not address feed- available, the partial draft is analysed in the light of the brief. back (depicted in Figure 1). The result of this analysis will be a diagnostic. Based on this diagnostic, the available draft is partitioned into sections, and each of these sections is marked as either to be left as it is (keep), to be regenerated (generate), to be revised (revise) or to be rejected (reject). Additionally, if the brief suggests sec- narrative Compose discoursive constraints constraints tions should be added to the partial draft, place holders for them are added to the partial draft tagged as to be generated. If only a draft is available, an interpretation process is run on ideas discourse it to reverse engineer a brief. Based on the resulting brief, Invent Tell the available draft is processed as above. If neither brief nor draft are available, the creative system may follow different procedures, depending on whether a brief or a partial draft is judgement story constructed first. social literary Once the initial reflection phase is over, the system would constraints constraints enter a phase of construction. In spite of the similarities, we ideas* do not refer to this stage as engagement, because engagement in the sense used by Sharples applies very specifically to a Validate Interpret process of production of new material, and the construction envisaged here may cover other processes such as revision, narrative constraints* editing, or omisin. During this phase, each of the sections into which the draft has been partitioned will undergo the op- eration for which it has been tagged. The draft will therefore be edited by the application of the four basic operations de- scribed above. Any sections of the draft that are rejected at Figure 2: Graphical representation of the ICTIVS model as this point are stored in a log of fruitless paths. described in this paper, taking feedback into account. The At the end of the construction phase, the system would en- feedback can trigger modifications of the brief (describing ter another phase of reflection. The first aim of this phase the constraints). The diagram represents draft generation would be to ascertain whether the creative process has been in clockwise direction, and feedback revision in counter- concluded satisfactorily. This would arise if the draft matches clockwise direction. the brief to perfection. If the draft does not match the brief, the system would pro- ceed with the rest of the reflection phase as described above, 4 Discussion and iterated over another reflection/construction cycle. Dur- The planning process that Flower and Hayes has a dual pur- ing reflection cycles other than the first one, the system may pose of generating ideas – which could correspond to addi- also consider modifications to the brief. These may arise from tions to the ongoing draft – and of generating goals for the three possible situations. First, if part of the brief has proven other processes. While the model that Flower and Hayes impossible to satisfy during the prior cycle, the system may propose is not necessarily focused on creative writing, we consider abandoning it. This would be plausible behaviour consider it to be fundamental for describing narrative com- for human creators and should therefore be considered a pos- position from the point of view of feedback. The ability to sibility for artificial models. It would also constitute a very generate additions to the ongoing draft would correspond to useful addition to allow creative system to steer themselves Sharples process of engagement, with the slight refinement out of unproductive regions of a conceptual space when the that the ideas generated in that case are restricted to the con- current brief constrains them to restrict the search so. Second, ceptual space defined by the initial constraints – which would if the reverse engineered brief shows positive features that correspond to our definition of the brief. These processes were not included in the original brief, the system may decide are covered by the generate option of our construction phase. to include them in the brief for the next iteration. This would The ability to set goals for other processes as established dur- allow such systems to incorporate the concept of serendipity ing the planning process that Flower and Hayes describe in into their computational models. Third, if the exploration of their model corresponds to the establishment of constraints the conceptual space during a prior construction phase has as described by Sharples. These processes are covered in our included an excessive number of choices between possible model by the part of the reflection stage where modifications candidate results, the system may decide to extend the brief to the brief are considered – in cycles beyond the initial one. to restrict the search to a subset of the conceptual space in The translating process described by Flower and Hayes is question. Extensions to the brief should be compatible with described as a process of transforming ideas into text. Ac- the rest of it, and may take into account information about cording to our definitions, this would correspond to the task prior attempts that have failed. of generating a new instance of a particular section of the draft Figure 2 depicts the reflective process in the proposed com- according to the corresponding brief, as carried out during a putational model for the tasks involved in literary creativity, construction stage. 37 Our reflection process combines features of the reviewing detail regarding the types of data involved and the types of process as described by Flower and Hayes – in as much as operation carried out on them, but phrased at a higher level it involves evaluating the material produced so far and de- of abstraction with respect to the type of artefact being con- termining which parts of it can stand and which ones need sidered. The ICTIVS model was designed for the specific further operations performed upon them – and the reflection domain of narrative, and because of this it included separate stage of Sharples’ model in the case of cycles beyond the ini- stages for the ideation of plot or fabula and the composition of tial one – where diagnostic leads not to revision of the draft such plots or fabulae into sequential discourses. This would but of the brief, which corresponds closely to Sharples’ con- correspond to having different narrative levels of representa- straints. These constraints may then affect subsequent pro- tion – fabula and discourse – for the material within the draft, cesses of revision but also of further construction or recon- and contemplating a specific process of conversion from one struction of material already in the draft. The detailed de- to the other. When we abstract away from these features scription of the reflection stage as described by Sharples can specific to narrative, we can consider that the stages of IN- be revisited using the proposed new terminology, which al- VENTION and COMPOSITION of the ICTIVS model would lows for finer consideration from a computational point of correspond to te construction phase that we have described view of the actual operations involved at each point. The step in the current model, the INTERPRETATION and VALIDA- that Sharples names reviewing involves a process of represen- TION stage would correspond to the reflection stage, and the tational redescription – following Karmiloff-Smith – which TRANSMISSION stage would correspond to the actual action clearly goes beyond comparing the results obtained with the of publishing or sending the final draft to an audience, which original brief. Sharples specifically describes how the re- would correspond to the fulfilling the stopping condition im- sults obtained are processed to make available for reasoning plicit in our current phrasing of the reflection stage. With within the system knowledge about “the procedures enacted respect to the low level details described above, the task of during composition”. Following Karmiloff-Smith, the raw reverse engineering a brief from a partial draft would match data received from the generation processes are interpreted closely the processing that is considered during the interpre- and internal representations are constructed describing valu- tation phase of the ICTIVS model. The task of comparing able properties of these data at a more abstract level. This such a reverse-engineered brief with the one actually used to may take several forms, but a simple solution is to consider drive the construction process would match the ICTIVS stage attempts to reverse engineer from the resulting draft a hy- of validation. pothetical brief that may have lead to it. This might be a The main advantage of this proposal is that two new reasonable match for the “explicit representations of specific sources of additional constraints are now included in the constraints” that Sharples describes as likely outcomes of the model of the creative process. First, the reverse-engineered process. The task of comparing such a reverse-engineered brief may contain valuable constraints that were not in the brief with the one actually used to drive the construction pro- original brief. This would correspond to the occurrence of cess would match Sharples step of contemplation, which op- serendipity: the constructive process employed leads to valu- erates on the results of the reviewing step. Sharples includes able features that were not in the original brief but which in the reflection stage an additional step of planning, where may be noticed during contemplation/interpretation of the re- the results of contemplation lead to the creation of plans or sult, and from then on added explicitly to the brief for sub- intentions to guide the next phase of engagement. To fully sequent iterations of the process. Second, the model allows capture the subtleties of Sharples analysis we have consid- for an explicit process for the generation of new constraints. ered that the validation stage may result in the application of This allows for the design of systems that can autonomously the four operations we have described – reject, generate, re- search for their own constraints, which would allow for a vise, keep – to the pair of briefs under consideration: the one broader range of creative process. With respect to Boden’s used to generate the draft under revision and the one reverse well known taxonomy of creative system [Boden, 2003], as engineered from the actual draft obtained. From this process the constraints being considered define the conceptual space a new brief will emerge, which can inherit constraints from that is being explored, a system capable of modifying the con- the original brief, or delete them. straints that drive it might be capable of achieving transforma- The reflection stage as we have described, in as much as it tional creativity. In this way the proposed models allows for includes procedures for deciding which operations are to be a more fine grained representation of data and processes that carried out next on which parts of the work in progress, in- may lead to the development of more expressive solutions. tegrates the task of monitoring the creative process to guide The importance of making the system able to reject se- the interaction and alternation between its constituent sub- lected parts of its original brief should not be underestimated. processes – as described in Flower and Hayes model. The The ability of human creators to depart – sometimes in very description provided in the present paper, given that it starts radical ways – from their original intentions in search for from a more fine-grained representation of both the data and new aesthetics experiences has long been considered a crit- the operations under consideration – allows for a more de- ical ingredient of creativity of the highest order. It ties in very tailed and expressive description. This should allow for easier closely with the concept of transformational creativity, and implementation of instances of creative systems that consider the implicit ability to shift into new paradigms rather than this type of behaviour. just explore the old ones. Although all such issues are cur- With respect to the ICTIVS model, the model proposed in rently beyond the state of the art of creative systems, it is this paper may be seen as a refinement at a lower level of important to enable our computational models to represent 38 the types of behaviour that may one lead to implementation [Flower and Hayes, 1981] L. Flower and J.R. Hayes. A cog- of similar behaviours. This would correspond to building ex- nitive process theory of writing. College Composition and plicitly into our computational models the idea of creativity at Communication, 32(4):365–387, 1981. the metalevel [Wiggins, 2006]. The issue of how such opera- [Gervás and León, 2014] P. Gervás and C. León. Reading tions might be profitably controlled is beyond the scope of the and writing as a creative cycle: The need for a computa- present paper and will need to be addressed in further work. tional model. In 5th International Conference on Com- Overall, a large proportion of the success of a creative system putational Creativity, ICCC 2014, Ljubljana, Slovenia, as described in the present proposal will depend on the im- 06/2014 2014. plementation of suitable strategies for the partitioning of the draft into sections requiring the different operations available, [Karmiloff-Smith, 1995] A. Karmiloff-Smith. Beyond Mod- and on the procedures for modifying the brief. These should ularity: A Developmental Perspective on Cognitive Sci- be the focus of further work along the lines described in this ence. A Bradford book. A Bradford Book, 1995. paper. [Pease et al., 2013] Alison Pease, Simon Colton, Ramin Ramezani, John Charnley, and Kate Reed. A discussion 5 Conclusions on serendipity in creative systems. In Proceedings of the Fourth International Conference on Computational Cre- The processes of literary creativity involve a complex web ativity, page 64–71, Sydney, Australia, jun 2013. of interacting procedures (generation from a brief, evaluating how a draft matches a given brief, revision of an intermedi- [Pérez y Pérez, 1999] R. Pérez y Pérez. MEXICA: A Com- ate draft to fit a given brief, identifying unexpected valuable puter Model of Creativity in Writing. PhD thesis, The Uni- features from a working draft, editing a brief to optimise the versity of Sussex, 1999. search for creative results,. . . ) and strategies for navigating [Sharples, 1996] M. Sharples. An account of writing as cre- between them. Existing cognitive models cover this space ative design. In C. M. Levy and S. Ransdell, editors, The of solutions, but tend to remain at a high level of abstrac- Science of Writing: Theories, Methods, Individual Differ- tion that leaves many of the features relevant for computa- ences, and Applications. Lawrence Erlbaum Associates, tion underspecified. The existing computational models of 1996. the writing task that have tried to take the cognitive models [Sharples, 1999] M Sharples. How We Write. Routledge, into account have focused on specific features of the process as their engineering mainstays, without trying to address the 1999. full complexity of the problem as a whole. The present paper [Wiggins, 2006] G. Wiggins. A preliminary framework for proposes a computational model of the writing task that con- description, analysis and comparison of creative systems. siders a broader set of ingredients than had been considered Knowledge-Based Systems, 19(7), 2006. before, represented at a lower level of granularity in terms of their computational nature, both in terms of data and in terms of operations. The resulting model shows a strong potential for capturing significant phenomena in the field of creativ- ity not often modelled computationally in the past, such as revision of drafts, working to a given brief, serendipity, and transformational creativity. A valuable contribution of the proposed model is that it opens for exploration a significant number of lines of research to explore how these various phenomena might be addressed either in terms of working implementations of the proposed computational model or refinements of its basic formulation. Acknowledgments This paper has been partially supported by the project WHIM 611560 funded by the European Commission, Framework Program 7, the ICT theme, and the Future Emerging Tech- nologies FET program. References [Boden, 2003] Margaret Boden. Creative Mind: Myths and Mechanisms. Routledge, New York, NY, 10001, 2003. [Corneli et al., 2014] Joseph Corneli, Alison Pease, Simon Colton, Anna Jordanous, and Christian Guckelsberger. Modelling serendipity in a computational context. CoRR, abs/1411.0440, 2014. 39