V¨
aster˚
as, Sweden
Thesis for the Degree of Master of Science in Engineering - Robotics
30.0 credits
EXPLORING HUMAN-ROBOT
INTERACTION THROUGH
EXPLAINABLE AI POETRY
GENERATION
Philippe Strineholm
pma14001@student.mdh.se
Examiner: Baran C
¸ ¨
ur¨
ukl¨
u
M¨
alardalen University, V¨
aster˚
as, Sweden
Supervisors: Annica Kristoffersson, Rikard Lindell
M¨
alardalen University, V¨
aster˚
as, Sweden
Industrial supervisor: Nina Bozic Yams,
Research Institutes of Sweden, V¨
aster˚
as, Sweden
May 19, 2021
Abstract
As the field of Artificial Intelligence continues to evolve into a tool of societal impact, a need of breaking its initial boundaries as a computer science discipline arises to also include different humanistic fields. The work presented in this thesis revolves around the role that explainable arti-ficial intelligence has in human-robot interaction through the study of poetry generators. To better understand the scope of the project, a poetry generators study presents the steps involved in the development process and the evaluation methods. In the algorithmic development of poetry gen-erators, the shift from traditional disciplines to transdisciplinarity is identified. In collaboration with researchers from the Research Institutes of Sweden, state-of-the-art generators are tested to showcase the power of artificially enhanced artefacts. A development plateau is discovered and with the inclusion of Design Thinking methods potential future human-robot interaction development is identified. A physical prototype capable of verbal interaction on top of a poetry generator is cre-ated with the new feature of changing the corpora to any given audio input. Lastly, the strengths of transdisciplinarity are connected with the open-sourced community in regards to creativity and self-expression, producing an online tool to address future work improvements and introduce non-experts to the steps required to self-build an intelligent robotic companion, thus also encouraging public technological literacy. Explainable AI is shown to help with user involvement in the process of creation, alteration and deployment of AI enhanced applications.
The list below presents the acronyms used throughout the paper along the figures and tables presented in the thesis.
Acronyms
AI Artificial Intelligence.
API Application Program Interface.
CBR Case Base Reasoning. CC Computational Creativity. CFG Context-free Grammar.
DIY Do It Yourself.
EA Evolutionary Algorithms.
HCI Human-Computer Interaction. HRI Human-Robot Interaction.
IDE Integrated Development Environment.
IoT Internet of Things.
MDH M¨alardalen Univerity.
NLP Natural Language Processing. NLTK Natural Language Toolkit.
OS Operating System.
POS Part-of-Speech.
RISE Research Institutes of Sweden. RNN Recurrent Neural Networks. RPi Raspberry Pi.
STT Speech-to-Text.
TTS Text-to-Speech.
List of Figures
1 An overview of the Design Thinking Methods employed and their interdependence. 12
2 Physical prototype main function flowchart. . . 33
3 Wiring diagram depicting the Arduino and the mechanical components. . . 34
4 Poe the poetry box. . . 35
List of Tables
1 Generator complexity categorisation. . . 162 Generator supported language . . . 17
3 User interaction . . . 17
4 Online availability . . . 17
Contents
1. Introduction 1
2. Problem formulation 3
2.1. Research questions . . . 3
2.2. Project background . . . 3
2.3. Constraints and limitations . . . 4
3. Literature review 5 3.1. State-of-the-art in poetry generators . . . 5
3.2. Experimental project framework . . . 6
3.3. Explainable artificial intelligence through open-source transparency . . . 8
4. Method 11 4.1. Research environment . . . 11 4.2. Research phases . . . 12 4.2.1 Discovery . . . 13 4.2.2 Interpretation . . . 13 4.2.3 Ideation . . . 14 4.2.4 Experimentation . . . 14 4.2.5 Evaluation . . . 15
5. Poetry generators study 16 5.1. Poetry generators benchmark . . . 16
5.2. Poetry generators development . . . 18
5.2.1 Algorithmic complexity . . . 18
5.2.2 Evaluation . . . 21
5.3. Empirical data . . . 22
6. Poe the poetry box physical prototype 24 6.1. Open-source pre-study . . . 24 6.2. Coding environment . . . 26 6.3. Code build . . . 29 6.4. Hardware schematics . . . 33 6.5. Implementation results . . . 34 7. Online platform 37 7.1. HRI ideation workshop . . . 37
7.2. Online platform . . . 37
8. Discussion 41 8.1. Theoretical findings . . . 41
8.2. Empirical data . . . 42
8.3. Physical prototype positioning . . . 43
8.4. Online project evolution . . . 43
9. Future work 45
10.Conclusions 46
11.Acknowledgements 47
References 51
Appendix B Generated poetry tests 58
Appendix C Physical prototype poetry generation 59
Appendix D HRI ideation workshop 62
Appendix E Online platform 63
1.
Introduction
The digital age is perceived as a time for creativity and reinvention and Virginia Dignum is a pioneer of responsible and explainable Artificial Intelligence (AI) [1]. An emerging field of developing ethical, transparent and accountable AI systems to be able to offer results that can be understood by a human expert. Dignum claims that the traditional separation between the academic branches is no longer suitable for the needs of the digital age. Future students are predicted to become transdisciplinary1 rather than multidisciplinary2 and possess capabilities outside the engineering
fields to properly assess AI and its implications.
Poetry generation is done with the help of AI enhanced methods to produce meaningful text. A stagnation has been identified in the development of generators only within the field of Computa-tional Creativity (CC) [2]. The latest state-of-the-art generators have incorporated funcComputa-tionalities from other fields to expand on the Human-Computer Interaction (HCI). The most complex gener-ator is idealised to be able to react to its environment [3], thus this thesis explores the possibility of adding verbal interaction in the form of Speech-to-Text (STT) and Text-to-Speech (TTS) open-ing the door to dynamically chainopen-ing the corpora to a topic that is discussed. Furthermore, the implementation of Human-Robot Interaction (HRI) is accompanied by an review of development history and evaluation of poetry generators as well as real-world test of different features to vali-date the functionalities that future poetry generators may present. The thesis aims to understand the influence of explainable AI in practical implementations. For that purpose the functionality of poetry generators is examined comparing algorithimc complexity and functionalities and the steps requiered to build an expalinable application are outlined along the benefits of real-world evaluation.
The research conducted revolves around AI and the role it plays in CC, a multidisciplinary field combining AI, arts, cognitive psychology and philosophy. The sub-field of CC concerning the development of poetry generators is studied both from literature and tested in practice, with regards to generator complexity and human interaction. Given the practical application, a pragmatic perspective is adopted as it offers the practitioner understanding of the situation, developed through the use of exploratory methods based on the Design Thinking methodology. [4]. The following sections represent the findings of the work performed within the thesis.
The premise of the thesis, the research questions as well as the constraints and limitations are presented in Section 2., while Section 3. presents an overview on poetry generators stated in the academic literature, along the experimental project framework and the attempts to achieve explainable AI through the transparency of open-source material. The methods employed in the thesis are covered in Section 4.
The next part focuses on the results of the thesis with Section 5. presenting a benchmark of poetry generators covered by research literature. Different poetry forms implemented, how AI algorithms employed and respective evaluation methods have changed over time is studied and documented in an explainable way. A selection of the two most complex generators covered by literature are tested in an office environment on a transdisciplinary group of researchers.
Section 6. introduces Poe the poetry box, the physical prototype built based on the user expec-tations and the information gathered in the benchmark. A practical implementation is conducted adding verbal HRI as well as the possibility to dynamicaly create new corpora to a poetry generator. The prototype is also documented with detailed explanations on the functions employed by the AI along the steps required to replicate the system to allow for non-experts to understand the process thus creating a framework for explainable AI. A demonstration of the interaction functionalities as well as a few examples of generated poems are also showed to assess the functionality.
A workshop is then conducted to understand real-life expectations upon creative AI systems both on their computational functionalities as well as their environmental interaction capacities. Using the learning of all the parts above an online platform is created to offer an open-sourced step-by-step introduction into the various modules involved in the process of building AI enhanced HRI in Section 7.
1Transdisciplinary is also an adjective that describes, “relating to more than one branch of knowledge.” 2Multidisciplinary is an adjective that describes, “combining or involving several academic disciplines or profes-sional specialisations in an approach to a topic or problem.”
The thesis is then wrapped up with discussions in Section 8., future work in 9., conclusions 10. and acknowledgements 11..For a clearer overview, all results from each stage are also structured in the appendices.
2.
Problem formulation
The main purpose of this thesis is to explore the HRI through creating both a product and a process of building explainable AI poetry generators, by translating the algorithms employed and providing future modularity implementations to non-experts.
Considering the developed field of CC, several poetry generators provide state-of-the-art algo-rithmic approaches in different system complexities. By understanding the evolution of the poetry generation field and the common key features that are present, a transdisciplinary perspective should be identified in AI development continuation. In order to validate the evaluation of the systems in a real-world setting, testing will be performed in an office environment.
The thesis aims next to showcase how a real application of poetry generation can be imple-mented in the form of a physical prototype, designed both based on the explainable AI objectives and limitations. The stagnation identified in the specific field of CC should be overcome by the introduction of sensors and new ways for the generator to interact with its environment.
By studying the desired HRI features, real-life needs can be identified and the user expectations understood on the steps required to achieve said features. The forth aim consists of applying an explainable model upon how AI projects are elaborated, in an open-sourced manner. The steps are intended to be easy and clear enough to be used by people outside the technical domain while at the same time varied enough to also offer value for someone with prior experience, thus aiming to empower ideas in non-experts to be put in practice.
2.1.
Research questions
The following research questions will be addressed:
• RQ1. How does explainable AI influence the development of technological applications? • RQ2. What denotes the complexity of poetry generators, given the algorithms employed and
interaction functionalities?
• RQ3. How can a poetry generation robot be designed based on explainable AI objectives? • RQ4. What are the benefits of real-world evaluation of poetry generators performance,
implemented in a reflective office environment?
2.2.
Project background
This section goes through the project collaboration where the stakeholders are presented and the purpose of implementing computer generated poetry is detailed. Bozic Yams, the industrial supervisor for this thesis has done previous work in the relation between creative methods and the well-being of employees, encouraging reflection and transdisciplinarity [5]. She has been previously experimenting with poetic methods in the V¨aster˚as office of Research Institutes of Sweden (RISE). The work presented in this thesis is part of a larger project where the research interest lies on how poetry can be used as a creative enabler and methodology that could support various aspects at work such as: creativity, new perspectives, reflection, feedback, as well as experiencing meaning, connection and well-being. This thesis takes a look at CC in the form of poetry generators, the methods utilised and the different interaction levels. Both the generated poems as well as the generation process itself are considered in order to identify how artistic practices perform when introduced into work environments. Some of the common interest themes are around mutual human and machine augmentation, identification of AI limitations and possible development areas in order to build AI to augment humans on different cognitive functions.
As Bozic Yams argues traditional simple cognitive tasks have been proven to be successfully delegated to AI though creativity is a more complex human process. The AI enhancement should be understood from the original perspective of mimicking human behaviour in order to conceptualise its capabilities for further human augmentation. The inspirations for the application of this project are drawn from her previous doctoral research in building human-centred organisations, where employees and their engagement are highlighted as drivers for innovation.
2.3.
Constraints and limitations
Unpredictable constraints, a pandemic outbrake3imposed the move to a strictly online environment for the second part of the thesis. All interaction both with the researchers from RISE (practice) and the researchers from M¨alardalen Univerity (MDH) (academia) was performed remotely.
Thanks to its flexibility Design Thinking was chosen as the methodology, offering several phases in conducting a project. The steps presented below are interdependent and the results from a step have influenced the limitations of the following one. A detailed connection between them is illustrated in Section 4. where the methods are elaborated.
For the Discovery and Interpretation phases only poetry generators from the field of CC present in literature and scientific articles are studied.
During the Experimentation phase, only online available generators are used that support the English language and offer distinct user interaction levels. Only the Haiku4 poetry style is tested
and the generators will be used in two predefined test scenarios, to assist and poetically rephrase reflection and meeting summary.
The second part of the Experimentation phase consists of Poe the poetry box, the implemen-tation of a physical prototype that generates poetry using verbal interaction. The program should perform continuous listening with a strict wake-up word and an always-on microphone. For the purpose of the test, Google Speech Recognition will be used for the STT part, as an online al-ternative given its accuracy and high development status. Other engines can be used from the SpeechRecognition[6] module to offer offline functionality and ensure a higher level of privacy. The TTS synthesis is strictly offline.
The following functionalities are desired: an open-sourced cross-platform coding environment, the ability to listen to a given amount of time or until user interaction, transcribe the audio to text and compose different types of poetry and lengths from the audio recording as well as other predefined different corpora.
The hardware limitations are set to open source code and affordable consumer electronics, thus the following three platforms are used:
1. Windows laptop with an Intel i7-7700HQ processor, 2.80GHz CPU and 16 GB of RAM 2. Raspberry Pi (RPi)54 model Cortex-A72 processor, 1.5GHz CPU and 8 GB of RAM
3. Arduino6Uno rev3 with an ATmega328 microcontroller, 16MHz and 2 kB SRAM
While a third part of the Experimentation phase consists of the creation of a future work alternative based on the empirical data collected in the form of an online platform. Properly licensed open-sourced material is showcased that can be implemented on affordable development boards, offering clear step-by-step non-expert level instructions.
3COVID-19 : https://www.who.int/emergencies/diseases/novel-coronavirus-2019/ 4Haiku is a traditional Japanese poem with three lines divided in 17 syllables. 5Series of small single-board computers
3.
Literature review
An overview is performed over the latest poetry generators documented within the academic field. The evaluation methods employed in the field of CC are studied in order to assess the progress in the application of poetry generators. To better understand the position of this thesis, the evolution of knowledge processing from disciplinarity to multi-, inter- and transdisciplinarity in higher academic education is studied. Since the thesis aims to aid in technological literacy through explainable AI, the representation of said knowledge in the public sector is identified in the form of the open-sourced community and the Do It Yourself (DIY) culture.
3.1.
State-of-the-art in poetry generators
To understand the field and identify common practices in poetry generators a few surveys and taxonomies7 are researched. To understand the base-level, one of the more complex poem def-initions is presented by Manurung et al. [7], who state that a poem should simultaneously be syntactically8 well-formed, offer a meaningful message, and have features that set it apart from the non-poetic text. This definition is developed as an argument against the abusing of the poetic license, as Boden[8] identified where the audience is more forgiving when consciously assessing a piece of information as art.
Lamb et al. [9], Oliveira [10, 11] and Ventura [3] taxonomies provide examples of recent poetry generators while Funkhouser’s book on ”Prehistoric digital poetry” [2] provides examples of poetry generators developed from 1959 to 1995.
With the focus of identifying the boundaries of generation categorisation Ventura [3] discusses the methods of mere generation as a term without a clear definition and the relevance of having this identification. A hierarchy is proposed from the least to the most creative ranking generators by their complexity and sophistication. Ventura presents seven approaches to Haiku generation, thanks to its permissive intelligibility over clear structure. The simplest approach is represented by (1) Randomisation of a given set of words, (2) Plagiarising by presenting parts of existing sets of po-etry, (3) Memorising and reproducing parts of popo-etry, (4) Generalisation by autonomously putting together parts of poetry from different sets, (5) Filtration where an additional self-evaluating step is incorporated to ensure a structure over the text, (6) Inception that adds a knowledge base to further constrain the poems and (7) Creation offering the system the ability to perceive its environment thus creating restraints by itself. The theoretical approach to generator creation is represented for each category and compared in terms of novelty, value and intentionality9.
Lamb et al. [9] take the categorisation presented by Ventura further by separating generators between mere, computer- and human-enhancement. Generators from different sources are intro-duced, augmenting the existing scientific fields with works developed by artists and hobbyists. The focus is also shifted from the technical process description to the purposes behind each generator, relying more on poetry and poetry forms instead of statistical CC performance. The taxonomy is written both for researchers within the field of computer science, acting as a brief introduc-tion to methods and goals, as well as a guideline to understanding the differences between trivial and enhanced generation methods. The inclusion of non-scientific-based generators is argued to enhance the potential of generative art as well their artistic relevance. Parallels between computer-generated poetry and music are drawn to strengthen the application of the proposed three-level categorisation in other CC fields as well.
A poetry generator named Co-PoeTryMe [10], presents the feedback that has influenced its development from an automated poetry platform, PoeTryMe, to a limited web interface, TryMe, to a co-creative web tool that enables user interaction with different functionalities. The direction of human-computer enhanced generators is also presented either by the process or the output that can be manipulated by the user, presenting new generators that were not covered by previous taxonomies.
In a previous study Oliveira [11], the developer behind Co-PoeTryMe, organises poetry gen-erators covered by literature with regards to form, features, techniques, reutilisation of material
7A taxonomy is a hierarchical classification where things are organised into groups. 8The way that words and phrases are put together to form sentences in a language. 9Performing a task on purpose
and evaluation. The focus relies on the performance and complexity of the generators as well as the ramifications that each generator presents. The paper also offers a detailed overview of poetic forms and the requirements for each poetry type as well as a comparison to song lyric generators where rhythm arises as a relevant feature. The reutilisation of human-created material is also stud-ied along with its integration in the building of new inspiration sets. The evaluation is perceived both from the obtained result but also by the process of generation itself. Different evaluation methods are presented based on several features, pointing to their respective generators.
Evaluation of computational creativity
The evaluation of poetry generators is a crucial part of this thesis, therefore it is relevant to understand the process of evaluation both within the field of CC and over the creative process itself. Boden [8] separates creativity in three forms, the most frequent being the creation of unfamiliar connections between familiar ideas, which relies on solid previous knowledge present in the interpreter, similarly to the basis upon which AI is built. The other two rely on the exploration and transformation of conceptual spaces.
Lamb et al. [12] evaluate CC dividing it into four categories. Creativity is identified to rely ei-ther on the person or agent that is performing the task, the process that is performed, the product as an artefact or the press, referring to the culture that affects the other categories to be seen as creative. The study is approached from an interdisciplinary perspective combining the CC research community with philosophy, psychology and other fields. The same division is also made by Maher [13] in her attempt to answer the question of “Who’s Being Creative?”. The roles of the computers are described as to support human creativity, to enhance by providing knowledge, or to change perception and generate creative artefacts for the human to interpret, evaluate or integrate. At the same time, the human can assume either the role of the user or developer of the computational model. The interaction has been scaled to take into consideration alongside individual interac-tion, group collaborative platforms and societal computational systems that encourage collective intelligence. The product perspective, explained by Lamb et al. [12] has the goal of producing something useful to humans. It should be noted that what is considered to be an artefact not only limits to classical artworks but also to mathematical theorems, engineering designs and the process of adapting to the environment. On the other hand, process evaluations are more of a qualita-tive analysis of systems capabilities and identification to existing categories. Surveys that contain domain-specific criteria can be used to identify aspects of the system. Perceived as beneficial to perform on non-experts if they are the target group of the system.
Several arguments against evaluating creativity are presented. Most prominent being domain-specific creativity where the question is redirected from how creative a system is, to identifying creative parts and the reason they are perceived as creative. A new measure is desired to evaluate CC in non-human standards. Whatever the evaluation method previously used an issue was repre-sented by expertise. The definition of creativity was accompanied by inconsistent knowledge and disagreeing opinions between experts and non-experts along with scepticism of applying different methods.
A list of general guidelines along with best practice suggestions regarding CC evaluation is presented, directed towards the person, process, product and press. For example, the process part emphasises generative evaluation and process division from inspiration to planning to creation, while as a whole the evaluation should use techniques appropriate to the domain and have a clear conceptualisation of the system either as being an artistic artefact, engineering product or scientific experiment. A reason is identified behind deviation from best practice relating to time and effort as the practical implementation of advanced theoretical methodologies require too many complicated steps and the difficult resourcing of experts, especially in experimental projects as it may even prove hard to identify an expert in some cases.
3.2.
Experimental project framework
Both the field of CC and the evaluation emphasise interdisciplinarity. The methodology shift from disciplinarity to interdisciplinarity can be observed in the traditional and agile methodologies, used in software development.
Software engineering is a continuously growing field that has transformed over the last few decades, even though every implementation requires its own modifications, in general, there have been two big approaches. The traditional or heavyweight methodologies that rely on stability and take on a predictive approach and the lightweight or agile methodologies that embrace fast changes and use adaptive development methods. A detailed side-by-side comparison between the development process of the two methodologies is performed by Stoica et. al. [14], where each step of the process is analysed from the fundamental hypothesis up to the delivery of systems to clients. Though no approach is considered superior, both present strengths and weaknesses depending on the implementation environment. Regardless of the methodology, the role of testing and validation is highly emphasised in both processes. Awad [15] identifies the limitations of the methodologies separately, then comparing their implementation processes based on project size, as well as people and risk factors. A heavy focus lies on traditional methodologies as they are more thoroughly documented and offer extensive studies. Even though heavyweight methodologies have a safe area of implementation such as larger corporations and the private sector, the environment is volatile and the increase of customer uncertainty forces agile methods even in these sectors.
The shift in the methodologies can be seen as traditional software development relies on well-documented knowledge, while agile uses individual skills and core team members interaction. The implication of experts from different fields iteratively in several steps of the process leads to more flexibility and reliability in a broader set of implementations.
The shift to interdisciplinarity can also be seen in educational practice where Eaton [16] takes a look into the current university approach of teaching AI as a general introductory course followed by specific courses on AI sub-fields such as vision, learning, reasoning, and planning. Furthermore, the sub-fields are identified to be studied alone in the research literature due to the need for domain-specific optimisation. The idea of combining different AI sub-fields is supported by analysing the problems presented by both research and industry. Proposed interdisciplinary project-driven courses take advantage of the diverse students’ backgrounds to develop mutual understandings and shared knowledge across the fields. A multidisciplinary team is perceived to encourage collaboration at the same time as it ensures specific individual responsibilities.
Transdisciplinarity encapsulates all the benefits of interdisciplinary knowledge sharing, empha-sising education developed around constructing meaning by organising and applying learning in specific real-world contexts.
Transdisciplinarity in higher technological education and its societal impacts is studied by Aneas [17], who identifies interdisciplinary interaction as the basis for the creation of a community of professionals while transdisciplinarity relying on acquiring the necessary skills and knowledge to fully participate in a complex community. Transdisciplinarity is able to open up new and critical perspectives on the way of understanding, representing and structuring knowledge as it focuses on all prerequisites of solving a complex world problem rather than within a specific field. Ph.D. students represent most of the transdisciplinary research as they search for new ways to enhance their discipline by working with different methods and data. A risk in transdisciplinary research is presented by the defense of the work in the disciplinary positions as academic achievement is based on publications in discipline-specific journals.
Transdisciplinary studies have shown to produce a sense of value and humility as they open the perspective and application beyond the field of expertise considering several factors, including the users and environment, thus redefining projects based on human-centered values.
The perspective of building AI from its engineering roots is shifted to a human-centric and society-grounded transdisciplinary field that alongside technological advances should also include contributions from the humanities, the cognitive and social sciences.
Dignum [1] in her book entitled ”Responsible Artificial Intelligence: How to Develop and Use AI in a Responsible Way” offers the development process of AI from its initial computer science and engineering background where the focus relied on building systems that show characteristics that may be associated with human intelligence, to the fields of sociology where AI is used to map human behaviour and philosophy where the focus relies on the actions, goals and consciousness of AI. A shared opinion across the fields is that AI ultimately is needed to solve real-world prob-lems. The basics of ethical decision-making are identified and the responsibilities are redefined to be taken both by the developers, systems and users of AI. A powerful acronym used in the new definition of AI is ART standing for Accountability, Responsibility and Transparency. By
acknowl-edging a system that is capable of understanding its environment and autonomously take action to complete its task, a need arises to complement the system characteristics of interactiveness with accountability, adaptation with transparency and autonomy with responsibility. Accountability refers to the abilities of the system to explain and justify its decisions, which in turn have been developed in accordance with moral values and societal norms. Responsibility refers to the roles in interaction with AI. Currently, two states are identified either the machine performs correctly thus the responsibility lies with the user, or the machine performs unexpectedly due to errors thus the responsibility lies with developers and manufacturers. Transparency refers to the ability of describing, examining and replicating the steps taken by AI. The people that are affected by the systems should have access to data and the decision-making process should be visible.
3.3.
Explainable artificial intelligence through open-source transparency
The same principles of transdisciplinarity apply outside of academia and industry into the public sector. The values apply to people without prior background in a domain who need to access the necessary knowledge to resolve a real-life problem they face.
Constructionism is a theory proposed by Papert [18] that focuses on the art of learning to learn. Importance is given on learners’ interaction with artefacts, and how these interactions influence self-learning. Equal importance is given towards tools, technologies and context in the process of knowledge development. Regarding the process of creation, Sadler-Smith [19] offers an overview of the stages involved in the creative process and interplay between conscious10 and unconscious, that are present in the work of Graham Wallas, Art of Thought. Creative individuals are identified in relation to the power of solving problems by both studying the problems domain but also relating information from other domains. The Wallas model depicts the first stage as Preparation which relies on conscious focused work in understanding the subject, followed by Incubation where other tasks are performed, letting the mind unconsciously process information, followed by Intimation where association occurs and a small moment of conscious intuition is developed that transforms into Illumination where by focal consciousness a solution is developed that is consciously implemented in the Verification stage.
The Maker Movement Manifesto [20] presents the process of transposing knowledge into practice as a fundamental human part alongside creation and self-expression. Along with the importance of having access to the right tools, seven principles are identified in the manifesto such as sharing, giving, learning, playing, participating, supporting and changing. The manifesto offers a set of sto-ries on how ordinary people have created products in a makerspace11that led to new development of new business. The stories are meant to encourage the general population to participate in the maker movement.
Fox [21] takes a look at domesticating AI and its applications in prosumption12 as a way of expanding human self-expression through the creation of original unique goods. The trade-off between economy and originality is taken into consideration as well as its implications in a value-sharing culture. The potential of AI integration is recognised in software embedded in digital cameras, web design tools and manufacturing equipment. The potential community inclusion and profit over own creativity is emphasised rather than a Marxian perspective of exploitation13.
The practices offered by the maker movement are related by Mazzilli-Daechse to Gilbert Simon-don’s philosophy [22], where the great importance of technical education is stressed for the general population to also understand the technology they use. Though the philosophy was built in the 1950s, the current generation is argued as technologically illiterate but not by the amount of techni-cal education that has increased yet by the rapid technologitechni-cal advancements and hyper-specialised jobs available. The apparition of the maker movement and makerspaces facilitates public access at an affordable price to tools and knowledge witch has previously only provided higher learning institutions.
The dominance of students and digital tools professionals present in makerspaces is identified as they utilise the opportunity to get affordable access to experiment with new technology, while
10Aware of and responding to the surroundings.
11A collaborative workspace for making, learning and exploring using high tech to no tech tools. 12A prosumer is an individual who both produces and consumes a product.
13Regarding exploitation, Karl Marx devised the society in two classes with the rich obliging the poor to sell their labour for less than the value of the product or service.
the wide population is held back by the feeling of technical incompetence. Free online resources and makerspaces are identified as crucial to attract new people that are willing and able to put in the effort. ”Educational Robotics in the Context of the Maker Movement” [23] is a collection of articles in the field of the maker movement where studies report on simple techniques being used to make robust robots utilising low-cost everyday materials like cardboard, wood and broken toys. In one of the articles present in the book, Sch¨on [24] also mentions that the target group of makerspaces has been adults, yet in recent years involving children has gained popularity. Sch¨on is behind one of the first makerspace event in Germany that has been created for children. The goal was to do something creative using new technologies. Participants, between the ages of 10 and 14, were offered upon entry a set of principles to follow, based on the makers mannifesto [20], which invited to creativity, curiosity, fun as well as failure. A core concept consisted in the sharing of ideas, materials and tools but also the emphasis of asking for support. Another collection of experiments is presented by the inclusion of Arduino Education within the field of robotics [25]. One DIY project consisted in the development of an engineering kit, where previous knowledge in developing modular systems was taken into consideration to design a set that could either be used individually or in groups to get an introduction in translating mathematical equations from physics to software such as motor control, computer vision and state machines.
Public communities
A culture is best represented by its communities. A large-scale survey with over 2600 participants from several DIY communities is presented by Kuznetsov [26]. The study aims to capture the size and the variety of the community as it encapsulates practitioners from fine artists to engineers and hobbyists sharing ideas through a variety of mediums. The motivation for the contribution of the users ranges from inspiration gathering, documenting and showcasing work, ideas and skills to receiving feedback on projects and finding employment opportunities. The strongest motivations however are identified as self-expression and learning, the second being both supported by dis-cussion forums and the underlying idea of getting a deeper understating by sharing and teaching others.
One of the communities from the survey [26] relevant for this thesis is Instructables [27]. A web-based documentation platform, where DIY projects are shared as step-by-step Instructables. Each step offering the possibility of attaching text as well as images and video. Similar to other forums users are able to rate and comment on projects but in the maker movement spirit users are encouraged to share pictures and opinions if they made, improved or remixed the project. Moreover, each instructable has a teachers note section where teachers can share how they incorporated the material into their classroom. The projects are divided into several categories from electrical circuits, workshops and crafts to cooking, living and outside. A separate teachers category provides projects divided by grades and subjects being either practical, technological or artistic. Another community is Adafruit [28], an online distributor of electronic parts and kits such as the RPi. Even though more focused on electronics it provides tutorials, with an emphasis on low-cost materials and easy-to-follow methods.
One online community that has eluded Kuznetsov’s survey is GitHub [29]. A platform for software development version control14 and online hosting. Project-specific collaboration features
are also available ranging from request and bug tracking to wikis15, offering both free online
repositories16 as well as professional and enterprise accounts. For example, some instructables
show a key portion of the code while pointing to the whole software project hosted on GitHub while other authors prefer to only post the description of the project on websites and redirect the viewer for source code on GitHub. It reports as of August 2019 over 100 million repositories hosted [29] and has been acknowledged for over a decade as the world’s leading and largest host of source code [30], and research literature deemed it underrepresented in scientific literature compared to similar sites such as Stack Overflow17. With an tenfold increase from 10M to 100M as of 2014
when the study performed by Gousios et al. to 2019 search results on Google Scholar18though now
14Management of changes between stages of development. 15Collaboratively flexible publication managed by own audience. 16Data structure storage for projects.
17Question and answers site created for developers by developers
present four times more articles when searches are performed utilising the name of the platforms as keywords.
Open-sourcing
One important aspect when posting content online is the legal right of the owner upon intellectual property. Licensing a public repository on GitHub [29] is required to truly be open source, so the content becomes free to use, allowing others to change the software. The Legal Side of Open Source19offers an overview of the importance of licensing software, different types of licenses as well
as answers to common questions regarding modifications and implications of licenses. There is an array of different licenses depending on the software used. The most common is the Massachusetts Institute of Technology (MIT) license. First devised in the late 1980s, it offers freedom over private use, modifications, distribution, sub-licensing and even commercial use as long as you include the copyright information of the original owner.
4.
Method
Due to the exploratory nature of the study, the Design Thinking methodology is adopted to identify and interpret the hidden needs of the thesis. The methods are presented following five development phases adapted from the chosen methodology, consisting of information gathering in the Discovery and Interpretation phases followed by the Ideation and Experimentation phases where concepts are tested and the Evaluation where new knowledge is created. The research enviroment is presented, then the research phases are detailed with the corresponding steps in the research process.
4.1.
Research environment
This section outlines the research approach, position and methodology.
Methodology
Compared to traditional and agile methods that are often employed in the field of product (soft-ware) development [14], the aim of the thesis is to develop a process of empowering people to acquire knowledge and shape their own products. Since the position of the thesis lies in a research and development environment, an overall Design Thinking methodology has been chosen thanks to its flexibility and human-centered focus. According to the head of IDEO20[31], Design Thinking is a discipline that uses the designer capabilities to match people’s needs with what is technologically feasible [31].
Different design categories according to Frankel et al. [32] divide knowledge into three sec-tions: (1) how-to-knowledge where someone demonstrates a skill, (2) knowing-that knowledge where someone learns about someones else skill and (3) knowledge-of where someone is learning about the existence of a skill. User engagement in research at any of the levels creates value in a variety of combinations, depending on the initial question or hypothesis. According to Tayal [33] there is traditionally a difference between design and engineering disciplines, where engineering disciplines tend to focus on design for technical stakeholders while the recurring elements of the design processes contain analysis, synthesis and prototyping. An aspect of the design field that is present in the thesis and might differ from science and engineering fields is the focus on satisfying the needs of the users that will use the product. Designers are identified to have the ability to put concepts together and developing new things into being.
Research position
As my position resides between academia and industry, I take the role of AI expert connecting research to implementation. Since there is extensive research done in the field of poetry generation, I also assume the role of facilitator, offering a chance to interact with different AI generators and test them in an office environment. The position of a robotics student with knowledge outside of the fields of AI into mechanical and electronic engineering is taken to create a tool to generate future physical interactions. Furthermore, I also had the role of participant in the transdisciplinary meetings.
Research approach
A pragmatic research approach is considered as a theoretical framework for this thesis. The pragmatic approach is described by Alvesson et al. [34], as the study on the theory’s practical utility. The participants are encouraged to take an active role in the AI experimentation as a bi-product helping them in taking control of AI, by understanding the ”magic” behind the technology proposed. This perspective is also supported by the use of explainable and responsible AI goals, aiding the process of answering the research questions on how the AI approaches developed over time and how the generators perform implemented in a different environment.
Though as Glaser and Strauss [4] argue, the pragmatic position should provide the under-standing and control of situations for the practitioners, as well as the possibility of prediction
20IDEO is a design and consulting firm that uses the Design Thinking approach to design products, services, environments, and digital experiences.
and explanation. In this research the practitioners are represented by the participants of the work-shops, the people interacting with the platform, as well as the readers, who will be able to apply the knowledge of AI provided by the study within their own work, with a minimum of pre-technical knowledge. James [35] describes that the pragmatic approach on philosophical topics such as knowledge and language is viewed in terms of their practical uses and successes. Reevaluation of poetry generators is done given the new implementation environment.
According to Sch¨on’s book, ”The Reflective Practitioner”, the only learning that significantly influences behaviour is self-discovered and self-appropriated [36]. In that regard, open-sourced projects provide the knowledge, but the learning belongs to the learner.
4.2.
Research phases
The premise of the thesis has been established as a research and development project instead of optimisation, thus the methods employed are closer to the Design Thinking iterative process, though always keeping an engineering point of view over the aspects studied. The methods of Design Thinking are taken from the ”Design Thinking for Educators Toolkit” [37] and the research process is illustrated in Figure 1.
Figure 1: An overview of the Design Thinking Methods employed and their interdependence.
The phases employed in this study are divided into their own paragraphs. On top of the information gathered in the transdisciplinary meetings, the Discovery phase consists of a theoretical overview of required knowledge in the form of a literature study within the development of poetry generation and the evaluation of computational creativity. Then in the Interpretation phase, a benchmark of 23 poetry generators is conducted in terms of system complexity, language, user interaction and online availability. An overview of poetry generator development is evaluated with
regards to algorithmic progression, poetry forms, user interaction and intended future development. The first part of the Experimentation phase gathers empirical data from the test of the two most complex poetry generators identified in the benchmark, in an office environment as parts of two transdisciplinary meetings. Based on the evaluation of the results both from the benchmark and the tests a practical implementation is conducted in the form of a physical prototype, adding verbal HRI as well as the possibility to dynamical create new corpora to an open-sourced poetry generator. After evaluating the prototype, a workshop is conducted to understand real-life expectations upon creative AI systems in a holistic manner both on their computational functionalities as well as their environmental interaction capacities. Using the feedback from the Ideation workshop as well as the perspectives developed during the literature study, an online platform is presented to offer a DIY step-by-step introduction into the various modules involved in the creation of a poetry robot companion.
4.2.1 Discovery
In order to identify the features of the poetry companion that were needed, the aim was discussed and sources of inspiration were identified from the interaction with the participants. Also, a learning process from research documentations on similar projects was conducted.
User needs
To define the users and their needs, a total of five weekly transdisciplinary meetings were attended before the pandemic influenced the move of the thesis to the online environment. A common interest for understanding the emerging field of explainable and responsible AI was identified. The interest in such AI development was most relevant to users without a technical background that wanted to integrate computer interaction within their respective fields.
Literature review
To identify the state-of-art on poetry generators, a literature study was conducted where the keywords ”AI poetry”, ”poetry generation” and ”computational creativity” were introduced in the Google Scholar database to access its collection of journals. The search was then limited to books and research articles containing several generators published within the last five years. The taxonomy of poetry generators by Lamb et al. [9] provided most search results and the newest publication in the field. Through the method of snowballing, where new articles are identified within the citations of the studied article, both other categorisations and individual generators were identified. Lastly, the poetry generators presented between several taxonomies were cross-referenced and chosen for review, ensuring one special feature present in each generator without repetition thus not overpopulating the benchmark.
4.2.2 Interpretation
The interpretation phase was to revise empirical and theoretical findings in the literature study. The research-covered generators are studied to identify the algorithmic development stages and features added.
Poetry generators benchmark
Comparing the generators between with each other, a poetry generator benchmark is created in order to define a suitable implementation. The benchmark is conducted as an extensive side-by-side comparison with regards to the language employed, user customisation level, meaningfulness, novelty, feature complexity, online accessibility, evaluation methods employed and future work intended by the developers.
Poetry generators development
A second purpose of the benchmark was to understand the development history of different poetry generators. The stages of the algorithms employed along the evaluation methods surrounding the generators is detailed, both following system complexity and a chronological order.
4.2.3 Ideation
A total of five RISE employees from different research fields have taken part in the HRI Ideation Workshop that was held during one of the transdisciplinary meetings. The workshop was also repeated with the two academic supervisors.
HRI ideation workshop
The workshop consisted of three reflection exercises upon which the participants should identify the desired interaction form with AI. The setting was around work and home environments so that the AI companion should help the user unwind and reflect on tasks that have been done and also help in the planning of future ones. Characteristics that the participants were asked to reflect upon were shape, functionalities and the interaction forms (input and output). The second exercises was inspired by the workshops held by Sch¨on [24] with children. The participants were asked to reflect on the one functionality they found most important and try to explain the steps required to achieve it, in a simple enough DIY step-by-step manner intended to be understandable by a 10-year old child.
4.2.4 Experimentation
The steps of Design Thinking employed during the experimentation phase revolved around making prototypes, where both top poetry generators were tested in an office, a physical robot with verbal interaction features on top of a poetry generator was created and a future tool for AI enhanced experiments was conceptualised. Feedback from one step was incorporated to make the next de-liverable.
Poetry generators test
The first part, consisted of two generators that were tested on a research office target group. The criterion that was taken into consideration were interactive input customisation, research-based po-etry generators, user-friendly online availability and different underlying AI algorithms employed. The purpose of the generators was to augment the tune-in exercise, where awareness and connec-tion to the moment are emphasised before a session, and the tune-out exercise, where reflecconnec-tion over the contents of the sessions is valued, following the method provided by Bozic Yams [5]. Two online workshops were held to tests different generators, a computer-enhanced Haiku generator [38] was used during the tune-in reflection exercise on one meeting and CoPoetry-Me [10], a human-enhanced generator, was used during the tune-out reflection during a second meeting. Feedback was collected, to be integrated into the next test, both directly through conversation sessions after the tests in form of generated user poems and notes from the discussion, as well as from an online survey sent afterward.
Physical prototype
Based on the results and feedback gathered from the first part, a physical implementation was conducted. A system capable of customise the corpora and produce coherent poems and direct output was desired. The option of understanding the process and having feedback at each step were also crucial parts of the build. Therefore a modular open-sourced solution was taken into con-sideration showcasing the ease of working with free material and put it together to fit the desired application. Smaller research was conducted on open-sourced generators equalling the complexity levels of the ones presented in the benchmark and on verbal interaction modules. A prototype
is built to run on several operating systems, both on Windows as an extension of a computer but also to be self-sustained on a RPi. In order to achieve explainable AI, both the installation steps and the functions employed are documented and thoroughly explained in such a manner to allow non-experts to understand each step of the build. The final product is able to record and transcribe audio to text, then analyse it and produce poetry that can be read out loud. Secondary functionalities are added to store and retrieve selected poems to a database as well as clean new corpora. A few motors and sensors are also implemented to emphasise the modularity and expan-sion capabilities of the system. The system is tested on both an online meeting and predefined corpora to showcase the quality of the generated poems.
Online platform
Lastly, the online platform contains a few entry-level options in each category to give a varied approach yet not overwhelm. Each category is also thought to offer both fast prototyping as well as more robust and complex solutions, though not over complicate with the use of specialised machinery. All solutions range from the use of household or office items to purchasing consumer-available goods. Using affordable development boards such as the RPi and 3D printers, both accessibility and flexibility were deciding parameters so that unavailability of parts or budget should not be barriers upon start. The selection contains free public available projects from specialised platforms such as Instructables [27], that provides DIY step-by-step guidance and GitHub [29] offering code and programming examples. To enhance HRI, the platform offers a few different cores of poetry generators (different levels of generation and poetry styles), and different interaction possibilities (different shapes, inputs and outputs), identified from the Ideation workshop.
4.2.5 Evaluation
The evaluation has been performed iteratively between the phases of the thesis, each phase influ-encing the definition of the next one. The poetry generators benchmark regarding AI methods, poetry forms and performance expectations is evaluated both individually, to some initial criterion, and discussed with the industrial supervisor before implementation. The criteria for the selection of poetry generators are research literature coverage, functionality in the the English language used in the meetings, the ability to manipulate different parts of the generation, and online availability. Feedback on the generator tests results was gathered in online meetings at the RISE office in a transdisciplinary environment, where user expectations and creative AI capacities were evaluated. The methods employed consisted of user observation during meetings and feedback sessions as well as a survey of poem evaluation by end-users. The physical prototype is tested both on HRI capa-bilities and generation accuracy on a recorded meeting and other corpora. The user feedback for the online platform was gathered only by a survey after individual tests to evaluate the usability of the platform.
5.
Poetry generators study
From the literature study a comparison has been made between 23 computer poetry generators presented in scientific literature based on complexity and user collaboration. Four taxonomies and surveys on poetry generation have been the foundation upon which a snowballing method has led to several generators’ discovery. Each generator has been examined to understand the need of the study, its outcomes, evaluation and future work intentions. The second scope of this study is to identify common practice as well as alternative approaches and their respective strengths and weaknesses. Poetry forms and other poetic characteristics will be briefly explained as footnotes by their description in the Oxfords Learner’s Dictionaries21.
Oliveira [10] presents related work in the human-computer co-creation process while mentioning Lamb et al.’s taxonomy [9] that creates the separation of generators in mere, human- and computer-enhancement. According to Lamb et al. taxonomy [9], mere generation is considered when there are no additional steps after the generation process. The category is a good place to start a simple implementation and provides the basis upon which interaction is built. If the generated text is then modified to create the final artefact, the generators are divided either into human- or computer-enhancement categories. For the scope of this thesis, all generators that allow human output interaction are placed in the human-enhancement category. Most enhanced generators have several iterations and publications backing up the progress, a filtration of the most relevant paper is performed, often depicting either the most recent generator build or unique feature implementation. Human-enhanced generators additionally provide the user with the ability to interact with both the input and output of the generator, thus allowing full user customisation.
Oliveira [11] presents a survey on intelligent poetry generation divided into different categories: features, techniques, reutilisation and evaluation. Ventura [3] discusses the methods that are considered mere and computer-enhancement by Lamb et al., dividing them in algorithm complexity and presenting a theoretical implementation of generators. Lastly, examples of poems and their respective poetry generators developed between 1959 and 1995 are presented in Funkhouser’s book, ”Prehistoric digital poetry” [2].
5.1.
Poetry generators benchmark
Since the benchmark overview is rather large in scale and would not fit as a table in the text, smaller section tables are presented below. A side-by-side comparison of all generators can be found in Appendix A. For future reference in the tables, generators are present by their reference. All generators are listed below by name and author(s), ordered by complexity and chronological deployment from low to high and past to present. The information below is also visualised in Table 1. In the mere generation category, Lutz [39] has the Stochastic Texts, and Morris [40] has Haiku -at random, Carpenter [41] has Electronic Text Composition, Gosper [42] has Disassoci-ated Press, Stross [43] has the Lovebible.pl, Bhatnagar [44] has the Pentametron and Harris [45] has the New York Times Haiku. The computer-enhancement category contains Gerv´as’s [46] ASPERA, D´ıaz-Agudo et al.’s [47] COLIBRI, Levy’s [48] Poevolve, Manurung’s [49] McGONAGALL, Gerv´as’s [50] redefined WASP, Toivanen [51] has an unnamed generator (that will later evolve into The Poetry Machine [52]), Rashel et al. [53] has Pemuisi, Zhang [54] has another unnamed generator, Yan [55] has iPoet while Ghazvininejad et al.[56] has Hafez. The human-enhancement category has the Flarf [57] generation techniques, Addad et al.’s [58] Gnoetry, Barbieri et al.’s [59] unnamed generator, Kantasolo et al.’s [52] The Poetry Machine and Oliveira et al.’s [10] Co-PoeTryMe.
Complexity Generator
Mere [39][40][41][42][43][44][45] Computer-enhanced [46][47][48][49][50][53][54][55][56][60] Human-enhanced [10][52][57][58][59]
Table 1: Generator complexity categorisation.
Test Selection
One of the primary features was language. As the communication language where the test would take place is English, that was the desired language. As seen in Table 2, most of the generators are by default written in English. Some offer even several languages such as Spanish, Chinese, Finnish and Indonesian. Some generators allow for any given input.
Language Generator English [10][40][41][43][44][45][56][57][58][59][60] Spanish [10][46][47][50] Chinese [54][55] Finnish [52][60] Indonesian [53] Non-language [57][58]
Table 2: Generator supported language
The second relevant feature of the benchmark was user customisation. While mere generation methods often do not present the user with any possibility of influencing the system, apart from manually selecting the best poems to present, computer-enhanced systems offer input customisa-tion, often in the form of seed words and interchangeable training corpus22. A great overview of
the majority of different data acquisition methods is performed by Lamb et al. [9], though it does not cover how much of the co-creation is performed by the user. Table 3 provides an overview of the generators in terms of the possibility of the user interacting with the generation process. Either no interaction is present, the information fed to the generator can be altered (input only) or the user has full control over the input and the output. The majority of the generators are built and tested for a specific scientific purpose, thus not offer user interaction. Only newer and a few of the more researched generators have evolved to offer user interaction.
User Interaction Generator
Input and Output [10][52][56][57][58][59] Input only [46][47][54][55][60]
None [39][40][41][42][43][44][45][48][49][50][53]
Table 3: User interaction
The third desired feature represented online accessibility. For this part, only generators that can be accessed are taken into consideration. For example, Hafez [56] is said to have an online platform though it proved unreachable upon several attempts. The generators are categorised whatever they present online platforms, publicly shared code or poems databases and presented in Table 4, while all their respective links are in Appendix A.
Online Availability Generator
Platform [10][52][57][58]
Code [41][43][58]
Poems [44][45][57][58]
Table 4: Online availability
Four platforms present both online availability and full user interaction though one of them is in Finnish [52]. Cross referencing all the requirements, Flarf poetry [57], Gnoetry [58] and Co-PoeTryMe [10] are suitable candidates for the office test.
Test generators description
The three generators that present the desired features differ in implementation. Below a summary of each generator is presented in regards to the interaction form and the literature support.
Flarf poetry is not a generator but an artistic method of interacting with computer programs. The method consists of mining the internet with odd search terms and rephrase the results. The user has full control over the generation and output. Deliberate shapelessness of content, form, spelling, and thought is perceived as assisting the mind in finding unfamiliar situations, novel language and unexpected combinations thus leading to new understandings.
Gnoetry offers a clean user interface with full control over the generation process’s input and output. The initial generation occurs randomly based on a set of three example poetry corpus that can be given different weights. The corpus can be modified to any user text in any language and several corpora can be added or removed. Once poetry has been generated, the platform offers the user the possibility to guide the generation cycles by choosing a few of the words and re-generating the rest whilst keeping the form of the poem. The poetry form can also be specified and tweaked by the user as well as the template for the poetry.
Co-PoetryMe offers an all in one platform with full user configuration both over the input and output. A modular architecture, allows the platform to take advantage of different functionalities depending on the task that is performed. The main module produces drafts of poetry after some user inputs such as language (English, Spanish or Portuguese), predefined poetry form, seed words that influence the semantic domain23 and a surprise factor that affects the relationship between the probability freedom upon generation. Both words and lines can be changed, each having its own module. The lines module can generate alternative lines with the possibility of using different seed words and a syllable counter. The words module can generate a new word based on one of the semantic relations of the selected word. A bank module exists to keep the selected words while a trash module stores the discarded words. An overall top module contains the utility buttons for guiding each user interaction. Finally, a visualisation tool also offers the possibility of comparing the changes made from the initial generated draft and the current state.
5.2.
Poetry generators development
From the discovery phase, the user needs were identified and the common interest lied in under-standing the emerging field of explainable and responsible AI. The interest in such AI development was most relevant to users without a technical background that wanted to integrate computer interaction within their respective fields. To achieve the goal of explainable AI systems, the de-velopment of the poetry generators covered in the benchmark is presented both with regards to algorithmic complexity and evaluation methods employed.
5.2.1 Algorithmic complexity
The comparison will be presented with regards to algorithmic progression, which will be highlighted in the text, poetry forms and user interaction as well as future work intended by the developers of the generators.
In 1959 Lutz [39] was writing the Stochastic Texts a Template Based free form poetry24
based on Kafka’s ”The Castle”25. Having a fixed template with a few words missing and randomly
changed from a list produced intelligible text though the poems were highly repetitive. Morris [40] (1973) then in an attempt to create novelty overcome the repetitiveness by introducing a set of shifting templates. The Haiku poetic form was introduced though the specific structure was not kept and the poems were mostly non-intelligible.
Flarf poetry [57] is a poetry form developed at the beginning of the 2000s inspired by dadaism26.
It presents a form of Found Poetry where poems are created by taking parts from different sources and re-framing them to obtain new meaning.
23Study of the relationship between words, phrases or symbols. 24Poetry without a regular rhythm or rhyme
25A 1926 novel by Franz Kafka
Natural Language Processing (NLP) emerged as the field under which interaction between computers and human language is studied. The methods above did not work with any larger amount of language data. The following generators work with a collection of texts, called corpus, either by themselves or in collaboration with an electronic lexical database. WordNet [61] was started in the mid-1980s and is the largest database of semantic relationships between words and has now grown to host over 200 languages. It can be perceived as a large electronic dictionary grouping words in nouns, adjectives, verbs, adverbs, simplex words27, phrasal28verbs and idioms29. In her book, Fellbaum, the developer of WordNet, a great introductory tool to the database [61]. The applications of the lexicon in NLP range from machine translation, word extraction, indexing and text categorisation to user support in formulation, word syntax, and relationships.
Similar to Flarf poetry, Case Base Reasoning (CBR) is a computer logic algorithm that allows the creation of poems by an intelligent adaptation of selected examples from a corpus of verses. It creates new poems from old ones by retrieving and reusing parts from the corpus and lexicon to produce new poetry fragments, then involves the user in revising the generated text and retains the new case, learning from it by performing a linguistic analysis. Gerv´as [46] (2001) takes a look into Spanish poetry using CBR to develop a forward reasoning rule-based system that can take user input parameters such as an intended message and a level of emotion. With the help of a knowledge-based pre-processor, the generator obtains metric30structure while translating the
given message from prose31 to poetry. D´ıaz-Agudo et al. [47] (2002) continues on the Spanish
poetry and develop an ontology32 that incorporates common CBR terminology with a
domain-independent framework. The evolution consists of using more than one case for adaptation and the incorporation of the ontology for the terms of the language being employed.
Another heavily used method for generalising structured text is Markov Chaining where the probability of prediction of the next entry is based on the previous entries sequentially. Gosper [42] (2011) uses characters instead of words to create novel words taking advantage of the free form poetry style. The intentional non-intelligible poems offer a high novelty factor. Stross [43] (2013) under the umbrella of dadaism builds poems from two contrasting sources, Bible verses and horror fiction. The outcome is intelligible and surprising. The code is also publicly available and the users are encouraged to change the sources to create new genre crossings. Stross planed to fine-tune the parameters of the Markov chain output, pick different seed words and possibly pre-filter the corpus data.
Carpenter [41] (2004) builds poems based on Stochastic Context-Free Grammar. Thought as a more coherent process than the Markov chain, it recursively builds sentences based on either terminal or non-terminal symbols. The project includes around 18,000 pages of free form poetry generated. The poems became both intelligible and grammatically correct. Addad et al. [58] (2006) offer an interactive online generator based on Markov chaining. It synthesises language randomly, any machine-readable text in any language can be fed and the human operator intervenes in the language generation cycle, helping to guide the artistic process. Worth mentioning is the possibility of changing poetry forms based on their templates. It is also the earliest generator to host its own online community, blog and chapbooks33.
The algorithmic progression complexity was represented by the introduction of Evolutionary Algorithms (EA) where the generation is further enhanced by iterative modifications of the draft with regards to different features and constraints. In general, changes are either produced by the crossover between the data sets, similar to CBR and Found Poetry, from combining two existing poems in the population or mutations, where only some words in the poem are replaced.
Levy [48] (2001) combines the syntax of one poem with the text from another. It is simulating an evolving system of creative expression. The poems, Limericks34, are declared interesting but the
27Shortest words depicting distinct meaning 28Part of or connected to a phrase
29A group of words creating different meaning than the meanings of the individual words 30Metre refers to the musicality rhythm and is symbolised by recurring patterns and accents. 31Language in its ordinary form.
32Ontology in computer science is a categorisation with precise naming and definition of the classes containing properties and relations between the concepts and entries of data.
33A small book, containing poems or fiction.
34Limerick is a humorous short poem, with two long lines that rhyme with each other, followed by two short lines that rhyme with each other and ending with a long line that rhymes with the first two.
