Human collaboration with systems within the Computational Creativity (CC)
field is often restricted to shallow interactions, where the creative
processes, of systems and humans alike, are carried out in isolation, without
any (or little) intervention from the user, and without any discussion about
how the unfolding decisions are taking place. Fruitful co-creation requires a
sustained ongoing interaction that can include discussions of ideas,
comparisons to previous/other works, incremental improvements and revisions,
etc. For these interactions, communication is an intrinsic factor. This means
giving a voice to CC systems and enabling two-way communication channels
between them and their users so that they can: explain their processes and
decisions, support their ideas so that these are given serious consideration by
their creative collaborators, and learn from these discussions to further
improve their creative processes. For this, we propose a set of design
principles for CC systems that aim at supporting greater co-creation and
collaboration with their human collaborators.
Abstract Human collaboration with systems within the Computational Creativity (CC) field is often restricted to shallow interactions, where the creative processes, of systems and humans alike, are carried out in isolation, without any (or little) intervention from the user, and without any discussion about how the unfolding decisions are taking place.
Fruitful co-creation requires a sustained ongoing interaction that can include discussions of ideas, comparisons to previous/other works, incremental improvements and revisions, etc.
For these interactions, communication is an intrinsic factor.
これらの相互作用にとって、コミュニケーションは本質的な要素である。
0.55
This means giving a voice to CC systems and enabling two-way communication channels between them and their users so that they can: explain their processes and decisions, support their ideas so that these are given serious consideration by their creative collaborators, and learn from these discussions to further improve their creative processes.
For this, we propose a set of design principles for CC systems that aim at supporting greater cocreation and collaboration with their human collaborators.
Introduction Although systems from the field of Computational Creativity (which we will refer to from now on as CC systems) – and more generally AI systems – have been successful in different application domains, a common challenge for users and researchers is that most of these systems behave as black boxes, limited to opaque interactions where processes and reasoning are completely unknown or obscure (Scherer 2016).
These limitations have raised the need for the development of models that offer more clarity and transparency in order to improve the potential for human-machine interactions with CC systems (Muggleton et al 2018; Bryson and Winfield 2017).
これらの制限により、ccシステムとのヒューマンマシンインタラクションの可能性を改善するために、より明確で透明性を提供するモデルの開発の必要性が高まった(muggleton et al 2018; bryson and winfield 2017)。
0.72
The field of Explainable AI (XAI) has grown in recent years with the goal of making black box systems more transparent and accountable through models of explanation that communicate the way decisions have been reached.
Current approaches to XAI, which are commonly associated with popular but opaque machine learning methods, centre on providing explanations as part of the output of the system; i.e. the focus is on delivering a final result to a user alongside a rationale of how this result was created.
XAIの現在のアプローチは、一般的なが不透明な機械学習手法と関連付けられており、システムの出力の一部として説明を提供することに集中している。 訳抜け防止モード: XAIの現状と展望 一般的な機械学習手法と 結びついています システムの出力の一部として説明を提供することの中心 i.e. 焦点は ユーザに対して最終的な結果を提供すると同時に、この結果がどのように作成されたかの根拠を提供します。
0.71
However, the creative process is often performed in isolation, with no place
しかし 創造的なプロセスは 孤立して行われることが多く
0.65
for intermediate explanations as the process progresses, let alone place for exchanges of information that can exploit human-machine co-creation.
The focus of this subfield is the study of bidirectional explainable models in the context of computational creativity – where the term explainable is used with a broader sense to cover not only one shot-style explanations, but also for co-creative interventions that involve dialogue-style communications.
More precisely, XCC investigates the design of CC systems that can communicate and explain their processes, decisions and ideas throughout the creative process in ways that are comprehensible to both humans and machines.
The ultimate goal of XCC is to open up two-way communication channels between humans and CC systems in order to foster co-creation and improve the quality, depth and usefulness of collaborations between them.
Designing and implementing this type of communication is extremely complex; however, we believe it is important to start a discussion towards what is needed for more fruitful and productive partnerships between CC systems and their users.
Creativity is not a solo act, it is a social activity that benefits from life experiences, from influences of people, and from the contributions of different collaborators.
This may result in increased novelty, or higher-value creative collaborations, however it is difficult to anticipate how successful and in which way any creative collaboration can emerge, but we believe the kind of active collaborations proposed through XCC would ultimately enhance the human-machine collaboration experience and increase the engagement of users with CC systems.
In the rest of the paper, we first provide relevant background literature.
論文の残りの部分では、まず、関連する背景文献を提示する。
0.56
Then we present an overview of the current state-of-the-art of co-creation and explainable AI in general and in CC in particular.
次に, 一般に, 特にccにおいて, 共同創造と説明可能なaiの現況について概観する。
0.62
We follow by identifying design principles that we believe CC systems with explainable capabilities should have.
私たちは、説明可能な能力を持つCCシステムが持つべき設計原則を特定します。
0.70
We finish with a discussion outlining challenges that need to be considered, opportunities that may arise, conclusions and directions for future work.
我々は、考慮すべき課題、起こりうる機会、今後の仕事の結論と方向性を概説する議論を終える。
0.60
英語（論文から抽出）
日本語訳
スコア
Figure 1: The set of annotations created during the evaluation of SpeakeSystem.
図1: SpeakeSystemの評価中に作成されたアノテーションのセット。
0.82
At the top of the image is the audio timeline.
画像の一番上には音声タイムラインがあります。
0.71
The coloured boxes below represent the annotations created by each of three annotators.
以下の色付きボックスは、3つのアノテーションによって作成されたアノテーションを表している。
0.52
Here the system cannot contribute to the conversation; i.e. the system does not have a voice.
ここでは、システムは会話に貢献できない;すなわち、システムは音声を持っていない。
0.83
Motivation and Related Work
モチベーションと関連する仕事
0.68
Co-creation, real-time interactions and collaboration are important topics in the CC community; however, the communication between CC systems and their users is limited to a few exchanges, or the rationale behind their individual actions is often unknown by either, or there is little or no opportunity for discussion of the ideas presented, and many system outputs are discarded without a second thought.
Take for instance the SpeakeSystem: a real-time interactive music improviser which takes as its input an audio stream from a monophonic instrument, and produces as its output a sequence of musical note events which can be used to control a synthesizer.
The BBC Radio 3 Jazz Line Up programme commissioned SpeakeSystem in 2015 for a live human and computer performance with alto saxophone player Martin Speake12.
BBC Radio 3 Jazz Line Upプログラムは2015年にSpeakeSystemに、アルトサックス奏者のMartin Speake12とのライブ人間とコンピュータのパフォーマンスを依頼した。
0.78
Figure 1 shows a recording of the performance that was uploaded to a timeline annotation system in order to carry out a focused discussion about it.
In a human-human interaction, the musician being ‘questioned’ would reply with a rationale about his/her decision at those points in the performance; however, in this conversation it is not possible to know what the rationale or motivations of the system were as the system does not have a voice in the conversation.
1https://www.bbc.co. uk/programmes/p033s4 gj 2A recording of the performance and the source code for the system is available in an open source repository (Yee-King 2016)
” Even though the algorithm designer provides information about how the system operates, he is unable to provide an explanation of what was going on at specific points in the performance as only the system “knows” the details of what went on at every point of the performance.
If the SpeakeSystem had been equipped with communication capabilities, not only it could have provided insights about its creative process in the conversation above, but the human musician could have communicated his intentions during the performance (perhaps through some kind of visual or haptic signal), giving the system a chance to respond during co-creation.
This could have resulted in a more engaging experience, for the musician and the audience alike.
これはミュージシャンや観客にとって、より魅力的な体験をもたらす可能性がある。
0.70
Co-creation, Real-Time Interactions and Collaboration in CC Models to improve collaboration and co-creation within CC have been explored in the community.
In (Davis et al 2015), for instance, an enactive model of creativity is proposed for collaboration and co-creation.
例えば(Davis et al 2015)では、コラボレーションと共同創造のために創造性の活発なモデルが提案されている。
0.71
The key principle for this model is that the system has a level of awareness and that co-creation happens through a real-time and improvised interaction with the environment and other agents.
Another approach is mixed-initiative co-creativity (Yannakakis, Liapis, and Alexopoulos 2014), which exploits a bi-directional communication based on the collective exploration of the design space and human lateral decisions that are used by the system to guide the creative task.
Although these models establish communication channels, these are limited to an action/reaction type of model; i.e. there is little opportunity for further introspection in addition that they do not enable the systems to further support their contributions.
これらのモデルはコミュニケーションチャネルを確立するが、これらはアクション/アクションタイプのモデルに限定される。 訳抜け防止モード: これらのモデルは通信チャネルを確立するが、それらはアクションタイプ/リアクションタイプに限られる。 I. E. さらなる検査の機会は ほとんどありません システムにさらなる貢献を 許さないのです
0.75
The You Can’t Know my Mind installation of The Painting Fool (Colton and Ventura 2014) presented real-time interactions with users that resembled some of the aspects we cover in this paper: the system provides explanations about its process, motivations, etc. (e g by providing a commentary alongside its output), the user provides content to guide the creative process (e g by expressing a particular emotion to be the inspiration for the painting), and the system utilizes visual cues to reveal what is going on (e g with the use of an on-screen hand while it paints the picture).
The You Can’t Know my Mind installation of The Painting Fool (Colton and Ventura 2014) presented real-time interactions with users that resembled some of the aspects we cover in this paper: the system provides explanations about its process, motivations, etc. (e g by providing a commentary alongside its output), the user provides content to guide the creative process (e g by expressing a particular emotion to be the inspiration for the painting), and the system utilizes visual cues to reveal what is going on (e g with the use of an on-screen hand while it paints the picture). 訳抜け防止モード: The You Ca n’t Know my Mind install of The Painting Fool (Colton and Ventura 2014 )では、この論文でカバーしているいくつかの側面に類似したリアルタイムインタラクションが紹介されている。 モチベーション等(例えば、その出力と共に注釈を提供することで) ユーザは創造的なプロセス(例えば、 by)をガイドするコンテンツを提供します 絵のインスピレーションとなる特定の感情を表現するさま このシステムは視覚的手がかりを利用して何が起きているのかを明らかにする(例えば、絵を描きながらスクリーンハンドを使う)。
0.78
Although these interactions simulate communication, it is mostly a one-way creative endeavour.
これらの相互作用はコミュニケーションをシミュレートするが、ほとんどは独創的な試みである。
0.61
The Beyond The Fence musical is also an important sample of human-machine collaboration in the field.
Beyond The Fence ミュージカルは、この分野における人間と機械のコラボレーションの重要なサンプルでもある。
0.67
The musical writers that were involved in this effort were enthusiastic about the possibilities of collaboration with CC systems; however, they highlighted some of the flaws they faced when working with them:
Musical writers comment in (Colton et al 2016) “Collaborating with computers is utterly unlike anything either of us have encountered before, and at times, it has been incredibly frustrating” Musical writers comment in (Colton et al 2016) These comments highlight that ultimately, for a longstanding (working) human-machine relationship to be sustainable in the context of creativity, there is a need for mechanisms that enable a more active partnership.
ミュージカルライターのコメント (colton et al 2016) “collaborating with computer isutterly unlike any who had ever ever encounter, and when it was been incredible フラストレーション” 音楽ライターのコメント (colton et al 2016) これらのコメントは、究極的には、長年(働く)人間と機械の関係が創造性という文脈で持続可能であるために、よりアクティブなパートナーシップを可能にするメカニズムが必要であることを強調している。
0.80
As the musical writers put it:
音楽作家は次のように述べている。
0.55
“I rather think that the future holds ways of allowing human artists to work with computers more comfortably, and with more control of their output, ultimately to support and perhaps shape their own creativity in ways they might not have been able to envisage” Musical writers comment in (Colton et al 2016)
将来的には、人間のアーティストがコンピュータでより快適に作業し、彼らのアウトプットをコントロールし、最終的には、彼らが想定できなかった方法で自分たちの創造性をサポートし、形作る方法があると思います」と、Musicalのライターはコメントしています(Colton et al 2016)。
0.69
The need for communication Nickerson et al (Nickerson, Elkind, and Carbonell 1968) described the increasing complexity of human-computer interactions based on the ability to communicate with one another “the thing that, above all others, makes the mancomputer interaction different from the interaction that occurs in other man-machine systems is the fact that the former has the nature of a dialogue”.
コミュニケーションの必要性 (nickerson et al) (nickerson, elkind, carbonell 1968) では、互いにコミュニケーションする能力に基づいて、人間とコンピュータの相互作用の複雑さが増大していることが述べられている。 訳抜け防止モード: the need for communication nickerson et al (nickerson, elkind, carbonell 1968)は、人間とコンピュータの対話の複雑さが増していることを説明している。 他の男性で起こる相互作用とは異なるマンコンピュータの相互作用 -機械システムとは 前者は対話の性質を持っている。
0.69
This thought has been echoed by other researchers, who have emphasised that computers are ‘comparable’ to humans in some dimensions when seen as collaborators, particularly as dialogue-partners (Kammersgaard 1988).
With the rapid growth of AI techniques, this discussion has gradually highlighted the existence of a higher level of intelligence when contrasting the views of ‘interaction as tool use’ and ‘interaction as dialogue’ (Hornbæk and Oulasvirta 2017), arguing that utility and usefulness are the main aspects for the first type of interaction to work, while having a constant, simple, direct and natural communication and understanding between human and computer, is key for the second type of interaction to work.
With the rapid growth of AI techniques, this discussion has gradually highlighted the existence of a higher level of intelligence when contrasting the views of ‘interaction as tool use’ and ‘interaction as dialogue’ (Hornbæk and Oulasvirta 2017), arguing that utility and usefulness are the main aspects for the first type of interaction to work, while having a constant, simple, direct and natural communication and understanding between human and computer, is key for the second type of interaction to work. 訳抜け防止モード: AI技術の急速な成長に伴い、この議論は、‘ツール使用としてのインタラクション’と‘対話としてのインタラクション’(HhornbækとOulasvirta 2017)の見解を対比して、より高度なインテリジェンスの存在を徐々に強調している。 実用性と有用性は、機能する最初のタイプのインタラクションの主要な側面である、と論じます。 人間とコンピュータの間の 絶え間なく 単純で 直接的で 自然なコミュニケーションと 理解を持ちながら 第二のタイプのインタラクションの鍵です。
0.73
This distinction of interacting with a machine considered as a creative intelligence rather than a tool is a key motivation for this work.
For this broader view, exposing the creative process is crucial.
このより広い視点では、創造的なプロセスを公開することが重要です。
0.58
Specifically we hypothesise that establishing two-way communication channels, within a co-creative human-machine partnership, where the creative process is transparent as well as discussed, would improve interactions, build up trust on CC systems and encourage human engagement.
Overview on the state-of-the-art of XAI in CC In (Zhu et al 2018) the authors defined a new area of XAI which they called Explainable AI for Designers (XAID).
CC In (Zhu et al 2018)におけるXAIの現状の概要 著者らは、XAIの新しい領域を定義し、その領域を Explainable AI for Designers (XAID) と呼んでいる。
0.80
The objective of XAID is to support games designers in specific design tasks.
XAIDの目的は、特定のデザインタスクでゲームデザイナーをサポートすることである。
0.71
In their work they identified three spectra
彼らの研究で 3つのスペクトルを特定しました
0.46
that describe co-creation in the setting of games design:
ゲームデザインの設定における共同創造を記述します
0.67
i) spectrum of explainability, which ranges from understanding of the underlying operation of AI techniques to understanding of the input-output pattern,
Relevant to this discussion is also the concept of framing as proposed in (Cook et al 2019).
この議論に関連するのは、提案されているフレーミングの概念である(Cook et al 2019)。
0.72
Framing, as has been applied to date within the CC community, is mostly intended as a ‘final interaction’; i.e. to accompany an output with the expectation that it will increase its perceptive value.
However, in (Cook et al 2019) the authors propose advocacy and argumentation as a potential purpose for framing.
しかしながら、著者らは(Cook et al 2019)において、フレーミングの潜在的な目的として擁護と議論を提案している。
0.54
Both of the works mentioned above are relevant to our work and are a step towards the vision of CC systems playing a more active role in creative collaborations; however, XCC focus spreads not only to co-creation but also to other interactions that occur when producing a creative act; such as setting up an initial goal, delivering the product to a final user, producing feedback, etc.
Moreover, we ground the interventions of a CC system not only on the current act of creation in which it is involved, but also in past experiences (i.e. we propose that CC systems should have a memory of their work).
Additionally, we also adopt the notion of argumentation and advocacy as a role for XCC; but our model, proposes this role not only as a way to support a creative artefact, but also as a way to increase their involvement since the conception of an idea towards the production of it.
In the next section we outline some design principles for systems with XCC capabilities.
次の節では、XCC機能を持つシステムの設計原則について概説する。
0.67
We use a running example to illustrate our ideas using linguistic communication as the primary medium for explainability; however, we consider communication in its broader sense, not just through linguistic forms.
The main objective of the design principles outlined next is to enable both CC systems and their human users to communicate with each other so that there is a common and clear understanding throughout their interactions.
We have drawn from a range of research that looks at human collaboration, teamwork, cognitive science and psychology, as well as from our experience on the development of computational creative systems.
Mental models: Are representations of key elements of the creative environment that help conceptualize, understand and construct expectations of how things work and how individuals interact within a creative collaboration (McCormack et al 2020; Mohammed, Ferzandi, and Hamilton 2010).
メンタルモデル(McCormack et al 2020; Mohammed、Ferzandi、Hamilton 2010)は、創造的な環境における重要な要素の表現であり、物事の仕組みや個人が創造的なコラボレーションの中でどのように相互作用するかを概念化し、理解し、構築するのに役立つ。 訳抜け防止モード: メンタルモデル : 創造的環境の重要な要素の表現 物事の仕組みに対する期待を概念化し、理解し、構築する そして、個人が創造的なコラボレーション(mccormack et al 2020; mohammed, ferzandi, hamilton 2010)の中でどのように相互作用するか。
0.74
The concept of a mental model comes from psychology and cognitive science research (Craik 1967; Johnson-Laird 1993) and has more recently been successful
Relevance to CC: The idea of mental models as a key aspect for the design of real-time co-creative systems has been highlighted previously in (McCormack et al 2020).
cc: リアルタイム共同創造システムの設計における重要な側面としてのメンタルモデルの概念は、以前にも強調されていた(mccormack et al 2020)。
0.75
Mental models consider a broad set of aspects of human interactions that would aid the understanding of essential elements within a collaboration and of the interactions throughout it.
CC systems can use these representations, not only to understand the operation of their co-creators, the environment and the domain, but also to reason about them and seek different (creative) ways of interaction by questioning them; for instance by playing a bossa nova style chord progression when improvising with a musician with a strong rock background (i.e. challenging preferences).
Shared mental models can increase CC systems’ awareness of processes within the collaboration.
共有精神モデルは、コラボレーション内のプロセスに対するccシステムの認識を高めることができる。
0.65
For instance, in order to achieve cognitive convergence when the mental models of the collaborators do not match (e g when they are not moving towards the same goal) (Fuller and Magerko 2010).
Equipping CC systems with mental models, of both themselves and of their co-creators, not only would enable better coordination when trying to come up with something new, but would also provide a valuable resource for CC systems to explain, justify and defend their contributions.
We believe this increases the capacity, of both CC systems and their users to generate appropriate and complementary output.
これにより、CCシステムとユーザの両方が適切な補完的な出力を生成する能力が向上すると考えています。
0.58
As pointed out in (Br¨uhlmann et al 2018), people are more motivated to use a specific technology when it is congruent with their personal values, goals, and needs.
氏が指摘した(br suhlmann氏ら2018年)ように、人々は、個人的な価値観、目標、ニーズに合致するときに、特定のテクノロジを使用するモチベーションが高い。 訳抜け防止モード: 指摘されているように(Br suhlmann et al 2018)。 人々はより動機づけられ 個人的価値や目標、ニーズと調和した特定の技術を使用すること。
0.72
Features: Relevant elements of mental models would involve team aspects such as goals, roles, capabilities, expectations, etc., domain aspects such as conventions governing the operation of particular domains, stakeholders profiles, relationships between elements of the domain, etc., as well as interpersonal aspects of individuals such as preferences and how they communicate.
For instance, improvisational settings are governed by implicit interactions with subtle signs and cues, while in an advertising campaign setting, for instance, participants can explicitly communicate their intentions, discuss their ideas and establish agreements.
Example: Imagine for instance the following interaction between an Advertising Executive (AE) and an XCC system (XCCS) when collaborating to design an advert for a toothpaste (goal) in which the clients would like to emphasise the ideas of teeth and decay (domain information): XCCS: How about the image in Fig 2?
Long-term memory: cessing information of past experiences and interactions.
長期記憶:過去の経験と相互作用の情報を停止する。
0.77
Is the capacity of storing and ac-
ストレージとACの能力は?
0.55
Relevance to CC: Studies in cognitive psychology have found that memory is a crucial element of creativity and that an important part of the creative process happens by drawing on previous experiences and the information we have in our memories (Madore, Addis, and Schacter 2015).
Being able to explain their decisions and support its ideas requires CC systems to have a memory of the processes, decisions and interactions they have undertaken in current and previous collaborations.
Equipping CC systems with this ability would enhance the creative capacity of their collaborations.
この能力でCCシステムを取得することで、コラボレーションの創造的能力が向上する。
0.65
Possessing a memory would also serve other purposes such as breaking habits and avoiding repetition or mistakes.
メモリの使用は、習慣を破る、繰り返しや間違いを避けるといった他の目的にも役立ちます。
0.61
Features: Enabling a bi-directional communication would provide opportunities for CC systems to store different types of information, such as failed attempts, successful artefacts, strategies used, temporal information, users’ reactions, etc.
How to store and access this information is an important aspect to consider here.
この情報を保存してアクセスする方法は、ここで考慮すべき重要な側面である。
0.64
For instance, one methodology could be as defined in (Davis et al 2015), where the authors use the concept of perceptual logic to classify information in a way that aids co-creation as follows: in local perceptual logic the system only considers specific details (such as a line in a drawing), in the regional perceptual logic the information is grouped into clusters (e g straight lines, lines that are close to each other, etc.), while in the global perceptual logic, the system considers relationships between regions (e g identifying that the left hand side of a drawing has fewer lines than the right hand side).
For instance, one methodology could be as defined in (Davis et al 2015), where the authors use the concept of perceptual logic to classify information in a way that aids co-creation as follows: in local perceptual logic the system only considers specific details (such as a line in a drawing), in the regional perceptual logic the information is grouped into clusters (e g straight lines, lines that are close to each other, etc.), while in the global perceptual logic, the system considers relationships between regions (e g identifying that the left hand side of a drawing has fewer lines than the right hand side). 訳抜け防止モード: 例えば、ひとつの方法論は(davis et al 2015)で定義することができる。 著者は知覚論理という概念を使い co - 生成を支援する方法で情報を分類する :局所知覚論理では、システムは特定の詳細(図中の線など)のみを考慮に入れる。 地域知覚論理では、情報はクラスタ(例えば直線)にグループ化される。 互いに近い線など) 大域的な知覚論理では、システムは領域間の関係を考える(例えば、図面の左辺が右辺よりも直線が少ないことを識別する)。
0.63
Depending on the domain and the purpose of a CC system in that domain, different mechanisms may use to handle such memory.
Example: Let us take for instance the example of the AE and the XCCS working on the toothpaste campaign.
例:歯磨き粉キャンペーンに取り組んでいるaeとxccsの例を例にとりましょう。
0.61
The AE did not like the idea for the ad even after the XCCS provided an explanation.
AEはXCCSが説明を提供した後も広告のアイデアを好まなかった。
0.74
The system can then review past experiences and find additional strategies to support its idea, in doing so the system finds that in a previous interaction another AE used a tag line to clarify an abstract concept for an ad.
The interaction then follows: XCCS: I think the connection is good.
XCCS: コネクションは良いと思います。
0.37
There is a strong
強い人がいます
0.74
英語（論文から抽出）
日本語訳
スコア
surface similarity between teeth and dice and the idea is surprising as the two concepts are not normally associated; maybe we could add a tagline to support the image?
Something related to the not apparent connection between teeth, gambling and money?
歯とギャンブルとお金の間に 明らかでない関係があるのか?
0.76
AE That is good.
AE それは良いことです。
0.79
I like the idea of using a tagline.
私はタグラインを使うアイデアが好きです。
0.74
Maybe something like “There are some things you just can’t afford to gamble with”.
おそらく“ギャンブルする余裕のないものがいくつかある”ようなものです。
0.65
AE: There may be something here, but I’m not sure yet...
AE: 何かあるかもしれませんが、まだわかりません。
0.67
Argumentation: Is the process of reasoning-about and supporting specific contributions within a creative collaboration.
議論:創造的なコラボレーションの中で特定の貢献を推論し支援するプロセスです。
0.74
Relevance to CC: A fruitful creative partnership enables participants to both explain and justify their ideas in order to unveil or clarify their creative value.
Features: As pointed out in (Cook et al 2019), argumentation provides a set of very valuable resources that can be used by CC systems to enhance communication with their users.
特徴: (Cook et al 2019)で指摘されているように、議論はCCシステムによってユーザとのコミュニケーションを強化するために使用できる非常に貴重なリソースセットを提供する。
0.78
Take for instance the theory of critical questions (Walton, Reed, and Macagno 2008), which helps anticipate questions or concerns that may arise in specific situations with different stakeholders, as well as the mechanisms to try and address possible conflicts that these questions or concerns may arise within a creative collaboration.
Ultimately this part of the process involves an attribution or cognitive process that requires different dimensions to be taken into account.
究極的には、この過程のこの部分は、異なる次元を考慮に入れる必要がある帰属または認知過程を含む。
0.66
Is it a one-shot argument, or is it the start of a dialogue?
ワンショットの議論なのか、それとも対話の始まりなのか?
0.72
Is the argument targeted for a co-creator, an end user, a member of the audience?
議論の対象は共同制作者、エンドユーザー、オーディエンスのメンバーなのか?
0.60
What elements of the mental model should most influence the selection of arguments?
メンタルモデルのどの要素が議論の選択に最も影響を与えるべきか?
0.78
When is the right moment to push or to stop?
プッシュや停止する適切なタイミングはいつですか?
0.71
What kind of language, visualisation or medium should be used to convey the argument?
議論を伝達するためにどのような言語、可視化、媒体を使うべきか?
0.74
What does the system want to achieve with it: clarification, persuasion, feedback, etc.?
明確化、説得、フィードバックなど、システムは何を達成したいのか?
0.62
Example: Take for instance the example of the AE and the XCC system in the toothpaste advert campaign.
例:歯磨き広告キャンペーンにおけるAEおよびXCCシステムの例を例に挙げる。
0.73
After the second interaction, when the system proposes the use of a tagline to clarify the image, the AE is willing to give a chance to the system’s idea.
A further interaction could go this way: XCCS: The idea and image are somewhat repulsive and repulsive adverts have a shock value which help people remember them.
com/watch? v=AOph5V78oxs (video showing an advert warning about the addictive nature of cocaine).
と? v=AOph5V78oxs(コカインの中毒性について警告するビデオ)。
0.51
AE: Hmmm... possibly. Without the system pushing back a little, it is easy to imagine that a user may stop engaging after the first generative act, thus missing out on the concept.
Exposing the creative process: consists of opening up the environment and exposing the steps, assessments, metrics, influences, etc. that constitute the processes and decisions within the operation of a domain.
Relevance to CC: Providing an explanation for a process or a decision is a useful way to obtain a better understanding of what is going on within a closed environment; however, descriptions or clues envisaged for this purpose may sometimes fall short in aiding that understanding.
Opening up the environment and exposing the steps, assessments, influences, etc. that constitute those processes and decisions would provide a clearer view of how a CC system works.
We believe that observing, feeling, or in some way sensing the underlying structures of a CC system, instead of being told how things work, may trigger thought processes in the mind of their co-creators that may result in benefits for the creative collaboration.
Features: Exposing the creative process requires the models and mechanisms used by a system to be interpretable so that they can be easily and unambiguously communicated to others.
The development of interpretable models is being encouraged in the AI community because of the inherent problems with unexplainable models, such as unfaithful accounts of their computations (Rudin 2019).
In other words, the way a CC system communicates must be simple and precise, yet the communication needs to be meaningful so that the new information helps progress the collaboration.
An example of this is provided in (McCormack et al 2019) where the authors equip an AI musician (whose underlying model consists of a neural network) with the ability to continuously communicate how confident it feels during an improvised performance.
この例は (McCormack et al 2019) で提供されており、著者らはAIミュージシャン(基礎モデルはニューラルネットワークで構成されている)に即興のパフォーマンスにおける自信を継続的に伝える能力を与えている。
0.77
Human performers also implicitly communicated their confidence to the computer via biometric signals.
また、人間のパフォーマーは生体信号を通じてコンピュータに自信を伝える。
0.68
The work showed that this type of simple, interpretable, communication increased the flow within the human-AI collaboration and the quality of the music produced.
Example: Let us take for instance the example of the AE and the XCC system working on the toothpaste campaign.
例:例えば、歯磨き粉キャンペーンに取り組んでいるaeとxccシステムの例を見てみましょう。
0.66
The AE is still not sure about the idea for the ad.
AEはまだ、この広告のアイデアについて定かではない。
0.79
The system can then expose its reasoning even further so that the AE has a better understanding where the ideas come from.
システムはさらに推論を公開し、AEがアイデアのどこから来たのかをよりよく理解できるようにします。
0.75
Imagine for instance the CC system has an interface that allows the user to investigate the underlying structures behind its ideas through the mean of a visual graph representation of the system’s knowledge base.
By seeing the connections among the concepts in the knowledge base, the user realises that the two related concepts; i.e. dice and teeth, are not actually directly connected and sees this as an opportunity to improve the concept of the ad.
How about blending the two similar concepts: dice and teeth?
ジスと歯の2つの類似概念を混ぜてみましょうか。
0.63
XCCS: Ok. I’ve made the images in Fig 3.
xccs: はい。 画像はFig 3で作成しました。
0.59
I prefer the first one, as the cube shape is clearer.
立方体がより鮮明になるので、私は最初の方を好む。
0.68
AE: I prefer that one too.
AE: 私もそれが好きです。
0.83
I’ll adapt it and add tagline, brand and product information...
私はそれに適応して、タグライン、ブランド、製品情報を追加します。
0.62
Here is the final advert, in Fig 4.
以下は、第4四半期の最後の広告だ。
0.57
英語（論文から抽出）
日本語訳
スコア
Figure 3: Second attempt.
図3: 第2の試み。
0.75
Figure 4: Final result.3 A framework for XCC Figure 5 summarises the kind of interactions of the proposed approach based on the four principles presented in this paper.
Instantiating these interactions naturally depends on the creative domain, the stakeholders involved, and the stage of the creative process where the interventions occur.
The domain is the most important factor to determine the kind of medium for communication.
ドメインはコミュニケーションのためのメディアの種類を決定する上で最も重要な要素です。
0.79
In music, for instance, communication may most likely occur through the music itself – imagine for instance a CC musical composer demonstrating through a virtual keyboard how a pianist should emphasise a particular phrase –, while in painting this communication may occur through brush strokes – imagine now a CC painter-collaborator that paints all over a section of a painting they think should be emphasised (McCormack et al. 2020).
例えば、音楽においては、コミュニケーションは音楽を通して起こる可能性が最も高い - 例えば、CC音楽作曲家が仮想キーボードを通してピアニストが特定のフレーズを強調すべきかどうかを実演するなど — と、このコミュニケーションは、ブラシストロークを通して起こる可能性がある – 現在では、彼らが強調すべきと思われる絵画のセクション全体にペンキを塗るCCパフォーマーが想像できる(McCormack et al. 2020)。
0.74
The stakeholders greatly influence the sources and type of information that these interactions would require.
利害関係者はこれらの相互作用に必要な情報源や種類の情報に大きく影響します。
0.72
A cocreator may need technical details of the operation of a system – an ideation system may provide a tree representation of the relevant concepts from the knowledge base from which an idea was produced (as in the advertising example), while for an end-user the intuition would probably be more useful – for instance a CC poetry composer that explains the mood reflected in the poem because it read sad news in the newspaper (the end-user does not need to know, for instance, the technical mechanism of sentiment analysis used).
A cocreator may need technical details of the operation of a system – an ideation system may provide a tree representation of the relevant concepts from the knowledge base from which an idea was produced (as in the advertising example), while for an end-user the intuition would probably be more useful – for instance a CC poetry composer that explains the mood reflected in the poem because it read sad news in the newspaper (the end-user does not need to know, for instance, the technical mechanism of sentiment analysis used). 訳抜け防止モード: コクリエータはシステムの操作の技術的な詳細を必要とするかもしれない。イデオレーションシステムは、知識ベースから関連する概念のツリー表現を提供するかもしれない。 アイデアが生み出された(広告の例のように)。 例えば、ccの詩人が詩に反映される感情を説明するのは、新聞で悲しいニュースを読んだからである(最後には、ユーザは必要ない)。 例えば、感情分析の技術的なメカニズムを知るには、)。
0.71
The stage of the creative process is most influential on how the elements of the design principles are managed.
創造的なプロセスのステージは、デザイン原則の要素がどのように管理されるかに最も影響します。
0.77
In a preparatory stage, when the collaboration is just starting, the interactions between CC systems and their users help set up the context of what the collaboration is about; i.e. a
3Original advert taken from http://bit.ly/2uucLq R
http://bit.ly/2uuclq rから取られた3つのオリジナル広告
0.34
shared mental model is agreed – take for instance the motivating example of the SpeakeSystem: this stage would have allowed the system to inform the human musician about its reset function, which could have avoided the human musician wondering why the system didn’t make sense at some points during the performance.
In a co-creation stage the interactions require an iterative process that involves a constant revision of the mental model in order to ensure that the collaboration is converging towards the same goal, the addition of new memories (which reflect the experience of the current interactions), access to old memories, and possibly various cycles of generative acts – here for instance, the human musician working with the SpeakeSystem could have signalled during the performance (through a visual cue or facial expression) when he wanted the system to challenge him instead of only responding to him.
In a co-creation stage the interactions require an iterative process that involves a constant revision of the mental model in order to ensure that the collaboration is converging towards the same goal, the addition of new memories (which reflect the experience of the current interactions), access to old memories, and possibly various cycles of generative acts – here for instance, the human musician working with the SpeakeSystem could have signalled during the performance (through a visual cue or facial expression) when he wanted the system to challenge him instead of only responding to him. 訳抜け防止モード: 共同創造の段階では、コラボレーションが同じ目標に向かって収束していることを保証するために、メンタルモデルの継続的な修正を含む反復的なプロセスが必要です。 新たな記憶の追加(現在の相互作用の経験を反映する) 古い記憶へのアクセス、そしておそらく様々な生成的行為のサイクル ― 例えば、 SpeakeSystemで作業する人間ミュージシャンは、パフォーマンス中に合図した可能性がある (視覚的なキューや表情を通して) 彼にだけ答えるのではなく システムに挑戦したかった時
0.79
Finally, a post-creative stage provides an outlet to present the artefact as well as for feedback and reflexion.
Such an outlet may also represent an opportunity for the revision of the mental models and for adding new memories to the system – here for instance the SpeakeSystem could have provided an explanation to the member of the audience who wanted to know what was going on at a certain point of the performance:
Human annotator: “I wonder what you are both thinking going into this section. The algorithm not a lot I suspect! Otherwise it would play notes” System: I was enjoying what my partner was playing here.
For example, in an improvised music performance, it would be inappropriate for participants to stop playing and start talking about why one made a particular musical decision in the previous bar.
Often visual, sonic, haptic or even olfactory communication may be the most suitable, and the form of communication can influence the creative outcomes in subtle, non-obvious ways.
For example, the smell of freshly baked bread or a spring meadow can trigger specific memories or support synaesthetic-like relationships to other sensory media (Ackerman 1990).
Discussion Challenges and Opportunities Different issues surround the idea of explainable models.
議論の課題と機会 異なる問題は、説明可能なモデルの概念を取り巻く。
0.69
A concern that has been raised in the XAI field is that explainable models are not accurate representations of the actual functionality of the system in some aspects of the feature space.
According to (Rudin 2019) “an inaccurate (lowfidelity) explanation model limits trust in the explanation, and by extension, trust in the black box that it is trying to explain”.
Although we believe that CC provides an outlet in which explanations do not have to be faithful to the intrinsic motivations/objectiv es of a system (as has been postulated through the notion of framing in CC), we need to be careful so as not to endanger the trust that co-creators, users and
Figure 5: XCC design principles: the creative process is shaped by relevant past experiences and the shared mental model, which is itself shaped by the arguments exchanged with the user as well as by relevant past experiences.
The long-term memory is updated with new experiences of the creative process, and the operation of the system is exposed through different interfaces that allow the user to understand underlying procedures of how the system works.
For instance, the CC advertiser in our example may be involved in pitching the concept of the ad to the client.
例えば、この例のCC広告主は、クライアントに広告の概念を売り込むのに関わっているかもしれない。
0.79
In the process it may provide intuitions behind the campaign but in doing so it should not take credit for aspects of it that were actually ideas of its co-creators (e g the idea of blending dice and teeth in the example).
Another common challenge is that explainable systems have a social responsibility.
もうひとつの一般的な課題は、説明可能なシステムが社会的責任を持つことだ。
0.50
This has been explored extensively in (Arrieta et al 2020) where the concept of Responsible AI is introduced together with some practical principles (such as fairness, transparency, and privacy).
責任あるaiの概念は、いくつかの実践的な原則(公正性、透明性、プライバシーなど)とともに導入されている(arrieta et al 2020)。
0.62
For instance, one thing is for the CC advertiser in our example to refer to gambling as an analogy to endangered oral health, and another (questionable possibility) is to use it to encourage gambling.
Finally, there must be clear limits about the extent of which a CC system may try to champion an idea; i.e. CC systems cannot disrupt or take control over a creative process.
Imagine for instance the CC advertiser in our example not backing down about using the first image of the poker table even tough its co-creator had already rejected it.
Despite these challenges, explainable models, as suggested here, may open up possibilities of new kinds of social interactions between CC systems and their users.
Work from the social sciences has shown that non-human entities can take up active roles in social practices and that imposing concrete boundaries or definitions between what roles these can or cannot play only limits their potential.
Strengers (Strengers 2019) illustrates this point by looking at roomba riding, a trend that refers to how pets enjoy ‘riding’ a robotic vacuum cleaner called the Roomba, and the potential of this type of technology to become a pet entertainment device.
Imagine for instance a CC collaborator that strongly suggests its human co-creator to stop working during the weekend to be with his/her family, or a human co-creator that opens up a co-creative collaboration with a CC system on social media, or a CC system that proposes its human cocreator to watch a film together in order to have a break and maybe get some inspiration.
Equipping CC systems with explainable capabilities may reveal a type of dynamics between humans and CC systems that is not possible with most current technologies.
Conclusions and Future Work We have presented Explainable Computational Creativity (XCC), as a sub-field of XAI that is focused on the applicability of explainable models in the area of CC and how these can be used to open up bidirectional communication channels between CC systems and their users.
We have outlined four design principles we believe are crucial for these types of models, namely mental models (individual representations of how things work), long-term memory (the capacity to store and access details of past experiences), argumentation (the ability to reason and support creative contributions), and exposing the creative process (revealing specific details about the operation of a system).
We believe that this framework will benefit human-CC interactions by:
我々は、この枠組みが人間とCCの相互作用に役立つと信じている。
0.47
i) enhancing their creativity (as a result of working together rather than in isolation),
一 創造性を高めること(孤立するよりも一緒に働くことによること)
0.71
ii) creating more fruitful and productive partnerships (by establishing bi-directional communication channels where creative contributions can be examined), and
The Nature of Explanation. Cambridge, UK: Cambridge University Press.
解説の性質。 ケンブリッジ大学出版局(Cambridge University Press)。
0.61
[Davis et al 2015] Davis, N.
Davis et al 2015] Davis, N.
0.32
; Hsiao, C. -P.
Hsiao, C。 -P。
0.36
; Popova, Y.
とPopova, Y。
0.72
; and Magerko, B. 2015.
and magerko, b. 2015年。
0.76
An Enactive Model of Creativity for Computational Collaboration and Co-creation.
計算コラボレーションと共同創造のための創造性に関するエンアクティブモデル
0.81
London: Springer London.
ロンドン:スプリンガー・ロンドン。
0.83
109–133. [Fuller and Magerko 2010] Fuller, D., and Magerko, B. 2010.
109–133. [Fuller and Magerko 2010] Fuller, D. and Magerko, B. 2010]
0.38
Shared mental models in improvisational perforIn Proceedings of the Intelligent Narrative Techmance.
インテリジェント・ナラティブ・テクマンスの即興的パーフォリン・プロシージャにおける共有精神モデル
0.58
nologies III Workshop, INT3 ’10, 15:1–15:6.
nologies III Workshop, INT3 ’10, 15:1–15:6。
0.34
ACM. and [Hornbæk and Oulasvirta 2017] Hornbæk, K., In ProOulasvirta, A. ceedings of the 2017 CHI Conference on Human Factors in Computing Systems, CHI ’17, 5040–5052.
acm。 そして[Hornbæk and Oulasvirta 2017] Hornbæk, K., In ProOulasvirta, A. ceedings of the 2017 CHI Conference on Human Factors in Computing Systems, CHI ’17, 5040–5052。
0.54
ACM. 2017.
acm。 2017.
0.53
What is interaction? 2019.
相互作用とは何か? 2019.
0.53
[Johnson-Laird 1993] Johnson-Laird, P. N. 1993.
ジョンソン=レイド(Johnson-Laird 1993) ジョンソン=レイド(Johnson-Laird, P. N. 1993)
0.41
Human and Machine Thinking. Lawrence Eribaum Associates.
人間と機械の思考。 ローレンス・エリバウム・アソシエイツ所属。
0.59
[Kammersgaard 1988] Kammersgaard, J. 1988.
Kammersgaard(1988年) Kammersgaard(1988年)
0.69
Four differInternaent perspectives on human-computer interaction.
人間とコンピュータの相互作用に関する4つの異なる視点。
0.53
tional Journal of Man-Machine Studies 28(4):343–362.
対人機械研究誌28(4):343–362。
0.75
[Krug 2014] Krug, S. 2014.
[Krug 2014] Krug, S. 2014
0.42
Don’t make me think, , Revisited.
考えさせてはいけない。
0.33
A Common Sense Approach to Web Usability.
web ユーザビリティに対する常識的なアプローチ。
0.69
New Riders, 3rd edition edition. [Madore, Addis, and Schacter 2015] Madore, K. P.
新版、第3版。 [Madore, Addis, and Schacter 2015]Madore, K.P.
0.45
; Addis, D. R.
アディス, d. r。
0.57
; and Schacter, D. L. 2015.
とschacter, d. l. 2015。
0.82
Creativity and memory: Effects of an episodic-specificity induction on divergent thinking.
創造性と記憶: エピソード特異性誘導が発散思考に及ぼす影響
0.75
Psychological Science 26(9):1461–1468.
心理学26(9)1461-1468。
0.75
[McCormack et al 2019] McCormack, J.
McCormack et al 2019] McCormack, J.
0.37
; Gifford, T. ; Hutchings, P.
ギフォード、t。 ; Hutchings, P。
0.45
; Rodriguez, M. T. L.
ロドリゲス(Rodriguez, M. T. L.)。
0.49
; Yee-King, M.
イエ・キング (Ye-King) M。
0.39
; and d’Inverno, M. In a silent way: Communication between AI and improvising musicians beyond sound.
サイレントな方法で:AIと音楽家とのコミュニケーション。
0.35
In Brewster, S. A.
ブリュースター、S.A。
0.49
; Fitzpatrick, G.
フィッツパトリック、G。
0.42
; Cox, A. L.
cox (複数形 coxs)
0.38
; and Kostakos, V., eds.
とkostakos, v., eds。
0.54
, Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems, CHI 2019, Glasgow, Scotland, UK, May 04-09, 2019, 38.
2019 CHI Conference on Human Factors in Computing Systems, CHI 2019, Glasgow, Scotland, UK, May 04-09, 2019, 38。
0.38
ACM. [McCormack et al 2020] McCormack, J.
acm。 McCormack et al 2020] McCormack, J.
0.50
; Hutchings, P.
; Hutchings, P。
0.40
; Gifford, T. ; Yee-King, M.
ギフォード、t。 イエ・キング (Ye-King) M。
0.44
; Llano, M. T.
Llano, M. T。
0.39
; and D’inverno, M. 2020.
とd’inverno, m. 2020。
0.77
Design considerations for real-time collaboration with creative artificial intelligence.
創造的人工知能とのリアルタイムコラボレーションのための設計考察
0.69
Organised Sound 25(1):41–52.
組織音25(1):41-52。
0.70
[Mohammed, Ferzandi, and Hamilton 2010] Mohammed, S.
〔mohammed, ferzandi, and hamilton 2010〕mohammed, s。
0.65
; Ferzandi, L.
フェルザンディ (Ferzandi, L。
0.66
; and Hamilton, K. 2010.
2010年、ハミルトン。
0.44
Metaphor no more: A 15-year review of the team mental model construct.
Metaphor No more: チームのメンタルモデル構築に関する15年間のレビュー。
0.77
Journal of Management 36(4):876–910.
行政誌36(4):876-910。
0.75
[Muggleton et al 2018] Muggleton, S. H.
[Muggleton et al 2018]Muggleton, S. H。
0.45
; Schmid, U.
英: Schmid, U。
0.41
; Zeller, C.
Zeller, C。
0.31
; Tamaddoni-Nezhad, A.
とTamaddoni-Nezhad, A。
0.36
; and Besold, T. 2018.
とBesold, T. 2018。
0.62
Ultra-strong machine learning: comprehensibility of programs learned with ilp.
Ultra-strong Machine Learning: ilpで学んだプログラムの理解性。
0.75
Machine Learning 107(7):1119– 1140.
機械学習107(7):1119–1140。
0.88
[Nickerson, Elkind, and Carbonell 1968] Nickerson, R. S.
[ニッカーソン、エルキンド、カーボネル1968]ニッカーソン、r.s.
0.58
; Elkind, J. I.
エルキンド (Elkind, J. I)。
0.60
; and Carbonell, J. R. 1968.
およびcar carbonell, j. r. 1968。
0.78
Human factors and the design of time sharing computer systems.
ヒューマンファクターと時間共有コンピュータシステムの設計
0.56
Human Factors 10(2):127–133.
ヒト因子10(2):127-133。
0.67
[Norman 1982] Norman, D. A. 1982.
1982年ノーマン、D.A.1982。
0.58
Learning and memory. Number xi, 129.
学習と記憶。 背番号は129。
0.43
San Francisco: W.H. Freeman.
サンフランシスコ - W.H.フリーマン。
0.68
[Rudin 2019] Rudin, C. 2019.
[Rudin 2019]Rudin, C. 2019。
0.39
Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead.