A DEVELOPMENT FRAMEWORK FOR A CONVERSATIONAL AGENT TO EXPLORE MACHINE LEARNING CONCEPTS

Machine Learning and Applications: An International Journal (MLAIJ) Vol.8, No.2/3, September 2021
DOI:10.5121/mlaij.2021.8301 1
A DEVELOPMENT FRAMEWORK FOR A
CONVERSATIONAL AGENT TO EXPLORE MACHINE
LEARNING CONCEPTS
Ayse Kok Arslan
Oxford Alumni of Northern California, Santa Clara, USA
ABSTRACT
This study aims to introduce a discussion platform and curriculum designed to help people understand how
machines learn. Research shows how to train an agent through dialogue and understand how information
is represented using visualization. This paper starts by providing a comprehensive definition of AI literacy
based on existing research and integrates a wide range of different subject documents into a set of key AI
literacy skills to develop a user-centered AI. This functionality and structural considerations are organized
into a conceptual framework based on the literature. Contributions to this paper can be used to initiate
discussion and guide future research on AI learning within the computer science community.
KEYWORDS
Machine learning, visual editing, construction, neural nets, artificial intelligence
1. INTRODUCTION
Online discourse in cyberspace (mainly textual data) raises methodological problems, including
data collection and coding (Vieira, Parsons, & Byrd, 2018) when it comes to analyzing it. It is
often not enough to measure users’ cognitive engagement only by observable indicators, but
combining quantitative data and qualitative content analysis will provide more reliable results
(e.g., Atapattu, Thilakaratne, Vivian, & Falkner, 2019). Therefore, this study presents an
automatic discourse analysis approach using a language modeling technique called ‘Neural Word
Embeddings’ (Word2vec) (Mikolov, Sutskever, Chen, Corrado, & Dean, 2013) to detect users’
active participation in discussions in cyberspace.
This study examines the use of a chat agent with four consecutive AI modules, as well as testing
machine learning and information representation. This study provides a design framework for a
program related to training content. While previous knowledge of chat agents and other factors
can affect a user’s ability to understand, the basic assumption is that all users can benefit greatly
from the interaction with the agent and chat. Within this context, the following ideas are based
on:
1. Collaborating with the agent helps to increase participation in the whole process.
2. Agent recognition and training content lead to user learning and understanding of how
machines learn.

2
2. RELATED WORK
When it comes to human-technology relations there are various theoretical frameworks
(Heersmink, 2012) which share the basic idea that technical objects are projections or extensions
of human organism “by way of replicating, amplifying, or supplementing bodily or mental
faculties or capabilities” (Lawson, 2010, p. 2).
With the growing scale of text data, discourse analysis has become a key application area of user
analytics and data mining. While the number of exact duplicates and shares of a message can be
used as a proxy for popularity, discovering and grouping together multiple messages making the
same claims in different ways can give a more accurate view of prevalence. The ability to group
claim-matched textual content in different languages would enable fact-checking organizations
around the globe to prioritize and scale up their efforts to combat misinformation.
Semantic textual similarity (STS) refers to the task of measuring the similarity in meaning of
sentences, and there have been widely adopted evaluation benchmarks including the Semantic
Textual Similarity Benchmark (STS-B) (2017; 2016; 2015; 2014; 2013; 2012) and the Microsoft
Research Paraphrase Corpus (MRPC) (Dolan and Brockett, 2005). The STS-B benchmark
assigns discrete similarity scores of 0 to 5 to pairs of sentences, with sentence pairs scored zero
being completely dissimilar and pairs scored five being equivalent in meaning. The MRPC
benchmark assigns binary labels that indicate whether sentence pairs are paraphrases or not.
Semantic textual similarity is a problem still actively researched with a dynamic state of the art
performance.
Also, given the recent developments in AI, for instance, today’s natural language processing
systems have come a long way toward solving many different problems, such as translation, text
generation, and question-answering on specific problems. Yet, the AI community still hasn’t
solved the problem of creating agents that can engage in open-ended conversations without losing
coherence over long stretches. Such a system requires more than just solving smaller problems; it
requires common sense, one of the key unsolved challenges of AI. To give a specific example,
the General Language Understanding Evaluation (GLUE) benchmark, developed by some of the
most esteemed organizations and academic institutions in AI, provides a set of tasks that help
evaluate how a language model can generalize its capabilities beyond the task it has been trained
for. Yet, if an AI agent gets a higher GLUE score than a human, it doesn’t mean that it is better at
language understanding than humans. According to the influential AI scholar Mitchell, capturing
such knowledge is the current frontier of AI research. As Mitchell states, while we still don’t
know the answers to many of these questions, a first step toward finding solutions is being aware
of our own erroneous thoughts to develop more robust, trustworthy, and perhaps intelligent AI
systems.
To give some examples, Google's Teachable Machine [4] provides a web page where users can
train the image classification system while TensorFlow provides a playground to collaborate,
train, and test an in-depth neural learning net.
 Learning with AI: While most AI educational programs are the latest in a long line of ideas,
the idea of introducing people to AI ideas began to work with Seymour Papert and Cynthia
Solomon using the LOGO program and the Turtle robot [3, 5, 7, 9]. It has served as the
basis for much of the current work. Many platforms teach AI by having an individual
program in block-based languages including Cognimates [23], Machine Learning for Kids
[29, 17] and eCraft2Learn [27]. Other platforms introduce AI within the context of robots,

3
such as Popbots [19, 21], as well as performance enhancements [3] and MIT App Inventor
AI Extensions [13, 17, 19].
 Learning with Conversational Agents: Conversation agents are also used for learning, often
as intelligent educators and learning friends [5]. The structure of these communication
systems varies greatly, from incorporations and results based on the text [6] to integrated
agents that can express emotions [7].
Researchers exploring how to engage younger students in design activities involving AI [4]
identified five "major ideas" of AI to guide the development of standards:
1) "Computers detect the earth using sensors";
2) "Agents maintain models / representation of land and use it for consultation";
3) "Computers can learn from data";
4) “Making social media agents is a major challenge for AI engineers”; and
5) "AI applications can affect society in both positive and negative ways" [13]].
The model development starts with the following questions about AI:
What is AI?
Explaining what AI can confuse even experts [16,12], as the term has changed over the years.
Nilson describes AI as "that work dedicated to making machines smarter ... [where] intelligence
is the quality that makes a business more efficient and foresight" [10]. However, Schank notes
that the definitions of intelligence may vary depending on the investigator and their approach to
understanding AI [16]. He suggests that there are two main objectives of AI research - "building
a smart machine" and "finding a kind of intelligence" [16]. He then proposed a set of features that
included common "intelligence" - communication, knowledge of the world, internal knowledge,
purpose, and art - emphasizing that the ability to read is the most important factor in intelligence
[16].
What Can AI Do?
While AI has been able to find patterns in the amount of data, perform repetitive tasks, and make
decisions in controlled environments, people now live better in many tasks that require art,
emotion, information transfer, and social interaction.
How does AI work?
A better understanding of how AI works can help people build more sensible types of programs
they work with. For this and other reasons, much of the research available on AI education in
university and K-12 environments focuses on informing how AI works.
Cognitive Systems
Cognitive systems - or AI systems are rated after ideas about the human mind [4] - are used in a
variety of application domains, including WordNet, IBM's Watson, expert systems, and cognitive
educators. Most comprehension program messages cover topics related to information
representation, planning, decision making, problem solving and learning.

4
Cognitive systems use many techniques for planning, decision-making, problem-solving and
learning. Users may not need to understand all of these strategies in detail, but a higher
understanding of how computers make decisions can help in interpreting and understanding the
algorithms [29].
Machine Learning
Many students think that computers think like humans and want to make connections between
human perceptions of understanding and machine learning [12]. Students are also often surprised
that ML requires human decision-making and is not self-inflicted.
Research suggests that one of the ways to eliminate students' misconceptions about ML is to get
involved in integrated integration. Sulmont et al. while others suggest that students develop
physical algorithms to understand them in a practical way [4,6,12]. This technique has also been
used in CS education [2]. In general, AI manual testing has been used as a means of
implementing a variety of AI education programs (e.g. [4]), including projects where students can
train ML models to analyze their movements and gestures [4,14].
Robotics
Understanding that AI agents can physically act on and react to the world is an important
prerequisite for understanding robotics. Learning about sensors and their capabilities (one of the
“big ideas” of AI [13]) can also aid in understanding how AI devices gather data and interface
with the world.
This study provides the development of a platform and curriculum around the three “Great Ideas”
in AI, as it directs ideas that have a major impact on making people understand about AI [3]. At
the same time, these ideas also allow people to explore the use of a visual chat interface. The
three ideas are:
 Representation and Consultation: People are expected to understand how the agent learns
and represents new information. The agent generates two different visuals to indicate the
representation of the information.
 Conceptualization: The agent also demonstrates the concept of how machines classify
ideas. Individuals witness instances when the agent might succeed or fail in its learning and
make attempts to correct it.
 Social Impact: The curriculum emphasizes the moral and social impact of the AI
community through structured dialogue and reflection on the impacts of larger image
environments.
Agent platform is a visual web connector that can be linked to a web browser and people engage
with the agent in a chat in a small setting led by a personal facilitator.
3. PROPOSED MODEL
This study examines the use of a chat agent with four consecutive AI modules, as well as testing
machine learning and information representation.

5
It takes into account what roles conversations per se appear to play as part of information
exchange.
 Conversations as Revealment: One significant role of conversation from an information
retrieval perspective is to allow the two parties to reach an understanding as to what is
required by the user, and what the answerer knows.
 Initiative and Engaging Behavior: A number of authors have studied how a “virtual
human” should behave [6, 38]. For instance, Traum et al. describe desirable aspects of a
system conversing with human-beings, such as being real-time and incremental as
utterances are formed over time [36]. Mixed initiative refers to both the human and the
system having initiative at different points in time [1, 10, 25]. For instance, the agent may
take initiative to clarify or elicit information from the user whenever appropriate, while
allowing the user to drive the conversation at other times.
Recently, Christakopoulou et al. [9] studied whether to ask absolute or relative questions,
comparing the utility of each for learning about users. They also asked questions contextually,
based on what is already known. that characterize the solution space and allowing the user to
express and modify their criteria. In each back and forth step in a conversation, the system
provides some information to the user, and the user responds. Existing approaches are
summarized in Figure 1.
Figure 1. Conversation action space
The simplest design would be for the user to provide either a binary or ordinal score in response
to a question, or a preference given two or more choices.
A more sophisticated feedback from the user would be a critique [23, 28] that indicates in what
way the item or partial item presented by the system does not represent the user’s information
need.

6
The most detailed level of feedback a user may provide would be free text. Each of this would
typically fall into a single cell as indicated in the Figure 1.
As seen in Figure 1, the system may provide three types of feedback, and expect three types of
responses in return. Each cell describes related work that falls into the appropriate category.
Many of the partial item field (F) or field+value (F+V) interaction approaches are often
considered variants of faceted search. Let’s now describe each of the labeled cells in turn:
Null System - Free Text User
This is the starting point for most information retrieval systems such as Web search engines, and
often for conversational systems where the user may specify many possible requests (such as
commercial intelligent agents including Siri). The user is simply presented with a search box into
which any query can be entered.
Partial Item System - Pref/Rating User
A user may be presented with partial information about matching items in various ways. The
most common approach is for a conversational system to confirm a slot that has been inferred,
such as "you are looking for a sushi place, correct?".
Some systems may also cluster items, asking for a preference. For instance, it might ask "would
you prefer to a fancy restaurant, or an inexpensive one?".
A third interaction mode, where a preference is elicited over a set of (feature,value) pairs would
for instance "would you prefer a laptop with a 12 inch screen for $1000, or a laptop with a 14
inch screen for $1200".
Partial Item System - Critique User
In the simplest case, fielded search provides users with a selection of known fields and users may
select or specify ranges for any property they desire. This is common in online shopping
scenarios, where often the allowed field values are pre-specified. In other settings a user is
presented with specific individual facet values. Some commercial intelligent agents allow users to
clarify in this way, rather requiring a simple yes/no.
Partial Item System - Free Text User
When a system asks a user to fill in a particular aspect of an information need, this is usually
referred to as slot filling. As an example, some travel systems require users to specify travel
details to complete a structured query over a public transit schedule.
Complete Item System - Pref/Rating User
Classic approaches to recommendation often request ratings of items to learn a user model for
further recommendations. These may be absolute rating requests ("how much did you enjoy the
movie Kill Bill?") or preference requests ("did you enjoy Kill Bill or Pride and Prejudice
more?”).

7
Complete Item System - Critique User
In this case, a system may select a given item, then allow the user to refine their information need
anchoring of the properties of the item. For instance, Reilly et al. [28] describe a system where
users are presented with an item and possible ways the information need can be refined. Users
may select a pre-defined rich critique that allows the system to move closer to the user’s goals.
Regarding the use of semantic analysis tools for feedback and conversation, the strength of Latent
Semantic Analysis (LSA) to measure text similarity has been demonstrated in many research
studies (e.g., Chen, Chen, & Sun, 2010; Sung, Liao, Chang, Chen, & Chang, 2016). Sung et al.
(2016) used some articles to construct a latent semantic space with 250 dimensions and the LSA
was used to compute the semantic similarity between pairs of sentences, such as user summary,
expert’s summary, or the source text. On account of the concept of new versus given information
is related but distinct from the concept of text similarity, Hu, Cai, Louwerse, Olney, Penumatsa,
and Graesser (2003) adapted the standard LSA and proposed the LSA based measure called a
span method to detect given and new information in written discourse.
3.1. Conceptual Framework
This paper explores a ML based method based on a reward model for discourse analysis which is
used to reveal the process of interaction and knowledge construction, investigate user behaviors
and discourse content, and acquire deeper understanding of engagement in collaborative learning
(e.g., Peng & Xu, 2020) by breaking the problem into two parts:
(1) learning a reward function from the feedback of the user that captures their intentions and
(2) training a policy with reinforcement learning to optimize the learned reward function.
In other words, we separate learning what to achieve (the ‘What?’) from learning how to achieve
it (the ‘How?’). We call this approach reward modeling.
Figure 2. Schematic illustration of the reward modeling setup
Figure 2 illustrates the basic setup in which a reward model is trained with user feedback and
provides rewards to an agent trained with RL by interacting with the environment.
As we scale reward modeling to complex general domains such as language modeling
techniques, there can be various challenges. A reward model is trained to learn user intentions by
means of feedback provided in online communities and provides rewards to a reinforcement
learning agent that interacts with the environment, in this case, the social media platform. Both
processes happen concurrently, thus the model is being trained with the user in the loop.

8
Consider this simple existence proof: let H be the set of histories that correspond to aligned
behavior- aka relevant post content- in a social media setting If the set H is not empty, then there
exists a reward function r such that any corresponding optimal policy π ∗ produces behavior from
H with probability 1.
A trivial example of such a reward function r rewards the model every few steps if and only if its
history is an element of the set H. In theory we could thus pick this reward function r to train our
RL model. However, in practice we also need to take into account whether our reward model has
enough capacity to represent r, whether r can be learned from a reasonable amount of data (given
the inductive biases of our model), whether the reward model generalizes correctly, and whether
the resulting behavior of the RL model produces behavior that is close enough to H.
3.2. Overview of AI Modules
Based on the reward modeling, the AI curriculum consists of four modules: "Do You Know the
Agent?", "Teaching a Lawyer", "Machine Witnessing", and "AI and Ethics". Before entering the
first module, individuals learn to communicate with the agent through an introduction, where the
agent greets and talks with them.
Part 1, “What Does the Agent Know?” introduces people to represent information and
consultation using mind maps. The representation method used is a mind map or “mind map”,
which people can analyze to find relevant attributes. Positive attributes (existing) are shown as
blue circles, and negative (missing) circles are shown as red circles. Users can also analyze the
corpus given by the agent to create mind maps, and thus draw connections between natural
language sentences and related mind map details.
In the next module, Module 2, "Teaching Agent", individuals are assigned the task of providing
the agent with data on selected subjects. People can give the agent any information about a topic
they would like to find. This enables them to put AI learning experience into their knowledge and
interests. Agent works as an AI with minimal knowledgeable information, and users help an
agent create mind maps on each topic. The concept map concept is presented in Module 1 and is
based on the following modules.
Module 3: "Machine learning", is where users look at the learning process and the agent's
thinking. They ask the agent to make a guess based on a previously taught concept and the agent
calculates the similarity of the words in each concept by using words that represent the words
representing the agent ideas and showing these schools using a bar graph. It is important for users
to understand why and why the agent may be incorrectly guessing by drawing a link between
similarities within Module 2 mind maps and Module 3 scores, and how to use this link.
In the final module, Module 4, "AI and Ethics", facilitators lead a discussion on how data
negotiating agents and data-learning agents are used in society with both positive and negative
outcomes. These discussion questions are divided into the Learning Resources section. Users can
think of situations where the agent makes mistakes, and then be asked questions such as, "Will
the agent know if what we are teaching is right or not?" and “How would you feel if an agent
found out something wrong with you?”. The purpose of this section is to empower users with AI
design tools with ethical principles. The study also aims to test users on the social consequences
of mistakes made by AI, and how they can reduce injuries (Payne 2019).

9
3.3. System Design
The system is designed to be easy to use. It can be used using any browser anywhere you have
internet access. The main features of the program are (1) a speech synthesizer, (2) a speech
identifier, (3) a semantic parser, (4) a word map classifier, and (5) a website visualizer, as
follows:
1. Speech amplifier: This section includes the voice of the agent using the Web Speech API. It
should be noted that a particular word from the visual interface of speech recognition should be
chosen so that it is not as gender neutral as possible. (A common advice is to refer to the agent as
“it” instead of “he” or “she”.)
2. Speech identifier: This section converts user speech to text using the Web Speech API.
3. Semantic parser: This section performs natural language processing using NLTK [29, 32, 17,
18] and CoreNLP tools (Manning et al. 2014). The NLTK toolkit works with words, while the
CoreNLP toolkit renders parts of speech [30] and processes sentences obtained from speech
recognition. The adjective performs the following three functions used in each function.
a) Name identification: As a job introduction, the agent asks people for their name and location.
The relay conveys the input received after each query, removing all nouns, proper nouns, and
foreign names. The specified name or location is the appropriate last name seen.
b) Subject Identification: In the first module, the tester identifies which user the question is
asking. Since the number of ideas that the agent is aware of is limited, the examiner searches for
these ideas. When a familiar concept is found, the parser retrieves the identified concept and a
pre-programmed mind map is displayed to users. These mind maps are created offline using a
third-party explorer capability.
c) Mind Map Builder: In Modules 1 and 2, the agent demonstrates his / her knowledge of ideas
and words on mind maps that are generated according to sentence structure and speech
components. An adjective identifies a field of concept for the interaction between descriptive
words and a topic (e.g., a negative interaction between water and desert), and then sends this
information to a web browser.
4. Classifier: This section classifies the word by comparing word representation using NLTK's
Wordline Interface (Loper and Bird 2002; Fellbaum 1998). In each sense, all the descriptive
words are compared to each other. Once the similarity between all pairs has been calculated, the
overall similarity between a concept and topic is its weighted average similarity score. The
classifier returns a normalized score for each topic, denoting the topic with the highest similarity
score to be the topic that the concept relates to.
5. Visualizer: This section creates mind maps and histograms based on pre-defined sentences and
user-defined sentences using D3.js [15, 25].
To show how the parts fit together, in the introduction, the system prompts the user to speak;
sends a user voice response to a browser-based speech recognition, which converts the response
into text; stores text response to user location session data; sends a text response to a server-based
server, processing the text for important details; returns the processed information to a browser-
based speech compiler, which causes the agent to speak, causing the user to speak again; and
repeats the process until it ends the conversation.

10
Figure 3. includes a view of the histogram, which shows similarities between the concept and the
sentence of the title.
Figure 3: (a) Mind map composed of sentences in (b). This is shown in Module 2 on the agent's website.
(b) Users of sentences have informed the agent during the study.
4. RESULTS AND DISCUSSION
To measure system performance, a small pilot study can be done to determine if people
understand the visual representations. Once hired people can hear about the three different
perspectives on information representation (Modules 1 and 2) and the three different machine
learning perspectives (Module 3) shown in Figure 4.
Figure 4: As shown in the orange bar in this histogram view, the agent can guess which
concept will affect it.

11
Testing can also be done even if the knowledge of the learning platform and the content of the
training improves people’s understanding of how machines represent information and learn.
Specifically, the level of user engagement can be assessed by measuring the number of sentences
used in the conversation with the agent, and whether it corresponds to their level of
understanding.
A consistent test protocol can be followed at different times. The size of the session can be from
one to four, each with a facilitator. First, researchers can randomly test participants' information
on voice assistants such as Siri and Alexa via icebreaker. After this, they can conduct a pre-
follow-up test for the agent (Figure 5). Modules 1 to 4 must then be completed respectively. All
work will last ∼60-80 minutes depending on the length of the conversation. People should do all
the tests individually on paper and without the slightest distraction from the investigators. In
addition to the measurement data from the test, researchers can also perform video recordings of
the time, participants in log sentences contribute and record users' responses to test questions.
Figure 5: Sample responses to pre- vs. post-examination
Assessments
To answer the idea that people can learn effectively and to understand how machines learn, the
following questions can be asked for assessment purposes:
1. What sentences can you say to the facilitator to create the following mind map in Figure 1?
(This tests their understanding of how information is represented.)
2. What can you tell the agent about a particular name so that he can correctly guess the
related topic? (This tests their understanding of how the agent reads.)

12
3. Why would an agent find it difficult to match another word to a particular topic even if he
knows everything that needs to be known about them? (This is an open-ended
questionnaire for examining potential mistakes an agent may make.)
4. Have you ever tried to mention the name ‘agent’ to the agent? If not, ask your tutor if you
can try. Does the agent recognize its own name? If not, why not? Can you think of another
name the agent would not see? (This is an open question in the agent's internal language
management.)
These pre- and post-test assessments can also be used to assess individual self-awareness as an
engineer and motivation to learn. People can be asked to rate how much they agree or disagree
with the various statements on a scale of 1 to 5.
1. I want to know how a machine learns.
2. I think I can teach it to the machine.
3. I trust voice assistants like Siri and Alexa.
4. I understand how the machine learns.
5. The activities we have done today have been helpful in learning how machines learn. (Post-
test only.)
6. The mind map of each concept and topic helped me think about the agent's brain. (Post-test
only.)
7. The histogram helped me to understand how the agent made decisions. (Post-test only.)
5. CONCLUSION AND FUTURE SCOPE
This work introduces a development framework for a conversational agency that educates
individuals about the representation of information and machine learning.
By training an agent, recognizing his mistakes, and retraining an agent, individuals can make
sense of the intelligence of a representative. In the future, the content of the agent training should
be expanded to address more topics in AI. Hopefully, this work will promote more AI training
content using conversational agents and viewing tools to help individuals understand the AI
algorithms.
The current framework of research has some limitations. First, it includes a small number of
participants. Further testing may reinforce the claim that the agent is operating. Also, there is a
need to create a future iteration of this study in which researchers can compare the performance
of the visual interface of another chat agent.
Computer scientists usually rely on statistical tools to demonstrate that a particular underlying
factor had a “causal” effect on the outcome of interest. While in the natural sciences, causal
effects are measured using lab experiments that can isolate the consequences of variations in
physical conditions on the effect of interest, more often than not, social life (including education
through means of AI) do not permit lab-like conditions that allow the effects of changes in the
human condition to be precisely ascertained and measured. This would merely provide evidence
on one of the causes, which may not even be one of the more important factors. In a quest for
statistical “identification” of a causal effect, scientists might often have to resort to techniques
that answer either a narrower or a somewhat different version of the question that motivated the
research. So, research can rarely substitute for more complete works of synthesis, which consider
a multitude of causes, weigh likely effects, and address spatial and temporal variation of causal
mechanisms.

13
For this study to be conducted, there is a need to control how participants interact with the agent
by reducing their knowledge required for participants to be able to train the agent in real time.
Last, but not least, the skills and design considerations outlined in this paper will need to be
expanded to keep pace with new discoveries, technologies, and rapidly changing social norms.
Researchers and educators of AI, and related technology and education communities should be
encouraged to both participate in intimate discussions and design considerations in this paper and
use them to lead and promote artistic and future research on AI learning.
REFERENCES
[1] Agre, P. E. (1994). Surveillance and capture: Two models of privacy. The Information Society, 10(2),
101–127.
[2] Allen, J. (2016). Topologies of power. Beyond territory and networks. Routledge.
[3] Bratton, B. (2015). The Stack: On software and sovereignty. MIT Press.
[4] Bucher, T. (2018). If...then: Algorithmic power and politics. Oxford University Press.
[5] Castañeda, L., & Selwyn, N. (2018). More than tools? Making sense of the ongoing digitizations of
higher education. International Journal of Educational Technology in Higher Education, 15(1).
[6] Decuypere, M. (2019a). Open Education platforms: Theoretical ideas, digital operations and the
figure of the open learner. European Educational Research Journal, 18(4), 439–460.
[7] Decuypere, M. (2019b). Researching educational apps: ecologies, technologies, subjectivities and
learning regimes. Learning, Media and Technology, 44(4), 414–429.
[8] Decuypere, M. (2019c). STS in/as education: where do we stand and what is there (still) to gain?
Some outlines for a future research agenda. Discourse: Studies in the Cultural Politics of Education,
40(1), 136–145
[9] Dieter, M., Gerlitz, C., Helmond, A., Tkacz, N., Vlist, F., Der, V., & Weltevrede, E. (2018). Store,
interface, package, connection : Methods and propositions for multi-situated app studies. CRC Media
of Cooperation Working Paper Series No 4.
[10] Drucker, J. (2020). Visualization and Interpretation: Humanistic Approaches to Display. MIT Press.
Journal of New Approaches in Educational Research, 10(1)
[11] Mathias, Decuypere The Topologies of Data Practices: A Methodological Introduction Fedorova, K.
(2020). Tactics of Interfacing. Encoding Affect in Art and Technology. MIT Press. Goriunova, O.
(2019). The Digital Subject: People as Data as Persons. Theory, Culture & Society, 36(6), 125–145.
[12] & Ruppert, E. (2020). Population Geometries of Europe: The Topologies of Data Cubes and Grids.
Science, Technology, & Human Values, 45(2), 235–261.
[13] Gulson, K. N., Lewis, S., Lingard, B., Lubienski, C., Takayama, K., & Webb, P. T. (2017). Policy
mobilities and methodology: a proposition for inventive methods in education policy studies. Critical
Studies in Education, 58(2), 224–241.
[14] Gulson, K. N., & Sellar, S. (2019). Emerging data infrastructures and the new topologies of education
policy. Environment and Planning D: Society and Space, 37, 350–366.
[15] Hartong, S. (2020). The power of relation-making: insights into the production and operation of
digital school performance platforms in the US. Critical Studies in Education, 00(00), 1–16.
[16] Hartong, S., & Förschler, A. (2019). Opening the black box of data-based school monitoring: Data
infrastructures, flows and practices in state education agencies. Big Data & Society, 6(1),
[17] Lash, S. (2012). Deforming the Figure: Topology and the Social Imaginary. Theory, Culture &
Society, 29(4-5), 261–287.
[18] Latour, B. (1986). Visualization and cognition: Thinking with eyes and hands. Knowledge & Society,
6, 1–40. Retrieved from http://hci.ucsd.edu/10/readings/Latour(1986).pdf
[19] Law, J. (2004). After Method: Mess in Social Science Research. Psychology Press.
[20] Lewis, S. (2020). Providing a platform for “what works”: Platform-based governance and the
reshaping of teacher learning through the OECD’s PISA4U. Comparative Education, 56(4).
[21] Lewis, S., & Hardy, I. (2017). Tracking the Topological: The Effects of Standardised Data Upon
Teachers’ Practice. British Journal of Educational Studies, 65(2), 219–238.
[22] Light, B., Burgess, J., & Duguay, S. (2018). The walkthrough method: An approach to the study of
apps. New Media and Society, 20(3), 881–900.

14
[23] Lindh, M., & Nolin, J. (2016). Information We Collect: Surveillance and Privacy in the
Implementation of Google Apps for Education. European Educational Research Journal, 15(6), Lury,
C., & Day, S. (2019). Algorithmic Personalization as a Mode of Individuation. Theory, Culture &
Society, 36(2), 17–37.
[24] Mathias, Decuypere The Topologies of Data Practices: A Methodological Introduction Lury, C.,
Fensham, R., Heller-Nicholas, A., & Lammes, S. (2018). Routledge Handbook of Interdisciplinary
Research Methods. Routledge.
[25] Lury, C., Parisi, L., & Terranova, T. (2012). Introduction: The Becoming Topological of Culture.
Theory, Culture & Society, 29(4-5), 3–35.
[26] Lury, C., Tironi, M., & Bernasconi, R. (2020). The Social Life of Methods as Epistemic Objects:
Interview with Celia Lury. Diseña, 16, 32–55.
[27] Lury, C., & Wakeford, N. (2012). Introduction: A perpetual inventory. Inventive Methods (pp. 15–
38). Routledge.
[28] Martin, L., & Secor, A. J. (2014). Towards a post-mathematical topology. Progress in Human
Geography, 38(3), 420–438.
[29] Piattoeva, N., & Saari, A. (2020). Rubbing against data infrastructure(s): methodological explorations
on working with(in) the impossibility of exteriority. Journal of Education Policy, 00(00), 1–21.
[30] Plantin, J. C., Lagoze, C., Edwards, P. N., & Sandvig, C. (2018). Infrastructure studies meet platform
studies in the age of Google and Facebook. New Media and Society, 20(1), 293–310.
[31] Prince, R. (2017). Local or global policy? Thinking about policy mobility with assemblage and
topology. Area, 49(3), 335–341.
[32] Ratner, H. (2019). Topologies of Organization: Space in Continuous Deformation. Organization
Studies, 1–18.
[33] Ratner, H., & Gad, C. (2019). Data warehousing organization: Infrastructural experimentation with
educational governance. Organization, 26(4), 537–552.
[34] Ratner, H., & Ruppert, E. (2019). Producing and projecting data: Aesthetic practices of government
data portals. Big Data & Society, 6(2), 1–16.
[35] Ruppert, E., Law, J., & Savage, M. (2013). Reassembling Social Science Methods: The Challenge of
Digital Devices. Theory, Culture & Society, 30(4), 22–46.
[36] Suchman, L. (2012). Configuration. In C. Lury & N. Wakeford (Eds.), Inventive Methods: The
Happening of the Social (pp. 48–60). Taylor and Francis.
[37] Thompson, G., & Cook, I. (2015). Becoming-topologies of education: deformations, networks and
the database effect. Discourse: Studies in the Cultural Politics of Education, 36(5), 732–748.
[38] Thompson, G., & Sellar, S. (2018). Datafication, testing events and the outside of thought. Learning,
Media and Technology, 43(2), 139–151.
[39] van de Oudeweetering, K., & Decuypere, M. (2019). Understanding openness through (in)visible
platform boundaries: a topological study on MOOCs as multiplexes of spaces and times. International
Journal of Educational Technology in Higher Education, 16(1).
[40] van de Oudeweetering, K., & Decuypere, M. (2020). In between hyperboles: forms and formations in
Open Education. Learning, Media and Technology, Advance online publication, 1–18.
[41] Williamson, B. (2017). Learning in the “platform society”: Disassembling an educational data
assemblage. Research in Education, 98(1), 59–82.
AUTHORS
Ayse received her MSc in Internet Studies in University of Oxford in 2006. She participated in various
research projects for UN, Nato and the EU regarding HCI (human-computer interaction). She completed
her doctorate degree in user experience design in Oxford while working as an adjunct faculty member at
Bogazici University in her home town Istanbul. Ayse has also a degree in Tech Policy from Cambridge
University. Currently, Ayse lives in Silicon Valley where she works as a visiting scholar for Google on
human-computer interaction design.

A DEVELOPMENT FRAMEWORK FOR A CONVERSATIONAL AGENT TO EXPLORE MACHINE LEARNING CONCEPTS

More Related Content

What's hot (20)

Similar to A DEVELOPMENT FRAMEWORK FOR A CONVERSATIONAL AGENT TO EXPLORE MACHINE LEARNING CONCEPTS (20)

Recently uploaded (20)

A DEVELOPMENT FRAMEWORK FOR A CONVERSATIONAL AGENT TO EXPLORE MACHINE LEARNING CONCEPTS