=Paper=
{{Paper
|id=Vol-2180/ISWC_2018_Outrageous_Ideas_paper_1
|storemode=property
|title=Can Children Teach AI? Towards Expressive Human-AI Dialogs
|pdfUrl=https://ceur-ws.org/Vol-2180/ISWC_2018_Outrageous_Ideas_paper_1.pdf
|volume=Vol-2180
|authors=Elena Demidova
|dblpUrl=https://dblp.org/rec/conf/semweb/Demidova18
}}
==Can Children Teach AI? Towards Expressive Human-AI Dialogs==
https://ceur-ws.org/Vol-2180/ISWC_2018_Outrageous_Ideas_paper_1.pdf
Can children teach AI?
Towards expressive human-AI dialogs
Elena Demidova
L3S Research Center, Leibniz Universität Hannover, Hannover, Germany
demidova@L3S.de
Abstract. AI-empowered dialog systems become increasingly widespread,
with over 40 million units of Amazon Alexa and Google Home installed
alone in the USA in 2017, according to recent surveys. Whereas this
adoption facilitates AI-based systems to engage in end user dialogs on
an unprecedented scale, the ability of AI to learn from their human dialog
partners is still substantially limited. In this paper we briefly discuss the
opportunities, challenges and risks on the interface of knowledge acqui-
sition and human computer interaction towards more expressive dialogs
between end users and AI. Our goal is to facilitate AI to directly gain
knowledge from end users, while ensuring mutually beneficial interaction.
1 Introduction and Motivation
Recent advances in AI-empowered dialog systems have led to their increasing
adoption by end users and thus to a significant raise of interaction opportunities
for AI with end users. Surveys indicate that over 40 million units of such systems
including Amazon Alexa / Echo and Google Home have been installed alone in
2017 in the USA, tendency increasing.1 These numbers also indicate that AI has
gained direct access to end user communications on an unprecedented scale.
Although the widespread adoption of AI dialog systems provides entirely
new opportunities for AI to communicate with end users, to request their input
and ultimately to learn from them at a scale that has not been possible ever
before, current learning algorithms are unable to take full advantage of rich and
expressive end user feedback that can be potentially provided during a human-
AI dialog. Existing AI algorithms mainly restrict the role of end users in their
learning to the provision of pre-defined patterns for specific tasks.
In fact, many recent developments demonstrate that end users are willing and
capable of creating new knowledge and sharing it with other agents. Examples
include large-scale collaborative development of open knowledge platforms on
the Web (e.g. Wikipedia, Wikidata and Freebase [3]) as well as numerous citizen
science and crowdsourcing platforms that evolved over the past decade.
However, the design of current human-AI interaction paradigm does not take
full advantage of rich end user knowledge. Interactive systems for Question An-
swering (e.g. [4]) typically aim to reduce interaction cost. Recent approaches for
1
https://searchengineland.com/analyst-firm-google-home-gains-ground-
amazon-echo-44-million-total-units-sold-290544
knowledge graph construction from conversational dialogs (e.g. [2]), mainly aim
to learn specific patterns (e.g. relations between concepts). Interaction paradigm
that empower end users to flexibly augment the system’s knowledge beyond the
foreseen patterns, to understand the sources of errors and to explicitly correct
them, or to directly teach and personalize dialog systems are to be developed.
In this paper we briefly present the key opportunities of expressive human-
AI dialogs with end users along with the challenges and risks that need to be
addressed to facilitate rich knowledge acquisition by AI systems while providing
benefits to end users.
1.1 Motivating Example
To illustrate how an AI-based dialog system can profit from an end user dialog,
let us consider a real-world dialog example, where a 10-year old end user talks
to the newly acquired Amazon Alexa dialog system.
Q1. Alexa, play me a cheerleader song.
– Alexa starts playing an unrelated song.
Q2. Alexa, stop. Alexa, What is cheerleading?
– Alexa reads the definition of the term “cheerleading” from Wikipedia.
Q4. Alexa, play me a cheerleader song.
– Alexa plays the same song again.
Q5. Alexa, stop.
As we can observe in this example, a 10-year old end user attempts to explain
the meaning of her request to an AI system by a sequence of requests clarifying
the context. As this example illustrates, the user is willing to clarify her intent
to obtain the desired result and explicitly builds up the context of the question
to facilitate this clarification in a dialog. Intuitively, such dialog context con-
tains several clues an AI system can use to optimize its response and, even more
importantly, its underlying knowledge representation. For example, the interrup-
tion of the response through the user reflects her dissatisfaction with the current
answer. The follow-up questions with a semantic overlap indicate user attempt
to build up the context connecting the individual requests. Therefore, the ques-
tion is how shall we design knowledge acquisition and interaction paradigm of
AI dialog systems and represent the resulting knowledge to facilitate better use
of such expressive end user input.
2 Opportunities, Challenges & Risks
An expressive human-AI dialog is a dialog between an AI system and an end
user that facilitates AI system to acquire knowledge.
In this paper we in particular consider dialog systems that rely on know-
ledge graphs. An example high-level framework for expressive human-AI dialogs
including selected components is presented in Fig. 1.
The opportunities of expressive human-AI dialogs from the perspective of an
AI system include:
Fig. 1: An example high-level framework for expressive human-AI dialogs.
– Direct acquisition of new knowledge from large-scale end user population.
– Construction and augmentation of the underlying knowledge graphs on a
large scale.
– Explicit verification and correction of knowledge.
Expressive human-AI dialogs can facilitate a wide range of applications that:
– Provide better quality results due to the enhanced knowledge representation.
– Enhance user experience through enhanced transparency.
– Offer personalized services.
These opportunities come along with several key challenges. In the area of
knowledge acquisition these challenges include:
– Designing flexible knowledge representations to directly incorporate user in-
put and feedback, in particular with respect to associative knowledge.
– Designing algorithms that actively acquire and utilize expressive user input.
Such algorithms can be inspired through modern pedagogical approaches,
e.g. peer instruction [1].
– Conducting quality control of the acquired knowledge.
– Performing learning process assessment.
– Facilitating transparency and explainability for end users.
The challenges in the area of human-computer interaction include:
– Design of incentives for end users to teach AI systems, while ensuring that
the dialogs are mutually beneficial for AI and users.
– Development of interaction paradigm for expressive human-AI dialogs (e.g.
through further development of approaches for active learning, games of
purpose, learning from errors, etc.).
– Providing feedback regarding learning process.
Finally, the associated risks include:
– Ethical risks including potential risks for end users in vulnerable groups.
– Data protection and privacy issues.
3 Related Work
Whereas initial approaches to dialog-based knowledge acquisition relied on ma-
nually defined dialog models and rules, recent works utilize more flexible conver-
sational dialogs without requiring strictly pre-defined dialog models. These ap-
proaches address specific tasks including data labeling, knowledge graph population
and relation extraction [2]. However, none of the existing dialog-based knowledge
acquisition approaches is capable of conducting flexible knowledge acquisition at
the scale envisioned in this paper. Learning through associations and learning
through dialogs (e.g. in modern peer instruction pedagogical approaches [1])
is a common way for humans to effectively acquire new knowledge. Recently,
applications utilized knowledge graphs to support human learning through asso-
ciations.2 In this work, we argue that new learning paradigm based on knowledge
graphs and end user dialogs can be mutually beneficial to AI and humans.
4 Summary and Outlook
In this paper we briefly summarized opportunities, challenges and risks in expres-
sive human-AI dialogs. Such dialogs can shift the interaction paradigm between
end users and AI systems towards more effective knowledge acquisition. This
paradigm can facilitate AI systems to take a direct advantage of their growing
end user base to improve the knowledge representation, provide personalized ser-
vices and enhance the overall user experience. These opportunities come along
with a number of research challenges regarding knowledge representation, algo-
rithms and human computer interaction as well as ethical risks to be addressed
in the next decade.
References
1. Crouch, C., Mazur, E.: Peer instruction: Ten years of experience and results. Am.
J. Phys. 69, 970–977 (2001)
2. Hixon, B., Clark, P., Hajishirzi, H.: Learning knowledge graphs for question answer-
ing through conversational dialog. In: Proc. of NAACL HLT 2015 (2015)
3. Ringler, D., Paulheim, H.: One Knowledge Graph to Rule Them All? Analyzing the
Differences Between DBpedia, YAGO, Wikidata & co. In: KI 2017 (2017)
4. Zafar, H., Dubey, M., Lehmann, J., Demidova, E.: IQA: Interactive SPARQL Query
Construction in Semantic Question Answering Systems. in press (2018)
2
https://noduslabs.com/cases/learning-associations-knowledge-graph