=Paper=
{{Paper
|id=Vol-3816/paper44
|storemode=property
|title=Explanatory Dialogues with Active Learning for
Rule-based Expertise
|pdfUrl=https://ceur-ws.org/Vol-3816/paper44.pdf
|volume=Vol-3816
|authors=Yao Yao,Horacio González–Vélez,Madalina Croitoru
|dblpUrl=https://dblp.org/rec/conf/rulemlrr/YaoGC24
}}
==Explanatory Dialogues with Active Learning for
Rule-based Expertise==
https://ceur-ws.org/Vol-3816/paper44.pdf
Explanatory Dialogues with Active Learning for
Rule-based Expertise
Yao Yao1 , Horacio González–Vélez1 and Madalina Croitoru2
1
Cloud Competency Centre, National College of Ireland, Dublin, Ireland
2
LIRMM, Faculty of Science, University of Montpellier, Montpellier, France
Abstract
Contemporary language models have enhanced the interaction capabilities of AI with users. The improved
understanding and processing abilities of AI with respect to the provided data, thanks to these language models,
have simplified related knowledge engineering tasks. In this research, we embed LLMs in computational agents
to reinforce the interaction between the system and expert users to improve knowledge engineering processes.
By combining explanatory dialogue and active learning into knowledge engineering pipelines, we provide a
framework that can help experts validate rule-based expertise in a specific domain. This validated expertise can
be represented in RuleML format and is available to support knowledge-driven AI applications in domain-specific
tasks. Our initial test indicates that such an integration is feasible and improves the overall usability of knowledge
engineering processes, using curriculum development scenarios from DIGITAL4Business, a four-year EU-funded
project to deliver a new European Master’s programme on the practical application of advanced digital skills
within European SMEs and companies.
Keywords
Multi-agent Systems, Active Learning, RuleML, Explanatory Dialogues, LLMs, DIGITAL4Business
1. Introduction
Artificial intelligence (AI) systems such as Large Language Models (LLMs) have exhibited proficiency in
comprehending, generating, and structuring large volumes of textual data. However, in sophisticated
and professional domains, such as high-level education, the practical deployment of AI applications
requires the support of integrated knowledge, and such effective knowledge integration demands
the continuous updating of relevant information within a specific context, along with the ability to
comprehend new data with expertise. This fact makes the question of how to efficiently gain and validate
sufficient knowledge to optimise the performance of AI models in domain-specific tasks increasingly
important today.
In this research, we provide an approach to assist in the above question by incorporating the
technologies of multiple agent systems, interactive AI, and rule language of the Semantic Web into
an integrated framework. In our proposed framework, we will use multiple computational agents to
provide explanatory dialogues to help the knowledge validation between the system and the expert
users. All selected knowledge will be verified with the active learning process and corresponding
experts in the next step. After knowledge validation is completed, new knowledge will be integrated
with the knowledge base which includes the expertise represented by the RuleML format.
This research is also part of the EU project DIGITAL4Business. DIGITAL4Business is a four-year
EU-funded project, one of the largest non-infrastructure projects awarded to date under the European
Commission’s flagship DIGITAL Europe programme. In DIGITAL4Business, our aim is to forge a new
European Master’s programme on the practical application of advanced digital skills within European
SMEs and companies [1]. By integrating customised requests from employers and students, we are
providing a flexible learning schema and microcredential service to support personalised learning for
RuleML+RR’24: Companion Proceedings of the 8th International Joint Conference on Rules and Reasoning, September 16–22, 2024,
Bucharest, Romania
Envelope-Open yao.yao@ncirl.ie (Y. Yao); horacio@ncirl.ie (H. González–Vélez); madalina.croitoru@lirmm.fr (M. Croitoru)
GLOBE https://www.ncirl.ie/cloud (Y. Yao); https://www.ncirl.ie/cloud (H. González–Vélez)
Orcid 0000-0002-5882-0058 (Y. Yao); 0000-0003-0241-6053 (H. González–Vélez); 0000-0001-5456-7684 (M. Croitoru)
© 2024 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
CEUR
ceur-ws.org
Workshop ISSN 1613-0073
Proceedings
1
Yao Yao et al. CEUR Workshop Proceedings 1–15
each student in their Master’s study. Through precise profile identification powered by the proposed
knowledge-driven AI modules and a comprehensive curriculum, this programme can deliver person-
alised learning processes to students. Such demand-orientated targeted learning increases students’
adaptability to future emerging market opportunities related to digital transformation.
We believe that this will help businesses achieve long-term competitiveness and growth through
digital transformation and innovation. This goal requires a customised deployment of generative
AI in domain-specific applications such as the curriculum development process with comprehensive
knowledge integration in multiple domains. In concrete use cases, concepts such as method and format
may have different meanings in different domains. The AI models can help the system to identify the
specific meaning of the given concept in the knowledge base based on a particular context and then
integrate the correct contents from different domains into the response of the particular application
by the collaboration of separated agents. The research in this paper aims to serve this goal better by
integrating the necessary interactive AI module into the corresponding agents to build and optimise
specific knowledge of concepts in the given domains.
This collective process requires intensive interaction, knowledge explanation, and knowledge valida-
tion between multiple domain experts and knowledge models. It usually poses a challenge in knowledge
engineering. Our method can facilitate the process above by using the interactive agents (powered with
LLM) to propose a specific context-orientated dialogue to each corresponding expert and integrate the
feedback of experts into knowledge validation and optimisation. The method could partially simplify
and automate the work of experts in the knowledge engineering process. We have demonstrated our
contribution by a detailed use case of the proposed framework inspired by curriculum development
scenarios at DIGITAL4Business.
2. Background
In this section, we will have a brief review of each of the relevant aspects of our framework and discuss
how these methods or techniques can help us improve the performance of AI in a domain-specific task.
2.1. Explanatory dialogue
Explanatory dialogues in AI refer to interactive conversations in which an AI system provides explana-
tions to users about certain phenomena, decisions, actions, or data. These dialogues are designed to
improve understanding, transparency, and trust between humans and AI systems. They play a crucial
role in making AI systems more interpretable and user-friendly, especially in complex high-stakes
applications such as healthcare, finance, and autonomous systems. The system of explanatory dialogues
is a formal dialogue system of explanation with two players turning the tables [2, 3]. It takes place
between an explainer and an explainee.
In our proposed framework, expert users would request the system to explain the newly learnt
knowledge and validate the quality of knowledge. By facilitating interactive communication between
AI systems and expert users, the explanatory dialogue provides a robust mechanism for knowledge
validation, where users can understand, verify, and refine the knowledge embedded within AI models.
The integration of explanatory dialogue into AI systems provides a significant advancement in our
knowledge validation process. Not only enhances the transparency and trustworthiness of AI systems,
it also promotes effective collaboration and continuous improvement of the knowledge models. The
knowledge validation and improvement process can request a collective contribution from experts
in different domains, and explanatory dialogues can help experts in diverse domains understand the
common topic efficiently and finally reach a consensus. In Arioua et al. [4], it also shows the potential
of explanatory dialogues in knowledge integration. As our work aims to leverage explanatory dialogues
for knowledge validation, we build on these foundational studies to enhance the interaction and
validation processes in our knowledge-driven framework. In addition, we also plan to continue to
extend our interactive module by introducing other dialogue systems [5, 6] based on the demand for
more sophisticated scenarios in the future.
2
Yao Yao et al. CEUR Workshop Proceedings 1–15
2.2. Active learning
Active learning is a special case of Machine Learning (ML) in which a learning algorithm can interactively
query a user (or some other source of information) to label new data points with the desired outputs
[7]. Figure 1 shows an illustration of a general active learning method. In our platform, we use active
learning to query experts for knowledge validation. Dubious or confusing concepts in knowledge that
may cause conflict or ambiguity, based on the result of previous training, will be sent to experts for
further investigation.
Figure 1: Illustration of Active Learning. The user, the data, the ML models and the predictions from these
models all work together to improve the final results. The user will first use some data to train a model; this
initial training set is generally either fully or partially labeled by the user. The model generated in this way will
then be used to classify or predict another set of data. The model can be examined by looking at its predictions
and, if possible, its internal structure. The user can then identify any shortcomings or blind spots which may
exist in the model and augment the training set used to train the model with data which can help the model to
train better. This loop continues until a model of desired quality is achieved. Through the interaction between
users and the learning algorithm, the system can effectively learn a model from the comparatively smaller data
set that is annotated by experts.
2.3. Semantic technology and Rule-based reasoning
The ultimate goal of semantic technology is to help machines understand data. To enable semantic
encoding with the data, well-known technologies were developed, such as the Resource Description
Framework (RDF) [8, 9] and Web Ontology Language (OWL) which is standard for RDF [10]. These
technologies formally represent the meaning involved in the information. In our research, we use
RuleML (Rule Markup Language) [11] as the standard format to store rule-based knowledge on the
platform while making the ontology concepts compatible with the RDF representation, and this knowl-
edge will be accessible to AI/ML models, as confirmed by Allemang and Hendler in integration in many
applications [12].
By using RuleML to represent the knowledge base, we can provide detailed and structured explana-
tions for the AI system’s decisions and actions. In this way, the knowledge will provide the necessary
resources and operational information to the AI/ML models to complete the corresponding tasks. Fur-
thermore, knowledge will also help the system explain the result of AI/ML models. These concepts in
knowledge will be connected by particular relations and constitute the knowledge base.
2.4. Rule-based knowledge representation
As an XML-based language for representing and exchanging different kinds of Horn rules (derivation
rules, reaction rules, integrity constraints) on the Semantic Web, RuleML has limited expressiveness
to represent some common situations that may arise in situations with partial knowledge or options.
3
Yao Yao et al. CEUR Workshop Proceedings 1–15
Typically, some semantic adaptation is needed to adopt options in an elective way. Moreover, traditional
RDF approaches require extensions to accommodate temporal constraints and specific time steps [13].
Consequently, we propose to introduce active learning to label intermediate knowledge points where
expert interaction is required using an agent-based interface, knowledge-base, and an LLM to underpin
the interactions over time.
3. Method
The framework discussed in this paper aims to perform a set of agents that implement the explanatory
dialogue and active learning in Human-Machine Interaction with the help of LLMs. Through the
interaction, each agent will individually extract and validate the corresponding knowledge from different
hierarchical levels under the same domain. All these agents working together present a systematic
strategy for efficiently collecting the relevant knowledge of the given concept from the expertise and
finally integrating the knowledge into the corresponding part of the knowledge base. In previous
research [14], the combination of computational agent and LLM showed impressive efficiency in
performing knowledge extraction tasks, and this fact encourages us to use LLM embedded in the agents
of our framework.
By applying the explanatory dialogue and active learning approach, we can efficiently annotate and
validate knowledge that is specific to the given use case context based on the user’s requirement. In
this process, we also applied LLMs to help the agent better understand the user input and the relevant
text-based data. Compared with predefined knowledge models and fine-tuned LLMs, our approach
focusses on interactively extracting knowledge from diverse contexts separately and uses a systematic
strategy to validate and refine the extracted knowledge models. This approach allows to use individual
enquiring for fragmental knowledge with a particular given context and integrating the context of
previous inquiries in the subsequent knowledge integration. By this, expert users can focus on the
specific context each time to provide more elaborate knowledge and then leave the agents systematically
to conclude the previously extracted knowledge on a higher level.
The proposed framework can be divided into three different parts: an agent-based interface, a rule-
based knowledge representation, and an LLM behind the system. Figure 2 shows a general example of
how the framework works in the knowledge engineering process. In the following, we will elaborate
on each of these parts separately.
Figure 2: The example of the agent-based framework pipeline in the knowledge engineering process
4
Yao Yao et al. CEUR Workshop Proceedings 1–15
3.1. Multi-agent system and explanatory dialogue
In this study, we apply a Multi-agent System (MAS) to implement the explanatory dialogue in the
corresponding knowledge extraction and validation process. MAS has been applied to collect the
context to help AI based on rules [15]. Figure 3 demonstrates that the Multi-agent system retrieves the
information from various data sources and integrates this information into a graph-based representation.
Figure 3: The multi-agent system extracts knowledge from diverse data sources
The extracted knowledge will be presented to the active learning agent, and this agent uses explanatory
dialogues to interact with expert users to optimise the extracted knowledge. In our framework, we
use the formalisation of explanatory dialogue that has been presented in the work of Arioua and
Croitoru [16]. The explanatory dialogue system is a formal, turn-taking dialogue between two players:
an explainer and an explainee. It includes dialogue moves to request and provide explanations. The
dialogue begins with the explainer making a statement, ASSERT(j). The explainee then requests an
explanation, EXPLAIN(j), with j being accessible and believed to be true by both parties. The explainer
(computational agent) can either provide an explanation or declare an inability to explain. Such an
interaction could be iterated and proceed with the hierarchy structure of the corresponding knowledge.
In the end, an expert user decision is introduced, allowing one to judge the success of the explanation
and corresponding knowledge validation. Expert users can also update the knowledge during the
process and restart the dialogue to verify the update.
3.1.1. The implementation of active learning support by agents
The active learning processes of our framework are supported by multiple agents. when new knowledge
has been extracted, the statement of the new knowledge will be presented to the active learning agent.
Active learning is a special case of machine learning in which a learning algorithm can interactively
query a user (or some other source of information) to label new data points with the desired output. In
our case, the corresponding experts will work as the user to annotate the given new knowledge and
complete the corresponding knowledge validation. The general active learning process is as follows.
• The active learning agent will search the relevant concepts based on the new knowledge statement.
The embedded LLM of the agent will determine the relevance between the knowledge statement
and the candidate concepts;
• After the relevant concepts are identified, each selected concept will be assigned to an independent
agent. The agent will go through the knowledge base retrieve the corresponding logical rules
5
Yao Yao et al. CEUR Workshop Proceedings 1–15
and constraints of the selected concepts (see more details in section 3.2) and send the information
to the active learning agent;
• The active learning agent collected all constraints of relevant concepts and checked if there were
any violations in the new given knowledge;
• If there is no violation, the given knowledge will be accepted in the knowledge base as the new
item automatically. Otherwise, the knowledge will be presented to the corresponding expert by
the active learning agent for further validation based on the violated constraints;
• Based on the annotation of experts, the updated knowledge will be rejected or checked again by
the active learning agent;
In this research, we embedded LLMs in agents to better identify the relevance between concepts
and facilitate the communication between agents and expert users. This is also one of the novelties
in our current research. Some implementation solutions of the agent-based active learning pipeline
introduced in this paper also draw on previous similar research which uses an embedded reinforcement
learning approach to identify the relevant patterns in knowledge. More details on the implementation
of active learning with agents can be found in previous research [17].
3.1.2. Knowledge validation with explanatory dialogue
In our framework, we use the explanatory dialogue to particularly facilitate knowledge validation. The
advantage of using an explanatory dialogue includes two aspects. First, explanatory dialogue helps to
clarify the context of concepts from different domains. Knowledge usually includes interdisciplinary
concepts, and an explanatory dialogue can explain the meaning of these concepts to all experts from
diverse domains. It allows for back-and-forth questioning, providing opportunities to elaborate on
information, clear up misunderstandings, and deepen understanding. This iterative process strengthens
the validity of the knowledge. Second, knowledge validation often involves the collaboration of multiple
experts where different perspectives come together to enrich understanding. When agents explain to
each other on behalf of different experts, they engage in collaborative learning, which can lead to the
discovery of new insights and more robust knowledge validation. An explanatory dialogue provides the
perfect foundation for continuing this collaborative learning in different times, domains, and experts.
Next, it will provide an example of how the agent uses explanatory dialogue to help experts in
knowledge validation. Assume that the active learning agent sent a suspicious knowledge statement (j)
to an expert for validation. This operation is regarded as ASSERT(j). For the given statement (j), the
expert could request the corresponding explanation of the statement which is regarded as EXPLAIN(j).
After explaining the statement, if the expert understands the statement well, the expert can accept the
statement as (POSITIVE(j)) or reject it as (INABILITY(j)) directly. Otherwise, the expert can request
more explanation with the new question(q) (NEGATIVE(q,j)) and iterate the request loop until all
questions have been well explained. In this step, the embedded LLMs will help the agent identify which
kind of operation the expert requested based on the textual input of the expert and extract the relevant
concepts from the input. After each EXPLAIN(j) or NEGATIVE(q,j) request, the agent will search the
relevant concepts in the knowledge base and try to provide an explanation to the expert. Each concept
(c) in the knowledge base has the attribute (Γ𝑐 ) of explaining its relations, constraints, and rules. These
predefined explanations in the attribute can be translated into the corresponding well-formed formulas
(wffs) by LLMs and the variables in wffs are the corresponding concepts or the attributes of the concepts.
The agent will continually extend these explanations based on comments from experts and annotate
them with particular relevant concepts and topics. Based on the matched concepts and topics, the agent
could identify the information in the knowledge base and integrate the information as the context in the
request prompt to LLMs. Finally, the LLMs will provide the explanation to the expert according to the
prompt and context of the request. If the expert chooses the status of (POSITIVE(j)) or (INABILITY(j)),
the expert will be asked to update the knowledge accordingly. If there is no information matching
in the knowledge base, the agent can request the corresponding explanation input from experts who
are related to the topic. After the inputted explanation has been accepted by other experts, the new
explanation will also be saved in the relevant concept in the knowledge base after knowledge validation.
6
Yao Yao et al. CEUR Workshop Proceedings 1–15
3.2. Rule-based expertise with RuleML format
The knowledge representation in this proposed framework follows RuleML format. RuleML provides a
robust and flexible framework for defining and sharing rules across different domains and applications.
Its XML-based structure ensures broad compatibility and ease of integration, making it a valuable tool
for implementing rule-based logic in the Semantic Web and beyond. RuleML is also used in expert
systems to encode the knowledge of human experts in a form that a computer can use to make decisions
or provide recommendations.
To implement the knowledge model, we employ a structured methodology that captures the expertise
through formalised logical rules and assertions. A knowledge statement K is a finite subset of L
(logic language L) precisely, K is a tuple (F, R , N ) of a finite set of facts F, rules R and constraints N.
These elements are then translated into logical facts and rules within RuleML. This process begins
with the identification of key concepts and relationships within the domain of expertise, such as
entities, attributes, and their interdependencies. These concepts are then encoded as logical facts using
elements in RuleML. Rules are developed to capture the dynamic relationships and constraints
between these facts, ensuring that specific conditions and dependencies are represented accurately
and all status meets necessary requirements before progressing. For example, rules might define
the prerequisites for certain actions, the conditions under which particular outcomes occur, or the
logical flow of decision-making processes. Constraints are special rules to limit the relation between
concepts, and these types of rules will be applied in the consequent active learning process to protect
the consistency, quality, and integrity of knowledge. In knowledge validation, violation of constraints
will request the review of expert users on the corresponding knowledge. This formalisation allows
automated reasoning and validation, enabling AI systems to infer new knowledge, validate existing
knowledge, and ensure consistency within the knowledge base. Using RuleML, we achieve a flexible
and extensible representation of expert knowledge that can be easily maintained and integrated into
various applications, enhancing decision support and automated reasoning capabilities across diverse
domains.
In the remainder of the section, we will explain how we define knowledge from different aspects in
RuleML.
3.2.1. Ontology Concepts
Ontology concepts in RuleML can be defined to represent structured knowledge and relationships
within a particular domain. RuleML’s XML-based format allows for the clear specification of ontology
components such as classes, properties, and relationships. Classes represent the primary concepts or
entities within the ontology. In RuleML, classes can be defined using the element. For example,
within a scenario of curriculum development, we could define the concept ”student” in RuleML as
follows:
student
we could also define the attributes and relationships of the concepts by using and
elements, Instances of concepts can be represented using the element. Facts
about instances can be asserted using the element. These assertions link instances with properties
and values. Using these elements, we can use agents to automatically define ontology concepts in
RuleML based on the given description and create a structured and formal representation of domain
knowledge.
7
Yao Yao et al. CEUR Workshop Proceedings 1–15
3.2.2. Logic Rules
Logical rules can be defined to express more complex relationships and constraints within the ontology
using the element. Defining rules in RuleML involves creating logical statements that express
relationships, conditions, and constraints within a knowledge domain. These rules are formulated
using elements that represent logical constructs, such as conditions (antecedents) and conclusions
(consequent). A RuleML typically consists of a head and a body. The body contains the conditions that
need to be satisfied, while the head specifies the conclusion that follows if the conditions are met.
3.2.3. Conceptual constraints
Conceptual constraints in RuleML are used to enforce specific rules and restrictions within a knowledge
domain to maintain data integrity and ensure logical consistency. It is typically represented as rules.
These constraints help to define the permissible states and relationships among the concepts in the
ontology. This ensures that specific undesirable conditions are not allowed within the knowledge base.
In our framework, we implement these constraints by running them with an active learning agent. The
system will extract the textual input as follows: The concept [curriculumIntroductionCreation] must
have a relation with the concept [curriculum](if this constraint has been violated, you need to reject
this knowledge statement, if it is uncertain, you need to inquiry expert) and automatically convert it
into the RuleML format with the help of the LLM as below:
constraint1
hasRelation
curriculumIntroductionCreation
curriculum
...
assertKnowledge
curriculumIntroductionCreation
The active agent will check if all refereed actions have been predefined properly and store this to the
knowledge base after necessary validation. With the given function calls, the agent will be able to
implement the constraint in the future active learning process.
8
Yao Yao et al. CEUR Workshop Proceedings 1–15
3.3. Interaction between users and agents supported by LLMs
In our framework, we use the LLM (Gemma [18]) and prompt engineering approaches to help the
interaction between users and agents. The agents are able to access the LLMs and knowledge base
of the framework and agents will retrieve the corresponding predefined prompt templates from the
knowledge base based on the given context of their task role. For example, assuming the user requests
agents to produce an introduction of curricula on the topic of Cloud computation. This request will be
processed by LLM and LLM will extract the relevant keywords to prepare the query in the Knowledge
Base. The relevant concepts K (introduction and CreationFunction) in the knowledge base will be
returned to the agent. If K (CreationFunction) includes a rule to produce a prompt template for LLM,
the relevant information will be retrieved from the knowledge base, and the agent will implement the
rule based on the given information. Figure 4 shows the example discussed above.
Figure 4: The illustration of knowledge supported agent working with LLM. Introduction and CreationFucntion
represent the concepts of knowledge and Templates represent the attributes of the concept CreationFucntion.
In fact, agents not only use LLMs to respond user’s requests but also interact with expert users for
knowledge validation and knowledge update by the same approach. For knowledge validation, as
we discussed in the previous sections, the input will be the suspicious or contradictory knowledge
assertion or the questions related to the knowledge assertion. The active learning agent will select
these assertions that violate the predefined rules and ask for the necessary validation or update about
the knowledge from experts in the format of explanatory dialogue. In this process, the agent uses LLM
and predefined prompt templates to implement each step in the dialogue until the knowledge has been
successfully validated. The main function of LLMs here includes two folders.
First, LLMs can help identify the relevance between the given textual statement and the concepts in
the knowledge base. For example, assume that the expert requested the agent to explain what means
of ”introduction” is in the context of ”curriculum development”. To perform this request, the agent
will access LLM and ask you to first list the essential keywords from the given input. After that, the
agent will search the knowledge base to identify the relevant concepts [introduction] and [curriculum
development] based on the given keywords. With the help of LLM, the agent can also identify the given
request type as an [explanation request] and retrieve the prompt template from the knowledge base
to respond to such type requests. Based on the information collected, the agent will search for the
marching attributes in the related concepts to fulfil the prompt template and then send the complete
prompt to provide the response for the given request. For sophisticated concepts with a hierarchy
structure, agents may need to run a recursive loop to retrieve all the information and iterate the above
process multiple times with LLMs.
Second, LLMs convert diverse formal language formats, such as wffs, to readable textual expressions.
9
Yao Yao et al. CEUR Workshop Proceedings 1–15
The computational agent processes the knowledge and the statement in diverse formats, but needs to
change back to readable dialogue format when the agent responds to the expert users. LLMs can be
used as an efficient tool to translate different format expressions into readable format expressions based
on the request. With LLMs, we can facilitate the interaction between experts and computational agents.
This function is also essential in knowledge update conversations between users and system agents.
In general, embedded LLMs help the interaction between agents and users in this framework and
play an essential role in the corresponding knowledge engineering process. Figure 5 shows an overview
of the framework on how to interact users with their knowledge base through computational agents
and LLM.
Figure 5: User interactions within the multi-agent system
4. Human-machine interaction
The goal of our research is to use LLMs and explanatory dialogue in the active learning process to help
validate and integrate knowledge. To be more concrete, our work in this paper focusses on using the
discussed approaches and modules to improve the domain-specific knowledge (expertise) extracted
in our framework. To explore the potential and efficiency of the proposed approach, we tested our
agents with an open-source LLM: Gemma model [18] to produce the relevant conversation to update
and validate the expertise in curriculum development based on human-machine interaction. This is
an ongoing study, and the test below aims to demonstrate and explore the potential of our proposed
solution and framework in practical scenarios. The main concern in this step is to examine how to
integrate all these discussed techniques into a common user interface and work well in the given
scenarios. Based on this test, we will continue to improve our approach and develop the corresponding
evaluation methods to measure the efficiency of the framework in our future work.
In our test, we use knowledge-driven agents to synthesise the LLM prompts based on the context
and organise the conversation between users and the system in an explanatory dialogue format for
the corresponding knowledge update or validation. In the framework, LLM is a tool for helping
agents understand the input of users and understandably respond to users. Therefore, the proposed
approach can be applied to any language model without knowing the particular internal structure of
the model. The following results show the different steps of the conversation in the knowledge update
and validation process.
10
Yao Yao et al. CEUR Workshop Proceedings 1–15
4.1. Knowledge update
In the framework, experts can choose input and predefined knowledge manually or request that the
system extract knowledge from a given textual data. Users can also choose to define the relationships
between two concepts. Our following examples show that the knowledge statements are represented
in a predefined standard format. Actually, user input can be in various textual formats, and the agent
will request LLMs to convert the input into the standard format given in the predefined samples for
knowledge update. If the input is not able to align with the standard format, the agent will explain
the format of the knowledge statement and require the user to rephrase their declaration. Users can
also extend the samples by adding conversion examples from their selfdefined formats to the standard
format to help LLMs better understand the input in the future. These customized samples will be
included in the user profiles in the knowledge base. Figure 6 shows an example of defining a concept
and a relationship through the interface.
Figure 6: Example of knowledge update
4.2. Knowledge validation
After the knowledge update, the system will try to integrate the new knowledge into the knowledge
base and perform an active learning process to check whether it is necessary to request an expert user
involving knowledge validation. Any violation of predefined rules or undefined concepts, relations,
and attributes in new knowledge updates will lead the active learning agent to report the case to an
expert user. Users could also add custom constraints to activate queries for expert users. For example,
for overwrite or deletion operations related to particular concepts, the user can define a constraint to
request the validation of expert users. The human-machine interaction in knowledge validation follows
the explanatory dialogue method. The active agent will start the conversation to provide an assertion
of the new knowledge and allow experts to request an explanation of the assertion and ask critical
questions during the process. In the event that the agent cannot provide a successful explanation, at
least one of the experts will manually review the conversation and then update the knowledge base.
After that, the agent will iterate the conversation again until the assertion has been considered and
explained to the experts. After all, the agent will extract the validated rule-based knowledge from the
conversation and apply them as new constraints in the future active learning process. Figure 7 shows
an example of the implementation of the explanatory dialogue in the knowledge validation process.
11
Yao Yao et al. CEUR Workshop Proceedings 1–15
Figure 7: Knowledge validation with explanatory dialogue
4.3. Review knowledge and define constraints for the active learning agent
Expert users can select concepts, relations, attributes, and rules from the knowledge base and review
them as necessary. Based on the feedback of a particular task, a user can trace the related knowledge
that has been applied to support the AI model in the task. For knowledge validation, expert users can
also request the agent to implement customised constraints through the interface. Figure 8 shows an
example of reviewing a concept related to the introduction creation task. After the review, the user
added a custom constraint on this concept.
Figure 8: Review and defining constraints
5. Conclusion
In this research, we proposed a novel approach using LLM and explanatory dialogues to improve the
active learning process within knowledge validation. Through the discussed use case of curriculum
development, we tested the potential of this approach and identified the steps of essential dialogues
12
Yao Yao et al. CEUR Workshop Proceedings 1–15
in this approach. The importance of relevant domain-specific knowledge and the necessity of expert
participation is critical to the performance of AI models, and our proposed framework could facilitate
the construction of a knowledge base underpinned by an efficient human-machine interaction. Despite
the surprising ability of LLMs in semantic processing and inference, many limitations and problems
revealed in practical scenarios tell us that LLMs still need input from human experts to construct
sophisticated models for practical tasks in the given domain. Building and validating such competent
knowledge models is a challenge because the knowledge is difficult to predefine by certain experts
at once. One reason is because of the necessary comprehensiveness and context awareness for these
models. Certain experts have difficulty covering all aspects of knowledge in all possible contexts in
knowledge engineering, especially for tasks requiring interdisciplinary knowledge.
Another reason is the extendability and compatibility of the knowledge models. For implementing
practical tasks, the required knowledge may need to be continually updated. Knowledge integration
and new knowledge validation are necessary, but implementation processes are difficult to pre-defined
in advance. Due to the reasons above, we believe that a constant and collective collaboration between
experts and the system would be a promising solution to cope with this challenge. Our current research
aims to provide a framework that constantly requires dynamic updates and optimisation based on the
feedback of experts to maintain competent knowledge models for domain-specific tasks.
There is a rational and practical concern about whether the system can have sufficient and adequate
input from experts to build the expected knowledge models. In our solution, we tried to use LLM-
powered agents and explanatory dialogue to mitigate this problem. With the help of LLMs, agents
can automatically extract knowledge statements from designated datasets and present them to the
system, reducing the need for manual inputs for knowledge updates. For uncertain or contradictory
knowledge statements, the agent will use explanatory dialogue to explain them to multiple related
experts and facilitate the corresponding knowledge integration and knowledge validation. Following
the accumulation of knowledge in active learning, the demand for expert knowledge validation will
quickly decrease. Experts only need to participate in the validation when an unknown case emerges.
To ensure the understanding of experts in the given topic during knowledge validation, we can provide
customised conversation samples as context into the prompts for corresponding experts to make the
response of LLM easy to understand for the corresponding experts.
Moreover, the explanatory dialogue can involve multiple experts and each expert will be consulted
with familiar topics based on the profile of the expert. The agent will explain the context of the given
topic based on validated knowledge so that it can minimise the knowledge requirement of experts on
other topics. We tested the idea with simple examples in our initial tests. Improving and enriching the
corresponding prompt templates and samples for this purpose is one of the important parts of our future
work. Through the initial tests that have been done, we realise that the development of the necessary
knowledge models in a practical task is complicated and challenging, and it requires more effort to
improve the current pipeline. Fortunately, previous tests have shown that, with proper dialogues and
prompt engineering templates, LLMs can help simplify the entire process, from knowledge validation
and integration to the deployment of knowledge models in specific tasks.
This research is ongoing. In this paper, we present our current progress by showcasing the initial
interface, introducing the agent-based framework through a basic use case, and assessing the feasi-
bility and efficiency of agents in human-machine interaction. The tests discussed previously show a
promising prospect and good feasibility in this direction. We developed the necessary agents, basic
knowledge models, and prompt templates to support these initial tests, but the framework still needs
more sophisticated knowledge models to cover a practical use case. In future studies, we will continue
to improve this approach and knowledge models discussed in the paper based on the experience that
we learnt from our initial use case test and try to improve the efficiency of the pipeline. In addition, we
will try to apply this framework in practical applications of DIGITAL4Business and improve it based on
feedback. The main plans that will be implemented in future research and development are listed below.
First, we will continue to introduce more prompt templates, concept sample samples, and constraints
to facilitate knowledge integration and support more sophisticated task applications. The pertinent
samples for particular tasks can positively enhance the response of LLMs, and we will reinforce the
13
Yao Yao et al. CEUR Workshop Proceedings 1–15
collection and annotation of these samples in the next version of our framework. Second, the interface
should be able to provide a more diverse knowledge representation to interact with expert users. For
example, a knowledge statement could be represented graphically in a tree-based structure, and a
more user-friendly interface should be developed for users to give feedback and review the previous
conversation context. Last but not least, the development of more comprehensive knowledge models
with the participation of experts. Through the iteration of active learning in the framework, we plan to
continue updating the initial knowledge model of curriculum development with lecturers in the relevant
master’s degree programmes. This learning process could gradually improve the quality and capacity
of the corresponding knowledge and make the interactive agents fit the given tasks better. At the same
time, we are developing benchmark tests to evaluate the effectiveness of expertise in improving the
performance of LLMs in curriculum development. This result will give a quantitative evaluation of the
impact of expertise in the framework.
Acknowledgments
This work has been developed under the auspice of 1) “DIGITAL4Business: Master’s Programme focused
on the practical application of Advanced Digital Skills within European Companies” URL: digital4business.
eu, a project funded from Dec/2022 to Nov/2026 by the European Commission’s DIGITAL programme
call: DIGITAL-2021-SKILLS-01 grant no.: 101084013; and 2) “SMARDY: Marketplace for Technology
Transfer of Research Data, Software, and Results” URL: smardy-project.eu/), a project funded from 2021
to 2024 by the European Eureka Network programme through The Ireland’s International Research
Fund of Enterprise Ireland (Ref#: IR20210058); and The Romania’s Ministry of Research, Innovation,
and Digitalisation CNCS/CCDI-UEFISCDI (Ref#: PN-III-P3-3.5-EUK-2019-0241) PNCDI III.
References
[1] C. Somers, D. Feenan, D. Fitzgerald, R. Henriques, M. Martignoni, D. A. Parletta, E. Cibin, A. E. Chis,
H. González-Vélez, Systematic needs analysis of advanced digital skills for postgraduate computing
education: The DIGITAL4Business case, in: 25th International Conference on Artificial Intelligence
in Education (AIED 2024), volume 2150 of Communications in Computer and Information Science,
Springer Nature, Recife, Brazil, 2024, pp. 179–191.
[2] D. Walton, Dialogical models of explanation, ExaCt 2007 (2007) 1–9.
[3] D. Walton, A dialogue system specification for explanation, Synthese 182 (2011) 349–374.
[4] A. Arioua, P. Buche, M. Croitoru, Explanatory dialogues with argumentative faculties over
inconsistent knowledge bases, Expert Systems with Applications 80 (2017) 244–262.
[5] A. Casciani, M. L. Bernardi, M. Cimitile, A. Marrella, Conversational systems for ai-augmented
business process management, in: International Conference on Research Challenges in Information
Science, Springer, 2024, pp. 183–200.
[6] N. Klievtsova, J.-V. Benzin, T. Kampik, J. Mangler, S. Rinderle-Ma, Conversational process modelling:
state of the art, applications, and implications in practice, in: International Conference on Business
Process Management, Springer, 2023, pp. 319–336.
[7] B. Settles, Active learning, Synthesis Lectures on Artificial Intelligence and Machine Learning,
Springer, Switzerland, 2012. ISBN:978-3-031-00432-2.
[8] O. Lassila, R. R. Swick, et al., Resource description framework (rdf) model and syntax specification
(1998).
[9] K. S. Candan, H. Liu, R. Suvarna, Resource description framework: metadata and its applications,
SIGKDD Explorations Newsletter 3 (2001) 6––19.
[10] D. L. McGuinness, F. Van Harmelen, et al., Owl web ontology language overview, W3C recom-
mendation 10 (2004) 2004.
[11] H. Boley, S. Tabet, G. Wagner, Design rationale for ruleml: A markup language for semantic web
rules., in: SWWS, volume 1, 2001, pp. 381–401.
14
Yao Yao et al. CEUR Workshop Proceedings 1–15
[12] D. Allemang, J. Hendler, Semantic web for the working ontologist: effective modeling in RDFS
and OWL, Elsevier, 2011. ISBN:978-1-4503-7617-4.
[13] D. Wu, H.-T. Wang, A. U. Tansel, A survey for managing temporal data in RDF, Information
Systems 122 (2024) 102368.
[14] J. R. Kirk, R. E. Wray, P. Lindes, J. E. Laird, Improving knowledge extraction from llms for task
learning through agent analysis, in: AAAI Conference on Artificial Intelligence, volume 38, 2024,
pp. 18390–18398.
[15] S. Yousaf, H. M. U. Haque, M. Atif, M. A. Hashmi, A. Khalid, P. C. Vinh, A context-aware multi-agent
reasoning based intelligent assistive formalism, Internet of Things 23 (2023) 100857.
[16] A. Arioua, M. Croitoru, Formalizing explanatory dialogues, in: International Conference on
Scalable Uncertainty Management, volume 9310 of LNAI, Springer, Québec City, 2015, pp. 282–297.
[17] Y. Yao, J. Liu, C. Ryan, Using active learning and an agent-based system to perform interactive
knowledge extraction based on the covid-19 corpus, The Knowledge Engineering Review 38 (2023)
e8.
[18] G. Team, T. Mesnard, C. Hardin, R. Dadashi, S. Bhupatiraju, S. Pathak, L. Sifre, M. Rivière, M. S.
Kale, J. Love, et al., Gemma: Open models based on gemini research and technology, arXiv preprint
arXiv:2403.08295 (2024).
15