On the Quest for an NLP-Driven Framework for Value-Based
Decision-Making in Automatic Agent Architecture
Alicia Pina-Zapata
a
and Sara Garc
´
ıa-Rodr
´
ıguez
b
CETINIA, Rey Juan Carlos University, Madrid, Spain
Keywords:
Automatic Agents, Value-Aware Engineering, Natural Language Processing.
Abstract:
As automatic agents begin to operate in high-stakes areas like finance and healthcare, the alignment of AI goals
with human values becomes increasingly critical, addressing the so-called “alignment problem”. To tackle this
challenge, the paper proposes the architecture of a Value-Based autonomous Agent capable of interpreting its
environment through the lens of human values and guiding its decission-making processes in accordance with
its own values. The agent utilizes a natural language processing (NLP) technique to detect and assess the
values associated with various actions, selecting those most aligned with its moral guidelines. The integration
of NLP into the agent’s architecture is crucial for enhancing its ability to make autonomous value-aligned
decisions, offering a framework for incorporating ethical considerations into AI development.
1 INTRODUCTION
Artificial Intelligence (AI) systems are already well-
integrated in our society. The scope of this technol-
ogy spans from automated agents that manage home
energy use or recommendation systems that help you
make the perfect choice for your next movie night to
autonomous vehicles, financial algorithms or medical
diagnosis assistance.
This highly innovative field has grown rapidly in
recent years and represent numerous advancements
and benefits that are part of a major technological
breakthrough. However, as autonomous agents be-
gin to make decisions in high-risk domains, such as
finance or healthcare, it becomes crucial to exercise
caution.
One of the primary risks of AI, according to Stu-
art Russell, is that autonomous agents can inadver-
tently cause harm by pursuing goals that conflict with
human values. This issue is often referred to as the
“alignment problem” (Russell, 2022). For instance,
an intelligent agent tasked with reducing pollution
might shut down entire industries without consider-
ing the social and economic consequences. His view
focuses on ensuring that AI systems are benefitial,
controllable and aligned with human well-being. This
perspective is closely related to the concept of value-
a
https://orcid.org/0009-0005-0412-4128
b
https://orcid.org/0009-0001-4880-605X
aware engineering, where the goal is to incorporate
ethical and social considerations into the design and
deployment of technology.
Within this framework arises the challenge of de-
veloping systems or intelligent agents capable of in-
terpreting the environment in terms of human val-
ues—referred to as value-aware agents (Osman and
d’Inverno, 2023). Once an agent can reason about
values, it is crucial that it also acts according to its
own moral or value guidelines, ensuring it can make
value-aligned decisions.
This paper presents the architecture of an au-
tonomous agent designed to guide its behavior based
on human values. The agent can infer which values
are promoted or demoted by a set of possible actions
and select the one most aligned with its own values.
To detect the values promoted or demoted by a partic-
ular option, an NLP technique is employed to extract
human values from the text description of the option.
This value-detection model consists of a pre-trained
text analyser and a neural network to determine which
values, ranked by their importance, are implicit in the
descriptions of the actions. Additionally, an aggre-
gation function is used to integrate the agent’s values
into the decision-making process, determining the de-
gree of alignment between the agent and each option.
The article’s content is structured as follows: Sec-
tion 2 presents related work concerning value con-
cepts and their computational extensions, along with
a brief review of existing value detection techniques.
Pina-Zapata, A. and García-Rodríguez, S.
On the Quest for an NLP-Driven Framework for Value-Based Decision-Making in Automatic Agent Architecture.
DOI: 10.5220/0013183300003890
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 17th International Conference on Agents and Artificial Intelligence (ICAART 2025) - Volume 3, pages 797-804
ISBN: 978-989-758-737-5; ISSN: 2184-433X
Proceedings Copyright © 2025 by SCITEPRESS – Science and Technology Publications, Lda.
797

In Section 3, the proposed Value-Based Agent is ex-
plained, with a particular focus on the integration of
the NLP model into the agent’s architecture. In Sec-
tion 4, a real-world domain is introduced where the
simulations will be conducted to observe the behav-
ior of diferent agents. Finally, conclusions and future
work are presented in Section 5.
2 STATE OF THE ART
A wide range of research across psychology, phi-
losophy and social sciences agree that values guide
human behaviour, playing a crucial role in human
decision-making processes. Related works regarding
the integration of human values in automatic agents
decision-making schemes include research on value-
based formal reasoning (VFR) frameworks (Wyner
and Zurek, 2024), value-based argumentation frame-
works (VAFR) (van der Weide et al., 2010) or the use
of LLM to generate responses that align with human
values (Abbo et al., 2024).
In state-of-the-art proposals, values are engi-
neered into the decision-making architecture of au-
tonomous agents as automatic behaviour, learned be-
haviour or through value-based reasoning (Noriega
and Plaza, 2024). When considering the implemen-
tation of a value-based computational framework, it
is essential to formally establish an explicit repre-
sentation of human values and their relations. This
involves defining sets of values and creating tax-
onomies. Among other theoretical frameworks that
structure value concepts, the Moral Foundations The-
ory (MFT) (Haidt, 2013) proposes six fundamental
moral values as universal across cultures: care, fair-
ness, loyalty, authority, sanctity, and liberty. An-
other well established framework is the Basic Hu-
man Values (BHV) (Schwartz, 1992), also known as
Schwartz’s Value Theory. This widely recognized
theory, which explores values and the relationships
between them, has been integrated into agents’ ar-
chitectures in various ways (Heidari, 2022) (Karanik
et al., 2024). In this article, Schwartz’s Value Theory
serves as the foundation for the proposed value-based
reasoning architecture, which will facilitate the im-
plementation of a value-driven agent.
According to Schwartz, values are beliefs that
relate to desirable end states or modes of conduct,
which go beyond specific situations and guide the
selection or evaluation of behaviour, people, and
events. In BHV, Schwartz proposes ten fundamen-
tal values, based on the motivational goal they ex-
press: self-direction, stimulation, hedonism, achieve-
ment, power, security, conformity, tradition, benevo-
lence and universalism. In addition to identifying ten
basic values, the theory explicates the structure of dy-
namic relations among them. One basis of this value
structure is the idea that pursuing the promotion of a
specific value will tipically be congruent with foster-
ing some values but will create conflict with others.
It is following this idea that he defines two bipo-
lar dimensions in which the 10 basic values are clas-
sified. One refers to the emphasis of values on per-
sonal interests, or on the well-being of others: so-
cial focus vs personal focus. The second dimension
captures the conflict between values that emphasize
personal growth and exploration and those that fo-
cus on maintaining stability and preventing potential
risks: anxiety-free vs anxiety-based. This classica-
tion leads to four main groups of values: openness
to change, conservation, self-enhancement and self-
transcendence.
The circular arrangement of the 10 basic values
following the previous dimensions leads to a moti-
vational continuum, as in Figure 1. Values located
closer together on the circle are motivationally re-
lated, while those farther apart tend to be motivation-
ally opposed.
Figure 1: Basic Continuum (Schwartz, 1992).
In an extension of his theory (Schwartz et al.,
2012), Schwartz refines the original 10 values into 19
to create a more comprehensive framework for under-
standing human motivations. This expansion allows
for greater specificity in capturing diverse human ex-
periences and highlights the complexity of value in-
teractions across different contexts. The resulting ex-
tended continuum can be seen in Figure 2.
When making a value-aligned decision, individu-
als evaluate and select behavior that maximizes har-
mony with their values. The first step is to identify
the degree to which an available action promotes or
demotes each of the values (ex: SVT 19 values). The
second step is to determine the degree of alignment
ICAART 2025 - 17th International Conference on Agents and Artificial Intelligence
798

Figure 2: Extended Continuum (Schwartz et al., 2012).
of the individual with every possible choice. It is then
that conflicts may arise, when trying to seek the pro-
motion of all values. For example, one action might
promote self-direction, while another action might
prioritize security. In such cases, the individual must
decide between the promotion of two diferent values.
According to van der Weide (van der Weide et al.,
2010), each person can determine the importance they
give to each of their values, creating a personal Value
System that guides their decision-making. This can
help resolve conflict between values, by stablishing
an order of preference between them.
Given this two concepts, the promotion magni-
tudes that relate every possible action with the values
and an individual’s own value system, the value align-
ment can be determined as a conjunction of both me-
assures. Based on this concept of calculating align-
ment by combining or aggregating the two magni-
tudes, an automatic agent can be built so that it repli-
cates the value-aligned decision-making process of an
specific profile of individual (Karanik et al., 2024).
In this state-of-the-art Value-Based architecture,
the agent receives a set of actions along with a pre-
computed list indicating the values that each action
promotes or demotes. This implies that the agent
needs someone to interpret the set of actions in terms
of values for them, which can be a major drawback
when the goal is to construct an autonomous Value-
Based Agent, as the agent depends on this value in-
terpreter. The automatization of the value-elicitation
process could enhance the agent’s autonomy while
also eliminating the bias that a human annotator might
introduce. If we consider the characterization of each
action through a textual description that reflects the
underlying motivations and objectives driving it, the
value-extraction problem can be reframed as a value-
detection in text problem.
This problem is still on ongoing challenge, with
approaches that go from simple word-count-based
methods (Fulgoni et al., 2016) to feature-based
methods utilizing word embeddings and sequences
(Kennedy et al., 2021). The used methods can be
classified in unsupervised or supervised. Some su-
pervised methods are based on The Frame Axis tech-
nique (Hopp et al., 2020), that projects words onto
micro-dimensions defined by two opposing sets of
words to analyze their semantic orientation without
labeled data, and others use the extended Moral Foun-
dation Dictionary (MFD), which includes words as-
sociated with virtues, vices, and moral aspects related
to the five dyads of Moral Foundations Theory (MFT)
(Mokhberian et al., 2020). Supervised techniques in-
clude performing multi-label classification, in which
each label corresponds to a specific human value. The
degree of association of the text with a category (hu-
man value) reflects the extent to which the text pro-
motes that value. In this group it can be highlighted
the use of NLP transformer-based models, such as
XLNET or BERT (Bulla et al., 2024).
The following section explains an extention of
the state-of-the-art Value-Based agent (Karanik et al.,
2024), focusing on the integration of a NLP model
into the agent’s architecture.
3 PROPOSED MODEL
The proposed model in this paper builds on the dis-
cussed Value-Based architecture concept. Along with
the value-detection model incorporation, further con-
tributions of this proposal include the use of the re-
fined SVT with 19 values, instead of the basic 10-
value theory used earlier or the inclusion of the con-
cept of negative promotion (i.e., demotion) of values,
which was absent in the base model, where only pos-
itive promotion was considered. This is an impor-
tant upgrade, as it better reflects real-world scenarios
where situations not only promote values positively
but can also demote them. It also allows for the ex-
pression of disagreement with the possible actions, in-
dicating a negative alignment with them. By consid-
ering both positive promotion and negative promotion
(demotion) of values, the model more accurately cap-
tures the dynamic and sometimes conflicting nature
of how values are influenced in real-life contexts. All
in all, the result is a value-aware and value-aligned
agent capable of perceiving its environment through
the lens of values and acting accordingly.
The proposed architecture for an autonomous
agent capable of making value-aligned decisions (see
Figure 3) consists of two main components: the
agent’s Value System, that represent its value pref-
erences, and a Decision Module, which simulates a
On the Quest for an NLP-Driven Framework for Value-Based Decision-Making in Automatic Agent Architecture
799

value-based decision-making process. The central
idea is that, given a set of options described in text,
the agent’s Decision Module is able to make decisions
guided by its Value System, and act in consequence.
Figure 3: Value-aligned Agent model.
The Decision module includes an NLP-based
value detection model that extracts the magnitudes of
value promotion and demotion from the text descrip-
tions of the actions. Based on these promotion and
demotion magnitudes, along with the agent’s Value
System, an alignment calculator computes the align-
ment magnitude for each option.
Once these magnitudes are calculated, the deci-
sion maker evaluates them to make a final decision.
This could involve selecting the action with the high-
est alignment if only one action is possible, or deter-
mining whether to proceed with each potential action
depending on the positivity or negativity of its align-
ment score. Finally, the agent will act according to
the decision made.
Next, a more in-depth analysis of the various com-
ponents will be presented.
3.1 Value System
As discussed in the previous section, the agent’s Value
System is designed to represent its preferences con-
cerning Schwartz’s 19 human values. These prefer-
ences can be captured by assigning importance mag-
nitudes to each individual value. However, following
the principles of Schwartz’s Value Theory (SVT), it
is crucial to not only consider values in isolation but
also the relationships and synergies between them.
This approach leads to evaluating the importance of
groups of values rather than solely focusing on indi-
vidual ones.
In line with prior models (Karanik et al., 2024),
a normalized fuzzy measure can be used to describe
the importance weights assigned to different value
groups. This allows for a more detailed representa-
tion, as it captures the interactions between values and
how they collectively influence the agent’s decision-
making process.
In this extension of the model, it is crucial to im-
pose a restriction on the values from the 19-refined-
values set that are result from the disaggregation of
one of the 10 Schwartz’s basic values. Specifically,
the sum of these sub-values’ individual importances
cannot exceed 1 (the maximum importance value al-
lowed for their corresponding higher-level Schwartz
value). This restriction is essential to preserve the
monotonicity property of the constructed fuzzy mea-
sure, which could otherwise be violated.
The computation of the fuzzy meassure (that is af-
terwards normalized) given a set of values is as fol-
lows:
ιω
i
({v
s
, ..., v
t
}) = ιω
i
({v
s
}) + ... + ιω
i
({v
t
})+
+
d p
∑
k=1
ic({v
1
, v
2
}
k
) × ιω
i
({v
1
}) × ιω
i
({v
2
}),
(1)
where d p is the number of distinct pairs within
the set and ic is the interaction coefficient used to
model the dynamic interaction of values. Following
Schwartz, three main interaction between values are
considered: (a) negative interaction between values
in the same wedge. Likely, an agent who prefers one
value will also prefer another of the same wedge, for
example, power and achievement, and the weight of
the importance of the group formed by both should
be less than the sum of their single weights (subaddi-
tive measure); (b) positive interaction between values
in opposite wedges. Due to it being unlikely that the
agent would prefer both values of opposite wedges,
such as power and universalism, the weight for this
group should be greater than the sum of their single
weights (superadditive measure) and (c) no interac-
tion between values in adjacent wedges. Values be-
longing to multiple wedges, such as hedonism, face
and humility, are considered to have a negative inter-
action with the values in the two wedges they are as-
sociated with and a positive interaction with the val-
ues in the other two wedges. The resulting expression
of the interaction coefficient is
ic({v
1
, v
2
}
k
) =





+0.25 v
1
, v
2
in opposite wedges
0 v
1
, v
2
in adjacent wedges
−0.25 v
1
, v
2
in the same wedge
(2)
In this way, the fuzzy measure constructed based
on the agent’s individual importance magnitudes over
each value will effectively capture and represent the
agent’s Value System.
ICAART 2025 - 17th International Conference on Agents and Artificial Intelligence
800

3.2 Decision Module
The Decision Module enables the agent to make deci-
sions aligned with its Value System. Firstly, the agent
should detect the promotions and demotions of values
for each possible action. Secondly, it needs to com-
pute its alignment with each option, using the promo-
tion magnitudes and the fuzzy measure that represents
the importance the agent assigns to each subset of val-
ues. Lastly, given the collections of all the alignment
magnitudes, the agent makes its final decision.
The value-detection model allows the agent to per-
ceive how each action promotes or demotes distinct
values. Given the text description of each available
action, this model outputs the promotion magnitudes
associated with each of them.
Figure 4: Value-detection model.
The proposed value-detection model, shown in
Figure 4, is structured into two fundamental compo-
nents: a pre-trained sub-model and a Classification
Layer. The former refers to the well-known BERT
(Bidirectional Encoder Representations from Trans-
formers) model, while the latter consists of a neu-
ral network that serves as the Classification Layer.
This layer allows to fine-tune the pre-trained model
to our specific task and is trained using a labeled
dataset (ValuesMLProject, 2024) consisting of text
fragments and the corresponding promotion or demo-
tion for each of the 19 Schwartz values.
The output of the model is a vector of promotion
magnitudes for the input action, representing the pro-
motion or demotion of each of the considered values.
Given the promotion magnitudes and the Value
System, the alignment of the agent with each action
is calculated using an aggregation function. The func-
tion selected in this paper is the signed Choquet’s in-
tegral (Choquet, 1954). The positive Choquet’s in-
tegral is computed considering the promoted values,
while the negative Choquet’s integral is calculated us-
ing only the demoted values. This two integrals repre-
sent the positive and negative alignment of the agent
with the action, respectively. Then, the alignment is
computed by subtracting the negative alignment from
the positive alignment.
Once the alignment with each possible action is
computed, the agent can make a decision, that ranges
from perceiving each action as desirable or undesir-
able to selecting the most aligned action.
4 CASE OF STUDY
4.1 Domain
Values not only guide human behaviour on an indi-
vidual level, but they also shape human behaviour
at a societal level. As several previous studies
(Caprara et al., 2006) (Barnea and Schwartz, 2008)
(Schwartz et al., 2010) have indicated, political voting
is strongly related to and driven by personal values.
This strong correlation between Schwartz’s theory of
values and political choices provides an ideal field of
application for our current research.
It was demonstrated that voters’ political choices
in Western democracies depend more on personal
preferences, especially values, than on other factors
such as voters’ social characteristics (Caprara et al.,
2006). This study not only proved the primacy of val-
ues among the factors that drive voters’ choices but
also highlighted their lasting influence over time.
It has already been established that human deci-
sions are driven by values, but the direct relation-
ship between values and politics appears to be even
more significant. Certain values are specifically re-
lated, both positively and negatively, to center-left and
center-right political parties (Caprara et al., 2006),
or more specifically to ideologies such as Classical
Liberalism and Economic Egalitarianism (Barnea and
Schwartz, 2008). Basic values are reflected in core
political values (Schwartz et al., 2010), such as law
and order, equality, or the acceptance of immigrants.
Taking into account these foundational studies
that demonstrated the direct correlation between vot-
ers’ values and their political choices, we propose
to extend this research by considering not only
Schwartz’s 10 basic values but also his extended set
of 19 values. One difficulty mentioned in previous
studies on political psychology (Schwartz et al., 2010)
was how to determine which values the political par-
ties are promoting or demoting. The NLP model we
proposed in Section 3 is the key element for overcom-
ing this problem.
4.2 Simulations
The simulations consist of analyzing the political vot-
ing process of several Value-Based Agents with dif-
ferent value preferences. To do so, different pro-
files of agents are implemented following the pro-
posed model, considering the relationship between
On the Quest for an NLP-Driven Framework for Value-Based Decision-Making in Automatic Agent Architecture
801

Schwartz’s values and their corresponding ideology
as discussed in the prior subsection.
The voting process simulates the elections to the
UK Parliament, in which the two major parties, the
Labour Party and the Conservative and Unionist
Party, present very strong ideologies (Social Democ-
racy and Conservatism/Economic Liberalism).
In this voting scenario, agents are given two alter-
natives: the Labour Party or the Conservative Party.
The agents will evaluate each party in terms of val-
ues, calculate their alignment with each of them and
decide on the one that best aligns with their own val-
ues. To facilitate this process, two texts summariz-
ing the ideology of each party (this is, the underlying
motivation of each voting alternative) are considered.
The promotion magnitudes extracted from these texts
can be observed in Figure 5, and represent the values
promoted and demoted by the two considered parties.
(a) Labour Party promotion magnitudes.
(b) Conservative Party promotion magnitudes.
Figure 5: Promotion magnitudes detected.
As a starting point, the voting process is tested
considering two agents with clearly defined ideolog-
ical profiles: one representing a left-wing voter and
the other representing a right-wing voter. The Value
Systems of these agents are constructed based on
Schwartz’s demonstrated correlations between human
values and the political preferences of center-left vot-
ers. Assuming the opposite correlations for center-
right voters, the fixed importances of each value for
both agents are illustrated in Figure 6. Note that these
importances are derived from the combination of sub-
values that constitute each primal value, which are
later broken down into their individual sub-values im-
portances. The Value System of each agent is built
according to this importances, computing the fuzzy
meassure described in Section 3.
(a) Left-wing voter importances.
(b) Right-wing voter importances.
Figure 6: Agent’s importances.
Each agent then computes its alignment with both
parties, by aggregating the promotion magnitudes of
each party (Figure 5) with the fuzzy measure that rep-
resents its Value System (Figure 6). The alignment
magnitudes results can be seen in Table 1.
Table 1: Voter alignments.
Labour Party Conservative Party
Left-wing voter 0.0727 0.0265
Right-wing voter 0.0636 0.0924
Following this, two simulations are carried out,
each considering a different number of agents and a
distinct method for constructing their Value Systems.
For the first simulation, we generate a population
of 200 agents, with 100 agents representing slight
variations of the left-wing profile and 100 represent-
ing variations of the right-wing profile defined earlier.
The Value Systems of these agents are constructed
by making small modifications to the importance val-
ues of the ideological profiles. These modifications
are introduced randomly, adjusting the importance
lw({v
i
}) of each value i by up to 0.1 + 0.1 · lw({v
i
}).
ICAART 2025 - 17th International Conference on Agents and Artificial Intelligence
802

The fuzzy measure is then constructed based on these
modified ideological profiles. Having generated the
population, the voting process is simulated.
To visualize the voting results (see Figure 7),
each agent is positioned within Schwartz’s two-
dimensional value space according to its preference
profile. As outlined in Section 2, this space is defined
by two bipolar dimensions: the x-axis represents the
continuum from social focus to personal focus, while
the y-axis from anxiety-based (protection) to anxiety-
free (growth). This two-dimensional division results
in four quadrants, each of them corresponding to a
wedge of Schwartz’s continuum (Figure 2).
The positioning of each agent reflects its orienta-
tion along both dimensions, and therefore across each
wedge, providing a clear graphical representation of
their Value System orientation. Each agent is repre-
sented in red or blue, indicating their choice to vote
for the Labour Party or the Conservative Party, re-
spectively, based on their alignment with each party.
Figure 7: Random population voting results.
For the second simulation, the voting process is
simulated for a population of 250 randomly generated
agents.
Figure 8: Random population voting results.
Instead of considering only agents with left-wing
or right-wing profiles, each agent’s preferences across
the 19 values are randomly assigned, but always
adhering to the structure established by Schwartz.
Specifically, the order of preferences follows the con-
tinuum of values (see Figure 2). The voting results
are shown in Figure 7.
5 CONCLUSIONS AND FUTURE
WORK
This paper proposes a model for a Value-Based Agent
able to interpret a set of options in terms of values
and make decisions based on its own values. By inte-
grating an NLP model for value detection, the agent
gains greater independence in its decision-making
processes and can evaluate real-time situations based
on relevant values, allowing it to respond dynamically
to changing conditions and leading to more informed
and effective decisions. Moreover, the inclusion of
NLP within the agents architecture paves the way
for future research, where each agent could develop
its own value detection capabilities, creating diferent
ways to perceive and interpret the environment.
The simulation results show the voting decisions
made by agents with different preferences over val-
ues (i.e., different Value Systems). It can be observed
(Table 1) that the two agents constructed with left and
right-wing profiles are more aligned with the Labour
Party (center-left) and the Conservative and Unionist
Party (right-wing), respectively. Moreover, the agents
whose profiles are generated as small variations of
these profiles vote in 98 % of the cases in alignment
with the profile from which they were generated. This
behavior is consistent with Schwartz’s value-based
characterization of a left and right-wing voter. The
graphical representation of the voting decisions of the
population of agents with random value preferences
(following Schwartz’s restrictions) shows that agents
with a tendency towards anxiety-free values, and more
notably, those with values in the Self-Transcendence
wedge, tend to vote for the Labour Party. In contrast,
agents oriented towards protection or anxiety-based
values, especially those with high preferences in the
Self-Enhancement wedge, show a tendency to vote
for the Conservative Party. These voting tendencies
are appropriate given the values associated with each
party’s ideology. Moreover, the voting patterns are
consistent with the agent’s Value Systems, as agents
with similar values tend to cast the same vote.
For future work, it is essential to explore improve-
ments to the NLP model or even investigate alterna-
tive models, as the precision of the model is crucial for
the agent’s interpretation of the environment based on
values. For instance, in the simulations conducted, it
was observed that while the demotion of some values
On the Quest for an NLP-Driven Framework for Value-Based Decision-Making in Automatic Agent Architecture
803

was detected for the Conservative party, no negative
promotion magnitudes were detected for the Labour
Party, which results into a predisposition for agents
with values not strongly associated with a specific ide-
ology to vote for the Labour Party (see Figure 8). Re-
fining the NLP model could lead to more reliable and
accurate results, ultimately influencing the agents’ de-
cisions in a meaningful way.
Additionally, a future line of research could in-
volve the integration of this Value-Based Agent archi-
tecture into an agent with practical applications, such
as an intelligent traffic light or a chatbot.
ACKNOWLEDGEMENTS
This work has been supported by grant VAE:
TED2021-131295B-C33 funded by MCIN/AEI/
10.13039/501100011033 and by the “European
Union NextGeneration EU/PRTR”, by grant
COSASS: PID2021-123673OB-C32 funded by
MCIN/AEI/ 10.13039/501100011033 and by “ERDF
A way of making Europe”, and by the AGROBOTS
Project of Universidad Rey Juan Carlos funded by
the Community of Madrid, Spain.
REFERENCES
Abbo, G. A., Marchesi, S., Wykowska, A., and Belpaeme,
T. (2024). Social value alignment in large language
models. In Osman, N. and Steels, L., editors, Value
Engineering in Artificial Intelligence, pages 83–97,
Cham. Springer Nature Switzerland.
Barnea, M. and Schwartz, S. (2008). Values and voting.
Political Psychology, 19:17 – 40.
Bulla, L., Gangemi, A., and Mongiov
`
ı, M. (2024). Do lan-
guage models understand morality? towards a robust
detection of moral content.
Caprara, G., Schwartz, S., Capanna, C., Vecchione, M., and
Barbaranelli, C. (2006). Personality and politics: Val-
ues, traits, and political choice. Political Psychology,
27:1 – 28.
Choquet, G. (1954). Theory of capacities. Annales de
l’Institut Fourier, 5:131–295.
Fulgoni, D., Carpenter, J., Ungar, L., and Preot¸iuc-Pietro,
D. (2016). An empirical exploration of moral foun-
dations theory in partisan news sources. In Calzolari,
N., Choukri, K., Declerck, T., Goggi, S., Grobelnik,
M., Maegaard, B., Mariani, J., Mazo, H., Moreno,
A., Odijk, J., and Piperidis, S., editors, Proceedings of
the Tenth International Conference on Language Re-
sources and Evaluation (LREC’16), Portoro
ˇ
z, Slove-
nia. European Language Resources Association.
Haidt, J. (2013). Moral psychology for the twenty-first cen-
tury. Journal of Moral Education, 42.
Heidari, S. (2022). PhD Thesis: Agents with Social Norms
and Values: A framework for agent based social sim-
ulations with social norms and personal values.
Hopp, F., Fisher, J., Cornell, D., Huskey, R., and Weber, R.
(2020). The extended moral foundations dictionary
(emfd): Development and applications of a crowd-
sourced approach to extracting moral intuitions from
text.
Karanik, M., Billhardt, H., Fern
´
andez, A., and Ossowski, S.
(2024). Exploiting Value System Structure for Value-
Aligned Decision-Making, pages 180–196.
Kennedy, B., Atari, M., Mostafazadeh Davani, A., Hoover,
J., Omrani, A., Graham, J., and Dehghani, M. (2021).
Moral concerns are differentially observable in lan-
guage. Cognition, 212:104696.
Mokhberian, N., Abeliuk, A., Cummings, P., and Lerman,
K. (2020). Moral Framing and Ideological Bias of
News, pages 206–219.
Noriega, P. and Plaza, E. (2024). On autonomy, governance,
and values: An agv approach to value engineering. In
Osman, N. and Steels, L., editors, Value Engineer-
ing in Artificial Intelligence, pages 165–179, Cham.
Springer Nature Switzerland.
Osman, N. and d’Inverno, M. (2023). A computational
framework of human values for ethical ai.
Russell, S. (2022). Artificial Intelligence and the Problem
of Control, pages 19–24.
Schwartz, S., Caprara, G., and Vecchione, M. (2010). Basic
personal values, core political values, and voting: A
longitudinal analysis. Political Psychology, 31:421 –
452.
Schwartz, S., Cieciuch, J., Vecchione, M., Davidov, E., Fis-
cher, R., Beierlein, C., Ramos, A., Verkasalo, M.,
L
¨
onnqvist, J.-E., Demirutku, K., dirilen gumus, O.,
and Konty, M. (2012). Refining the theory of basic
individual values. Journal of Personality and Social
Psychology, 103:663–88.
Schwartz, S. H. (1992). Universals in the content and struc-
ture of values: Theoretical advances and empirical
tests in 20 countries. volume 25 of Advances in Ex-
perimental Social Psychology, pages 1–65. Academic
Press.
ValuesMLProject (2024). Valuesml dataset. Data pro-
vided as tab-separated values files with one header
line. In addition to the original files in nine languages,
a machine-translated version in English is available.
van der Weide, T. L., Dignum, F., Meyer, J. J. C., Prakken,
H., and Vreeswijk, G. A. W. (2010). Practical rea-
soning using values. In McBurney, P., Rahwan, I.,
Parsons, S., and Maudet, N., editors, Argumentation
in Multi-Agent Systems, pages 79–93, Berlin, Heidel-
berg. Springer Berlin Heidelberg.
Wyner, A. and Zurek, T. (2024). Towards a formalisation
of motivated reasoning and the roots of conflict. In Os-
man, N. and Steels, L., editors, Value Engineering in
Artificial Intelligence, pages 28–45, Cham. Springer
Nature Switzerland.
ICAART 2025 - 17th International Conference on Agents and Artificial Intelligence
804