Fine-Grained Clustering of Social Media: How Moral Triggers Drive

Preferences and Consensus

Emanuele Brugnoli

1,2 a

, Pietro Gravino

3 b

, Donald Ruggiero Lo Sardo

1,2 c

Vittorio Loreto

1,2,4 d

and Giulio Prevedello

3 e

Sony Computer Science Laboratories Rome, Joint Initiative CREF-SONY, Centro Ricerche Enrico Fermi, Rome, Italy

Centro Ricerche Enrico Fermi, Rome, Italy

Sony Computer Science Laboratories Paris, Paris, France

Physics Department, Sapienza University of Rome, Rome, Italy

Keywords:

Moral Foundation Theory, LLMs, Social Media Analysis, Political Polarization, Clustering Comparison.

Abstract:

The increased access to online information provided by social media platforms allows individuals to form and

convey their beliefs regarding events in their daily lives. The wide range of interactions carried out in these

virtual environments has the power to impact the decisions and behaviours of others, also creating conﬂict,

polarisation, misinformation, and toxic content. When individuals engage in public debates about topics tied

to signiﬁcant societal concerns, these discussions often regard or imply moral values. By analyzing ﬁve years

of Italian Twitter/X debate on immigration, we show how a language model aware of moral values detects

community structures more accurately, better depicting the actual political scenario in Italy.

1 INTRODUCTION

Social media platforms are pivotal in the modern in-

formation landscape, acting as robust channels for the

swift dissemination of a vast array of information to a

global audience. This rapid and extensive online dis-

tribution of content is crucial in shaping public opin-

ion on various issues. On these platforms, users play

dual roles: they are not only recipients of informa-

tion but also active contributors, constantly shaping

and reshaping the online narrative. This process ap-

parently democratizes information sharing, enabling

diverse voices to participate in the collective dialogue,

profoundly inﬂuencing the formation and propagation

of opinions. However, this diversity of interactions

on social media can sometimes result in the creation

of echo chambers (Cinelli et al., 2021). In such en-

vironments, individuals primarily encounter opinions

and information aligning with their beliefs (Brugnoli

et al., 2021). This phenomenon can deepen ideolog-

ical divides as users become increasingly entrenched

https://orcid.org/0000-0002-5342-3184

https://orcid.org/0000-0002-0937-8830

https://orcid.org/0000-0003-3102-6505

https://orcid.org/0000-0002-2506-2289

https://orcid.org/0000-0002-9857-2351

in their viewpoints, potentially intensifying disagree-

ments and societal divisions. While the role of In-

formation Technologies in echo chamber dynamics

is still not fully understood (Gravino et al., 2019;

De Marzo et al., 2023), this aspect of social media

interaction is particularly evident in discussions sur-

rounding social and moral issues.

Building on the principles of the Moral Founda-

tion Theory (Graham et al., 2013), it is evident that

moral beliefs, which are not uniform but are based

on a diverse range of “irreducible basic elements”

described by ﬁve moral dyads: care/harm, fair-

ness/cheating, loyalty/betrayal, authority/subversion,

and purity/degradation, play a signiﬁcant role in these

discussions. Social media platforms thus emerge

as critical arenas for the debate, challenge, and re-

shaping of societal norms and values. The relative

anonymity and lack of face-to-face interaction on

these platforms can sometimes lead to more polarized

and extreme expressions of moral values, escalating

discussions into conﬂicts.

Despite the recognition of these dynamics, there

remains a gap in understanding how the expression of

moral values on social media inﬂuences group forma-

tion and affects user reactions. Our current research

addresses this gap by examining the complex inter-

Brugnoli, E., Gravino, P., Lo Sardo, D., Loreto, V. and Prevedello, G.

Fine-Grained Clustering of Social Media: How Moral Triggers Drive Preferences and Consensus.

DOI: 10.5220/0012595000003636

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 16th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2024) - Volume 3, pages 1405-1412

ISBN: 978-989-758-680-4; ISSN: 2184-433X

1405

play of moral beliefs in the Italian Twitter/X discourse

on immigration. In this study, we integrate moral do-

main analysis into social network analysis, revealing

nuanced patterns of user interactions and group dy-

namics related to moral beliefs.

Our methodological approach employs advanced

computational techniques, including a BERT-based

model speciﬁcally ﬁne-tuned to classify Italian tweets

according to the expressed moral dyads (Brugnoli

et al., 2023b). This allows us to effectively map

tweet content onto a framework of moral beliefs, pro-

viding insights into how these beliefs inﬂuence user

preferences and lead to the formation of consensus

within online communities. Through this innovative

approach, our work contributes signiﬁcantly to un-

derstanding social media dynamics, particularly in the

context of morally charged debates.

In our comprehensive analysis of the Italian immi-

gration debate on social media, we employed a novel

approach to delineate the intricate relationships be-

tween leaders and followers within the network. By

“leaders” we mean the main inﬂuential actors of the

social debate: news outlets or political entities (Brug-

noli et al., 2023a), while the “followers” represent

the rest of the community of users. By focusing

on retweets (RTs) as a primary measure of endorse-

ment, we identiﬁed followers’ engagement with lead-

ers’ content, thus revealing the patterns of inﬂuence

and information dissemination in the network.

We constructed a bipartite network, segregated

into leaders and followers, connected by retweets.

Analyzing this conﬁguration through a monopartite

projection on the leader layer, we utilized an opti-

mized Louvain algorithm (Blondel et al., 2008) to dis-

cern communities based on political leanings. This

analysis was further extended by labelling these com-

munities and mapping their inﬂuence on the follow-

ers, thus inferring followers’ political orientations

based on their retweet behaviour.

Our study deepened by examining two distinct

datasets: one encompassing all tweets from leaders

and the other exclusively containing tweets that ex-

pressed moral values. This dual approach unveiled

critical insights, particularly in distinguishing polit-

ical groups and ideologies when moral content was

considered. Notably, a more nuanced political spec-

trum emerged from the moral tweets dataset, high-

lighting the signiﬁcance of moral values in discerning

political afﬁliations.

Additionally, we observed a consistent cluster-

ing of questionable sources among the news outlets

within speciﬁc communities, indicating potential bi-

ases or alignments with certain political ideologies.

We conducted a rigorous comparison against a ran-

dom benchmark to validate our ﬁndings, assessing the

signiﬁcance of the overlap between community struc-

tures derived from both datasets.

Crucially, we quantiﬁed the moral orientations of

individuals within these communities using a novel

method. By mapping each node onto a 4-dimensional

probability simplex, we assigned a “moral vector”

to every individual, indicating their likelihood of en-

gaging with content related to speciﬁc moral dyads.

This mapping, normalized using z-scores and soft-

max functions, revealed distinct moral conﬁgurations

within communities, showcasing varied ethical under-

pinnings across the political spectrum.

2 RESULTS AND DISCUSSION

2.1 Unveiling Real and Nuanced

Ideological Divisions with a Moral

Focus

Distinguishing accounts between leaders and follow-

ers in social networks, particularly in the context of

discussions surrounding immigration, offers a valu-

able framework for understanding how public opin-

ions and attitudes are shaped and how consensus is

reached within digital communities. Leaders in social

networks play a crucial role in inﬂuencing the beliefs

and attitudes of others, thus driving the collective con-

sensus on various topics (Dong et al., 2017).

In our study, we focus on identifying these leaders

within the context of the immigration debate in Italy

over ﬁve years from 2018 to 2022. To accomplish

this, we analyze a comprehensive set of entities com-

prising news media outlets and political ﬁgures ac-

tively participating in this discourse. These news me-

dia outlets are further classiﬁed based on assessments

by independent third-party organizations, which cate-

gorize them as either questionable or reliable sources.

This categorization is pivotal as it helps in under-

standing the quality and potential bias of the infor-

mation disseminated by these outlets, particularly in

terms of their reputation for spreading misinforma-

tion.

To capture the discourse around immigration,

we collected tweets containing speciﬁc immigration-

related keywords (see Section 3). This keyword-

based approach ensures that our dataset is focused and

relevant to the topic of interest.

Followers in the immigration debate are then iden-

tiﬁed through their engagement with the content pro-

duced by these leaders. Speciﬁcally, we consider

retweets as a key metric for this identiﬁcation, being

AWAI 2024 - Special Session on AI with Awareness Inside

1406

retweeting unanimously regarded as an endorsement

of the content created by others (Becatti et al., 2019).

By analyzing retweet patterns, we can discern the fol-

lowers in the network, and understand how they inter-

act with and propagate the leaders’ messages. This in-

teraction is crucial, as it reﬂects the extent of inﬂuence

leaders have in shaping the discourse and how follow-

ers contribute to the dissemination and reinforcement

of these narratives.

With this node conﬁguration, we built a bipar-

tite network whose layers are leaders and follow-

ers and whose links represent retweets. If a partic-

ular group of followers retweets two different lead-

ers, it suggests that these leaders are likely convey-

ing similar messages or viewpoints. To further an-

alyze these relationships, we employed a monopar-

tite projection on the leader layer (for further infor-

mation on the construction of the monopartite net-

work refer to Section 3.3). This projection simpli-

ﬁes the network by focusing only on the leaders and

the connections inferred through their shared follow-

ers. For the analysis of this projected network, we

utilized an optimized version of the Louvain algo-

rithm (Blondel et al., 2008). Namely, the nodes’ or-

der undergone a random shufﬂe 1, 000 times, and the

conﬁguration with the highest modularity value was

selected. Once these communities were identiﬁed,

we assigned labels to each community based on the

political leanings of the leaders within them (for a

detailed list of leaders and their identiﬁed commu-

nity membership see https://github.com/Sony-CSL-

Rome/Italian information leaders). The next step in

our analysis involved propagating these community

labels to the followers in the retweet network (Ragha-

van et al., 2007). This process allows us to infer the

political leanings of the followers based on the leaders

they retweet. We assume that followers who predom-

inantly retweet leaders from a particular community

are likely to share similar political leanings.

To deepen our analysis, we repeated this labelling

process twice with two different datasets. The ﬁrst

dataset included all tweets collected from the leaders,

providing a general overview of the network dynam-

ics. The second dataset was more focused, including

only those tweets that expressed values aligned with

at least one of the moral dyads, as identiﬁed by our

BERT ﬁne-tuned model. This allowed us to under-

stand not only the political leanings but also the moral

underpinnings of the interactions within the network.

Panels (A) and (B) of Figure 1 show a representa-

tion of the monopartite RT network of leaders aggre-

gated in terms of communities identiﬁed. In these vi-

sualizations, the node positions remain constant, pro-

Figure 1: (a) Monopartite RT network of leaders obtained

considering all tweets they produced. (B) Monopartite RT

network of leaders obtained considering only moral tweets.

Node positions are preserved. Node colours refer to com-

munities. Nodes frame colours refer to the different types of

leaders: political entities (azure), questionable news sources

(dark red), and reliable news sources (dark blue).

viding a consistent framework for comparison. The

community labels, assigned a posteriori, reﬂect the

political leanings of the leaders within each commu-

nity. This approach allows us to visually and analyti-

cally discern the political landscape within the debate.

The comparison between the two representations

– one leveraging all retweets in the followers-leaders

interaction network and the other limiting retweets

to those expressing moral values – reveals critical

insights into the structure of the debate. When all

retweets are considered, the network appears to reﬂect

a two-party system. This system is accompanied by a

minority of news sources that do not align with any

political faction, indicating a certain level of unbiased

or neutral reporting within the media landscape.

However, when the analysis is reﬁned to include

only retweets that express moral values, a more nu-

anced and realistic separation of political entities

emerges. This approach successfully identify the dis-

unity of the Italian left, distinguishing the Far Left

from the Left, as well as the Five Stars Movement

(M5S) from the right-wing bloc (simply called Right),

offering a clearer understanding of the complex polit-

Fine-Grained Clustering of Social Media: How Moral Triggers Drive Preferences and Consensus

1407

ical spectrum within the debate. It underscores the

signiﬁcance of moral values in discerning ﬁner dis-

tinctions between political groups and ideologies.

An interesting observation from both network

conﬁgurations is the placement of questionable

sources. In both cases, these sources tend to cluster

within the same community. This consistency sug-

gests a potential alignment or afﬁnity of questionable

sources with speciﬁc political leanings or ideologies

within the debate.

Then, we exploit the contingency table (Brier,

1980) associated with the the two network represen-

tations to compute the community overlap. This cal-

culation is designed to demonstrate how the commu-

nity structure emerging from the moral tweet analysis

represents a more detailed, ﬁne-grained version of the

broader community structure obtained by analyzing

all tweets.

In Figure 2, we present a visualization that illus-

trates this relationship. This panel speciﬁcally shows

the maximum percentage of nodes from each commu-

nity in representation (B) of Figure 1 - which is based

on moral tweets - that correspond to a single commu-

nity in representation (A) of the same ﬁgure, derived

from all tweets. This comparison is crucial as it re-

veals the degree of alignment or divergence between

the two community structures, providing insights into

how the inclusion of moral content reﬁnes our under-

standing of the network’s dynamics.

To ensure the robustness of our ﬁndings, we

benchmark these results against a random model.

This is achieved by shufﬂing the order of the nodes in

representation (A) 10, 000 times, thereby generating

a wide range of random community structures. For

each shufﬂed version, we calculate the community

overlap with the structure of representation (B). By

comparing our observed overlaps to this conﬁdence

interval derived from the random benchmark, we can

assess the signiﬁcance of our ﬁndings.

This comparison against a randomized benchmark

is essential for two main reasons. First, it allows us

to determine whether the observed overlap between

the two community structures is statistically signif-

icant or merely a product of random chance. Sec-

ond, it provides a baseline to understand the extent

to which focusing on moral tweets enhances the res-

olution of community detection in social media net-

works. If the observed overlap signiﬁcantly exceeds

the random benchmark, it underscores the value of in-

cluding moral dimensions in network analyses, partic-

ularly in understanding complex social and political

discussions.

Figure 2: Community overlap between the Leader RT net-

works of all tweets (A) and moral tweets (B), respectively

(see Figure 1). Radar shows the maximum percentage of

nodes of a community of (B) that fall in exactly one commu-

nity of (A). Results are also compared with the conﬁdence

interval of a random benchmark (10,000 reshufﬂing).

2.2 Diverse Ideologies Align with Varied

Moral Conﬁgurations

To characterize the moral conﬁguration of the

emerged communities, we developed a method to

quantify the moral orientation of both leaders and fol-

lowers. This method involves mapping each individ-

ual in the network onto a 4-dimensional probability

simplex, a mathematical space used to represent prob-

abilities of different outcomes. In our context, the

vertices of the simplex (the 5 standard unit vectors

in R

) correspond to the ﬁve moral dyads as outlined

in the Moral Foundations Theory. These dyads are

care/harm, fairness/cheating, loyalty/betrayal, author-

ity/subversion, and purity/degradation. Namely, we

assigned to each individual in the retweet network its

“moral vector”, whose i-th component represents the

probability to tweet (if the node is a leader) or retweet

(if the node is a follower) content related to the i-th

moral dyad. For example, if a leader frequently tweets

about issues related to fairness/cheating, this moral

dyad will have a higher probability in their moral

vector compared to the other dyads. To ensure that

our analysis of these moral vectors is consistent and

comparable across all components, we further process

the data using a statistical normalization technique.

Speciﬁcally, we convert the raw probabilities into z-

scores. This transformation allows us to compare how

strongly an individual aligns with a particular moral

dyad relative to the average alignment within the net-

work. Finally we normalize using softmax in order to

AWAI 2024 - Special Session on AI with Awareness Inside

1408

Figure 3: Moral conﬁguration of both leaders (diamonds) and followers (circles) of the retweet network, divided by commu-

nity membership. The position of a node represents how the propensity to share moral content is distributed between the ﬁve

moral dyads. The node size indicates the corresponding activity regarding tweets if the node is a leader or retweets if the node

is a follower.

better appreciate the behaviours in which moral val-

ues are over-represented.

Figure 3 clearly shows how nodes within the same

community share similar moral beliefs and how dif-

ferent communities exhibit very different moral con-

ﬁgurations. Namely, the Right community is almost

exclusively represented in the region to the right of the

diagonal identiﬁed by the vertices fairness/cheating

and purity/degradation. Conversely, the Far Left,

Left, and Centre communities, each following dif-

ferent moral patterns, mainly occupy the region to

the left of this diagonal. In other words, care/harm

seems to be not a pivotal dyad in the moral conﬁgura-

tion of Right, as well as loyalty/betrayal and author-

ity/subversion for Far Left, Left, and Centre. Instead,

nodes belonging to the M5S community are mainly

distributed along the diagonal authority/subversion

- care/harm. Beyond visual inspection we assess

how signiﬁcantly the distributions differ through a

MANOVA test (Stevens, 2012). The results show the

differences are extremely signiﬁcant, p-value ≪ 0.05

(see Appendix for further information on the results

of the MANOVA).

3 METHODS

3.1 Data Collection

Our study focuses on the social discussion around im-

migration in Italy, particularly its representation on

Twitter/X. To capture a comprehensive view of the

online immigration debate, we combined lists from

external organizations, encompassing a wide range

of news media and political groups active in Italy

from 2018 to 2022. We then utilized the Twitter/X

API to search for tweets from these sources, ﬁlter-

ing for content containing speciﬁc keywords (Poletto

et al., 2017), i.e. immigrat* (immigrant*), immi-

grazion* (immigration*), migrant*, stranier* (for-

eigner*), profug* (refugee*), ong (ngo). These terms

were chosen for their neutrality, avoiding bias towards

any particular stance on immigration. We also gath-

ered data on the retweeters of these tweets, focus-

ing on tweets with signiﬁcant interaction. Addition-

ally, we used a binary classiﬁcation from prior studies

on disinformation spread (Gravino et al., 2022; Pen-

nycook and Rand, 2019) to determine the credibil-

ity of Twitter/X accounts, categorizing them as either

questionable or reliable sources. Unlike news outlets,

whose reliability can be more commonly agreed upon

in academic literature, political entities do not have

a widely accepted measure of reliability. Therefore,

we did not apply the reliability label to political en-

tities in our research. Table 1 shows a breakdown of

the dataset, which is the same used in (Brugnoli et al.,

2023b).

Table 1: Breakdown of the Twitter dataset.

Category Accounts Tweets Retweets

Questionable 76 23,033 345,624

news outlets (14.7%) (13.1%) (20.5%)

Reliable 403 130,398 362,595

news outlets (78.1%) (74.1%) (21.5%)

Political entities

37 22,507 976,033

(7.2%) (12.8%) (58.0%)

Total

516 175,938 1,684,252

(100.0%) (100.0%) (100.0%)

3.2 Modelling Morality

To explore the inﬂuence of moral beliefs on content

creation and user engagement on Twitter, we ﬁne-

tuned a BERT-based model (Devlin et al., 2018) to

classify, limited to the topic immigration, tweets in

Italian according to the moral dyad expressed, as de-

Fine-Grained Clustering of Social Media: How Moral Triggers Drive Preferences and Consensus

1409

ﬁned by the Moral Foundation Theory (Graham et al.,

2013). Of note, moral values are considered as dyads

of opposing poles (e.g., care/harm), instead of consid-

ering the poles as two separate labels (e.g., care and

harm). Then, every post was annotated by which one

of the ﬁve dyads was mostly expressed, or if no dyad

was present. The ﬁne-tuned model is the same used

in (Brugnoli et al., 2023b) to which the readers can re-

fer for further details about the training set (Stranisci

et al., 2021) and the performance scores.

3.3 Leader Networks

By distinguishing between the selected Twitter/X

accounts (leaders) and the general audience who

retweeted their content (followers), we naturally de-

ﬁne a biadjacency matrix A whose entries a

f l

indi-

cate the number of times the follower f retweeted

the leader l (Holme et al., 2003). Then, to make

the connections between different leaders explicit, the

bipartite network of followers and leaders, which is

completely deﬁned by its biadjacency matrix A, is

projected on the corresponding layer (Saracco et al.,

2017). In other words, by exploiting the relations es-

tablished between followers and leaders when the for-

mer retweet the latter, we can obtain an adjacency ma-

trix for each layer. The adjacency matrix for the lead-

ers tells us how similar are the population of follow-

ers between each couple of leaders. This operation

is straightforwardly implemented through the matrix

product A

A · A. If we indicate with L the to-

tal number of users and with F the total number of

accounts, the dimensions of A and its transpose are

F × L and L × F, respectively. This implies that A

results in a symmetric L ×L matrix whose generic el-

ement a

′

, with l ̸= l

′

, represents the strength of the

link between the leaders l and l

′

. In addition, to pro-

vide a fair representation of all the leaders, thus re-

ducing popularity bias and size effects, we consider

the following normalization procedure. Let T be the

vector of total retweets per page, namely the column

sums of A

. We set T

= (T

· T )

◦−1

, where X

◦−1

denotes the Hadamard inverse of X (Horn and John-

son, 2012). Then, we consider A

= T

⊙ A

the normalized adjacency matrix of the correspond-

ing leader network, where ⊙ denotes the Hadamard

product. The diagonal of A

is set to zero in order to

discard loops.

3.4 Clustering Comparison

A clustering C refers to a way of dividing a set of

data points D into pairwise disjoint non-empty subsets

, . . . ,C

called clusters. Namely,

C = {C

, . . . ,C

} such that











̸=

∩C

i=1

= D.

Let the cardinality of D and C

be n and n

, respec-

tively. Let a second clustering of the same set D be

′

= {C

′

, . . . ,C

′

}, with |C

′

| = n

′

. It is straight-

forward to observe that

n =

∑

i=1

′

∑

′

Most methods for comparing clusterings can be ex-

plained through the use of a contingency table asso-

ciated with the pair of clusterings C and C

′

(Meil

2007). This table is essentially a matrix N with di-

mensions I ×I

′

, where each element n

′

represents the

count of data points belonging to the intersection of

clusters C

from C and C

′

from C

′

. The cluster sizes

in respective clusterings are the row and column totals

of N, that is,

| = n

′

∑

′

and |C

′

| = n

′

∑

i=1

′

In simpler terms, the contingency table provides a

systematic way to compare how data points are dis-

tributed across the clusters of the two different clus-

terings.

4 CONCLUSIONS

Our comprehensive analysis of the Italian immigra-

tion debate on social media has underscored the crit-

ical role of moral values in shaping the dynamics of

digital discourse. Through a novel approach that com-

bined Network Analysis with the Moral Foundations

Theory and Natural Language Processing, we were

able to dissect the intricate relationships and inﬂuence

patterns between leaders (news outlets and political

entities) and followers in this debate.

Our key ﬁndings reveal that the structure of the de-

bate on social media is not only inﬂuenced by political

afﬁliations but is also deeply rooted in moral values.

By examining retweet patterns, we were able to map

out a complex landscape where leaders and follow-

ers formed distinct communities based on shared po-

litical and moral orientations. The use of a bipartite

network, coupled with a monopartite projection and

the application of the optimized Louvain algorithm,

allowed us to identify and label these communities ef-

fectively.

AWAI 2024 - Special Session on AI with Awareness Inside

1410

The distinction between the datasets - one encom-

passing all tweets and the other focused on tweets

with moral content - brought to light the nuanced na-

ture of the discourse. The more granular analysis of

moral tweets led to a clearer differentiation of polit-

ical entities and ideologies, highlighting the pivotal

role of moral values in distinguishing between seem-

ingly similar political groups.

Furthermore, our observation of the consistent

clustering of questionable sources within speciﬁc

communities sheds light on the potential biases in in-

formation dissemination and the echo chambers that

can arise as a result. This ﬁnding emphasizes the need

for critical examination of source credibility in social

media discourse.

By mapping individuals onto a 4-dimensional

probability simplex and assigning them moral vec-

tors, we were able to characterize the moral land-

scape of the debate quantitatively. This approach illu-

minated the varied moral underpinnings within each

community and revealed how different communities

prioritize different moral dyads.

Our study’s visualization, particularly in Fig-

ure 3, effectively illustrates these moral conﬁgura-

tions, clearly representing how different communities

align with speciﬁc moral values. This visual evidence

reinforces the importance of considering the moral di-

mension in understanding social media interactions

and community formation.

In conclusion, taking the moral domain into ac-

count seems to be crucial not only for inferring the

community structure of social networks at a ﬁner res-

olution, but also for understanding where preferences

and the resulting consensus are rooted and differ-

entiated. Our ﬁndings highlight the need for com-

munication strategies that recognize and leverage the

moral dimensions, particularly in politically and so-

cially charged discussions. This approach holds the

potential to mitigate polarization and the formation of

echo chambers.

As we navigate the complexities of digital dis-

course in an increasingly interconnected world, our

ﬁndings offer valuable insights for researchers, pol-

icymakers, and communicators alike. They empha-

size the importance of a holistic approach to under-

standing social media dynamics, one that goes beyond

political leanings and takes into account the under-

lying moral values that drive human interactions and

consensus-building in the digital age.

ACKNOWLEDGEMENTS

This work has been supported by the Horizon Eu-

rope VALAWAI project (grant agreement number

101070930).

REFERENCES

Becatti, C., Caldarelli, G., Lambiotte, R., and Saracco, F.

(2019). Extracting signiﬁcant signal of news con-

sumption from social networks: the case of Twitter in

Italian political elections. Palgrave Communications,

5(91).

Blondel, V. D., Guillame, J.-L., Lambiotte, R., and Lefeb-

vre, E. (2008). Fast unfolding of communities in large

networks. Journal of Statistical Mechanics: Theory

and Experiment, 10008(6).

Brier, S. S. (1980). Analysis of contingency tables under

cluster sampling. Biometrika, 67(3):591–596.

Brugnoli, E., Cinelli, M., Quattrociocchi, W., and Scala, A.

(2021). Recursive patterns in online echo chambers.

Scientiﬁc Reports, 11(1):5909.

Brugnoli, E., Galletti, M., Lo Sardo, R., Prevedello, G.,

Di Canio, M., and Gravino, P. (2023a). Decoding po-

litical social media posts. Nature Italy.

Brugnoli, E., Gravino, P., and Prevedello, G. (2023b).

Moral values in social media for disinformation and

hate speech analysis. In Proceedings of the VALE

workshop at ECAI, (to appear).

Cinelli, M., De Francisci Morales, G., Galeazzi, A., Quat-

trociocchi, W., and Starnini, M. (2021). The echo

chamber effect on social media. Proceedings of the

National Academy of Sciences, 118(9):e2023301118.

De Marzo, G., Gravino, P., and Loreto, V. (2023). Recom-

mender systems may enhance the discovery of novel-

ties. arXiv preprint arXiv:2312.08824.

Devlin, J., Chang, M.-W., Lee, K., and Toutanova, K.

(2018). Bert: Pre-training of deep bidirectional trans-

formers for language understanding. arXiv preprint

arXiv:1810.04805.

Dong, Y., Ding, Z., Mart

ınez, L., and Herrera, F. (2017).

Managing consensus based on leadership in opinion

dynamics. Information Sciences, 397-398:187–205.

Graham, J., Haidt, J., Koleva, S., Motyl, M., Iyer, R., Wo-

jcik, S. P., and Ditto, P. H. (2013). Moral founda-

tions theory: The pragmatic validity of moral plural-

ism. In Advances in experimental social psychology,

volume 47, pages 55–130. Elsevier.

Gravino, P., Monechi, B., and Loreto, V. (2019). Towards

novelty-driven recommender systems. Comptes Ren-

dus Physique, 20(4):371–379.

Gravino, P., Prevedello, G., Galletti, M., and Loreto, V.

(2022). The supply and demand of news during covid-

19 and assessment of questionable sources produc-

tion. Nature Human Behaviour, 6(8):1069–1078.

Holme, P., Liljeros, F., Edling, C. R., and Kim, B. J. (2003).

Network bipartivity. Phys. Rev. E, 68:056107.

Fine-Grained Clustering of Social Media: How Moral Triggers Drive Preferences and Consensus

1411

Horn, R. A. and Johnson, C. R. (2012). Matrix analysis.

Cambridge University Press.

Meil

a, M. (2007). Comparing clusterings—an informa-

tion based distance. Journal of Multivariate Analysis,

98(5):873–895.

Pennycook, G. and Rand, D. G. (2019). Fighting misin-

formation on social media using crowdsourced judg-

ments of news source quality. Proceedings of the Na-

tional Academy of Sciences, 116(7):2521–2526.

Poletto, F., Stranisci, M., Sanguinetti, M., Patti, V., and

Bosco, C. (2017). Hate speech annotation: Analy-

sis of an italian twitter corpus. In Proceedings of the

Fourth Italian Conference on Computational Linguis-

tics CLiC-it 2017, pages 263–268.

Raghavan, U. N., Albert, R., and Kumara, S. (2007). Near

linear time algorithm to detect community structures

in large-scale networks. Phys. Rev. E, 76:036106.

Saracco, F., Straka, M. J., Di Clemente, R., Gabrielli,

A., Caldarelli, G., and Squartini, T. (2017). Infer-

ring monopartite projections of bipartite networks: an

entropy-based approach. New Journal of Physics,

19(053022).

Stevens, J. P. (2012). Applied multivariate statistics for the

social sciences. Routledge.

Stranisci, M., De Leonardis, M., Bosco, C., and Patti, V.

(2021). The expression of moral values in the twit-

ter debate: a corpus of conversations. IJCoL. Italian

Journal of Computational Linguistics, 7(7-1, 2):113–

132.

APPENDIX

Table 2: Signiﬁcance of moral value distribution difference in identiﬁed clusters. Summary statistics of the MANOVA

regression.

Intercept Value Num DF Den DF F Value Pr > F

Wilks’ lambda -0.00 5.00 19369.00 −1.76 × 10

1.00

Pillai’s trace 1.00 5.00 19369.00 −1.76 × 10

1.00

Hotelling-Lawley trace −4.55 × 10

5.00 19369.00 −1.76 × 10

1.00

Roy’s greatest root −4.55 × 10

5.00 19369.00 −1.76 × 10

1.00

Group Value Num DF Den DF F Value Pr > F

Wilks’ lambda 0.51 20.00 64240.66 732.29 0.00

Pillai’s trace 0.54 20.00 77488.00 602.08 0.00

Hotelling-Lawley trace 0.89 20.00 42604.36 861.67 0.00

Roy’s greatest root 0.78 5.00 19372.00 3026.45 0.00

AWAI 2024 - Special Session on AI with Awareness Inside

1412