This study introduces a novel computational framework to analyze multi-modal antagonisms-semantic, emotional, and relational-in dramatic literature, specifically focusing on Shakespeare's Julius Caesar. Employing natural language processing (NLP) techniques, text embeddings, emotion classifiers, and network-based character analyses, we systematically extract and quantify antagonistic relationships within the play. Semantic antagonisms are identified through hierarchical clustering and dimensionality reduction of character embeddings, revealing rhetorical groupings aligned closely with narrative functions. Emotional antagonisms, captured via emotion distribution profiles and variance analysis, illuminate characters' afective dynamics and their alignment with dramatic roles. Relational antagonisms are explored through co-occurrence networks, highlighting unexpected centrality of minor characters as critical mediators of conflict. Integrating these modalities with Hegelian dialectics and Nietzschean interpretations, our tri-modal analysis provides fresh insights into ideological tensions, character motivations, and narrative structure. This interdisciplinary approach demonstrates the efectiveness of AI-driven tools in enriching literary criticism opening new avenues for exploring conflict dynamics in canonical texts.
How can computational methods uncover and analyze multi-modal antagonisms—semantic, emotional, and relational—in dramatic texts, and what does this reveal about the narrative structure and ideological tensions in canonical literature? This question anchors our study at the intersection of computational methods and literary criticism, where advanced methods probe the complexities of narrative conflict in dramatic texts [ 1, 2, 3, 4]. By focusing on antagonism, we employ natural language processing (NLP) and network-based techniques to extract and analyze semantic, emotional, and relational dimensions of conflict [ 5, 6, 7], ofering fresh insights into narrative dynamics.
We apply these methods to Shakespeare’s Julius Caesar, a text rich in antagonistic relationships [8]. The play’s central conflict—between Caesar’s autocratic ambition and the republican ideals of Brutus and the conspirators—drives a dialectical progression of political ideologies, making it an ideal case study for computational analysis of antagonisms.
As a philosophical analyses, from a Hegelian perspective, the clash between Caesar’s power (thesis) and republican resistance (antithesis) resolves in the rise of Octavius and the Roman Empire (synthesis) [9]. Nietzschean lenses further illuminate the characters’ actions as expressions of the will to power and a transvaluation of moral values, with Brutus’s moral ambiguity challenging conventional notions of good and evil [10]. These philosophical frameworks, combined with computational methods, reveal how Julius Caesar navigates individual agency, societal norms, and historical transformation [11].
This study bridges computational techniques and literary analysis to uncover latent patterns in Julius Caesar, advancing our understanding of narrative structure and ideological tensions in canonical literature. Our main contributions are: • Tri-modal Framework: We propose a novel framework to analyze literary antagonism through semantic, emotional, and relational dimensions, leveraging NLP and network-based techniques. • Computational Reading of Julius Caesar: We apply this framework to Shakespeare’s play, revealing hidden patterns of conflict across characters, emotions, and ideologies.
We review key related works, organized by their methodological and thematic contributions to computational literary studies (CLS).
The conceptual foundation of our tri-modal antagonism framework draws directly from prior work that operationalized computational techniques to explore conlfict in dramatic literature. Semantic antagonism originates [12], who applied statistical inference methods to reveal ideological and conceptual oppositions within literary texts, highlighting how contrasting thematic elements can be quantified. Emotional antagonism is rooted [13], where character emotions were analyzed using the EmoLex lexicon, enabling the detection of afective dissonance and mood-based tension across narrative arcs. Relational (or social) antagonism stems from the graphbased analysis of character interactions [14], in which social dynamics and conflict structures were mapped through co-occurrence networks, revealing underlying power struggles and interpersonal oppositions. These three modes—semantic, emotional, and relational—not only capture distinct facets of dramatic conflict but also provide complementary lenses through which narrative antagonism can be systematically modeled and interpreted.
Recent studies have applied Information Theory to characterize writing styles and compare authors quantitatively. For instance, Rosso et al. introduced complexity quantifiers combining Jensen-Shannon divergence with entropy variations computed from word frequency distributions [15]. Their analysis of 30 English Renaissance texts, including works attributed to Shakespeare, revealed distinct entropy clusters for Shakespeare’s corpus, highlighting the homogeneity of his writing style compared to contemporaries. This approach informs our semantic analysis, as entropy-based methods could quantify stylistic markers of ideological conflict in Julius Caesar. However, their focus on stylometry lacks the multi-modal perspective of our framework, which integrates emotional and relational dimensions.
Emotion and sentiment analysis have become central to CLS, ofering insights into narrative emotional arcs and character dynamics. Kim and Klinger surveyed computational approaches to sentiment and emotion analysis, emphasizing their role in tracking plot development and modeling character relationships [16]. Their proposed
mentary dimensions—semantic, emotional, and relational—each of which we operationalize in our computational analysis of Julius Caesar.
Semantic Antagonism: This modality addresses the linguistic and conceptual dimensions of conflict. It encompasses opposing ideas, contradictory statements, or conflicting narratives, where clashes arise from diferences in meaning, interpretation, or framing.
Emotional Antagonism: This modality highlights the afective dimension of conflict. It involves incomderived from character dialogueß. Let = {1, . . . , } denote the set of all speeches and let denote the total number of distinct characters.
embedding for each character is computed as Given a textual encoder : Text → R, the semantic e = 1 || ∈
∑︁ () ∈ R, where ⊆
denotes the speech set for character .
Pairwise semantic similarity between characters and =
e⊤e ‖e‖‖e ‖
Hierarchical clustering (Ward linkage) is then applied on the distance matrix to form semantic clusters {}.
Dimensionality reduction via t-SNE is performed by minimizing
KL( ‖ ), ∝ exp − ︂( ‖e − e ‖2 )︂ chunks {,1, . . . , , }. An emotion classifier emo : eters, and software dependencies necessary to reproduce An undirected weighted graph = (, , ) is con- rate 200, and 1,000 iterations.
Each speech is segmented into overlapping textual Text → ∆ − 1 assigns a probability distribution , ∈ R over emotional categories for each chunk. The aggregate emotional representation for character is computed as
1 p¯ = ∑︀ ∈ =1 ∑︁ ∑︁ ,, with corresponding emotion covariance
1 Σ = ∑︀ ∈ =1 ∑︁ ∑︁
(, − p¯)(, − p¯)⊤.
The emotional distance between characters is defined as = ‖p¯ − p¯ ‖2, and hierarchical clustering is applied on to generate emotion-based character groupings. Emotion volatility for each character is analyzed through diag(Σ ). 4.3. Graph-Based Relational Algorithm ℓ=1 structed with vertices = {1, . . . , } representing characters. Edge weights represent co-occurrence in scenes:
= ∑︁ 1{, co-occur in scene ℓ}, where is the total number of scenes. The following metrics are computed: • Degree centrality: = ∑︀ . • Betweenness centrality: = ∑︀ where is the number of shortest paths from to . • Community detection: Communities obtained by maximizing modularity: () are ̸≠= () , = 1 ∑︁ 2 , ︂( − ︂) 2 via the Louvain algorithm. ( () , ()),
Semantic clusters, emotional clusters, and graph-based communities are integrated to identify and analyze characters’ thematic, afective, and structural roles within the dramatic narrative, highlighting both convergent and divergent patterns.
This section details all engineering choices, hyperparamour multi-perspective analysis. All relevant code is accessible in the companion repository1.
XML parsing of the Shakespeare corpus for Julius Caesar was performed using xml.etree.ElementTree 2, extracting speech nodes and discarding stage directions. Speaker aliases were standardized using a predefined lookup table. Speeches were tokenized and segmented into overlapping chunks of 200 tokens with a 50-token stride using SpaCy 3.7. Semantic Embedding Pipeline. Semantic embeddings were generated using Qwen1.5–Embedding–0.6B (2,048-dimensional output) as the state-of-the-art and most comprehensive comprehensive data embedding model, accessed via sentence-transformers. Speech embeddings exceeding 8,096 tokens were truncated. Mean embeddings per speaker were calculated and cosine similarity was used to create a distance matrix. Ward linkage hierarchical clustering was applied, and embeddings were visualized using t-SNE with perplexity 30, learning
Emotional analysis utilized the j-hartmann/emotion-english-distilroberta-base classifier, predicting probabilities for Ekman’s 6 basic emotions, plus a neutral class. Inference was performed in batches of 32 chunks per GPU pass with gradients disabled via torch.no_grad(). Mean emotion vectors and covariance matrices were computed per speaker. Hierarchical clustering was conducted separately on mean emotion vectors and emotion variance vectors.
A co-occurrence graph was constructed by connecting characters appearing together within each scene. Edge weights represented shared scenes.
Degree and betweenness centralities were computed using NetworkX’s parallel brandes algorithm. Louvain community detection identified stable relational communities (resolution parameter 1.0), and a Fruchterman-Reingold layout (with default parameters) was cached for reproducible visualization. 1https://github.com/convergedmachine/narrative-conflicts 2The XML-encoded version of Julius Caesar used in this study is derived from the public domain edition prepared by Jon Bosak as part of the Moby Lexical Tools project, with SGML and XML markup dating from 1992–1998. The full text is freely available and widely used for computational literary studies. https://www.ibiblio.org/xml/examples/shakespeare/j_caesar.xml Cross-View Integration. Semantic, emotional, and re- speaker and performed unsupervised clustering based lational cluster assignments were integrated into a com- on pairwise cosine distances. The resulting groups were bined character-by-view matrix. Adjusted Rand Index visualized using both a two-dimensional t-SNE projection (ARI) and Normalized Mutual Information (NMI) metrics and a hierarchical dendrogram (see Figures 2 and 3). were calculated pairwise to quantify alignment. The t-SNE plot reveals five coherent clusters:
To uncover latent rhetorical patterns among characters in Julius Caesar, we extracted sentence embeddings for each • Cluster 1 (blue): This small, isolated group includes Varro, Claudius, and Volumnius, all servants of Brutus. Their compact position in the lower-left quadrant suggests a tightly constrained lexical field, largely limited to practical and obedient speech. • Cluster 2 (green): Popilius Lena appears as a lone semantic outlier. His brief but thematically loaded line foreshadowing the assassination gives him a unique lexical profile, detached from any dominant rhetorical faction. • Cluster 3 (brown): This dominant cluster encompasses nearly all central political actors—Caesar, Brutus, Cassius, Antony, and others. Their discursive similarity stems from shared themes of persuasion, honour, and betrayal. Sub-clusters within this group reflect localized interactions, such as the conspirators’ planning or Caesar’s dialogue with Calpurnia and Decius. • Cluster 4 (grey): Characters appearing primarily in Acts IV–V, such as Octavius, Lepidus, and Lucilius, group together due to their military and strategic vocabulary. Their speeches diverge semantically from the courtroom rhetoric of earlier acts. • Cluster 5 (cyan): This group includes Strato, Clitus, Dardanius, and Ghost, unified by themes of death, loyalty, and moral hesitation—especially in the context of Brutus’ final scene.
The dendrogram complements these findings by revealing the relative semantic distances between speakers. The early separation of the servant characters (Cluster 1) from the rest confirms their rhetorical distinctiveness. The clustering of the battlefield and ghostly figures (Clusters 4 and 5) at greater hierarchical distances further illustrates their deviation from the political core.
Overall, these unsupervised methods yield a linguistically grounded stratification of Shakespeare’s dramatis personae, aligning semantic similarity with dramatic function and narrative arc.
To explore the emotional landscape of Julius Caesar, we conducted hierarchical clustering of the main characters using two complementary feature sets: (i) mean scores for seven canonical emotions (fear, anger, sadness, disgust, surprise, joy, and neutrality), and (ii) the variance of each emotion across all speeches. The resulting dendrogram- tional scores, producing interpretable groupings that mirheatmaps reveal distinct patterns of both afective tone ror narrative roles: and emotional dynamism, enabling nuanced insights into dramatic function (see Figures 4 and 5).
ysis clusters characters according to their average emo• Fear-Dominant Cluster: Calpurnia, Cinna, Cicero, and Trebonius display uniformly high fear and minimal joy or anger. These characters voice premonition, anxiety, and the foreboding atmosphere that precedes the play’s central conspiracy. • Anger-Dominant Cluster: Artemidorus, Cato, and Clitus are marked by extreme anger and negligible fear, representing moral outrage and rhetorical resistance within the narrative. • Political-Conspirator Cluster: Central figures such as Caesar, Cassius, Decius Brutus, Marullus, Casca, and Octavius exhibit a balance of moderate fear and anger, with sporadic elevations in disgust. Their emotional complexity aligns with their roles as plotters and statesmen, navigating both ambition and trepidation. • Peripheral and Tragic Clusters: Secondary characters are divided into subgroups reflecting neutrality, disgust, or sadness. For instance, Brutus, Titinius, and the Ghost cluster on high sadness and disgust, encapsulating the play’s tragic undercurrents.
ond analysis leverages the variance (rather than the mean) of each emotional score to capture the dynamic range of afect displayed by each character: • High-Variance Oscillators: Decius Brutus and Marullus show pronounced swings in fear and disgust, indicating characters who are especially reactive to dramatic shifts and moments of crisis. • Steady Strategists: The principal conspirators and leaders (Cassius, Caesar, Casca, Antony, Brutus, Portia, Octavius) exhibit moderate, balanced variance—demonstrating emotional adaptability but avoiding extremes. • Volatile Grievers: Messala and Titinius are distinguished by their high variation in sadness, relfecting the erratic and volatile mourning present in the aftermath of Caesar’s death. • Emotionally Static Roles: Calpurnia and Lucius exhibit near-zero variance across all emotions, reflecting their dramatically narrow and functionally consistent roles. • Citizenry Clique: A tight public-voice community expressing collective opinion. • Conspiratorial Circle: An insular revolutionary faction (Brutus, Cassius, Decius Brutus, Casca, Trebonius, Metellus Cimber) united by shared secrecy and action. • Military-Political Group: A post-assassination alliance (Cato, Strato, Octavius, Clitus, Pindarus, Titinius) reflecting battlefield loyalties and emerging power structures. • Peripheral Actors: Figures such as Lepidus and Cicero occupy network fringes, marking episodic involvement and rhetorical interventions. character is throughout the play. Together, these analyses provide a layered map of afective structure: highlighting both the tonal “centers” of each character and the degree of their emotional mobility. This dual approach uncovers not only who is most fearful or angry, but also who remains steadfast, who wavers, and who undergoes the most dramatic emotional transformations on stage.
Relational antagonism emerges from our co-occurrence network analysis (Fig. 6), which models characters as nodes and shared scene adjacency as edges. Node size reflects degree (number of unique co-occurrences), and spatial proximity indicates stronger relational ties. Two These relational patterns mirror the play’s thematic unexpected hubs—the Servant and Lucius—play outsized tensions—populism versus aristocracy, secrecy versus roles in mediating conflicts across social strata. spectacle—and demonstrate that antagonism in Julius
The Servant, located at the network’s geometric center, Caesar is as much a product of mediated interactions links the citizen-cluster (First–Fourth Citizens, All, Cinna among minor characters as it is of head-on clashes bethe Poet) to private councils (Calpurnia, Artemidorus, tween leading figures.
Decius Brutus). This bridging function highlights how subordinate figures sustain information flow between public assemblies and clandestine plots, driving antag- 6. Discussion onism through mediated exchanges rather than direct confrontation. Lucius, with high betweenness, connects The tri-modal analysis sheds light on the multifaceted naPortia, Ligarius, and the core conspirators. His interme- ture of antagonism in Julius Caesar. Semantic clustering diary position underscores familial and servant-master (Experiment 1) aligned tightly with dramatic function: dynamics that both facilitate and fracture alliances. central conspirators and statesmen coalesced into a coheDistinct clusters reveal competitive factions: sive cluster, while servants and battlefield figures formed distinct outliers. This stratification confirms that lexical choices map onto ideological and role-based divisions within the play. Moreover, Popilius Lena’s isolation underscores how brief but thematically charged utterances can create semantic singularities (Figure 2).
Emotional antagonism (Experiments 2 and 3) further nuances these patterns. Mean-based clustering distinguished afective archetypes—fearful, angry, or neutral—consistent with character motivations and plot turns. Variance-based clustering, by contrast, captured dynamic emotional trajectories: Decius Brutus and Marullus emerged as high-variance oscillators, reflecting their reactive roles during crisis moments, whereas figures like Calpurnia exhibited emotionally static profiles.
Taken together, these two views reveal not only “what” emotions characters express but also “how” flexibly they traverse afective states, deepening our understanding of dramatic tension.
Relational network analysis uncovered hidden mediaFCiageusarre c6h:arFaocrtceers-d(iprreucnteedd csou-boscectu).rNreondcee snizeetwreofrlkecotsf sJcuelniues- tors of conflict. Contrary to expectations that leading figadjacency degree; edges indicate shared scenes. The Servant ures dominate network centrality, minor characters such serves as the central mediator linking the citizenry clique to as the Servant and Lucius emerged as high-betweenness conspirators, while Lucius and Messala act as secondary hubs. hubs (Figure 6), facilitating information flow between Distinct clusters correspond to citizens, the conspiratorial cir- political and popular spheres. This finding highlights the cle, a military-political faction, and peripheral actors. importance of subordinate roles in sustaining narrative antagonism and suggests that relational antagonism of- //ieeexplore.ieee.org/abstract/document/9882117/. ten operates through mediated interactions rather than doi:https://ieeexplore.ieee.org/ direct confrontations. abstract/document/9882117/.
Across modalities, we observe significant intersections. [3] M. Escobar Varela, Theater as data: Computational Characters central in the relational graph also tend to journeys into theater research, University of Michioccupy semantically intermediate positions and exhibit gan Press, 2021. moderate emotional variance, indicating a balance of dis- [4] M. Riedl, R. Young, Narrative planning: Balancing course, afect, and connectivity. This interplay suggests plot and character, Journal of Artificial Intelligence that multi-modal antagonism is not merely the sum of Research 39 (2010) 1–40. URL: https://www.jair.org/ its parts but a synergistic network of linguistic, afective, index.php/jair/article/view/10669. and social forces. [5] S. Kusal, S. Patil, J. Choudrie, K. Kotecha, D. Vora, I. Pappas, A systematic review of applications of Limitations Our emotion classifier relies on modern natural language processing and future challenges lexica and may not fully capture Early Modern English with special emphasis in text-based emotion deafective nuance. Minor characters with limited lines also tection, Artificial Intelligence Review 56 (2023) pose challenges for embedding stability. 15129–15215. [6] S. Anwar, M. O. Beg, K. Saleem, Z. Ahmed, Social relationship analysis using state-of-the-art em7. Conclusion beddings, Information Processing (2023). URL: https://dl.acm.org/doi/abs/10.1145/3539608. doi:10.
We have presented a comprehensive computational study 1145/3539608. of antagonism in Shakespeare’s Julius Caesar, introduc- [7] M. Grandjean, Network visualization: mapping ing a tri-modal framework that unites semantic, emo- shakespeare’s tragedies (2015). tional, and relational analyses. Key contributions include: [8] M. E. Hartsock, The complexity of julius • A systematic methodology for extracting and clus- caesar, PMLA 81 (1966) 777–799. URL: tering semantic embeddings, emotion profiles, https://www.cambridge.org/core/journals/ and co-occurrence networks. pmla/article/complexity-of-julius-caesar/ • Empirical demonstrations of how each modality 3FA05E829345562E33A0F8B8371972D7.
illuminates distinct facets of narrative conflict. [9] E. Koomson, Rhetoric and political power in the • Theoretical integration with Hegelian dialectics last century of the Roman Republic, Ph.D. thesis, and Nietzschean will-to-power, enriching inter- University of Cape Coast, 2020. pretive claims about ideological tensions. [10] B. Leiter, Nietzsche’s moral and political philosophy, 2004. URL: https://plato.stanford.edu/ENTRIES/ Our findings underscore the potential of AI-driven tools nietzsche-moral-political/. to augment literary criticism by revealing latent struc- [11] S. O’Dair, Social role and the making of identity in tures of conflict. julius caesar, Studies in English Literature, 15001900 33 (1993) 289–307.
Future Directions Extensions of this work could ex- [12] M. C. Yavuz, Analyses of literary texts by using plore temporal dynamics of antagonism (e.g., sliding- statistical inference methods., in: CLiC-it, 2019. window embeddings across acts), cross-play comparisons [13] M. C. Yavuz, Analyses of character emotions in to identify genre-specific conflict patterns. dramatic works by using emolex unigrams (2020). [14] M. C. Yavuz, Analyses of dramatic network simulations by using markov chains, in: 2021 8th InReferences ternational Conference on Behavioral and Social Computing (BESC), 2021, pp. 1–6. [1] I. Mani, Computational modeling of narrative, [15] O. A. Rosso, H. Craig, P. Moscato, Shakespeare and books.google.com, 2022. URL: https://books.google. other english renaissance authors as characterized com/books?hl=en&lr=&id=w4hyEAAAQBAJ& by information theory complexity quantifiers, Physoi=fnd&pg=PP1&dq=advanced+computational+ ica A: Statistical Mechanics and its Applications 388 tools+probe+the+complexities+of+narrative+ (2009) 916–926. conflict+in+dramatic+texts&ots=JbD3gjWNUj& [16] E. Kim, R. Klinger, A survey on sentiment and sig=LeDYu9IlBQwJPvyZbVlgvncEbyw. emotion analysis for computational literary studies, [2] P. Ranade, S. Dey, A. Joshi, T. Finin, Computational Digital Humanities Quarterly 13 (2019). understanding of narratives: A survey, IEEE [17] H. Makhdom, R. Li, H. F. Wu, K. Lui, S. Zarrieß, Access 10 (2022) 119872–119888. URL: https:
A comprehensive discussion of Julius Caesar through Hegelian and Nietzschean philosophical lenses reveals Table 1 provides a structured summary of the principal several key insights into the play’s exploration of power characters in Shakespeare’s Julius Caesar, annotated with dynamics, morality, and historical progress. their primary narrative roles. The categorization is based
From a Hegelian perspective, Julius Caesar can be in- on their function within the play’s central conflict and terpreted as a dialectical progression of political ideolo- their relationship to the main ideological and emotional gies. The initial thesis of Caesar’s growing autocratic currents. power is met with the antithesis of republican ideals embodied by Brutus and the conspirators. Their conflict ultimately results in a synthesis - the rise of Octavius and the establishment of the Roman Empire. This dialectical movement aligns with Hegel’s view of history as a process of continual development through conflict and
Often considered the tragic hero, Brutus struggles with loyalty to Caesar and duty to Rome.
The key instigator who persuades Brutus to join the conspiracy against Caesar.
The first to strike Caesar; a conspirator against him.
Conspirator who persuades Caesar to ignore omens and attend the Senate.
A conspirator against Caesar.
One of the conspirators against Caesar.
A conspirator against Caesar.
A conspirator who joins late due to his admiration for Brutus.
Assassinated early, but his ambition and legacy drive the play’s events.
Loyal to Caesar, he becomes the primary antagonist to the conspirators after the assassination.
Caesar’s adopted son and heir; member of the Second Triumvirate who wages war on the conspirators.
Member of the Second Triumvirate with Antony and Octavius.
Brutus’s wife.
Caesar’s wife, who warns him against going to the Senate.
Friend and soldier in Brutus’s army.
Friend of Cassius and soldier in the conspirators’ army.
Soldier in Brutus’s army.
Soldier in Brutus’s army.
Soldier who assists in Brutus’s suicide.
Brutus’s young servant.
Servant of Cassius who assists in his suicide.
Servants and soldiers of Brutus.
Represent the Roman populace, easily swayed by the rhetoric of both Brutus and Antony.
Warns Caesar to "beware the Ides of March".
Tries to give Caesar a letter warning him of the conspiracy.
Tribunes punished for removing decorations from Caesar’s statues.
A respected senator who is not part of the conspiracy and is later killed by the Triumvirate.
A senator who frightens the conspirators by wishing them well just before the assassination.
Mistaken for Cinna the conspirator and killed by the angry mob.
The Ghost of Caesar, who appears to Brutus as a manifestation of his guilt.
Declaration on Generative AI During the preparation of this work, the author(s) used ChatGPT (OpenAI) and Copilot (Microsoft) in order to: Drafting content, Text translation, Paraphrase and reword, Improve writing style, Grammar and spelling check, Citation management, and Content enhancement. After using these tool(s)/service(s), the author(s) reviewed and edited the content as needed and take(s) full responsibility for the publication’s content.