This paper introduces the first step towards a computational method for detecting semantic textual reuse in Ancient Greek literature. While existing tools focus primarily on exact or nearlexical matching, our approach leverages the semantic capabilities of contextual LLMs, aiming to finetune a pretrained encoder via contrastive learning to recognize textual reuse even when expressions are paraphrased and/or morphologically altered. To build a suitable dataset, we developed an automatic pipeline that generates positive samples by extracting paraphrases for each sentence using the Ancient Greek Wordnet and a customtrained morphological re-inflection model. Negative samples, or “confounders”, are selected through topic modeling to ensure thematic relevance while preserving semantic dissimilarity. The model is evaluated through a curated case study on Homeric formulae. We retrieve the top ten most similar sentences in a corpus of Ancient Greek authors from the classical age, assessing model outputs using both standard metrics and comparison with established philological studies. The outcomes demonstrate that contrastive fine-tuning, paired with linguistically informed data augmentation, offers promising directions for identifying non-literal textual reuse in historical corpora. This work contributes a framework for philological discovery, combining deep learning with interpretive scholarship in classical studies. Ancient Greek, intertextuality, contrastive learning, paraphrase generation, topic modeling, morphological inflection, synonyms extraction, computational philology.1
Reuse and, more generally, intertextuality have
always been peculiar lens through which literary
works can be analyzed. It has been the focus of literary
critics and philologists such as Gerard Genette
Existing computational tools for reuse detection in
classical languages are primarily based on lexical
similarity. Among them, the most prominent is the
Tesserae project
Another widely used tool is Diogenes3, a desktop application that enables exact lexical searches across a large corpus of classical texts.
Another significant tool in this domain is TRACER (Büchler et al., 2014), a flexible framework for automatic detection of text reuse that supports multiple similarity measures.
Despite their usefulness, these systems are focused on surface-level matches and fail to capture semantic paraphrases or allusive reuse.
To detect such deeper forms of intertextuality,
recent approaches have turned to distributional
semantics. A key challenge, however, is the scarcity of
annotated and homogeneous corpora in ancient
languages, which makes training large language
models (LLMs) difficult
A seminal contribution in this direction is
The paper by Burns et al. also frames intertextuality as a form of anomaly detection, using the embeddings created with the corpus of a specific author (in this case Livy) as input for a SVM: with this model, the goal is to predict the “Livianess” of each work, so as to find instances in which the authors have alluded to Livy’s works.
Following the parallel between intertextuality and
anomaly detection, similar methods have been
explored in the context of authorship attribution. In
While this approach demonstrates the feasibility of adapting general-purpose models to low-resource historical languages, it suffers from the limitations of using a tokenizer and vocabulary not optimized for Ancient Greek.
A common obstacle encountered in our research pertains the shortage of digitized Ancient Greek texts. The main source would be the Thesaurus Linguae Grecae4, but its policy is against using the data for machine learning purposes.
Nonetheless, the work by
A similar strategy is adopted by
This paper makes the following contributions: • •
We propose an automated pipeline for generating paraphrases of Ancient Greek sentences, combining resources such as the Ancient Greek WordNet with a custom-trained morphological re-inflection model based on annotated Ancient Greek data.
We conduct a qualitative assessment of different contextual encoders for 3 https://d.iogen.es/web/?ver=1.003&user=stud 4 https://stephanus.tlg.uci.edu/
To fine-tune a model for semantic reuse detection in Ancient Greek, we first selected a suitable encoder.
We then constructed a contrastive dataset consisting of 11,305 triplets, each composed of a query sentence, a positive sample (paraphrase), and a negative sample (confounder). The query sentences were randomly extracted from a subcorpus of works by Homer, Thucydides, and Herodotus, taken from the Opera Graeca Adnotata. Positive and negative samples were generated automatically through the paraphrase and confounder generation pipeline described in Sections 2.1 and 2.2.
Although Ancient Greek remains a low-resource language, recent years have seen the development of several contextual language models tailored to its linguistic properties. For our task, the encoder must be able to encode semantic contextual information, particularly the similarity between lexically and morphologically varied expressions.
To evaluate model performance in capturing semantic relationships, we designed a synonym retrieval task, which will be described in detail in Section 2.2.
The models considered include: •
& Graziosi, 2023): A BERT-based architecture pre-trained on modern Greek and fine-tuned on Ancient Greek texts from First1KGreek5, Perseus Digital • •
Library6 and data obtained from fellow scholars. The training corpus comprises approximately 70 million words. In its 50K version, a WordPiece tokenizer was trained on the same corpus, resulting in a vocabulary of 50,000 subword units tailored to Ancient Greek.
GreBERTA
Word2Vec: A non-contextual baseline model, included for comparison.
As will be further explained in section 2.2, lemmatization was necessary for synonym extraction. We therefore compared the two main lemmatization libraries available for Ancient Greek: CLTK11 and greCy12. 5 https://opengreekandlatin.github.io/First1KGreek/ 6 https://www.perseus.tufts.edu/hopper/ 7 https://opengreekandlatin.org/. 8 https://inventory.clarin.gr/corpus/890. 9 https://patristica.net/graeca/. 10 https://archive.org/. 11 http://cltk.org/ 12 https://github.com/jmyerston/greCy Meaning “To go up” “To go up” Predicted Model and synonym lemmatization διαβαίνω Logion 50K
with CLTK διαβαίνω Logion 50k
with greCy στείχω Logion BASE “To go”
with CLTK στείχω Logion BASE “To go”
with greCy στείχω GreBerta with “To go”
CLTK στείχω GreBerta with “To go”
greCy διαβαίνω Word2Vec
with CLTK διαβαίνω Word2Vec with greCy “To go up” “To go up”
Similarity
Score
0.34
13 https://github.com/gcelano/LemmatizedAncientGreekXML
14 https://github.com/sigmorphon/2023InflectionST
Since the objective of our model is to detect semantic
reuse, positive samples must exemplify cases of
nonliteral reuse. For this purpose, we developed an
automated pipeline for paraphrase generation
through targeted lexical substitution, following data
augmentation techniques such as those described in
Specifically, we focused on substituting
semantically salient tokens—nouns, verbs, and
adjectives—with suitable synonyms. To identify
these, we combined lexical information from the
Ancient Greek WordNet
For each semantically relevant word in a sentence, we queried the WordNet to retrieve its synsets (i.e., sets of synonyms grouped by sense). For each offset (individual sense), we collected a candidate list of synonyms. We then computed the cosine similarity between the contextual embedding of the original word and four contextual embeddings of each synonym, obtained by extracting four different sentence contexts in which that synonym appears. The sentences were extracted from the corpus Lemmatized Ancient Greek Texts13 by Giuseppe Antonio Celano.
This method allowed us to select the most semantically coherent synonym among candidates, accounting for the high degree of polysemy in Ancient Greek vocabulary. 2.1.1. Re-inflection Model As mentioned above, the synonym selection pipeline outputs the lemma of the best synonym. However, to generate a valid paraphrase within the Ancient Greek sentence, it is necessary to re-inflect the selected lemma according to the morphological features of the word it replaces.
To this end, we developed a morphological reinflection model, which takes as input the lemma and a set of morphological features (e.g., case, number, tense) and returns the inflected form.
The model was trained on a corpus constructed by merging and normalizing data from multiple resources: • SIGMORPHON 2023 – UniMorph Shared Task14: 5,572 inflected forms annotated with morphosyntactic features. • •
linguistically annotated forms originally
produced by the Morpheus parser and
generator of Ancient Greek inflected forms
Opera Graeca Adnotata15
After removing defective entries and applying standard normalization procedures (e.g., Unicode harmonization, feature unification), we trained a sequence-to-sequence model composed of an LSTM layer, a dropout layer, and a Bidirectional LSTM decoder. This architecture was chosen for its balance between simplicity and effectiveness in characterlevel morphological generation tasks.
To create negative samples for the contrastive learning task we introduced the notion of lexical confounders: these are sentences that share semantically relevant words with the target sentence but express a different meaning. This technique allows us to create “hard negatives”, capable of aiding the model in identifying sentences with no lexical overlap but semantically similar, teaching it to disentangle lexical similarity from semantic equivalence.
To automatically select these confounders, we applied topic modeling with the goal of identifying sentences that differ in thematic content. The underlying assumption is that sentences on distinct topics are unlikely to convey the same meaning, even if they share lexically similar elements.
The topic modeling process was carried out on the Opera Graeca Adnotata corpus, leveraging lemmatized tokens to improve generalization. We first applied the Hierarchical Dirichlet Process (HDP) to estimate the optimal number of latent topics (resulting in k = 10), and then trained a Latent Dirichlet Allocation (LDA) model accordingly. The resulting LDA model achieved an average UMass topic coherence score of −0.68, indicating a moderate level of interpretability suitable for the identification of semantically distinct negative samples. 15 https://github.com/OperaGraecaAdnotata/OGA
In this section, we present the results obtained from the evaluation of the two main components of our pipeline: the re-inflection model and the contrastive sentence encoder.
To generate grammatically coherent paraphrases, we trained a sequence-to-sequence model to perform morphological inflection from lemma + features to surface form. The architecture consists of a singlelayer LSTM followed by dropout and a bidirectional LSTM.
The model was trained for a maximum of 120 epochs with early stopping (patience = 10), halting at epoch 77. We used the Adam optimizer with a learning rate of 0.001.
The learning curves of accuracy and loss for the training and validation set can be seen in Figure 1 and Figure 2.
The model reached 0.90 accuracy on both the validation and test set. While performance on frequent forms is consistent, rare accented forms remain problematic. For instance, characters such as “ΐ” and “ΰ”, which appear only 398 and 48 times respectively in the validation set, obtained F1-scores as low as 0.39 and 0.21. This imbalance affects the macro average, which is significantly lower than the weighted average, as shown in Table 3 (test set results).
Nonetheless, performance on frequent cases is sufficient to support the generation of realistic paraphrastic samples.
To fine-tune the Logion 50k model, we used the HuggingFace SentenceTransformers library, representing each sentence with its [CLS] embedding. The model was trained for 7 epochs, reaching its optimal performance at epoch 6.18. We used the AdamW optimizer with a learning rate of 5e-6 and a weight decay of 0.01.
The contrastive dataset was split into 80% training, 10% validation, and 10% testing, with sentence triplets shuffled prior to the split to ensure distributional uniformity across subsets.
Figures 3 and 4 illustrate the training and validation curves for loss and accuracy, showing a stable convergence pattern.
The final accuracy on the test set is 0.81, marking a notable improvement over earlier experiments. In a preliminary run using only 5,000 triplets, the model reached an accuracy of 0.71, highlighting its sensitivity to the amount of training data.
Due to the computational complexity of the pipeline used to generate positive and negative samples, we limited the dataset to ~11,000 triplets. However, we hypothesize that a larger dataset— enabled by scaling the paraphrasis and confounder generation—would likely lead to further performance improvements. The model shows strong generalization capabilities despite the relatively limited dataset size.
To evaluate the model’s ability to detect semantic reuse, we selected Homeric formulas from the Odyssey and retrieved their most similar counterparts from the prose corpora of Herodotus and Thucydides using cosine similarity.
We first performed a general comparison by encoding all sentences from Homer, Herodotus, and Thucydides. For each Homeric sentence, we computed the most similar sentence from both historians. Figure 5 reports how often the most similar match came from each author. Herodotus consistently emerged as the “most Homeric” in style.
We then zoom in on the top matches for a handful of Homeric formulas. Table 4 reports the top-3 most similar matches (with cosine similarity) from Herodotus and Thucydides.
In Herodotus, the top match for “ἄσμενοι ἐκ θανάτοιο, φίλους ὀλέσαντες ἑταίρους” (“Glad to have escaped death, having lost dear companions”) is: κομισθεὶς ἄρα ἐς τὰς Ἀθήνας ἀπήγγελλε τὸ πάθος (V.87) “Back in Athens, he reported the terrible news.” (CosSim: 0.73)
Though the sentences are lexically unrelated, the narrative context aligns: both recount survival from disaster followed by the emotional burden of reporting it. In the Herodotean passage, the warrior coming home is the only survivor: he, too, has “lost dear companions”. The model appears to capture these semantic and narrative parallels, ignoring surface forms.
On the other hand, the matching Thucydidean
phrase “καὶ τροπαῖον στήσαντες ἀνεχώρησαν ἐς τὸ
Ῥήγιον” (IV.25) refers to a commemorated but
marginal victory: as noted by
Sim: 0.73
Across both historians, the model demonstrates sensitivity to semantic and narrative similarities even in absence of direct verbal overlap. This reinforces the notion that the contrastive objective, paired with linguistically-informed data, enables detection of nonliteral textual reuse. Herodotus tends to reuse Homeric motifs to elevate the narrative or align with epic tradition, while Thucydides may repurpose similar forms to subvert or problematize epic conventions.
The results of our evaluation show that the proposed model is capable of identifying semantic similarity in Ancient Greek texts with a significant degree of accuracy. The performance of the contrastive model— reaching 0.81 accuracy on the test set—suggests that even with a relatively limited dataset, it is possible to fine-tune contextual embeddings for a low-resource language such as Ancient Greek.
Importantly, our qualitative case study demonstrates that the model does not rely solely on lexical overlap, but is able to capture semantic connections grounded in context. This capability is particularly relevant for supporting scholarly analysis of textual relationships, where surface variation and thematic connections require careful interpretation.
Our analysis of Herodotus' proximity to Homer in the similarity distributions aligns with established literary hypotheses about thematic continuity and shared motifs between these authors. However, it is important to note that the semantic similarities detected by our model represent connections that merit further philological investigation rather than definitive instances of literary allusion. The distinction between shared themes, common literary topoi, and intentional intertextual references requires expert scholarly judgment that goes beyond computational analysis.
The matches found in Thucydides, while semantically related to Homeric passages, illustrate this distinction clearly: while our model identifies thematic connections, determining whether these represent ironic reuse, coincidental similarity, or genuine allusion requires deeper interpretive knowledge of the historical and literary context. The contrastive learning objective appears well-suited to identifying such semantic connections as potential candidates for scholarly investigation.
Our approach faces several important limitations that should be acknowledged: • Methodological limitations: The generation of paraphrastic and confounding samples, while linguistically motivated, is computationally expensive and depends on the quality of available lexical resources. The method relies heavily on the accuracy of synonym lists from Ancient Greek WordNet and morphological re-inflection models. • Evaluation constraints: Our evaluation remains primarily qualitative and impressionistic. A more rigorous assessment would require comparison with known allusions identified in scholarly literature, which represents a significant challenge for future work. • Scope of detection: Our model identifies semantic similarities and thematic connections, but cannot distinguish between coincidental similarity, shared literary tradition, and intentional allusion. This distinction requires expert philological knowledge and cultural context that computational methods cannot currently provide. • Dataset limitations: The relatively small dataset limits the model's generalizability, and further work is needed to expand coverage across different genres, time periods, and authors to explore cross-genre or diachronic reuse phenomena.
These limitations do not invalidate our approach but rather define its appropriate scope: as a tool for identifying semantically related passages that warrant scholarly attention, rather than as an autonomous detector of literary allusions.
This paper presented a novel approach to the detection of semantic reuse in Ancient Greek literature through the use of contrastive learning and contextual language models. We developed a pipeline for generating paraphrastic sentence pairs and lexically confounding negatives, enabling the finetuning of an encoder model specifically trained for Ancient Greek.
Our method demonstrates the feasibility of identifying thematic connections and semantic relationships in ancient texts, providing a foundation for future work in computational intertextuality detection.
While promising, this system is not meant to replace human judgment. In many cases, interpretation requires close reading and contextual insight that go beyond the scope of automated retrieval. Rather, our model should be seen as an exploratory aid, offering novel perspectives and candidate matches for scholarly validation.
Looking ahead, our goal is to scale the dataset by including larger portions of Herodotean, Thucydidean, and Homeric corpora, and to refine the model further through application to other authors and genres. In particular, we aim to focus on specific thematic domains such as the lexicon of the sacred. Future work should also include more rigorous evaluation against annotated corpora of known literary allusions identified in scholarly literature as well as an evaluation of the paraphrases and confounders by scholarly experts,
Ultimately, this study shows that the intersection of artificial intelligence and philology is not only feasible, but capable of generating innovative and promising contributions to the study of ancient textual reuse.
We thank Federico Boschetti for sharing annotated corpora and the Ancient Greek WordNet, Sebastian Padò for his contribution to the initial phases of this work, and Barbara Graziosi for her kind willingness to discuss the use of large language models for Ancient Greek.
The implementation code is available from the corresponding author.
In the following image, in green the correct synonyms, in yellow those semantically similar to the original word and in red the results stemming from an incorrect lemmatization, while the synonyms considered wrong are not underlined. During the preparation of this work, the author(s) used ChatGPT (OpenAI) in order to: Text translation and Improve writing style. 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.