results. This paper explores the employment of LLMs, specifically of Mistral-Nemo, in the semi-automatic population of the Ancient Greek WordNet synsets. Several approaches are investigated: zero-shot, few-shots, and fine-tuning. The results are compared against an English baseline. Zero-shot approach yields the highest accuracy, while fine-tuning leads to the highest number of potential synonyms. Our analysis also reveals that polysemy and PoS play a role in the model's performance, as the highest scores are registered for polysemous words and for verbs and nouns. The results are encouraging for the application of such approaches in a human-in-the-loop scenario, since human validation still proves crucial in ensuring the accuracy of the ∗Corresponding author.
CEUR Workshop ISSN1613-0073
guage Models (LLMs) for populating the synsets of the
In this paper, we explore the application of Large Lan- short definition and an ID-number ([
groups of cognitive synonyms, each associated with a designed to represent both synonymy and polysemy, via meanings by groups of quasi-synonymous words con- riphéggeia, augasmós, bolē ́, kiéllē 1 all belong to the synset tent to which these models can support such a task.
WordNet is a lexical resource that organizes word
nected to each other in a network structure ([
Drawing from a previous collaboration with the
Uniwas developed in 2014 as the result of an international
collaboration between the Institute of Computational
Linguistics “Antonio Zampolli” (Pisa), the Perseus Project,
the Open Philology Project, and the Alpheios Project. It
1Note that in the experiment both the inputs and the outputs of the
model were written in the Greek alphabet. In this paper, however,
all Ancient Greek lemmas are transliterated and provided with
translations supplied by the LSJ lexicon [
that are interchangeable in all possible contexts, but also words that
are similar in meaning limited to certain contexts ([
Princeton WordNet ([
The work on the AGWN continues in the framework ine how cross-linguistic diferences might influence its of the PRIN project Linked WordNets for Ancient Indo- performance.
European Languages, whose aim is to harmonize three The pretrained model used in all stages of the
experiWordNets for Ancient Greek, Latin, and Sanskrit, and ex- ment is Mistral-NeMo4, a multilingual open source model
pand their coverage in terms of the number of annotated selected because of its balance between performance and
words and populated synsets ([
While various methods have been proposed for the
automatic population of synsets, their outputs typically 2.1. Datasets
still require substantial manual validation. For instance,
word embeddings have been employed to identify lexi- The testing data used in the experiment consists of two
cal relations absent from existing WordNets for Ancient datasets, one made up of (chiefly) monosemous lemmas
Greek ([
The first experiment with LLMs, conducted on Latin the same approach: lemmas possessing just one meaning
([
To validate the results against a benchmark, an English onyms each (minimum 1, maximum 315 for the lemma
baseline (EB) dataset was created. Considering that the peribállō (throw around)), thus divided across PoS: 2946
English baseline serves as a benchmark to highlight difer- nouns (54%), 1372 verbs (25%), 955 adjectives (18%) and
ences in performance between a high-resource modern 185 adverbs (3%)5.
language such as English and Ancient Greek, a substan- The aim of the experiment with Latin WordNet ([
The fine-tuning dataset was created by extracting data as previously discussed. from back-translation dictionaries, based on the assumption that such dictionaries provide, for any given entry 2.2. Zero-Shot Approach in a modern language, a list of Ancient Greek words that can be used in context to translate that entry, that is, contextual synonyms. An example of a back-translation dictionary entry is ofered below: The first approach of the experiment is zero-shot (ZS) learning. This strategy tests the generalization potential and performance of models in tasks for which they were not specifically trained, since “no demonstrations are • Accusation (subs.): P. katēgoría, hē, katēgórēma, allowed, and the model is only given a natural language tó, P. and V. aitía, hē, aitíama, tó, énklēma, tó, V. instruction describing the task” ([25]: 7). Indeed, models epíklēma, tó ([22]). pre-trained on various and general datasets are usually able to generalize across new tasks, thus saving resources needed to create labeled data for additional training or demonstrations ([26]).
Compared to other approaches, zero-shot learning presents several drawbacks, including dificulty with complex tasks and lower accuracy, as outputs may lack precision or contextual relevance. Moreover, it is highly sensitive to prompt framing, which plays a crucial role in this setting ([27]).
As the first stage of the experiment, the zero-shot strategy was applied for both the Ancient Greek dataset and the English baseline. The prompts were tailored to each language and followed the best practices of prompt engineering, such as assigning a persona, specifying the desired output format, and organizing assertions as a bullet list ([28]; [29]). For the complete prompts, see A.1 and A.2.
a dictionary of Ancient Greek synonym sets was extracted from the English-Greek Dictionary ([22]) and the Deutsch-Griechisches Wörterbuch ([23]), merging the results obtained from each dictionary to avoid overlap.
It is important to note that the digital versions of these back-translation dictionaries were obtained through OCR (Optical Character Recognition), which - while generally accurate for modern languages written in the Latin script - yields sub-optimal results for Ancient Greek, often producing incorrectly digitized data and, consequently, inexact outputs. To address this problem, a series of cleaning operations was performed, from encoding normalization to checking the lemmas against the entries of the Brill Dictionary ([24]) to exclude incorrect or non-existent words.
Such cleaning procedures ensure that the assembled dictionary only contains existing Ancient Greek words in their lemmatized form and that each set of synonyms exclusively features lemmas pertaining to the same PoS. 2.3. Few-Shot Approach An example of the synonym sets resulting from the data collection procedure is presented below: In the few-shot (FS) setting, some examples demonstrating the expected output, its format, and style are given • phrikō ́dēs (awe-inspiring): ouránios (heavenly), 5The data collected for fine-tuning will be imported in the AGWN, theîos (divine), deinós (wondrous). to help with the automatic population of the resource. to the model to enhance performance, helping it under- a task-specific model. This was achieved by fine-tuning stand the reasoning required for the new task ([25]). This the quantized Mistral-NeMo model, which was loaded in approach has been proven to generally outperform zero- 8-bit format to optimize computational eficiency, using and one-shot learning ([25]; [30]), especially in structured the previously described fine-tuning dataset on a GPU and complex tasks, such as synonym generation. Com- node of an HPC cluster. LoRA was used to optimize finepared to fine-tuning, this method proves cost-efective tuning, setting the low-rank matrix dimension to 8 and because the weights of the model are left unchanged, the scale factor lora_alpha to 16, with a dropout of 10%. sparing a computationally intensive process, and only a The dataset was split into training (80%) and validation small set of labeled items is needed, which is convenient (20%), and the training was set for five epochs with a in cases of scarcity of data ([27]: 24). However, this strat- learning rate of 1e-4. An early stopping mechanism with egy is strongly dependent on careful prompt engineering a patience of one epoch was established to avoid overfitand on suitable and verified examples. Therefore, par- ting, and a parameter was set to save the model with the ticular attention is needed when designing the prompts lowest value of validation loss, which corresponded to ([31]: 3). As for prompt engineering best practices, per- the output of the fourth epoch. The metrics calculated formance has been proven to increase the more similar during fine-tuning over the five epochs of training are the examples are to testing data. The choice of examples presented in Table 1. also seems to have a great efect on the output ([ 27]: 16).
To test this approach on the Ancient Greek dataset, Table 1 an ad-hoc prompt was created by maintaining the basic Fine-tuning metrics over the five epochs of training. For each structure of the zero-shot prompt and adding a set of metric, the best value is highlighted in bold type. eight examples featuring the same structure of the desired output. The examples are equally divided into roughly 1 2 3 4 5 monosemous and polysemous word sets and are balanced for PoS, so that for each of the four PoS, two lemmas Training loss 1,2943 1,4099 1,1478 1,2232 1,1855 are provided, that is, one monosemous, the other one polysemous. The examples added to the few-shot prompt are listed in A.3. Validation loss 1,4814 1,4366 1,4137 1,4087 1,4100
Training mean token accuracy 0,6587 0,6262 0,6597 0,6720 0,7206 A recent trend with demonstrated advantages is to adapt large-scale pre-trained language models to specific downstream tasks. Indeed, a first stage of generative pre- The overall loss trend is descending, even if gradually, training leads to gaining a greater world and language both in training and in validation, and the accuracy values knowledge and, consequently, to an improved perfor- are increasing. Overall, the metrics show that the training mance. Then, the following fine-tuning (FT) on domain- was conducted successfully and without overfitting. specific labeled data updates the pre-trained parameters with a new training cycle to adapt the model to the task 3. Results and Discussion at hand. This combination of unsupervised pre-training and supervised fine-tuning results in a semi-supervised The validation of the results took place in two steps. The approach able to construct a universal representation, first step was to automatically lemmatize each word uswhich can be applied to a wide array of tasks ([32]: 2). ing greCy ([34]), so that even inflected forms generated
Although fine-tuning greatly enhances model per- by the model are traced back to the corresponding lemma. formance, it is very resource-intensive. Some strate- Notably, this pre-processing step is pointless in the case gies were explored to mitigate this issue, such as LoRA of hallucinations or incorrect forms (for a more detailed (Low-Rank Adaptation), which is a PEFT (Parameter- discussion, see 3.2.1 and 3.2.2). It is worth pointing out Eficient Fine-Tuning) method that makes fine-tuning that the lemmatization, while correct in most cases, was more parameter- and compute-eficient by freezing the not always impeccable (e.g., theoí (gods, masculine nompre-trained model’s parameters and adapting only a sub- inative plural) > theoí (FS)). set of weight matrices. This method proves to be highly After lemmatization, three human annotators 6 valieficient compared to traditional fine-tuning, especially dated the results, determining for each generated item if with regard to memory and storage ([33]: 5), meeting and it constituted a potential synonym of the input word. In sometimes surpassing the baselines, without increases in inference times ([33]).
The final step of the experiment involved fine-tuning
guistics at the University of Pavia with a BA Degree in Classics. cases of disagreement between the annotators, the mat- As for the similarity metric, cosine similarity was comter was resolved through discussion until an agreement puted using pre-trained Word2vec embeddings based was reached. The inter-annotator agreement, measured on a skip-Gram model for both English7 and Ancient with Fleiss’ Kappa ([35]), reached a value of 0.71 on the Greek8. In a task such as synonym generation this metAncient Greek data and 0.66 on the English data, both of ric is useful in determining if the output might be a valid which fall under the label of good to substantial agree- synonym to the target word based on semantics and ment. For the purposes of this work, the concept of syn- distribution. However, one limitation is represented by onymy is interpreted in a shallow and contextual sense, out-of-vocabulary (OOV) terms, meaning that in some consistent with the framework upon which the WordNet cases, for both English and Ancient Greek, the metric architecture is based (see footnote 2). Thus, words whose fails to capture the actual similarity between the genermeaning is similar enough that they might be assigned ated output and the input lemma, as one or both of the to the same synset are considered potential synonyms, two words are not contained in the embedding dictionary, as in 1. such as in 2.a and 2.b: 1 anankázō: rule, hold sway.
kratéō: force, compel.
2.a gourmand: epicure. Similarity: 0. 2.b katasparássō (tear in pieces): katagnúō (break in pieces). Similarity: 0. y EB emZS s on FS o
MFT The results are analyzed both from a quantitative and a qualitative perspective, and the analysis is carried out by comparing the diferent approaches employed, which are bench-marked against the English baseline. Regarding the quantitative data discussed in Section 3.1, the performance of each of the approaches is evaluated through the metrics of accuracy, similarity, number of generated outputs, and potential synonyms.
cient Greek, the latter is more severely impacted by this problem due to the more limited size of the embedding dictionary, thus the similarity values for Ancient Greek tend to be underestimated compared to the English baseline.
As shown in Table 2, the two datasets of the English baseline score the highest values in accuracy, similarity, 3.1. Quantitative Analysis total, and mean of potential synonyms. The results highThe results of the quantitative analysis are shown in Ta- light that the model reaches a high performance in the ble 2, which displays the values of the metrics for each task at hand, even in a zero-shot setting without taskof the approaches, both providing the overall scores and specific demonstrations or training. This result indicates distinguishing between the polysemous and the monose- that the generalization potential of the model is quite mous datasets. high for a high-resource language such as English. As for the zero-shot approach, the first step of the Table 2 experiment shows a much lower performance compared Metrics comparison (acc: accuracy, sim: similarity, n_gen: to the English baseline, across all metrics. Considering number of generated outputs, p_syn: number of potential that pre-trained models have much less data available for synonyms). For each row, the best scores, excluding those of Ancient Greek compared to modern languages such as the EB, are highlighted in bold type to facilitate comparison English, the drop in performance and in the number of across approaches for Ancient Greek synonym generation. generations is to be expected.
acc sim n_gen p_syn Considering now the few-shot approach, the results
EB 90% .377 167 151 show an unexpected drop in performance compared to
l the zero-shot strategy. Indeed, the instructions given in
lra ZS 30% .261 116 34 the prompt apparently do not help the model, but rather
ev FS 5% .099 169 9 afect the outputs negatively. However, it is important to
OFT 11% .077 403 43 point out that the number of generated outputs increases
y EB 98% .407 85 83 compared to the zero-shot approach, reaching the same
emZS 40% .296 63 24 value as the English baseline.
lsy FS 7% .066 61 4 Finally, the results of the fine-tuned model register an
oP FT 13% .113 288 38 overall increase in performance compared to the
few83% .347 68 shot approach. Compared to zero-shot learning, this
19% .226 10 approach scores lower accuracy and similarity, but
registers a higher number of validated potential synonyms.
This is because the number of generated outputs in- of [
Our findings show that the results of the English base- model is provided with a broader semantic basis from
line greatly outperform those of the other approaches which to draw suitable candidates.
across all metrics but the number of generations, which
is highest for the fine-tuned model. Considering the pro- 3.2. Qualitative Analysis
gression of the approaches adopted in the experiment,
one can note that the scores of accuracy and similarity Examples of generations across approaches divided for
drop along every stage of the experiment, contrary to the monosemous and polysemous datasets are shown in
the expectations discussed in Section 2.2-2.4, and to the Table 3.
results of [
4 homôs (similarly): hómoios (similar) (FS). several typologies of orthographic errors and
inconsisAnother type of task misalignment that was frequently tencies were observed. Across approaches, some outputs
observed in Santoro et al. [
A second typology of orthographic errors that was observed is closely tied to the internal conventions of 3.2.1. Non-Ancient Greek Generations Ancient Greek. Across all three training settings, lemmas were generated lacking either the accent (7.a) or the initial breathing mark (7.b). In other cases, the lemmas were generated with an incorrect accent (7.c).
of hallucinations, a term that refers to ‘generated content that is nonsensical or unfaithful to the provided source content’ ([41]). It has been observed in previous literature that hallucinations are amplified by the scarcity of data when dealing with low-resource languages ([42], [43]). Hallucinations are far more frequent in the FS and FT approaches than in ZS. In some cases, the hallucinations share features with the input words, such as the root (see 5.a) or the prefix ( 5.b). In other cases, no such formal relationship seems to exist (5.c).
(FT). 7.b hellēnikós (Greek): ellēnēios (vs hellēnēios) (Greek) (FS).
7.c kritē ́s (judge): brabeûs (vs brabeús) (arbiter) (FS).
5.a plêthos (multitude): poluplēstía (ZS). quent in the zero-shot setting. One may hypothesize that 5.b diakrínō (distinguish): dialúeimi, diēkribállēn these errors are related to the fact that Modern Greek (FS). lacks the initial breathing mark and the iota subscript, and retains a single accent type. A similar type of ortho5.c eupetôs (easily): tlēmatikós (FT). graphic inconsistency, afecting only two generations, is Notably, some of the outputs are generated in languages the use of the iota adscript instead of the iota subscript. other than Ancient Greek, namely English and Modern For the target word kléptēs (thief), the few-shot and fineGreek, even though the prompt specifically instructs to tuning outputs are respectively lēistē ́s (robber) and lēïstē ́s. avoid this behavior (see A.1 and A.2). The inability of While such instances are linguistically and philologically LLMs to consistently generate text in a user’s desired correct, they were not validated as potential synonyms language is widely known in NLP and is referred to as since they are not compatible with the AGWN graphic language confusion ([44]). Examples of language confu- standard regarding the iota subscript. sion in the model’s generations are presented in 6.a and
6.a arktikós (northern): psēlóten/flutter/tall (FT). 6.b éris (strife): antagōnismós (competition) (ZS). Notably, Mistral models have been found to exhibit high degrees of language confusion ([44]), so the presence of languages other than Ancient Greek in the model’s output is not surprising. The problem of English generations also impacted the results of Santoro et al., even though such instances are quite rare in our study. On the contrary, the outputs in Modern Greek are much more numerous, which could depend on an interference efect of the target language’s script. This is because the model likely tends to produce outputs in a higherresource modern language with the same script, as for Latin and English on the one hand, and Ancient Greek and Modern Greek on the other.
as potential synonyms, some interesting observations emerged from the results. One interesting phenomenon that was observed is the generation of rare lemmas or lexical items dating to the Postclassical stages of Ancient Greek (e.g., the Roman or Byzantine period, [45]: 3-6).
For example, as a synonym for kritē ́s (judge) the model generates lutē ́r (arbitrator), a rare lemma that occurs only 6 times in the Thesaurus Linguae Graecae (TLG)9. Only three of such instances are found in Classical texts, while the remaining occurrences come from texts belonging to the Imperial and Byzantine period. Furthermore, the meaning ‘arbitrator’ associated with lutē ́r is rare, as it is attested only for one of its occurrences (A.Th.940), while 9Accessed July, 2025 it usually means ‘deliverer’. An example of a generation distribution of the training data used for fine-tuning, in consisting of a Postclassical lemma is boreinós (northern), which nouns and verbs constituted the majority classes, generated as a synonym for arktikós (northern), which making up, respectively, 54% and 25% of the dataset (see is attested 7 times in the TLG, all in Imperial Greek and Section 2.1), possibly resulting in a bias of the fine-tuned later, and eventually gives rise to the Modern Greek term model. Furthermore, another possible explanation is vorinós. While unexpected, these phenomena do not connected to the diference in performance between the impact the potential for the automatic population of the (roughly) polysemous and monosemous datasets already AGWN proposed in this work, since the AGWN collects discussed in Section 3.1: independently of the PoS of lemmas independently of their frequency or the language the input word, the performance of the model is better stage in which they are attested. for polysemous input words across all approaches but
Focusing now on the diference in performance de- FS. Indeed, verbs are generally considered more polysepending on the PoS of the input lemma, Table 4 shows mous than other PoS as their meanings are thought to for each approach the number of generations and the be more flexible, thus encoding richer semantics ([ 46], number of validated synonyms across PoS, both divided [47]). Nouns also exhibit a high degree of polysemy ([48]). for datasets and overall. Since, as already discussed, polysemous words tend also to be more frequent, the increase in performance for Table 4 these PoS may be linked both to a higher frequency in Model performance across PoS (Tot: generations for PoS; Syn: the training data and to their greater polysemy, which potential synonyms for PoS). For each cell, the highest value provides a broader semantic basis for the generation task is presented in bold type to facilitate comparison. at hand.
This work has explored the potential of LLMs in the semiautomatic population of the AGWN, evaluating and comparing multiple approaches. The first approach tested was zero-shot, which, despite the lack of examples, generated numerous potential synonyms and achieved considerable accuracy and similarity scores, given the task at hand. Contrary to expectations, the few-shot setting marked a decline in results across all evaluation metrics, except the number of generations. Finally, fine-tuning outperformed the few-shot setting, but scored lower accuracy and similarity values compared to zero-shot prompting. However, this approach scored the highest number of generated outputs and potential synonyms.
The divergence between our results and the outcomes
of Santoro et al.’s analysis [
carried out on the High Performance Computing DataCenter at IUSS, co-funded by Regione Lombardia through the funding programme established by Regional Decree No. 3776 of November 3, 2020. The authors wish to express their sincere gratitude to Cristiano Chesi for granting access to the HPC cluster.
Research for this study was funded through the European Union Funding Program – NextGenerationEU S. Wang, L. Wang, W. Chen, Lora: Low-rank adap- Big Data Analytics, Decision-Making, and Pretation of large language models, in: ICLR 2022 - dictive Modelling Systems 8 (2024) 17–24. URL: 10th International Conference on Learning Repre- https://polarpublications.com/index.php/JABADP/ sentations, 2022. article/view/2024-12-10. [34] J. Myerston, J. López, grecy: Ancient greek spacy [44] K. Marchisio, W.-Y. Ko, A. Bérard, T. Dehaze, models for natural language processing in python, S. Ruder, Understanding and mitigating language 2023. confusion in llms, 2024. doi:10.48550/arXiv. [35] J. L. Fleiss, Measuring nominal scale agreement 2406.20052.
among many raters, Psychological Bulletin 76 [45] G. D. Bartolo, D. Kölligan, Postclassical Greek: Prob(1971). doi:10.1037/h0031619. lems and Perspectives, De Gruyter, 2024. [36] S. Stopponi, N. Pedrazzini, S. Peels-Matthey, [46] C. Fellbaum, English Verbs as a Semantic Net, InterB. McGillivray, M. Nissim, Natural language national Journal of Lexicography 3 (1990) 278–301. processing for ancient greek, Diachronica 41 URL: http://dx.doi.org/10.1093/ijl/3.4.278. doi:10. (2024) 414–435. URL: https://www.jbe-platform. 1093/ijl/3.4.278. com/content/journals/10.1075/dia.23013.sto. [47] D. Gentner, I. M. France, The verb mutabildoi:https://doi.org/10.1075/dia. ity efect: Studies of the combinatorial seman23013.sto. tics of nouns and verbs, 2013. doi:10.1016/ [37] H. Nguyen, K. Mahajan, V. Yadav, J. Salazar, P. S. B978-0-08-051013-2.50018-5.
Yu, M. Hashemi, R. Maheshwary, Prompting with [48] A. A. Freihat, F. Giunchiglia, B. Dutta, A taxonomic phonemes: Enhancing llms’ multilinguality for non- classification of wordnet polysemy types, in: Prolatin script languages, in: L. Chiruzzo, A. Ritter, ceedings of the 8th Global WordNet Conference, L. Wang (Eds.), Proceedings of the 2025 Conference GWC 2016, 2016. of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), Online Resources Association for Computational Linguistics, 2025, pp. 11975–11994. URL: https://aclanthology.org/ • Thesaurus Linguae Graecae® Digital Library. Ed. 2025.naacl-long.599/. doi:10.18653/v1/2025. Maria C. Pantelia. University of California, Irvine naacl-long.599. (accessed May 31 2025). [38] O. Shliazhko, A. Fenogenova, M. Tikhonova, A. Kozlova, V. Mikhailov, T. Shavrina, mgpt: Few-shot learners go multilingual, Transactions of the As- A. Prompts Used in the sociation for Computational Linguistics 12 (2024). Experiment doi:10.1162/tacl_a_00633. [39] G. K. Zipf, The meaning-frequency relationship of This appendix contains the full prompts used in the exwords, Journal of General Psychology 33 (1945). periment for both Ancient Greek and English. doi:10.1080/00221309.1945.10544509. [40] T. Fu, R. Ferrando, J. Conde, C. Arriaga-Prieto, P. Reviriego, Why do large language models (llms) strug- A.1. Ancient Greek Prompt gle to count letters?, 2024. doi:10.48550/arXiv. zs_prompt = f"""You are a powerful 2412.18626. AI assistant trained in semantics and [41] Z. Ji, N. Lee, R. Frieske, T. Yu, D. Su, Y. Xu, E. Ishii, Classics.
Y. J. Bang, A. Madotto, P. Fung, Survey of hal- You are an Ancient Greek native lucination in natural language generation, ACM speaker. The only language you speak Comput. Surv. 55 (2023). URL: https://doi.org/10. is Ancient Greek.
1145/3571730. doi:10.1145/3571730. Your task is to provide a bullet list [42] N. M. Guerreiro, D. M. Alves, J. Waldendorf, B. Had- of Ancient Greek synonyms for a userdow, A. Birch, P. Colombo, A. F. Martins, Hal- chosen word. lucinations in large multilingual translation mod- Your response must contain the els, Transactions of the Association for Computa- generated synonyms as comma-separated tional Linguistics 11 (2023). doi:10.1162/tacl_ values.
a_00615. Observe the following instructions [43] M. Abdelrahman, Hallucination in low-resource very closely: [INST] languages: Amplified risks and mitigation strate- - Generate only Ancient Greek gies for multilingual llms, Journal of Applied synonyms. - Provide single-word expressions synonyms for each lemma. ONLY. -- ABSOLUTELY AVOID including any - Do NOT generate long phrases. additional explanations or comments - Make sure to provide numerous in your output. synonyms for each lemma. - VERY IMPORTANT: Make sure the -- ABSOLUTELY AVOID including any outputs are spelled correctly. additional explanations or comments - IMPORTANT: Generate words with the in your output. same part of speech as the input word, - VERY IMPORTANT: DO NOT translate the for example if the input word is a words. verb you must generate only verbs as - VERY IMPORTANT: Use ANCIENT GREEK synonyms. exclusively. -- List each English word separately - VERY IMPORTANT: Generate ANCIENT with proper formatting. GREEK lemmas in the original script """ with accurate diacritics (accents, breathing marks, and vowel quantity A.3. Examples for the Few-Shot Prompt for long vowels indicated by macrons or other notations). word: ’nouthetē ́seis’ - VERY IMPORTANT: Make sure the synonyms: [’paramuthía’, ’protropē ́’, ’parakéleusis’, outputs are spelled correctly. ’parórmēsis’, ’paroksusmós’, ’peithō ́’, ’pístis’, ’kéntron’, - IMPORTANT: Do NOT generate any word ’múōps’, ’paraínesis’] in Modern Greek. - IMPORTANT: Generate words with the word: ’atimázō’ same part of speech as the input synonyms: [’kataiskhúnō’, ’aischúnō’, ’atimóō’, word, ’atimáō’] for example if the input word is a verb you must generate only verbs as word: ’theosebē ́s’ synonyms. synonyms: [’deisidaímōn’, ’eusebēē ́s’, ’eúphēmos’, -- For NOUNS generate only the ’pistós’] NOMINATIVE CASE, as shown in the examples below. word: ’autoû’ -- For VERBS generate only the FIRST- synonyms: [’entaûtha’, ’entháde’, ’autóthi’, ’éntha’, PERSON SINGULAR of the INDICATIVE. ’ekeî’] -- List each Ancient Greek word separately with proper formatting. word: ’trophē ́’ """ synonyms: [’deîpnon’, ’edōdē ́’, ’sîtos’, ’édesma’]
en_prompt=f"""You are a powerful AI assistant trained in semantics. You are an English native speaker. Your task is to provide a bullet list of English synonyms for a user-chosen word.
Your response must contain the generated synonyms as comma-separated values.
Observe the following instructions very closely: [INST] - Generate only English synonyms. - Provide single-word expressions ONLY. - Do NOT generate long phrases. - Make sure to provide numerous word: ’iskhurós’ synonyms: [’drastē ́rios’, ’karterós’, ’energē ́s’, ’rhōmaléos’, ’krataíos’, ’óbrimos’, ’sthenarós’, ’kraterós’]