This paper investigates the performance of two models on Cavalcanti's Rhymes: a supervised neural model (Stanza) trained on the Italian-Old treebank (comprising Dante's Divine Comedy), and an unsupervised generative Large Language Model (LLM) accessed via the ChatGPT API (o3 version). This study highlights the crucial role of textual edition in processing historical texts, illustrating this through examples from diferent editions. It also presents a manual error analysis of the models' outputs, focusing on both the most frequent and the most linguistically nuanced errors.
tion (2.3), and discusses annotation practices and inter- and similar lyric types.9 Moreover, alongside the classical annotator agreement (2.4). Section 3 outlines the training poems, some rhymes belong to the “sonetti di risposta” processes (3.1 for the neural model and 3.2 for the LLM) (stilnovist reply sonnets addressed to other poets) and as well as the evaluation of the models’ performance (3.3). thus difer in their compositional purpose. In Section 4, we provide a manual error analysis of the It is important to recall that, when working with output of the LLM (4.1) and the supervised model (4.2), texts transmitted through manuscript traditions, editoalong with a discussion of the errors common to both rial choices play a critical role. This is also the case for models (4.3). Finally, Section 5 presents the conclusions. the Rhymes, whose textual variants and structure depend heavily on editorial interpretation, as it is discussed in Subsection 2.3. By selecting the edition of the Rhymes 2. Data edited by Rivalta, we adhere to his reconstruction of the texts, along with all the implications that such a choice This Section presents the data used in the study. entails.10
As of now, the Italian-Old Universal Dependencies
treebank [
For the present work, we selected the Rhymes of Guido
Cavalcanti edited by Ercole Rivalta [
Rivalta’s edition of the Rhymes divides the poems into
three chronological groups: those composed before 1290,
those of uncertain date, and those composed after 1290.
The corpus includes a total of 63 poems, distributed as
follows: 25 in the first group, 23 in the second, and 15
in the third. Various metrical forms are attested among
the poems, ranging from ballads and sonnets to canzoni
6The critical edition of the Divine Comedy used while building the
treebank is Petrocchi (1994) [
obj chè morte ’l porta in man that death it holds in hand nsubj acl obj ch’ e’ morto ’l porta ’n man that he being dead it holds in hand
10Rivalta’s edition is not the most up-to-date edition of the Rhymes,
which have since been the subject of further scholarly studies
and have led to other editions curated by other scholars (such as
Contini [
In this case, the main divergence is exegetical: in Ri- to incontrare (to meet). As seen previously, even though valta’s version, morte (death), personified, is the subject this variation does not pose a syntactic problem in this (nsubj) who holds the object (obj) ’l (it) in her hand. case, the use of two diferent forms can impact other By contrast, in Rea’s edition, the personified morte is levels of linguistic analysis and obviously yield diferent replaced by an actual agent, e’ (he), who is morto (dead). results depending on which edition was used as the basis While the logical relation, namely, that there is a subject for annotation. who “holds” something in their hand, remains consistent In conclusion, editorial choices play a crucial role, across both versions, Rea’s interpretation, as shown in as they can influence tokenization, interpretation, and the Figure 1, introduces an additional token. This addi- lemmatization, as demonstrated by the three examples tion occurs since morte is replaced by the nominal mod- discussed above. ifier morto, in the tree acl (adnominal clause) and the modified referent e’ must therefore be explicitly included. 2.4. Manual Annotation and From a strictly syntactic perspective, this diference results in a clearly distinct sentence subject, as shown in Inter-Annotator Agreement Example I, and in the presence of an additional token, Once the reference edition was established, we manuwhich inevitably yields a slightly more complex syntactic ally annotated 22 sonnets out of the 63 poems included structure.12 in Rivalta’s edition of Cavalcanti’s Rhymes (2135 token,
Secondly, in verse 7 of Sonnet 10, we observe a case in excluding punctuation marks). We chose to focus exwhich the diference between editions afects only the to- clusively on sonnets, as this metrical form is the most ken count, without any consequences for interpretation: frequently attested throughout the collection. The manual annotation was carried out by two expert annotators.
Perchè Sospiri e Dolor mi pigliaro (Rivalta ed.) Among the 44 sonnets13 in the corpus, we selected 22, Per che Sospiri e Dolor mi pigliaro (Rea ed.) distributing them across the three periods identified in (garnodanws)ere delighted to hear my sighs and Rivalta’s edition: 6 sonnets from the first period, 7 from the second, and 6 from the third (totaling 19 sonnets14).
In addition, 3 sonnets were annotated by both annotators to calculate inter-annotator agreement.
To evaluate consistency between the two annotators and alignment with the annotation style of the Italian-Old treebank, we also performed inter-annotator agreement on three Canti of the Divine Comedy. The selected Canti were the 13th of each Cantica, namely Inferno, Purgatorio and Paradiso, in order to account for potential stylistic variation across the three parts of the poem (3296 token excluding punctuation marks).
Table 1 reports the inter-annotator agreement results
for both the three 13th Canti of the Divine Comedy
(Dante) and the three sonnets by Cavalcanti.
Interannotator agreement was assessed using Fleiss’ kappa
[
In Rivalta’s edition, the causal conjunction appears as perché (because), which is a single token. In Rea’s edition [19, p.62], the same causal meaning is conveyed through the two-word form per che, resulting in two token. This variation afects not only the token count, but also the structure of the tree, as it requires an additional syntactic dependency and dependency relation label. Moreover, frequent variation in tokenization may lead to measurable diferences in other types of metrics (such as Type/Token Ratio analysis) depending on the reference edition adopted.
Finally, the third case examines an example that difers from the first two previously discussed. In this instance, the editorial variation does not result in a diferent number of token, but rather in the presence of lemmas with diferent meanings:
Veder poteste quando vui scontrai (Rivalta ed.) Veder poteste quando v’inscontrai (Rea ed) (one seldom sees him as if in flesh)
As can be observed, the diference in verse 1 of sonnet 35 between the two editions concerns the use of scontrai versus inscontrai. When these token are lemmatized, they result in diferent lemmas: scontrai is the conjugated form of scontrare (to collide), whereas inscontrai corresponds 12For reasons of space and clarity, we report only the dependency relations discussed in the text. To view the full trees, please refer to the GitHub page: https://github.com/CIRCSE/ CavalcantiRepository.git. 13This total includes one “ritornellato” sonnet (i.e., a sonnet composed of more than the canonical 14 lines and featuring a specific metrical pattern; for details, refer to [15, p. 284]) and one “rinterzato” sonnet (i.e., a composition structured as a duplex sonnet; see [15, p. 283]). 14The 19 sonnets were split between the two annotators: 10 for one and 9 for the other.
The overall inter-annotator agreement is very high for both authors. For Cavalcanti, Fleiss’ kappa on dependency edges reaches 0.94, and the agreement on dependency labels assigned to correctly matched edges is even higher, at 0.98. For Dante, the values are similarly strong, with 0.95 for edges and 0.92 for labels. These results indicate a high level of consistency between annotators and confirm the overall reliability of the annotation process across both corpora.
internal evaluation on that test split and pertain to
complete annotation from scratch, rather than being limited
to syntactic analysis alone.19 In contrast, the scores in
the Cavalcanti columns, both for full parsing and
syntaxonly (OS), reflect the evaluation of the sonnets annotated
by Dante model, compared against the gold-standard
annotations produced by the two annotators (as described
in Section 2.4).
In this experiment, we evaluate the parsing performance
of two diferent models in order to identify the most
suitable for this task. Specifically, we assess the performance Interestingly, the performance of the Dante-trained
of (i) a Stanza retrained model [
To this end, we tested the ChatGPT API using a
tai3.1. Dante Model lored prompting technique. We report the prompt in
We train a Stanza model [
We evaluate only syntactic metrics,17 specifically the ent sonnet. We set the temperature to a minimum value
Unlabeled Attachment Score (UAS) and the Label At- (0.05) and set top_p20 to 1 in order to make the model as
tachment Score (LAS).18 Table 2 reports the model’s per- deterministic as possible.
formance. Specifically, the scores in the Dante column Since our aim is to compare the performance of the
reflect the performance of Dante model on a Divine Com- LLM with that of Dante model, we tested the LLM in two
edy test set. These scores derive from Stanza’s automatic settings, mirroring the evaluation setup used for Dante
model: (i) generating the full CoNLL-U file from scratch,
15While preliminary results suggest that such models yield an
improvement in syntactic accuracy for Italian (see [
development, and test sets with an 80-10-10 split. 17To perform evaluation we use eval.py script, available at https:
//github.com/UniversalDependencies/tools/blob/master/eval.py.
18Refer to [
In light of the obtained scores, we conducted a samplebased manual error analysis, which is described in Section 4.
To compare the two models, we evaluate their syntactic performance on eight randomly selected sonnets22, specifically, sonnets 9, 10, 28, 29, 30, 31, 35 and 54. 23 4. Manual Errors Analysis
The corresponding UAS and LAS scores are reported in Table 3, showing the performance of the LLM for the syn- In this Section , we analyse the models’ errors to provide tactic task (under the LLM column) and that of Dante’s useful insights for potential improvements and further model in the syntax-only setting (D_OS). Although, as considerations. We conduct a manual error analysis on mentioned in Subsection 3.2, we were unable to fully eval- the three sonnets exhibiting the most significant discrepuate the LLM across all annotation levels, we nonetheless ancies in scores, namely, sonnets 10 and 31, which perreport the scores of Dante model when performing full formed poorly with the LLM (see Subsection 4.1), and annotation (D_All) for reference. The Dif column indi- sonnet 29, whose score was comparable for both the LLM cates the diference in performance between D_OS and and Dante model (see Subsection 4.2). For the sake of clarthe LLM. ity, we provide both sonnets along with their translations
Notably, the LLM achieves relatively high performance in Appendix A. (see average row (av.) in Table 3), demonstrating its potential for syntactic annotation. However, its results remain 4.1. LLM Errors below those of the neural model (Dante model), which was specifically trained on data closely aligned with the target material. As shown in Table 3, these findings suggest that, at least in this experiment and at the current stage of development, at the current stage of development, the neural model trained on a domain-relevant corpora ofers a more reliable solution for high-quality linguistic annotation.
The two sonnets that received the lowest scores from the LLM are sonnet 10 and sonnet 31.
Interestingly, a detailed inspection of Sonnet 10, the one that received the lowest score, reveals that one of the model’s errors was the incorrect identification of the sentence’s root. In this case, the LLM mistakenly analyses a subordinate clause as the main clause, assigning the verb vider (saw) in example II the status of root instead of identifying it as the head of an adverbial clause (advcl). 21A cycle in a dependency tree occurs when a word ends up de- This misassignment results in a series of incorrect synapnennodtinatgioonninitvseallfid,.violating the tree structure and rendering the tactic attachments throughout the sentence. To illustrate 22The number of sonnets was limited to eight due to funding con- this, we present two syntactic trees: the gold-standard straints, as the ChatGPT API is a paid service. tree (Figure 3) and the one produced by the LLM (Figure 23Refer to Appendix A for the titles of the selected sonnets. 2). Incorrect attachments in the LLM output are highlighted in red, while incorrectly assigned dependency labels are shown in yellow.
Quando mi vider, tutti con pietanza/ dissermi (Their somber welcome to me I still shudder at) root nsubj obl parataxis
Tu m’ài sì piena di dolor la mente (you have filled my mind with so much sorrow) root root nsubj
xcomp iobj nsubj iobj aux tu m’ ài sì piena you me.DAT have so full tu m’ ài sì piena you me.DAT have so full mi vider , tutti con pietanza disserme.ACC saw , all with pity said 24Refer to: https://universaldependencies.org/u/dep/aux_. 25Dante model correctly annotates the tree, identifying piena (full) as the root and correctly attaching its dependents. The only mistake made by the model is the labeling of the auxiliary ài (to have), which is incorrectly assigned the label cop (copula) instead of the correct label aux (auxiliary).
The incorrect identification of the clause head, assigned to the copula è (is), leads to a structurally inconsistent analysis, accompanied by further errors. For instance, the noun vita (life) is incorrectly labeled as obl (oblique), instead of being identified as the head of a relative clause (acl:relcl), a role mistakenly assigned to the copula è (is). In addition to the incorrect labeling, the subject che (who) (marked as nsubj) and the adverb fuori (out) (advmod) are also wrongly attached.
In sum, this close examination of the two lowestscoring sonnets highlights how a single error, especially one involving a crucial dependency relation from which other subtrees depend, such as the incorrect identification of the root, can compromise both annotation quality and evaluation scores. Auxiliaries also pose a particular challenge for the LLM, as they are often incorrectly annotated as heads rather than leaves.
to both models (Dante-All and Dante-OS). One such error is the incorrect identification of the root, as shown in Figures 7 and 8, in contrast to the gold standard shown in Figure 6. This error is reported here due to its significant impact on the overall sentence structure. More specifically, as indicated in the gold annotation (Figure 6), the root of the sentence is giriano (wander), token 60. Nevertheless, both Dante models fail to recognize it, selecting instead diletti (delights), token 35, as the root.
nmod nsubj
root amod angosciosi diletti miei sospiri ... giriano anguished delights my sighs ... wander During the evaluation of the models’ performance, Sonnet 29 emerges as a particularly problematic case, in which both Dante models (All and OS) perform poorly, Figure 6: Gold dependency tree of Sonn. 29 v. 5 showing no substantial improvement over the LLM’s performance.
Upon comparing the gold annotation with the output root xcomp of the Dante-All model, it becomes immediately evident obl nsubj that one of the main reasons behind this unexpectedly low performance lies in an incorrect token split of the word angosciosi (anguished), at verse 5: expl:pv angoscio si diletti miei sospiri ... giriano anguish REFL delights my sighs ... wander Examples V - Sonn. 29 v. 5 angosciosi diletti miei sospiri ... giriano angoscio si diletti miei sospiri ... giriano (my sighs would not only subside but turn into hosannas of praise [...] gives) angosciosi 34 angoscio 34 diletti 35 si 35 miei 36 diletti 36 sospiri . . . giriano 37 . . . 60 miei sospiri . . . 37 38 . . .
giriano 61
The Dante-All model splits the word into two distinct token: angoscio and si, interpreting si (originally part of the adjective angosciosi) as a reflexive clitic. As a result, it is annotated as a separate token. As already observed for the LLM in 3.2, this error brings to light a range of phenomena that, while not entirely incorrect, are influenced by issues stemming from improper tokenization. In fact the addition of a token leads to a misalignment of token indices (i.e., the id field in the CoNLL-U format; see footnote 3), such that even when syntactic dependencies are correctly labeled and attached to the correct token, they are still considered incorrect due to the wrong numbering with respect to the gold standard (see Table 4).
Indeed, this error can only be observed in the DanteAll model, which is required to perform all annotation tasks independently, including tokenization. Despite this specific tokenization issue, there are also errors common root
xcomp nsubj obl angosciosi diletti miei sospiri ... giriano anguished delights my sighs ... wander
Following this error, subsequent dependency relations are misassigned. The adjective angosciosi is annotated as an oblique (obl) dependent of diletti. In the Dante-All model, this is further complicated by incorrect tokenization, which results in si being erroneously split of and assigned the expl:pv (expletive: pronominal) relation.
The noun sospiri, although correctly labeled as a subject (nsubj), is attached to the incorrect root (diletti) rather than to giriano (the gold root). Finally, giriano itself is annotated as an open clausal complement (xcomp) depending on another token.
We hypothesize that this error may be caused by the length of the sentence, in which the token identified as the root appears in the 60th position, thus making the parsing of the sentence more complex. However, further experiments in this direction are needed to confirm this hypothesis.
While performing the manual error analysis, we identiifed a set of specific errors shared by both models (Dante(All and -OS) and and LLM), highlighting a common dif- p.63], along with an ellipsis of the noun donna (woman), ifculty in analysing such constructions. Notably, these that is, di tal (donna). As a result, the preposition di (of) errors both arise from and reflect the complexity of an- should be attached as a case marker (case) to tal, and the notating these texts. We report two cases. entire phrase di tal should be analysed and annotated as
In Sonnet 10, both models fail to correctly annotate a nominal modifier ( nmod) of the noun servente (servant), a construction that poses dificulties even for human yielding the interpretation servente di una tale donna (serannotators. This issue is found specifically in verses vant of such a woman). This full phrase functions as an 10–11. open clausal complement (xcomp) of the main verb fatto (made). We report the correct syntactic tree in Figure 10.
Example VI - Sonn. 10 vv. 10-11: In this case, both models incorrectly annotate di and in una parte là ’v’i’ trovai gente / tal as modifiers of the noun servente, assigning them the (cthoeacpialascceunwhsiedreolneovabldem’Aemnograftohretere.d who also labels case and det (determiner), as shown in Figure 10. sufered in the thrall of Love.) Neither model captures the ellipsis of donna, and both attach di (of) and tal (such) directly to servente. However, this annotation leads to a diferent interpretation: di tale servente (of such a servant). Based on this incorrect interpretation, servente is assigned the syntactic role of oblique (obl), and is correctly attached to the verb fatto.
In v. 11, the che (who) is a specific instance of the "neutrum relative pronoun che" [27, p.193], used to replace a pronoun governed by a preposition. This construction was typical of familiar Tuscan and is also attested in other texts of the same period [27, pp.193-194]. As noted by the editors Rea and Inglese in their commentary on this verse [19, p.62], this instance of che represents a case of preposition elision, namely, standing in for the form dei quali, (of whom), which functions as a nominal modifier of the actual subject ciascun (each) (ciascun dei quali) (each of whom). The gold annotation is reported in Figure 9. xcomp
case nmod Fatto sè di tal servente Made himself of such a servant obl case
det Fatto sè di tal servente
Made himself of such a servant acl:relcl
nsubj
nsubj trovai gente che ciascun si doleva
I found people that each one REFL was grieving
5. Conclusion and Future Work
Both models misanalyse che, incorrectly identifying This study has evaluated the syntactic parsing perforit as the subject of a relative clause. At the same time, mance of two distinct models on a corpus of Old Italian they correctly assign the nsubj relation to ciascun (each); poetry: Stanza, a supervised neural model trained on however, this leads to an inconsistent syntactic analysis, a domain-specific corpus, and an unsupervised autoreresulting in a spurious double subject, as shown in Figure gressive large language model accessed via the ChatGPT 9. API (o3 version). The results show that while the LLM
Another error made by both models that is worth demonstrates acceptable accuracy scores, the Stanzacommenting on, as it reflects a stylistic peculiarity, is the based model trained on the Italian-Old treebank conone reported in Example VII (Sonnet 10, v. 13): sistently outperforms it in syntactic annotation tasks, particularly when provided with gold-standard tokenization and morphology. Although this experiment involved Example VII - Sonn. 10 v. 13: only a single LLM , the results suggest that using a neural Fatto sè di tal servente model trained on domain-specific data to pre-parse texts (gYeooun)are a denizen of the kingdom or the dun- as a basis for human annotation could be advantageous.
Moreover, while this may appear self-evident, the ex
This verse presents a fronting of the genitive comple- periment reinforces the importance of involving expert
ment di tal (of such), as noted by Rea and Inglese [19, annotators, especially in cases like those shown in this
study, where stylistic nuances escape the models. Future
directions include conducting experiments with other
LLMs, such as the aforementioned UDify [
Manual error analysis was crucial in highlighting the challenges encountered by both models, revealing not only more mechanical errors, such as incorrect root identification often accompanied by mistaken head and dependency assignments of the dependent nodes, but also more specific issues, particularly in handling poetic constructions, ellipses, and auxiliary verbs. While in the ifrst scenario some workarounds can be found to mitigate such errors (for example, by adopting rule-based integrations to guide model performance), when it comes to more specific errors, human skill and expertise become decisive.
Another crucial element that emerged from this study is the influence of editorial choices on syntactic annotation. As demonstrated in Section 2.3, diferent editions of the same text can result in substantial variation in tokenization, lemma interpretation, and syntactic structure. This observation underscores the need for openly accessible, high-quality digital editions, particularly for historical texts, which often lack standardized resources.
Future work will extend the evaluation to include other critical editions of Cavalcanti’s Rhymes to analyse stylistic distance and to assess model performance across distinct editorial variants. Moreover, further investigations will also aim to explore syntactic variation across diferent poetic genres and authors within the same historical period, as well as to examine prose texts in order to assess whether significant syntactic diferences emerge.
Sonnet 3126 1 Tu m’ài sì piena di dolor la mente che l’anima si briga di partire, e li sospir che manda il cor dolente mostrano a li occhi che non pon sofrire . 5 Amor, che lo tuo grande valor sente, dice: — mi duol che ti convien morire per questa fera donna, che neente par che pietade di te voglia udire. 9 Io vo come colui ch’è fuor di vita, che pare, a chi lo sguarda, c’omo sia fatto di rame o di pietra o di legno, 12 che sè conduca sol per maestria, e porti ne lo core una ferita che sia, com’egli è morto, aperto segno.
Sonnet 10 1 Li miei foll’occhi, che prima guardaro vostra figura piena di valore, fuor quei che di voi, donna, m’acusaro nel fero loco, ove ten corte amore. 5 E mantenente avanti lui mostraro ch’io era fatto vostro servidore; perchè sospiri e dolor mi pigliaro 8vedendo che temenza avea lo core. 9 Menarmi tosto senza riposanza in una parte, là ’v’i’ trovai gente che ciascun si doleva d’amor forte. 12 Quando mi vider, tutti con pietanza dissermi: — fatto se’ di tal servente che mai non dei sperare altro che morte. —
Sonnet 29 1 Se mercè fosse amica a’ miei disiri e ’l movimento suo fosse dal core di questa bella donna e ’l suo valore mostrasse la virtute a’ miei martiri, 5 d’angosciosi diletti miei sospiri, che nascon de la mente ov’è amore e vanno sol ragionando dolore e non trovan persona che li miri, 9 giriano a gli occhi con tanta vertute che ’l forte e duro lagrimar che fanno ritornerebbe in allegrezza e ’n gioia. 12 Ma sì è al cor dolente tanta noia e a l’anima trista è tanto danno che per disdegno uom non dà lor salute. 26See footnote 7 for details on the translation edition. 1 You have filled my mind with so much sorrow that the soul itself is assaulted and tries to flee.
Heartsick, my body sighs from my bones’ marrow and I’ve reached my limit, as anyone can see. 5 Even Love is sympathetic and says, “It is hard that the cruel lady for whom you pine gives you no pitiful glance or comforting phrase.
This was never a part of my design.” 9 I am unmanned and have to wander through the world like an intricate figure of wood or brass produced by some toymaker to amuse. 12 Strangers who pause to stare at me as I pass can’t tell that I sufer and haven’t a clue that I am dead, a victim of her abuse. 1 It was my reckless eyes that first beheld your inefable worth and condemned me to live in that wasteland over which the bold master, Love, holds court as tyrants do. 5 They welcomed me there, a new captive, a slave, and were delighted to hear my sighs and groans.
We are taught as little boys to try to be brave, but I felt an icy fear deep in my bones. 9 They led me to a place where noblemen gathered who also sufered in the thrall of Love. Their somber welcome to me I 12 still shudder at: “You are a denizen of the kingdom or the dungeon that holds us all and from which there is no escape until you die.” 1 If luck could look with favour on my desire and if it came from my lady’s heart with the power to encourage me and let me thrive in a shower of hope from heaven that knows how to admire 5 devotion of any kind, I do believe my sighs would not only subside but turn into hosannas of praise as grey brightens to blue when angry weather grants us a reprieve. 9 My squalls of tears would cease and joy at last would be what gives my eyes their special shine, each teardrop like a jewel delighting to be 12 dug from the earth to glitter and be free ...
But it hasn’t happened, and the pain that has been mine deforms me so that I am an oucast.
Correspondence between sonnet numbering and the first verse of each sonnet:
Certe mie rime a te mandar volendo Dante ai poeti Vedeste, al mio parere, ogni valore Se vedi amore assai ti prego, Dante Avete ’n vo’ li fiori e la verdura Chi è questa che ven ch’ ogn’om la mira Beltà di donna di piagente core Io vidi li occhi dove Amor si mise Li miei folli occhi che prima guardaro Dante a Guido S’io fossi quelli che d’amor fu degno Dante, un sospiro messagier del core Sonetto dell’Orlandi La bella donna dove amor si mostra Guido Orlandi a Guido Amore e monna Lagia e Guido ed io L’Orlandi a Guido Di vil matera mi conven parlare L’Orlandi a Guido The poems of uncertain date Un amoroso sguardo spirituale Voi che per li occhi mi passaste ’l core Perchè non furo a me gli occhi dispenti Se mercè fosse amica a’ miei disiri L’anima mia vilment’è sbigotita Tu m’ài sì piena di dolor la mente S’io prego questa donna che pietade Io temo che la mia disaventura Certo non è de lo ’ntelletto accolto Veder poteste quando vui scontrai De! spiriti miei, quando mi vedete Pe’ gli occhi fere un spirito sottile A me stesso di me pietate vene Gianni Alfani a Guido The poems composed after 1290 Io vengo il giorno a te infinite volte Una figura de la Donna mia Guido Orlandi a Guido Una giovane donna di Tolosa O tu che porti ne li occhi sovente Donna mia non vedestu colui Noi sian le triste penne isbigotite Novelle ti so dire, odi, Nerone Farinata degli Uberti a Guido Se non ti caggia la tua Santalena Guata, Manetto, quella scrignotuzza
Prompt LLM: The prompt was given in Italian; the original version is shown in italics, with the English translation provided in parentheses. role user: dato il testo in formato txt, produci un’annotazione completa secondo lo standard di Universal Dependencies. Produci un file in formato CoNLL-U, assicurandoti che abbia 10 colonne. (Given a plain text file (.txt), produce a complete annotation according to the Universal Dependencies standard. Generate a CoNLL-U formatted ifle, ensuring that it includes all 10 required columns.) role user: produci l’annotazione eseguendo i task di tokenizzazione, lemmatizzazione, part of speech tagging, morphological features e dependency parsing. (Perform the annotation by carrying out the tasks of tokenization, lemmatization, part-of-speech tagging, morphological feature annotation, and dependency parsing.) role user content: "file: "file_raw (raw sonnet)" role user assistant: "file: "file_gold (gold sonnet)" role user content: assicurati che non ci siano dei cicli nella sintassi e che ogni frase abbia soltanto una root. (Ensure that the syntactic structure contains no cycles and that each sentence has exactly one root.) role user content: assicurati anche che tutti i nodi dell’albero sintattico siano raggiungibili fra di loro. (Also ensure that all nodes in the syntactic tree are mutually reachable.’) role user content: procedi con l’altro file. Assicurati di annotare la colonna 7 e 8 e di avere 10 colonne per linea. (Proceed with the other file. Make sure to note columns 7 and 8 and to have 10 columns per line.)
• Italian-Old, • Cavalcanti Repository.
Declaration on Generative AI During the preparation of this work, the author(s) used ChatGPT (OpenAI) in order to: Paraphrase and reword, Improve writing style, and Grammar and spelling check. 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.