This paper presents a data-driven study focused on the automatic simplification of in-domain texts for specific target readers, which is “controlled” through data collected from behavioral analysis. We used these data to create Admin-It-L2, a parallel corpus of original-simplified sentences in the Italian administrative language, in which simplifications are aimed at Italian L2 speakers. Then, we used this corpus to test controllable models for text simplification based on Transformers. Although we obtained a high SARI score of 39.24, we show that this datum alone is not fully reliable in evaluating text simplification.
ATS aims at simplifying a text while maintaining its meaning [5]. Since the spread of Neural Networks, models’ architectures are mostly taken from “Neural Machine 1https://github.com/Unipisa/admin-it-l2 2https://openai.com/blog/chatgpt Translation” (NMT) [6]. For example, to build their sim- verbs, and a lower number of indicative verbs. Their anplification model, [ 7] started from OpenNMT [8], consist- swer error rate also increased with a higher number of ing of an encoder-decoder with two layers of LSTM. [9] multi-words and entities. followed a similar approach and trained a model for the Admin-It-L2 counts 134 pairs of sentences extracted Italian language thanks to data augmentation techniques. from the texts used in this study (34 sentences) and from
In some of the most recent models, simplification Admin-It (100 sentences), a subsection of Admin-It, can be “controlled” to generate texts for specific groups a parallel corpus of Italian administrative texts [3]. The of readers. CROSS [10] proposed a Transformer-based original sentences of this subset were selected from webmodel able to control the level of simplicity and the type sites of Italian municipalities, and from the longest senof applied simplification. To this aim, at a lexical level, tences of the PaWaC Corpus [21]. [3] manually rewrote they marked the token to replace, whereas, at a syntactic the sentences simplifying them both at lexical and syntaclevel, they leveraged templates. ACCESS [11] introduced tic levels. This simplification as well was only based on the use of control tokens to bind the simplification with the typical traits of the administrative language [2, 20]. specific attributes, such as the amount of paraphrased The similar nature of the simplification applied, alcontent and lexical and syntactic complexity. [12] used lowed us to further simplify the simple sentences of ACCESS to simplify Italian texts and trained it on the Admin-It and the simple sentences employed in the automatic translation of NewsEla [13] (see Sec. 5.1), comprehension test (together referred to as Administobtaining promising results. MUSS [14] presented an semp in Table 1) by considering the results of the study unsupervised method to collect parallel data to train a related to Italian L2 speakers. The annotation was conmodel based on ACCESS adopting BART and its multi- ducted by a single annotator among the authors, and lingual version, mBART [15]. The authors’ goal was to validated through a quantitative analysis shown in Tadeal with the paucity of available parallel corpora in lan- ble 1.3 guages diferent from English. To this aim, we leverage cross-language and fully translated data, by exploiting the ability of multilingual models to use linguistic knowl- 4. The Controllable Simplification edge from diferent languages. [ 16] also used ACCESS Model but replaced BART with T5 [17]. With this controllable model, [18] achieved SOTA performance for SARI on Spanish texts.
Diferently from the described work, we focus our simplification on in-domain texts written in the Italian bureaucratic language. Furthermore, we leverage data on Italian L2 speakers to create a test set to control the simplifications generated by Transformer-models. Such a dataset is described in the next section.
Our models are based on the implementation released by [16].4 The two authors used ACCESS and replaced BART with T5, a model with an encoder-decoder architecture pre-trained on various tasks through supervised and unsupervised approaches [17].
We adopted the multilingual version of T5, mT5 [22], for all our experiments. [23] observed that multilingual models proved to be better for tasks like text summarization or switching from a formal to a more informal language. The authors suggest that this is probably due to the diferent distributions of linguistic data used for training the model. In particular, the multilingual model would perform better with automatically translated texts.
Unlike [16], due to hardware constraints, we employed the basic mT5 model and fine-tuned it on a batch of size 16. We retained the values of all other parameters, including the number of epochs (5). Fine-tuning was performed on an NVIDIA A100 Tensor Core GPU, with 40 GB of RAM.
In this section, we present Admin-It-L2, a parallel corpus of complex-simplified sentences in the Italian administrative language for ATS aimed at Italian L2 speakers.
We manually simplified 134 sentences by focusing on the linguistic traits that emerged from a comprehension test conducted over 86 participants [19], involving also Italian L2 speakers (30,2%), and elderly Italian native speakers (33,72%). The participants were asked to answer questions about an original text and a simplified version of another text, both from the administrative domain. 5. Experimental Settings Such simplification focused only on the typical traits of the bureaucratic language [20]. Given the limited availability of parallel corpora in Italian,
By analyzing the participants’ answers to the com- we created training sets for fine-tuning by adopting two prehension test, the authors observed that L2 struggled more when reading simple texts with long sentences, long prepositional chains, a high number of participle 3Details about the simplification process are presented in Appendix A.
4https://github.com/KimChengSHEANG/TS_T5 main strategies. The first involves the full translation into Spanish NewsEla (NewsEs) sentence-aligned version Italian of parallel English corpora with Google Translate.5 was created by [9]. Documents were translated from We translated only from English since translation systems English, and then simplified manually. for this language generally perform better, thanks to Terence contains short stories in Italian for children the amount of data available for model training. The rewritten by a group of experts [20]. second strategy involves the creation of cross-language Teacher contains teaching materials manually simplidatasets, in which the original sentences are in English, fied by a teacher for Italian L2 learners [20]. whereas their simplified counterparts were translated PaCCSS-IT was presented by [30], who collected seninto Italian. On the one hand, translating simple texts tences in Italian from the web through a semi-supervised is an easier task for machine translation systems [24, method. 25]. On the other hand, we intended to take advantage Simpitiki is the portion of Simpitiki composed of of the capabilities of multilingual generative models to Italian Wikipedia articles selected from the edits labeled succeed in applying linguistic knowledge from diferent as simplifications. The simplified sentences are obtained languages. This potential has previously been tested for by applying one simplification operation at a time [31]. summarizing documents in diferent languages [26]. Admin-It and Admin-It are subsections of Admin-It. The former corresponds to another section of
administrative texts manually simplified by applying a We used Admin-It-L2 as the test set, to assess whether single operation, whereas in the latter, the simplification the models succeed in producing simplifications close to is applied at a document level, then the sentences were the needs of Italian L2 speakers. To train the controllable manually aligned [3]. models we combined the following datasets:6 As for the automatic translations, firstly we used
NewsEla (NewsEng) consists of newspaper articles Google Translate only on the simple sentences of Enmanually simplified by experts according to diferent glish corpora. The resulting corpora are then referred degrees of complexity. NewsEla was then automatically to as NewsEn-Ita, OSE-EnIta, and SimPA-EnIta. Then, we sentence-aligned by [27]. automatically translated datasets including both simple
OneStopEnglish (OSE) contains articles from the and complex sentences. In this case, the resulting corpora
in three levels of readability for English L2 learners [28]. We extracted two random samples from NewsEn,8
SimPA is a parallel corpus of sentences in English instead of fine-tuning models on the whole dataset, administrative language [29]. The authors applied two since [18] achieved SOTA performance for SARI on Spansteps of simplification: lexical and syntactic. ish on a relatively small amount of data.9 5https://translate.google.com/ 6From these datasets, pairs containing the same sentence in their original and simplified versions were filtered out. Additional processing was applied to Terence and Teacher, which were provided in XML format. Datasets statistics are reported in Appendix B.
7The simplified sentences in these corpora show variations in the two translation versions, as the statistics in Appendix B show.
8One for each kind of translation.
9They fine-tuned their model on a sample of Spanish NewsEla that counted about 7k sentences. 5.2. Evaluation metrics Three metrics were used to evaluate the models: BLEU [32], SARI [33] and BERTScore [34].
The first, BLEU, measures the n-gram overlap between gold and generated sentences. It is inherited from Machine Translation, and [35] observed that BLEU is less reliable when sentence splitting is applied in the simplified sentence. SARI is specific to evaluating simpliifcation models and measures the efectiveness of copy (KEEP), insertion (ADD), and deletion (DEL) operations applied [36]. This metric, therefore, is widely used to evaluate simplification at the lexical level. By design, SARI takes as input several gold references. Since Admin-ItL2 contains only one reference per sentence, this metric might not be fully reliable. The third evaluates texts generally created by generative models. BERTScore sums up the cosine similarity of token pairs with the highest similarity. Specifically, BERTScore measures the similarity of tokens in the predicted sentence with respect to the tokens in the gold sentence. [35] observed that BERTScore correlates with the simplicity values given by human annotators when such values are low.
We applied these metrics by using the implementation of the EASSE evaluation tool [37] provided by [16]. As for BERTScore, since there is no variant of BERT available for the Italian language for this metric, we employed a multilingual model, xlm-roberta-large,10 which performs better than mBERT on the Italian administrative language [38]. 5.3. Control tokens For these experiments, we used the same control tokens as [16]: • NbChars. Synthesis: the ratio of the length in characters between source and simplified sentences; • LevSim. Paraphrasing: the normalized Levenstein
distance between source and simplified sentences; • WordRank. Lexical complexity: the ratio between word frequencies in the two sentences, original and simplified; • DepTreeDept. Syntactic complexity: the ratio between the maximum depth of the syntactic tree of source and target sentences; • NbWords. Lexical complexity: the number of words in the simplified sentence divided by that of the complex sentence.
Although WordRank leverages English language data, we included this control token since [12] drew some advantage from it on Italian text simplification. To extract 10https://huggingface.co/xlm-roberta-large these linguistic features, we added a specific function to the implementation of [16] to detect the language of the examined sentence.11 Given the results obtained through behavioral analysis [19], we employed all available control tokens for these experiments. A more suitable simplification for Italian L2 speakers can be obtained by controlling word frequency through WordRank and NbWords. This afects morphological aspects because verbs in the indicative present tense are generally more frequent than those in other verb tenses and modes.
LevSim and NbChars operate more on sentence length whereas the depth of the syntactic tree can be reduced with DepTreeDept, also afecting propositional chains average length. Regarding the number of complex entities and terms, this trait was traced back to aspects related to sentence length, as emerged from the behavioral analysis.
Diferently from [ 11] and [16], who conducted their experiments by extracting training, validation, and test sets from the same corpus, the sentences used for our experiments come from diferent domains and languages.
The goal here is to modify the values of the control tokens to optimize the performance of the model on the test set.
For this purpose, we extracted the values of the controlled variables from each pair of sentences in Admin-It-L2, and then we computed the average.12 We fed the models with such values to condition the simplification generated during evaluation. 5.4. Baselines We selected two baseline models. Admin-It is obtained by fine-tuning mT5 using the subsection of Admin-It.
This subset corresponds to a subsection of Simpitiki, the largest Italian dataset of parallel sentences in the administrative language. We also evaluated this model by setting the values of the control tokens with the average values on Admin-It-L2. The second baseline is a hypothetical model that generates a copy of the complex sentences of the test set. This baseline is reported in Table 2 as AdminIt-L2 . We assume that models obtaining results close to this baseline are very conservative in their simplification process.
The results of our experiments are shown in Table 2. The baseline obtained by fine-tuning mT5 only on AdminIt reached results close to zero for both BLEU and BERTScore indicating that this dataset might be too small for the model to learn any simplification rule. Also,
11We used the language identification module provided by fastText [39, 40]. In this case, as to compute the corpora statistics we used: it_corenews_sm and en_coreweb_sm.
12Table 11 in Appendix C contains the values given to the controllable tokens. the low BERTScore is indicative of incorrect simplifi- ment with All-Ita+SimPA-Ita. This model scores similarly cation [35]. Admin-It-L2 , obtained a lower SARI score to All-Ita+SimPA-EnIta for BLEU (-0.34) and BERTScore and higher BLEU and BERTScore than the first baseline. (-0.01), but the SARI score is significantly worse (-7.35),
The first model we tested, All-Ita, was obtained by fine- generating rather conservative simplifications, with a tuning mT5 on all available Italian corpora. However, SARI even higher than Admin-It-L2 (+0.17). this model obtained a lower SARI score than Admin- Albeit considering the variations in the translations, It (-6.66), and even though it equals Admin-It-L2 for the contribution of the cross-language data for speBERTScore, it reached a lower BLEU score (-0.43). cific corpora would seem to make a diference com
We built a second model by adding SimPA-EnIta to All- pared to the texts fully translated into Italian (e.g., OSEIta. This model scores slightly better than Admin-It-L2 EnIta+SimPA-En-Ita vs. OSE-Ita and SimPA-Ita). Even on both BLEU (+0.61) and BERTScore (+0.02). We also though they remain low, the best results were obtained included OSE-EnIta and NewsEn-Ita in the fine-tuning by All-Ita+SimPA-EnIta for BERTScore (0.52) and BLEU of a third model, All-Ita+All-EnIta. However, adding (25.34), while the highest score for SARI is reached by all the cross-language data worsened the performance. OSE-EnIta+SimPA-En-Ita (39.24), not too far from the Therefore, we ran further experiments focusing on the SOTA for English.13 characteristics of each corpus used.
OSE-EnIta and SimPA-En-Ita, and although this model achieves the worst results for BLEU (12.02), its Then, we manually analyzed a random sample of 50 sim
EnIta (-0.01), and above all, it surpasses this model for tends to copy the content of the original sentences. The 11.73 points for SARI, obtaining the best score for this model merely eliminates some portion of text, but in sevmetric: 39.24%. This model also achieved the highest eral cases, this happens with relevant information. In value for SARI (3.99) - although it remains low - OSE-EnIta+SimPA-EnIta simplifications, the deletion opand SARI (77.82), and the lowest for SARI erations are even more massive, and much of the source (36.72). Such scores may indicate how much less con- text is usually removed, producing inconsistent sentences. servative this model is than the others in generating In other cases, OSE-EnIta+SimPA-EnIta produces “artifisimplifications. Our next experiment involved the full cial hallucinations”. However, OSE-EnIta+SimPA-EnIta translation of these two corpora, OSE-Ita and SimPA- succeeds in producing better simplifications when the Ita. Although the model achieves the same BERTScore complex sentences are shorter, even when All-Ita+SimPA(0.50) as OSE-EnIta+SimPA-En-Ita, the performance is EnIta still reproduces them entirely (see Table 3).14 significantly worse for SARI (-11.67). Finally, for BLEU, The results confirmed what was pointed out by [ 35]: although it outperforms All-Ita (+0.13), the best model for BERTScore is a reliable metric in detecting mostly lowthis score remains All-Ita+SimPa-EnIta. Next, we decided quality simplifications, and metrics like BLEU and SARI to fine-tune mT5 using all the corpora fully translated cannot be analyzed in isolation. Besides the low quality from English. The resulting model is reported in Table 2 of the generated simplification, the low values obtained as All-TransIta and improves over the previous one only by BERTScore may also find an explanation both in the for SARI, but without reaching the score obtained by OSE-EnIta+SimPA-En-Ita. We conducted a final experi13Martin et al. [14] obtained a SARI score of about 42 on NewsEn. 14Other examples of models output can be seen in Appendix D.
Table 3 loses sight of the identity of the complainants, namely citThe Table shows an example of the generated simplifications. izens, and instead refers to a generic “person”. This way, All-Ita+SimPA-EnIta exactly reproduces the text of the origi- ChatGPT shows to be able to simplify administrative text nal sentence in Admin-It-L2, while OSE-EnIta+SimPA-EnIta in Italian at both the syntactic and the lexical level and eliminates the more complex portion of text contained within to tune the simplification according to the prospected an aside (in italic in the original sentence). scenario, i.e., when addressing Italian L2 speakers. NeverOriginal Admin-It-L2: È, inoltre, possibile richiedere, per non theless, the model is not entirely accurate when inferring più di 3 giorni consecutivi, un pasto “in bianco”, in presenza di that “the uncertainty as to the antecedence of the power disturbi gastroenterici, senza la presentazione di certificazione of attorney” led to the “confusion about who is represento prescrizione medica. ing the citizen in the process”. Indeed, this latter case could juridically give rise to defects of a diferent nature aScimopoliaflile’idntAedstminion,-Iitl-gLe2n:iSteorielbpaumòbirnicohhieaddeirsetuurbni palalostsoto“min- than the one depicted by the Province of Pistoia.17 Chatbianco” senza presentare una certificazione o prescrizione GPT seems not to grasp here the intrinsic and necessary medica per massimo 3 giorni di fila. complexity of the content, which is closely linked to the complexity of the subject matter described [41].
3 giorni consecutivi, un pasto “in bianco”, in presenza di dis- like ChatGPT are not open-source and are available only turbi gastroenterici, senza la presentazione di certificazione o via web APIs. This leads to a lack of transparency in prescrizione medica. the data used for training, limiting the reproducibility of experiments and quality control of the results [42].
OSE-EnIta+SimPA-EnIta: È possibile richiedere un pasto “in bianco per non più di 3 giorni consecutivi, senza la presentazione di certificato o prescrizione medica.
In this paper, we presented the first benchmark for the BERT model employed for the extraction of the vector automatic simplification of Italian administrative texts representations, which is not specific to Italian (Sec. 5.2), aimed at Italian L2 speakers. We employed data collected and in the composition of the test set, Admin-It-L2. Its from behavioral analysis to build Admin-It-L2, a parallel sentences are more complex and longer than the other corpus of original-simplified sentences. Then, we used corpora and underwent a deeper simplification process. this corpus to test controllable models for text simplification based on Transformers. Although we obtained high
cannot be used to evaluate text simplification and that a We conducted a preliminary and exploratory experiment fully reliable automatic metric for the Italian language is to get a brief overview of the capabilities of ChatGPT still missing. Finally, we asked ChatGPT to simplify one when asked to simplify Italian administrative texts for of the sentences in Admin-It-2, obtaining a higher quality Italian L2 speakers in a zero-shot setting. simplification than mT5, although the model presents
In the conversation shown in Table 4,15 when asked to some limitations, especially in reproducibility, and in simplify a sentence contained in Admin-It-L2, ChatGPT some cases, in answer accuracy. provides a short and syntactically accurate text, even For future work, we aim at validating Admin-It-L2 though the long period in the original sentence has not involving human annotation, and, more specifically, dobeen broken down into shorter sentences. The model main experts and Italian L2 speakers, to properly evaluate shows that it can provide the meaning of “procura alle the employed models and the quality of their generations. liti” (power of attorney) and can insert a short definition As for the models, we aim to conduct an ablation study into the simplified sentence. However, when asked to to assess the contribution of each control token. We simplify the sentence further by phrasing it as if it were also would like to adopt a control token to encode the addressing a person who is not fluent in Italian, the model language used in the sentences and employ the corpora adopts a very informal register: the sentence sounds al- in their original versions. Moreover, we aim to create most ungrammatical and less coherent in some parts.16 a specific automatic metric to evaluate Italian text simWhen prompted to use a more polished language, Chat- plification. Finally, we would like to test open-source GPT raised the register, although apparently, the model LLMs on administrative text simplification focused on specific target readers, with diferent prompts, in zero and few-shot settings. 15We tested ChatGPT between December 2022 and January 2023.
16The absence of the pronoun "lo" referring to “documento” makes less explicit what the expression “e non doveva essere” is referring to.
Semplifica questa frase: La Provincia di Pistoia eccepisce anche l’inammissibilità del ricorso per nullità della procura alle liti rilasciata dai cittadini ricorrenti in quanto apposta, a diferenza di quella dell’associazione attrice, su appositi fogli aggiunti e spillati all’atto introduttivo del giudizio, in diformità della previsione contenuta nell’art. 83 c.p.c. e ciò in quanto, in tal caso non vi sarebbe certezza in ordine alla anteriorità del conferimento del mandato rispetto alla notifica dell’atto di resistenza o di impugnazione (cfr. Cass. civ., sez. iii, 14 marzo 2006, n. 5443). Simplify this sentence: The Province of Pistoia also objects to the inadmissibility of the appeal due to the nullity of the power of attorney issued by the citizen plaintifs in that it was afixed, unlike that of the plaintif association, on special sheets added and stapled to the writ of summons, contrary to the provision contained in Art. 83 c.p.c. and this because, in such a case, there would be no certainty as to the anteriority of the conferment of the mandate concerning the notification of the act of resistance or appeal (see Cass. civ, sec. iii, March 14, 2006, no. 5443). La Provincia di Pistoia contesta l’ammissibilità del ricorso per nullità della procura alle liti in quanto è stata apposta su fogli aggiunti e non nell’atto introduttivo del giudizio, violando l’articolo 83 c.p.c. e causando incertezza sulla anteriorità del mandato rispetto alla notifica dell’atto di resistenza o di impugnazione.
The Province of Pistoia contests the admissibility of the appeal on the grounds that the power of attorney was null and void because it was afixed on supplementary sheets and not in the writ of summons, violating Article 83 of the Code of Civil Procedure and causing uncertainty as to the antecedence of the grant of power of attorney with respect to the service of the notice of resistance or appeal. cos’è la procura alle liti? what is the power of attorney? La procura alle liti è un documento formale con cui una persona (il mandante) conferisce a un’altra (il mandatario) il potere di rappresentarla e agire in suo nome in una causa legale. La procura alle liti deve essere redatta in forma scritta, sottoscritta dal mandante e autenticata da un notaio o da un avvocato. In essa vengono specificati i poteri che il mandatario ha in relazione alla causa, come ad esempio il potere di depositare documenti, di comparire in giudizio e di transigere. A power of attorney is a formal document by which one person (the principal) grants another (the agent) the power to represent and act on his behalf in a legal case. The power of attorney must be in writing, signed by the principal, and notarized by a notary public or lawyer. It specifies the powers that the agent has in relation to the lawsuit, such as the power to file documents, to appear in court, and to settle.
Puoi inserire questa definizione nella semplificazione che hai fornito poco fa in modo che il testo sia chiaro anche alle persone che non sanno il significato di procura alle liti. La frase non deve risultare quindi troppo lunga You can include this definition in the simplification you provided earlier so that the text is clear even to people who do not know the meaning of power of attorney in litigation. The sentence should therefore not be too long La Provincia di Pistoia contesta l’ammissibilità del ricorso per nullità della procura alle liti, un documento in cui una persona conferisce a un’altra il potere di rappresentarla in una causa legale, in quanto è stata apposta su fogli aggiunti e non nell’atto introduttivo del giudizio, violando l’articolo 83 c.p.c. e causando incertezza sulla anteriorità del mandato rispetto alla notifica dell’atto di resistenza o di impugnazione.
The Province of Pistoia challenges the admissibility of the appeal on the grounds that the power of attorney, a document in which one person confers on another the power to represent him or her in a legal case, is null and void because it was afixed on additional sheets of paper and not in the writ of summons, violating Article 83 of the Code of Civil Procedure and causing uncertainty as to the antecedence of the power of attorney from the service of the notice of resistance or appeal.
Adesso semplifica ancora di più la frase come se stessi parlando a qualcuno che conosce poco l’italiano Now simplify the sentence even more as if you were speaking to someone who is not fluent in Italian La Provincia di Pistoia dice che il documento che permette a qualcuno di rappresentare un’altra persona in un processo legale è sbagliato, perché è stato scritto su fogli aggiunti e non doveva essere e questo causa confusione su chi rappresenta la persona nel processo.
The Province of Pistoia says that the document that allows someone to represent another person in a legal process is wrong because it was written on added sheets and should not have been, and this causes confusion about who is representing the person in the process.
Potresti migliorare un po’ lo stile? il tuo interlocutore conosce poco l’italiano ma il testo deve essere un po’ più curato Could you improve the style a little? your interlocutor is not fluent in Italian, but the text needs to be a little more polished La Provincia di Pistoia sostiene che il documento chiamato “procura alle liti" che consente a una persona di rappresentare un’altra persona in un processo legale è stato scritto in modo errato. Infatti, è stato scritto su fogli aggiunti anziché nell’atto uficiale del processo, e questo può causare confusione su chi rappresenta la persona nel processo.
The Province of Pistoia claims that the document called "power of attorney" that allows a person to represent another person in a legal process was written incorrectly. In fact, it was written on added sheets instead of in the oficial document of the trial, and this can cause confusion about who represents the person in the trial.
derstanding italian administrative texts: A reader- Resources and Evaluation (LREC 2018), European oriented study for readability assessment and text Language Resources Association (ELRA), Miyazaki, simplification, in: Proceedings of the 1st Workshop Japan, 2018, pp. 4333–4338. URL: https://aclantholo on AI for Public Administration (AIxPA), 2022, pp. gy.org/L18-1685.
71–87. [30] D. Brunato, A. Cimino, F. Dell’Orletta, G. Venturi, [20] D. Brunato, F. Dell’Orletta, G. Venturi, S. Monte- Paccss-it: A parallel corpus of complex-simple senmagni, Design and annotation of the first italian tences for automatic text simplification, in: Proceedcorpus for text simplification, in: Proceedings of ings of the 2016 Conference on Empirical Methods The 9th Linguistic Annotation Workshop, 2015, pp. in Natural Language Processing, 2016, pp. 351–361. 31–41. [31] S. Tonelli, A. P. Aprosio, F. Saltori, Simpitiki: a [21] L. C. Passaro, A. Lenci, Extracting terms with ex- simplification corpus for italian, in: Proceedings of tra, Computerised and corpus-based approaches CLiC-it, 2016. to phraseology: Monolingual and multilingual per- [32] K. Papineni, S. Roukos, T. Ward, W.-J. Zhu, Bleu: a spectives (2016) 188–196. method for automatic evaluation of machine trans[22] L. Xue, N. Constant, A. Roberts, M. Kale, R. Al- lation, in: Proceedings of the 40th annual meeting Rfou, A. Siddhant, A. Barua, C. Rafel, mT5: of the Association for Computational Linguistics, A massively multilingual pre-trained text-to-text 2002, pp. 311–318. transformer, in: Proceedings of the 2021 Confer- [33] W. Xu, C. Napoles, E. Pavlick, Q. Chen, C. Callisonence of the North American Chapter of the As- Burch, Optimizing statistical machine translation sociation for Computational Linguistics: Human for text simplification, Transactions of the AssociLanguage Technologies, Association for Computa- ation for Computational Linguistics 4 (2016) 401– tional Linguistics, Online, 2021, pp. 483–498. URL: 415. https://aclanthology.org/2021.naacl- main.41. [34] T. Zhang, V. Kishore, F. Wu, K. Q. Weinberger, doi:10.18653/v1/2021.naacl-main.41. Y. Artzi, Bertscore: Evaluating text generation with [23] G. Sarti, M. Nissim, It5: Large-scale text-to-text bert, in: International Conference on Learning Reppretraining for italian language understanding and resentations, 2019.
generation, arXiv preprint arXiv:2203.03759 (2022). [35] F. Alva-Manchego, C. Scarton, L. Specia, The (un) [24] S. Štajner, M. Popović, Can text simplification help suitability of automatic evaluation metrics for text machine translation?, in: Proceedings of the 19th simplification, Computational Linguistics 47 (2021) Annual Conference of the European Association 861–889.
for Machine Translation, 2016, pp. 230–242. [36] E. Sulem, O. Abend, A. Rappoport, Bleu is not [25] K. Mishra, A. Soni, R. Sharma, D. M. Sharma, Ex- suitable for the evaluation of text simplification, ploring the efects of sentence simplification on arXiv preprint arXiv:1810.05995 (2018). hindi to english machine translation system, in: [37] F. Alva-Manchego, L. Martin, C. Scarton, L. Speproceedings of the workshop on automatic text cia, Easse: Easier automatic sentence simplification simplification-methods and applications in the mul- evaluation (2019) 49–54.
tilingual society (ATS-MA 2014), 2014, pp. 21–29. [38] S. Auriemma, M. Miliani, A. Bondielli, L. C. Passaro, [26] L. Perez-Beltrachini, M. Lapata, Models and A. Lenci, Evaluating pre-trained transformers on datasets for cross-lingual summarisation, in: Pro- italian administrative texts, in: Proceedings of 1st ceedings of the 2021 Conference on Empirical Meth- Workshop on AI for Public Administration, 2022, ods in Natural Language Processing, 2021, pp. 9408– pp. 54–70.
9423. [39] A. Joulin, É. Grave, P. Bojanowski, T. Mikolov, Bag [27] C. Jiang, M. Maddela, W. Lan, Y. Zhong, W. Xu, Neu- of tricks for eficient text classification, in: Proral CRF model for sentence alignment in text sim- ceedings of the 15th Conference of the European plification, CoRR abs/2005.02324 (2020). URL: https: Chapter of the Association for Computational Lin//arxiv.org/abs/2005.02324. arXiv:2005.02324. guistics: Volume 2, Short Papers, 2017, pp. 427–431. [28] S. Vajjala, I. Lučić, Onestopenglish corpus: A new [40] A. Joulin, E. Grave, P. Bojanowski, M. Douze, H. Jécorpus for automatic readability assessment and gou, T. Mikolov, Fasttext.zip: Compressing text text simplification, in: Proceedings of the thirteenth classification models, ArXiv abs/1612.03651 (2016). workshop on innovative use of NLP for building URL: https://api.semanticscholar.org/CorpusID: educational applications, 2018, pp. 297–304. 16196524. [29] C. Scarton, G. Paetzold, L. Specia, SimPA: A [41] A. Fioritto, Manuale di stile. strumenti per semplifisentence-level simplification corpus for the pub- care il linguaggio delle amministrazioni pubbliche lic administration domain, in: Proceedings of the (1997).
Eleventh International Conference on Language [42] A. Liesenfeld, A. Lopez, M. Dingemanse, Opening up chatgpt: Tracking openness, transparency, and accountability in instruction-tuned text generators, in: Proceedings of the 5th International Conference on Conversational User Interfaces, 2023, pp. 1–6.
Table 11 shows the values associated with each controllable token in the prediction phase of the fine-tuned models. We computed the controllable variables from each sentence pair in Admin-It-L2 and then we calculated their average.
In this section, we present further details about the re- D. Examples of simplification alization of Admin-It-L2. As described in Section 3, the behavioral analysis shows that Italian L2 speakers strug- In this section, we reported some examples of the gle in reading simplified texts with long sentences, long sentences generated by two of the created models: prepositional chains, a high number of participle verbs, All-Ita+SimPA-EnIta and OSE-EnIta+SimPA-EnIta. Alland a lower number of indicative verbs. Their answer Ita+SimPA-EnIta tends to copy most of the content from error rate also increased with a higher number of multi- the original sentence (Table 13) or delete portions of the words and entities, but also these lexical features were as- texts regardless of their relevance (Table 12). In the other sociated in the behavioral analysis with sentence length. two reported tables, there are interesting examples of
Regarding morphological aspects, we applied transfor- the sentences returned by OSE-EnIta+SimPA-EnIta. In mations from the subjunctive (6) or the infinitive (5) to Table 13, the model tends to delete most of the content the indicative, and verbal periphrases were replaced with and generates inconsistent text. Such inconsistencies can single verb forms of similar meaning, as in the example also lead to artificial hallucinations, as in the example in 8. Concerning sentence length reduction, references to Table 14. laws whose citation in the text was not crucial to sentence comprehension were removed (6, 9). The idea is that the automatic simplification operated by the neural models is always flanked by the original text, which complements the absence of some information that might lower text readability [2]. The depth of the syntactic tree was also reduced (-0.80) by limiting redundant expressions, from an average sentence length in tokens of 29.86 to 24.83 (see Table 1), without afecting the cohesion of the text (6). A ifnding also emerges from Levensthein’s distance from the original sentences to the simplified sentences (170.44) and the simplified sentences in Admin-It-L2 (202.73). We simplified the sentences syntactically by intervening on prepositional chains, which in some cases were eliminated (7) or reduced (9). Finally, 27 sentences were not further simplified.
Original: La gestione della pesa sarà interamente informatizzata, collegando l’apparato di pesatura ad un sistema informativo; tutte le operazioni di pesatura verranno così memorizzate all’interno di un server, per poter essere storicizzate, consultate ed elaborate con la massima comodità e sicurezza e la gestione principale dei dati verrà svolta attraverso l’impiego di un software commerciale leader nel settore, in grado di soprintendere opportunamente all’intero iter di trattamento del rifiuto, dal momento del conferimento fino allo stoccaggio o smaltimento finale.
Simplified: La gestione della pesatura sarà informatizzata, collegando l’apparato di pesatura ad un sistema informativo. I dati sulle operazioni di pesatura verranno così conservati all’interno di un server, per poter essere consultati ed elaborati con la massima comodità e sicurezza. La gestione dei dati verrà svolta attraverso l’impiego di un software commerciale leader nel settore, in grado di monitorare l’intero processo di trattamento dei rifiuti: dal conferimento fino allo stoccaggio o smaltimento. Admin-It-L2: La gestione della pesatura sarà informatizzata, collegando l’apparato di pesatura ad un sistema informatico. I dati sulle operazioni di pesatura verranno così conservati all’interno di un server, e potranno essere consultati ed elaborati con la massima comodità e sicurezza. La gestione dei dati verrà svolta attraverso l’impiego di un software che monitora l’intero processo di trattamento dei rifiuti.
Original: La domanda di abitabilità o agibilità di quanto edificato, deve essere proposta all’Uficio Tecnico Comunale, allegando a questa il certificato di collaudo delle opere edili, ove necessario, la dichiarazione presentata per l’iscrizione al catasto dell’immobile, restituita dagli ufici catastali con l’attestazione dell’avvenuta presentazione e la dichiarazione del direttore lavori che certifichi, sotto la propria responsabilità, la conformità di quanto realizzato rispetto al progetto approvato, l’avvenuta prosciugatura dei muri e la salubrità degli ambienti, ai sensi del d.p.r. 22 aprile 1994, n. 425. Simplified: La domanda di abitabilità o agibilità dell’immobile deve essere presentata all’Uficio Tecnico Comunale. Alla domanda va allegato il certificato di collaudo delle opere edili. Se necessario, va allegata anche la dichiarazione dell’iscrizione dell’immobile al catasto (certificata dall’Uficio del Catasto) e la dichiarazione del direttore dei lavori che attesti sotto la propria responsabilità: che quanto realizzato sia conforme al progetto approvato; che i muri siano asciugati; che l’ambiente sia salutare (d.p.r. 22 aprile 1994, n. 425).
Admin-It-L2: La domanda di abitabilità o agibilità dell’immobile va presentata all’Uficio Tecnico Comunale. Alla domanda va allegato il certificato di collaudo della costruzione. Se necessario, va allegata anche la dichiarazione dell’iscrizione dell’immobile al catasto e la dichiarazione del direttore dei lavori. Il direttore dichiara sotto la propria responsabilità che l’immobile è conforme al progetto approvato, che i muri sono asciugati e che l’ambiente è sano. Original: Per l’espletamento dei Servizi oggetto del presente contratto, nient’altro è dovuto dalla Provincia alla Società oltre a quanto previsto nel presente Contratto, salvo il reintegro delle somme relative ad agevolazioni tarifarie e a riduzioni imposte unilateralmente rispetto a quanto previsto nell’allegato 6 "sistema tarifario", salvo eventuali richieste giudicate ammissibili da parte del Comitato Tecnico di cui al successivo art. 49 e salvo provvedimenti di Autorità comunitarie, nazionali, regionali e locali destinati direttamente o indirettamente ai servizi oggetto del presente contratto, senza alcuna decurtazione. Simplified: Per l’esecuzione dei Servizi oggetto di questo contratto, la Provincia deve alla Società solo quanto previsto dal contratto stesso, e l’eventuale rimborso di: somme relative a riduzioni di tarifa e a sconti non concordati nell’allegato 6 "sistema tarifario"; eventuali richieste giudicate ammissibili da parte del Comitato Tecnico (elencate al successivo art. 49); provvedimenti di Autorità comunitarie, nazionali, regionali e locali destinati direttamente o indirettamente ai servizi oggetto di questo contratto.
Admin-It-L2: La Provincia deve alla Società che esegue i servizi solo quanto previsto dal contratto, e l’eventuale rimborso di somme legate a: riduzioni non presenti nel "sistema tarifario" (allegato 6); richieste approvate dal Comitato Tecnico (elencate all’art. 49); provvedimenti di enti comunitari, nazionali, regionali e locali destinati ai servizi oggetto del contratto.
Original: I coniugi che intendono procedere alla separazione personale consensuale, allo scioglimento o cessazione degli efetti civili del matrimonio e modifica delle condizioni di separazione o di divorzio dinanzi all’Uficiale di Stato Civile devono compilare una richiesta (ENTRAMBI I CONIUGI, vedi modello allegato) ed inviarla all’Uficio di Stato Civile via mail all’indirizzo uficio.anagrafe@comune.it, alla PEC comune@postacert.it oppure via fax al numero 000/0000000 insieme alle copie dei documenti di identità.
Simplified: I coniugi che intendono procedere presso l’Uficiale di Stato Civile a: separazione personale consensuale, scioglimento o cessazione degli efetti civili del matrimonio, modifica delle condizioni di separazione o di divorzio, devono compilare il modulo allegato e inviarlo all’Uficio di Stato Civile. Il modulo e la copia dei documenti di identità possono essere inviati per mail all’indirizzo uficio.anagrafe@comune.it, alla PEC comune@postacert.it oppure via fax al numero 000/0000000. Admin-It-L2: I coniugi che chiedono la separazione consensuale, lo scioglimento o la cessazione degli efetti civili del matrimonio, o la modifica delle condizioni di separazione o di divorzio, devono compilare il modulo allegato e inviarlo all’Uficio di Stato Civile. Il modulo e la copia dei documenti di identità possono essere inviati per mail all’indirizzo uficio.anagrafe@comune.it, alla PEC comune@postacert.it oppure via fax al numero 000/0000000. Original: In attuazione dell’articolo 6 delle norme tecniche di attuazione del piano del parco di cui alla deliberazione del Consiglio Regionale 12.12.1989 n. 515, per le zone costituenti riferimento ambientale dell’assetto del parco, il piano di gestione è finalizzato al mantenimento dell’esistente, al ripristino di morfologie preesistenti dei luoghi ed alla costruzione di elementi di connessione tra le varie connotazioni naturalistiche del territorio; il piano di gestione definisce anche le funzioni ricreative e produttive compatibili con il mantenimento, il restauro ed il ripristino dell’assetto paesaggistico e lo disciplinano nel tempo. Simplified: Secondo l’articolo 6 della delibera del Consiglio Regionale 12.12.1989 n. 515 relativo alle norme tecniche di attuazione del piano del parco, il piano di gestione delle zone ambientali ha l’obiettivo di: tutelare il paesaggio esistente, di ristabilire l’aspetto originario di questi luoghi e di costruire connessioni tra gli elementi naturalistici del territorio. Il piano di gestione definisce e disciplina nel tempo anche le funzioni ricreative e produttive compatibili con gli obiettivi del piano stesso. Admin-It-L2: Il piano del parco tutela il paesaggio esistente, ristabilisce l’aspetto originario di questi luoghi e costruisce connessioni tra gli elementi naturalistici del territorio. Il piano definisce e disciplina nel tempo anche le funzioni ricreative e produttive compatibili con gli obiettivi del piano stesso. Original Admin-It-L2: Per quanto riguarda i permessi temporanei è possibile richiederli attraverso il servizio online per il rilascio dei permessi oppure al comando di Polizia Municipale - solo permessi temporanei giornalieri con denominazione "T". Simplified Admin-It-L2: L’interessato può richiedere i permessi temporanei giornalieri di tipo "T" attraverso il servizio online apposito oppure alla sede della Polizia Municipale.
All-Ita+SimPA-EnIta: Per quanto riguarda i permessi temporanei è possibile richiederli attraverso il servizio online per il rilascio dei permessi oppure al comando di Polizia Municipale.
OSE-EnIta+SimPA-EnIta: Per quanto riguarda i permessi temporanei, è possibile richiedere i permessi temporanei tramite il servizio online per il rilascio dei permesso temporanei giornalieri con denominazione "".""."
Original Admin-It-L2: Il Comune può imporre limitazioni all’apertura di nuovi esercizi limitatamente ai casi in cui ragioni, non altrimenti risolvibili, di sostenibilità ambientale, sociale e di viabilità, rendano impossibile consentire ulteriori flussi di pubblico nella zona senza incidere in modo gravemente negativo sui meccanismi di controllo, in particolare per il consumo di alcolici, e senza ledere il diritto dei residenti alla vivibilità del territorio e alla normale mobilità.
Simplified Admin-It-L2: Il Comune può ostacolare l’apertura di nuovi negozi solo se ci sono problemi di sostenibilità ambientale, sociale e di viabilità, che non consentono ulteriori flussi di pubblico nella zona. Questi limiti servono a non ostacolare i controlli, in particolare sul consumo di alcolici, e a rispettare il diritto dei residenti a vivere tranquillamente e a muoversi normalmente.
All-Ita+SimPA-EnIta: Il Comune può imporre limitazioni all’apertura di nuovi esercizi limitatamente ai casi in cui ragioni, non altrimenti risolvibili, di sostenibilità ambientale, sociale e di viabilità, rendano impossibile consentire ulteriori flussi di pubblico nella zona senza incidere in modo gravemente negativo sui meccanismi di controllo, in particolare per il consumo di alcolici e senza ledere il diritto dei residenti alla vivibilità del territorio e alla normale mobilità.
OSE-EnIta+SimPA-EnIta: Il Comune può imporre limiti di apertura di nuovi esercizi entro i casi in cui le ragioni della sostenibilità ambientale e della viabilità non siano risolvibili, in particolare per il consumo di alcolici e alla mobilità normale. Original Admin-It-L2: Requisiti di residenza e soggiorno Semplified Admin-It-L2: Caratteristiche di residenza e soggiorno All-Ita+SimPA-EnIta: Il Requisiti di residenza e soggiorno sono questi.
OSE-EnIta+SimPA-EnIta: Il requisito di locazione è l’afitto della città.