Add 1 files

2203/2203.03759.md

@@ -0,0 +1,249 @@
+Title: Text-to-text Pretraining for Italian Language Understanding and Generation
+
+URL Source: https://arxiv.org/html/2203.03759
+
+Markdown Content:
+###### Abstract
+
+We introduce IT5, the first family of encoder-decoder transformer models pretrained specifically on Italian. We document and perform a thorough cleaning procedure for a large Italian corpus and use it to pretrain four IT5 model sizes. We then introduce the ItaGen benchmark, which includes a broad range of natural language understanding and generation tasks for Italian, and use it to evaluate the performance of IT5 models and multilingual baselines. We find monolingual IT5 models to provide the best scale-to-performance ratio across tested models, consistently outperforming their multilingual counterparts and setting a new state-of-the-art for Italian language generation.
+
+Keywords: Italian NLP, Natural Language Generation, Language Modeling
+
+\NAT@set@cites
+
+IT5: Text-to-text Pretraining for 
+
+Italian Language Understanding and Generation
+
+Gabriele Sarti, Malvina Nissim
+Center for Language and Cognition (CLCG), University of Groningen
+{g.sarti, m.nissim}@rug.nl
+
+Abstract content
+
+1.Introduction
+--------------
+
+The text-to-text paradigm introduced by T5(Raffel et al., [2020](https://arxiv.org/html/2203.03759v2#bib.bib34)) has been widely adopted as a simple yet powerful generic transfer learning approach for most language processing tasks(Sanh et al., [2022](https://arxiv.org/html/2203.03759v2#bib.bib39); Aribandi et al., [2022](https://arxiv.org/html/2203.03759v2#bib.bib1)). Although the original T5 model was trained exclusively on English data, the same architecture has been extended to a massively multilingual setting covering more than 100 languages by mT5 and ByT5(Xue et al., [2022](https://arxiv.org/html/2203.03759v2#bib.bib47), [2021](https://arxiv.org/html/2203.03759v2#bib.bib48)), following recent advances in the multilingual pre-training of large language models such as mBERT, XLM, XLM-R and mDeBERTa(Devlin et al., [2019](https://arxiv.org/html/2203.03759v2#bib.bib14); Conneau and Lample, [2019](https://arxiv.org/html/2203.03759v2#bib.bib8); Conneau et al., [2020](https://arxiv.org/html/2203.03759v2#bib.bib7); He et al., [2023](https://arxiv.org/html/2203.03759v2#bib.bib15)). Multilingual language models were shown to excel in cross-lingual and low-resource scenarios. Still, multiple studies have highlighted their suboptimal scale-to-performance ratio when compared to monolingual counterparts for language-specific applications in which data are abundant(Nozza et al., [2020](https://arxiv.org/html/2203.03759v2#bib.bib29); Rust et al., [2021](https://arxiv.org/html/2203.03759v2#bib.bib38)).
+
+For this reason, monolingual T5 models have recently been pretrained to serve specific language communities, covering languages such as Arabic, Portuguese, Vietnamese and Slovenian(Nagoudi et al., [2022](https://arxiv.org/html/2203.03759v2#bib.bib28); Carmo et al., [2020](https://arxiv.org/html/2203.03759v2#bib.bib5); Phan et al., [2022](https://arxiv.org/html/2203.03759v2#bib.bib31); Ulčar and Robnik-Šikonja, [2023](https://arxiv.org/html/2203.03759v2#bib.bib44)). These models improve over multilingual baselines on language understanding and generation tasks such as news summarization, headline generation and natural language inference in their respective languages.
+
+In this work, we follow a similar approach to pre-train and evaluate four Italian T5 models of different sizes, which we identify as IT5. In Section[3](https://arxiv.org/html/2203.03759v2#S3 "3. Data and Model Pretraining ‣ IT5: Text-to-text Pretraining for Italian Language Understanding and Generation"), we present the cleaning procedure for the Italian portion of the mC4 corpus(Xue et al., [2021](https://arxiv.org/html/2203.03759v2#bib.bib48)) used in pre-training the IT5 models. Section[4](https://arxiv.org/html/2203.03759v2#S4 "4. Evaluation ‣ IT5: Text-to-text Pretraining for Italian Language Understanding and Generation") describes multilingual baselines and downstream tasks used to evaluate fine-tuned IT5 models and presents the obtained results. Finally, our findings and future directions are summarized in Section[5](https://arxiv.org/html/2203.03759v2#S5 "5. Conclusion ‣ IT5: Text-to-text Pretraining for Italian Language Understanding and Generation"). We make the following contributions:
+
+*   •We introduce a large-scale cleaned version of the Italian mC4 corpus and use it to pre-train four IT5 models of various dimensions. 
+*   •We introduce ItaGen, a benchmark for Italian language understanding and generation tasks. 
+*   •We evaluate IT5 on ItaGen, showing improvements over multilingual baselines and previous state-of-the-art work. 
+*   •We publicly release all the code, data, and pre-trained/fine-tuned checkpoints for further experimentation by the research community. 
+
+To the best of our knowledge, our IT5 models are the first publicly available encoder-decoder models pre-trained exclusively on the Italian language. IT5 constitutes a significant contribution to Italian NLP, as evidenced by its prompt adoption by the research community upon its release(La Quatra and Cagliero, [2023](https://arxiv.org/html/2203.03759v2#bib.bib20); Papucci et al., [2022](https://arxiv.org/html/2203.03759v2#bib.bib30); Mousavi et al., [2023](https://arxiv.org/html/2203.03759v2#bib.bib27)inter alia), especially in the context of the latest evaluation campaign of Italian NLP tools(Leonardelli and Casula, [2023](https://arxiv.org/html/2203.03759v2#bib.bib22); Hromei et al., [2023](https://arxiv.org/html/2203.03759v2#bib.bib16)). This paper serves as the prime reference for IT5, providing all relevant details regarding training data and parameters, a battery of experiments on a collection of tasks, which can be further used as a reference benchmark, especially for Italian generation, and a discussion of its limitations.1 1 1 Resources: [https://github.com/gsarti/it5](https://github.com/gsarti/it5).
+
+2.Related Work
+--------------
+
+### 2.1.Text-to-text Transfer Transformers
+
+The Text-to-text Transfer Transformer (T5) model(Raffel et al., [2020](https://arxiv.org/html/2203.03759v2#bib.bib34)) adapts the original Transformer architecture proposed by Vaswani et al. ([2017](https://arxiv.org/html/2203.03759v2#bib.bib45)) by reformulating multiple natural language processing tasks into a unified text-to-text format and using them alongside masked span prediction for semi-supervised pre-training. The encoder-decoder architecture of T5 is especially suited for sequence-to-sequence tasks(Sutskever et al., [2014](https://arxiv.org/html/2203.03759v2#bib.bib42)), which cannot be performed by encoder-only models like BERT(Devlin et al., [2019](https://arxiv.org/html/2203.03759v2#bib.bib14)) and can prove to be challenging for decoder-only models like GPTs(Radford et al., [2019](https://arxiv.org/html/2203.03759v2#bib.bib33); Brown et al., [2020](https://arxiv.org/html/2203.03759v2#bib.bib4)) due to the lack of explicit conditioning on source context. The same architecture can be easily applied to natural language understanding tasks by using a text-to-text format, making the T5 model highly versatile in most NLP settings.
+
+### 2.2.Pre-trained Language Models for Italian
+
+The high technical expertise and heavy computational resources required for developing state-of-the-art models recently exacerbated inequalities in access to state-of-the-art systems for non-English languages. Despite the good amount of linguistic resources currently available, the Italian NLP community can currently count on a small set of publicly available pre-trained language models based mostly on the BERT architecture – AlBERTo(Polignano et al., [2019](https://arxiv.org/html/2203.03759v2#bib.bib32)), UmBERTo 2 2 2[https://github.com/musixmatchresearch/umberto](https://github.com/musixmatchresearch/umberto) and GilBERTo 3 3 3[https://github.com/idb-ita/GilBERTo](https://github.com/idb-ita/GilBERTo)inter alia, see Miaschi et al. ([2022](https://arxiv.org/html/2203.03759v2#bib.bib26)) for a survey – and most notably on a single decoder-only model for text generation, GePpeTto(De Mattei et al., [2020a](https://arxiv.org/html/2203.03759v2#bib.bib11)). Our IT5 models fill the current gap in the availability of sequence-to-sequence models, providing natural choices for monolingual tasks such as summarization, question answering and reformulation.
+
+3.Data and Model Pretraining
+----------------------------
+
+The original T5 models were pre-trained on the 750GB web-scraped English C4 corpus(Raffel et al., [2020](https://arxiv.org/html/2203.03759v2#bib.bib34)). C4 authors cleaned the corpus with heuristics to remove templated fillers, text deduplication, Javascript code, slurs and non-English texts. The multilingual counterpart of T5 adopts a similar procedure to create mC4(Xue et al., [2021](https://arxiv.org/html/2203.03759v2#bib.bib48)), a multilingual version of C4 including 107 languages. While mC4 authors adopted a similar procedure, the language detection threshold is lowered to 70% and other useful heuristics are omitted due to their brittleness across various character systems. As a result, the resulting corpus has an overall lower quality, with recent work finding 16% of sampled mC4 examples having the wrong language tag, and 11% not containing any linguistic information(Kreutzer et al., [2022](https://arxiv.org/html/2203.03759v2#bib.bib18)). For this reason, we perform a thorough cleaning of the Italian portion of mC4 before pre-training IT5.
+
+### 3.1.Cleaning the Italian mC4 Corpus
+
+The original Italian mC4 Corpus includes approximately 359GB of raw text data and is one of the largest public Italian corpora. To perform a more thorough cleaning of this data, we use a public implementation 4 4 4[https://gitlab.com/yhavinga/c4nlpreproc](https://gitlab.com/yhavinga/c4nlpreproc) reproducing and improving the original C4 data cleaning pipeline. Specifically, we sentence-tokenize documents and remove sentences containing either (i) words from a manually selected subset of the Italian and English List of Dirty Naughty Obscene and Otherwise Bad Words;5 5 5[https://github.com/LDNOOBW](https://github.com/LDNOOBW) (ii) less than three words, or a word longer than 1000 characters; (iii) an end symbol not matching standard end-of-sentence punctuation for Italian; or (iv) strings associated to Javascript code, lorem ipsum, English and Italian privacy policy/cookie disclaimers. We finally keep only documents containing more than five sentences, having between 500 and 50k characters, and having Italian as the main language.6 6 6 We use the [langdetect](https://github.com/Mimino666/langdetect) toolkit. The resulting Clean Italian mC4 Corpus 7 7 7[https://hf.co/datasets/gsarti/clean_mc4_it](https://hf.co/datasets/gsarti/clean_mc4_it), contains roughly 215GB of raw Italian text, corresponding roughly to 103M documents and 41B words.
+
+### 3.2.Model and Training Parameters
+
+The first 10M documents sampled from the cleaned corpus are used to train a SentencePiece unigram subword tokenizer Kudo ([2018](https://arxiv.org/html/2203.03759v2#bib.bib19)) with a vocabulary size of 32k words. The full cleaned corpus is then used to pre-train three models following the canonical small, base and large sizes(Raffel et al., [2020](https://arxiv.org/html/2203.03759v2#bib.bib34)). Moreover, a fourth model adopting the efficient small EL32 architecture by Tay et al. ([2022](https://arxiv.org/html/2203.03759v2#bib.bib43)) is also pre-trained and evaluated.
+
+Table 1: Summary of datasets composing ItaGen.
+
+4.Evaluation
+------------
+
+### 4.1.The ItaGen Benchmark
+
+We propose a selection of seven representative tasks, collectively referred to as ItaGen, to evaluate the downstream performances of fine-tuned IT5 and mT5 models. ItaGen aims to provide a comprehensive overview of canonical conditional text generation applications such as summarization, style transfer and question generation, and is constrained by the limited availability of Italian corpora for such tasks. Moreover, we also include a direct comparison of IT5 performances against encoder-based extractive systems for extractive question answering. Table[1](https://arxiv.org/html/2203.03759v2#S3.T1 "Table 1 ‣ 3.2. Model and Training Parameters ‣ 3. Data and Model Pretraining ‣ IT5: Text-to-text Pretraining for Italian Language Understanding and Generation") provides an overview of ItaGen tasks and datasets.
+
+#### Wikipedia Summarization
+
+We evaluate encyclopedic summarization on the Wikipedia for Italian Text Summarization (WITS) corpus(Casola and Lavelli, [2022](https://arxiv.org/html/2203.03759v2#bib.bib6)), containing 700k articles extracted from a cleaned dump of the Italian Wikipedia alongside their leading sections used as approximated summaries. We adopt the original evaluation setup using a 10k examples test set.
+
+#### News Summarization
+
+We evaluate news article summarization by concatenating Fanpage.it and IlPost newspapers articles collected by Landro et al. ([2022](https://arxiv.org/html/2203.03759v2#bib.bib21)). We refer to this concatenated corpus as NewsSum-IT. We fine-tune our systems on the training set, including roughly 100k articles and respective short summaries and evaluate them separately on the two test sets defined by the dataset creators. We report the averaged metrics across the two newspapers in the results section.
+
+#### Question Answering
+
+We evaluate extractive question answering using the SQuAD-IT dataset(Croce et al., [2018](https://arxiv.org/html/2203.03759v2#bib.bib10)), containing 50k paragraph-question-answers triplets automatically translated from the English SQuAD dataset(Rajpurkar et al., [2016](https://arxiv.org/html/2203.03759v2#bib.bib35)). We frame the QA task as a text-to-text problem aimed at generating responses given a source text using the format <CONTEXT> Domanda: <QUESTION>. We use the original evaluation script and splits.
+
+#### Question Generation
+
+We use SQuAD-IT to evaluate question generation capabilities by reordering the text triplets, making the model predict a plausible question given a source text in the format <CONTEXT> Risposta: <ANSWER>, where the answer is the first among the available per-example answers, using the same train-test splits of QA.
+
+#### Headline Style Transfer
+
+We evaluate style transfer abilities in the news domain on the CHANGE-IT shared task(De Mattei et al., [2020b](https://arxiv.org/html/2203.03759v2#bib.bib12)), containing 60k newspaper articles and headlines from the left-leaning Italian newspaper la Repubblica and the right-leaning Il Giornale, respectively. We train and validate our models on author-defined splits using the original cross-source article-to-headline generation. We report average scores for the two style transfer directions (Il Giornale ↔↔\leftrightarrow↔ la Repubblica).
+
+#### Headline Generation
+
+To evaluate news headline generation, we combine the two CHANGE-IT subsets to create a corpus of roughly 120k news articles and headlines pairs, which we refer to with the name HeadGen-IT. Original CHANGE-IT test sets are preserved.
+
+#### Formality Style Transfer
+
+We evaluate the formality style transfer capabilities of our models on the Italian subset of the XFORMAL dataset(Briakou et al., [2021](https://arxiv.org/html/2203.03759v2#bib.bib3)), containing a training set of 115k forum messages automatically translated from the GYAFC corpus(Rao and Tetreault, [2018](https://arxiv.org/html/2203.03759v2#bib.bib36)) and covering the topics of entertainment, music, family and relationships, and a small test set of 1000 formal-informal pairs obtained directly in Italian from four crowd workers via Amazon Mechanical Turk. We evaluate our models in both style transfer directions (Formal ↔↔\leftrightarrow↔ Informal).
+
+### 4.2.Evaluation Metrics
+
+We use a combination of common lexical and trainable metrics across all available tasks. We use the language-independent ROUGE metric(Lin, [2004](https://arxiv.org/html/2203.03759v2#bib.bib23)) in its R1, R2 and RL variants to evaluate lexical matches, and BERTScore( [2020](https://arxiv.org/html/2203.03759v2#bib.bib49); BS) to evaluate correspondence at the semantic level.8 8 8 We use an Italian BERT model to obtain baseline scores to broaden the metric range and remove noise, following authors’ recommendations. For QA, the canonical exact-match (EM) and F1-score (F1) metrics are used. Finally, for the news headline style transfer task, we use trained classifiers provided by the authors 9 9 9[michelecafagna26/CHANGE-IT](https://github.com/michelecafagna26/CHANGE-IT) to ensure comparable headline-headline (HH) and headline-article (HA) coherence performances.
+
+### 4.3.Baselines
+
+Besides baselines available from previous studies using the selected datasets, we also adopt the same fine-tuning procedure for fine-tuning two sizes (small and base) of the multilingual T5 model (mT5)Xue et al. ([2021](https://arxiv.org/html/2203.03759v2#bib.bib48)). These multilingual models are used to assess the validity of our pre-training procedure and to observe whether the monolingual setting improves performance.10 10 10 mT5 models are bigger than T5s due to larger vocabularies and embedding matrices, making their usage on consumer accelerators more challenging.
+
+WITS R1 R2 RL BS TextRank([2022](https://arxiv.org/html/2203.03759v2#bib.bib6)).302.076.197-LexRank([2022](https://arxiv.org/html/2203.03759v2#bib.bib6)).269.059.175-SumBasic([2022](https://arxiv.org/html/2203.03759v2#bib.bib6)).206.048.140-mT5 Small.347.200.316.517 mT5 Base.348.200.315.520 IT5 Small (ours).337.191.306.504 IT5 EL32 (ours).346.196.314.513 IT5 Base (ours).369.217.333.530 IT5 Large (ours).335.191.301.508 CHANGE-IT HH HA RL BS PointerNet([2020b](https://arxiv.org/html/2203.03759v2#bib.bib12)).644.874.151-BiLSTM+Att+Att{}_{\texttt{+Att}}start_FLOATSUBSCRIPT +Att end_FLOATSUBSCRIPT([2020c](https://arxiv.org/html/2203.03759v2#bib.bib13)).744.846.155-mT5 Small.777.807.211.372 mT5 Base.795.799.236.398 IT5 Small (ours).898.882.231.392 IT5 EL32 (ours).822.786.244.406 IT5 Base (ours).904.868.247.411 IT5 Large (ours).895.861.237.390 Size SQuAD-IT QA#F1 EM DrQA-IT([2018](https://arxiv.org/html/2203.03759v2#bib.bib10))-.659.561 mBERT([2019](https://arxiv.org/html/2203.03759v2#bib.bib9))110M.760.650 BERT-IT 11 11 11[bert-base-italian-uncased-squad-it](https://huggingface.co/antoniocappiello/bert-base-italian-uncased-squad-it)([2019](https://arxiv.org/html/2203.03759v2#bib.bib14))110M.753.638 XLM-R Large+st+st{}_{\texttt{+st}}start_FLOATSUBSCRIPT +st end_FLOATSUBSCRIPT([2021](https://arxiv.org/html/2203.03759v2#bib.bib37))560M.804.676 mT5 Small 300M.660.560 mT5 Base 580M.757.663 IT5 Small (ours)60M.716.619 IT5 EL32 (ours)143M.747.645 IT5 Base (ours)220M.761.663 IT5 Large (ours)738M.780.691 NewsSum-IT SQuAD-IT QG HeadGen-IT XFORMAL-IT F→→\rightarrow→I XFORMAL-IT I→→\rightarrow→F R1 R2 RL BS R1 R2 RL BS R1 R2 RL BS R1 R2 RL BS R1 R2 RL BS mBART Large([2022](https://arxiv.org/html/2203.03759v2#bib.bib21)).323.150.248-----------------mT5 Small.340.161.262.375.306.143.286.463.277.094.244.408.651.450.631.666.638.446.620.684 mT5 Base.330.155.258.393.346.174.324.495.302.109.265.427.653.449.632.667.661.471.642.712 IT5 Small (ours).354.172.278.386.367.189.344.505.287.100.253.414.650.450.631.663.646.451.628.702 IT5 EL32 (ours).339.160.263.410.382.201.357.517.299.108.264.427.459.244.435.739.430.221.408.630 IT5 Base (ours).251.101.195.315.382.199.354.516.310.112.270.433.652.446.632.665.583.403.561.641 IT5 Large (ours).377.194.291-.383.204.360.522.308.113.270.430.611.409.586.613.663.477.645.714
+
+Table 2: IT5, mT5 and baseline models performances on ItaGen datasets. Best scores are highlighted.
+
+### 4.4.Results and Discussion
+
+Table[2](https://arxiv.org/html/2203.03759v2#S4.T2 "Table 2 ‣ 4.3. Baselines ‣ 4. Evaluation ‣ IT5: Text-to-text Pretraining for Italian Language Understanding and Generation") present the results of our fine-tuning experiments. Given the broad scope of our analysis, we limit ourselves to comment on salient trends we observe across tasks and model categories.
+
+#### IT5 models provide state-of-the-art performances for language generation and understanding tasks in Italian.
+
+The IT5 models outperform multilingual models and previous systems in 6 out of 8 evaluated tasks, with noticeable improvements over mT5 systems, particularly for question answering and generation and for headline-headline coherence on the news headline style transfer task. For QA, the IT5 Large model outperforms most extractive systems, including the XLM-R Large by Riabi et al. ([2021](https://arxiv.org/html/2203.03759v2#bib.bib37)), despite its ad-hoc synthetic data augmentation procedure.
+
+#### Multilingual models can still be helpful in specific applications and when using translated data.
+
+We observe that multilingual language models perform best in the news summarization and the formal-to-informal style transfer tasks. In the case of news summarization, we attribute the performance gap in large part to the scale of the mBART baseline model. For the formality style transfer task, after a preliminary error analysis, we conjecture that translation errors and English acronyms present in the noisy training split of XFORMAL act as out-of-distribution samples in the monolingual setting, disrupting the performances of IT5 systems but are captured more easily by multilingual systems which were exposed by multiple data distributions by design. This would indicate a better fit of multilingual pre-trained models for such settings if verified. We leave a more thorough analysis of these patterns to future work.
+
+#### Scaling model size does not guarantee an increase in performance if not supported by an increase in computational resources.
+
+Contrary to common scaling trends for Transformers Brown et al. ([2020](https://arxiv.org/html/2203.03759v2#bib.bib4)), we do not observe a systematic increase in downstream performances for IT5 models when increasing their size, despite lower loss scores and higher accuracies achieved by larger models during pre-training. While recent work highlighted how better pre-training performances do not always correspond to better downstream scores for T5 models(Tay et al., [2022](https://arxiv.org/html/2203.03759v2#bib.bib43)), we hypothesize that our results might be related to a bottleneck in the maximal batch size for large models, set to 128 examples instead of the 2048 reported by Raffel et al. ([2020](https://arxiv.org/html/2203.03759v2#bib.bib34)) due to lack of resources. We observe that the EL32 architecture can frequently outperform larger model variants, suggesting efficient model design as a promising direction for monolingual model development.
+
+5.Conclusion
+------------
+
+This paper introduced IT5, the first family of large-scale encoder-decoder models pre-trained in Italian. We presented a detailed overview of the overall training and evaluation procedure, including comparisons with multilingual counterparts on a broad set of Italian language generation tasks. We obtained new state-of-the-art results across most evaluated tasks and concluded by discussing the shortcomings of large-scale monolingual language modeling when dealing with automatically translated data and limited computational resources.
+
+In light of our results, we deem a further investigation of time and quality trade-offs between pre-trained monolingual models and a language-specific continued pre-training of multilingual models as a future step to further narrow the gap in modeling performances for less-resourced languages.
+
+6.Acknowledgements
+------------------
+
+We thank the Google TPU Research Cloud program for providing us with free access to TPU v3-8 machines used in pre-training the IT5 models and the Center for Information Technology of the University of Groningen for providing access to the Peregrine high-performance computing cluster used in fine-tuning and evaluation experiments. We also thank the Huggingface team for creating the original template for TPU-compatible pre-training and fine-tuning scripts for T5 models and making them available during the Huggingface JAX/Flax Community Week.
+
+7.Ethics Statement
+------------------
+
+Despite our thorough cleaning procedure aimed at removing vulgarity and profanity, it must be acknowledged that models trained on web-scraped contents such as IT5 will inevitably reflect and amplify biases present in Internet blog articles and comments, resulting in potentially harmful content such as racial or gender stereotypes and conspiracist views. In light of this, we encourage further studies to assess the magnitude and prevalence of such biases. Model usage should ideally be restricted to research-oriented and non-user-facing endeavors.
+
+Due to our limited computational resources, we could not conduct an exhaustive hyperparameter search for pre-training and fine-tuning IT5 models. For this reason, reported scores should not be treated as the best achievable results, and further improvements can undoubtedly be achieved with additional benchmarking effort.
+
+Finally, despite using standard metrics capturing lexical and semantic similarity to evaluate our models, we do not explicitly evaluate the factual consistency of generated outputs. For this reason, the real-world effectiveness of our models should be further assessed in future studies, especially for tasks prone to hallucination, such as abstractive summarization.
+
+8.Bibliographical References
+----------------------------
+
+\c@NAT@ctr
+
+*   Aribandi et al. (2022) Vamsi Aribandi, Yi Tay, Tal Schuster, Jinfeng Rao, Huaixiu Steven Zheng, Sanket Vaibhav Mehta, Honglei Zhuang, Vinh Q. Tran, Dara Bahri, Jianmo Ni, Jai Gupta, Kai Hui, Sebastian Ruder, and Donald Metzler. 2022. ExT5: Towards extreme multi-task scaling for transfer learning. In _Proceedings of the Tenth International Conference on Learning Representations (ICLR’22)_, Online. OpenReview.net. 
+*   Bradbury et al. (2018) James Bradbury, Roy Frostig, Peter Hawkins, Matthew James Johnson, Chris Leary, Dougal Maclaurin, George Necula, Adam Paszke, Jake VanderPlas, Skye Wanderman-Milne, and Qiao Zhang. 2018. JAX: composable transformations of Python+NumPy programs. 
+*   Briakou et al. (2021) Eleftheria Briakou, Di Lu, Ke Zhang, and Joel Tetreault. 2021. Olá, bonjour, salve! XFORMAL: A benchmark for multilingual formality style transfer. In _Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies_, pages 3199–3216, Online. Association for Computational Linguistics. 
+*   Brown et al. (2020) Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners. In _Advances in Neural Information Processing Systems_, volume 33, pages 1877–1901. Curran Associates, Inc. 
+*   Carmo et al. (2020) Diedre Carmo, Marcos Piau, Israel Campiotti, Rodrigo Nogueira, and Roberto de Alencar Lotufo. 2020. PTT5: Pretraining and validating the T5 model on brazilian portuguese data. _ArXiv Computation and Language_, arXiv:2008.09144(v2). 
+*   Casola and Lavelli (2022) Silvia Casola and Alberto Lavelli. 2022. WITS: Wikipedia for italian text summarization. In _Proceedings of the Eight Italian Conference on Computational Linguistics (CLiC-it’21)_, Milan, Italy. CEUR.org. 
+*   Conneau et al. (2020) Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. 2020. Unsupervised cross-lingual representation learning at scale. In _Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics_, pages 8440–8451, Online. Association for Computational Linguistics. 
+*   Conneau and Lample (2019) Alexis Conneau and Guillaume Lample. 2019. Cross-lingual language model pretraining. In _Advances in Neural Information Processing Systems_, volume 32. Curran Associates, Inc. 
+*   Croce et al. (2019) Danilo Croce, Giorgio Brandi, and Roberto Basili. 2019. Deep bidirectional transformers for italian question answering. In _Proceedings of the Sixth Italian Conference on Computational Linguistics (CLiC-it’19)_, Bari, Italy. CEUR.org. 
+*   Croce et al. (2018) Danilo Croce, Alexandra Zelenanska, and Roberto Basili. 2018. Neural learning for question answering in italian. In _AI*IA 2018 – Advances in Artificial Intelligence_, pages 389–402, Cham. Springer International Publishing. 
+*   De Mattei et al. (2020a) Lorenzo De Mattei, Michele Cafagna, Felice Dell’Orletta, Malvina Nissim, and Marco Guerini. 2020a. GePpeTto carves italian into a language model. In _Proceedings of the Seventh Italian Conference on Computational Linguistics (CLiC-it’20)_, Online. CEUR.org. 
+*   De Mattei et al. (2020b) Lorenzo De Mattei, Michele Cafagna, Felice Dell’Orletta, Malvina Nissim, and Albert Gatt. 2020b. CHANGE-IT @ EVALITA 2020: Change headlines, adapt news, generate. In _Proceedings of Seventh Evaluation Campaign of Natural Language Processing and Speech Tools for Italian. Final Workshop (EVALITA 2020)_, Online. CEUR.org. 
+*   De Mattei et al. (2020c) Lorenzo De Mattei, Michele Cafagna, Huiyuan Lai, Felice Dell’Orletta, Malvina Nissim, and Albert Gatt. 2020c. On the interaction of automatic evaluation and task framing in headline style transfer. In _Proceedings of the 1st Workshop on Evaluating NLG Evaluation_, pages 38–43, Online (Dublin, Ireland). Association for Computational Linguistics. 
+*   Devlin et al. (2019) Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In _Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers)_, pages 4171–4186, Minneapolis, Minnesota. Association for Computational Linguistics. 
+*   He et al. (2023) Pengcheng He, Jianfeng Gao, and Weizhu Chen. 2023. DeBERTaV3: Improving DeBERTa using ELECTRA-style pre-training with gradient-disentangled embedding sharing. In _Proceedings of the Eleventh International Conference on Learning Representations (ICLR’23)_, Kigali, Rwanda. OpenReview.net. 
+*   Hromei et al. (2023) Claudiu D. Hromei, Danilo Croce, Valerio Basile, and Roberto Basili. 2023. ExtremITA at EVALITA 2023: Multi-task sustainable scaling to large language models at its extreme. In _Proceedings of the Eighth Evaluation Campaign of Natural Language Processing and Speech Tools for Italian. Final Workshop (EVALITA 2023)_, volume 3473 of _CEUR Workshop Proceedings_, Parma, Italy. CEUR-WS.org. 
+*   Ji et al. (2023) Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Ye Jin Bang, Andrea Madotto, and Pascale Fung. 2023. Survey of hallucination in natural language generation. _ACM Comput. Surv._, 55(12). 
+*   Kreutzer et al. (2022) Julia Kreutzer, Isaac Caswell, Lisa Wang, Ahsan Wahab, Daan van Esch, Nasanbayar Ulzii-Orshikh, Allahsera Tapo, Nishant Subramani, Artem Sokolov, Claytone Sikasote, Monang Setyawan, Supheakmungkol Sarin, Sokhar Samb, Benoît Sagot, Clara Rivera, Annette Rios, Isabel Papadimitriou, Salomey Osei, Pedro Ortiz Suarez, Iroro Orife, Kelechi Ogueji, Andre Niyongabo Rubungo, Toan Q. Nguyen, Mathias Müller, André Müller, Shamsuddeen Hassan Muhammad, Nanda Muhammad, Ayanda Mnyakeni, Jamshidbek Mirzakhalov, Tapiwanashe Matangira, Colin Leong, Nze Lawson, Sneha Kudugunta, Yacine Jernite, Mathias Jenny, Orhan Firat, Bonaventure F.P. Dossou, Sakhile Dlamini, Nisansa de Silva, Sakine Çabuk Ballı, Stella Biderman, Alessia Battisti, Ahmed Baruwa, Ankur Bapna, Pallavi Baljekar, Israel Abebe Azime, Ayodele Awokoya, Duygu Ataman, Orevaoghene Ahia, Oghenefego Ahia, Sweta Agrawal, and Mofetoluwa Adeyemi. 2022. Quality at a Glance: An Audit of Web-Crawled Multilingual Datasets. _Transactions of the Association for Computational Linguistics_, 10:50–72. 
+*   Kudo (2018) Taku Kudo. 2018. Subword regularization: Improving neural network translation models with multiple subword candidates. In _Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)_, pages 66–75, Melbourne, Australia. Association for Computational Linguistics. 
+*   La Quatra and Cagliero (2023) Moreno La Quatra and Luca Cagliero. 2023. Bart-it: An efficient sequence-to-sequence model for italian text summarization. _Future Internet_, 15(1). 
+*   Landro et al. (2022) Nicola Landro, Ignazio Gallo, Riccardo La Grassa, and Edoardo Federici. 2022. Two new datasets for italian-language abstractive text summarization. _Information_, 13(5). 
+*   Leonardelli and Casula (2023) Elisa Leonardelli and Camilla Casula. 2023. DH-FBK at HODI: multi-task learning with classifier ensemble agreement, oversampling and synthetic data. In _Proceedings of the Eighth Evaluation Campaign of Natural Language Processing and Speech Tools for Italian. Final Workshop (EVALITA 2023)_, volume 3473 of _CEUR Workshop Proceedings_, Parma, Italy. CEUR-WS.org. 
+*   Lin (2004) Chin-Yew Lin. 2004. ROUGE: A package for automatic evaluation of summaries. In _Text Summarization Branches Out_, pages 74–81, Barcelona, Spain. Association for Computational Linguistics. 
+*   Loshchilov and Hutter (2019) Ilya Loshchilov and Frank Hutter. 2019. Decoupled weight decay regularization. In _Proceedings of the Seventh International Conference on Learning Representations (ICLR’19)_, New Orleans, LA, USA. OpenReview.net. 
+*   Maynez et al. (2020) Joshua Maynez, Shashi Narayan, Bernd Bohnet, and Ryan McDonald. 2020. On faithfulness and factuality in abstractive summarization. In _Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics_, pages 1906–1919, Online. Association for Computational Linguistics. 
+*   Miaschi et al. (2022) Alessio Miaschi, Gabriele Sarti, Dominique Brunato, Felice Dell’Orletta, and Giulia Venturi. 2022. Probing linguistic knowledge in italian neural language models across language varieties. _Italian Journal of Computational Linguistics (IJCoL)_, 8(1):25–44. 
+*   Mousavi et al. (2023) Seyed Mahed Mousavi, Simone Caldarella, and Giuseppe Riccardi. 2023. Response generation in longitudinal dialogues: Which knowledge representation helps? In _Proceedings of the 5th Workshop on NLP for Conversational AI (NLP4ConvAI 2023)_, pages 1–11, Toronto, Canada. Association for Computational Linguistics. 
+*   Nagoudi et al. (2022) El Moatez Billah Nagoudi, AbdelRahim Elmadany, and Muhammad Abdul-Mageed. 2022. AraT5: Text-to-text transformers for Arabic language generation. In _Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)_, pages 628–647, Dublin, Ireland. Association for Computational Linguistics. 
+*   Nozza et al. (2020) Debora Nozza, Federico Bianchi, and Dirk Hovy. 2020. What the [MASK]? making sense of language-specific BERT models. _ArXiv Computation and Language_, arXiv:2003.02912(v1). 
+*   Papucci et al. (2022) Michele Papucci, Chiara De Nigris, Alessio Miaschi, and Felice Dell’Orletta. 2022. Evaluating text-to-text framework for topic and style classification of italian texts. In _Proceedings of the Sixth Workshop on Natural Language for Artificial Intelligence (NL4AI 2022)_, volume 3287 of _CEUR Workshop Proceedings_, pages 56–70, Udine, Italy. CEUR-WS.org. 
+*   Phan et al. (2022) Long Phan, Hieu Tran, Hieu Nguyen, and Trieu H. Trinh. 2022. ViT5: Pretrained text-to-text transformer for Vietnamese language generation. In _Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies: Student Research Workshop_, pages 136–142, Hybrid: Seattle, Washington + Online. Association for Computational Linguistics. 
+*   Polignano et al. (2019) Marco Polignano, Pierpaolo Basile, Marco de Gemmis, Giovanni Semeraro, and Valerio Basile. 2019. Alberto: Italian BERT language understanding model for NLP challenging tasks based on tweets. In _Proceedings of the Sixth Italian Conference on Computational Linguistics (CLiC-it ’19)_, volume 2481 of _CEUR Workshop Proceedings_, Bari, Italy. CEUR-WS.org. 
+*   Radford et al. (2019) Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. Language models are unsupervised multitask learners. _OpenAI Blog_. 
+*   Raffel et al. (2020) Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. _Journal of Machine Learning Research_, 21(140):1–67. 
+*   Rajpurkar et al. (2016) Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. 2016. SQuAD: 100,000+ questions for machine comprehension of text. In _Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing_, pages 2383–2392, Austin, Texas. Association for Computational Linguistics. 
+*   Rao and Tetreault (2018) Sudha Rao and Joel Tetreault. 2018. Dear sir or madam, may I introduce the GYAFC dataset: Corpus, benchmarks and metrics for formality style transfer. In _Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers)_, pages 129–140, New Orleans, Louisiana. Association for Computational Linguistics. 
+*   Riabi et al. (2021) Arij Riabi, Thomas Scialom, Rachel Keraron, Benoît Sagot, Djamé Seddah, and Jacopo Staiano. 2021. Synthetic data augmentation for zero-shot cross-lingual question answering. In _Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing_, pages 7016–7030, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics. 
+*   Rust et al. (2021) Phillip Rust, Jonas Pfeiffer, Ivan Vulić, Sebastian Ruder, and Iryna Gurevych. 2021. How good is your tokenizer? on the monolingual performance of multilingual language models. In _Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers)_, pages 3118–3135, Online. Association for Computational Linguistics. 
+*   Sanh et al. (2022) Victor Sanh, Albert Webson, Colin Raffel, Stephen H. Bach, Lintang Sutawika, Zaid Alyafeai, Antoine Chaffin, Arnaud Stiegler, Arun Raja, Manan Dey, M Saiful Bari, Canwen Xu, Urmish Thakker, Shanya Sharma Sharma, Eliza Szczechla, Taewoon Kim, Gunjan Chhablani, Nihal V. Nayak, Debajyoti Datta, Jonathan Chang, Mike Tian-Jian Jiang, Han Wang, Matteo Manica, Sheng Shen, Zheng Xin Yong, Harshit Pandey, Rachel Bawden, Thomas Wang, Trishala Neeraj, Jos Rozen, Abheesht Sharma, Andrea Santilli, Thibault Févry, Jason Alan Fries, Ryan Teehan, Teven Le Scao, Stella Biderman, Leo Gao, Thomas Wolf, and Alexander M. Rush. 2022. Multitask prompted training enables zero-shot task generalization. In _Proceedings of the Tenth International Conference on Learning Representations (ICLR’22)_, Online. OpenReview.net. 
+*   Shazeer (2020) Noam Shazeer. 2020. GLU variants improve transformer. _ArXiv Machine Learning_, arXiv:2002.05202(v1). 
+*   Shazeer and Stern (2018) Noam Shazeer and Mitchell Stern. 2018. Adafactor: Adaptive learning rates with sublinear memory cost. In _Proceedings of the 35th International Conference on Machine Learning (ICML’18)_, volume 80 of _Proceedings of Machine Learning Research_, pages 4596–4604. PMLR. 
+*   Sutskever et al. (2014) Ilya Sutskever, Oriol Vinyals, and Quoc V Le. 2014. Sequence to sequence learning with neural networks. In _Advances in Neural Information Processing Systems_, volume 27. Curran Associates, Inc. 
+*   Tay et al. (2022) Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, and Donald Metzler. 2022. Scale efficiently: Insights from pre-training and fine-tuning transformers. In _Proceedings of the Tenth International Conference on Learning Representations (ICLR’22)_, Online. OpenReview.net. 
+*   Ulčar and Robnik-Šikonja (2023) Matej Ulčar and Marko Robnik-Šikonja. 2023. Sequence-to-sequence pretraining for a less-resourced slovenian language. _Frontiers in Artificial Intelligence_, 6. 
+*   Vaswani et al. (2017) Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Ł ukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In _Advances in Neural Information Processing Systems_, volume 30. Curran Associates, Inc. 
+*   Wolf et al. (2020) Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander Rush. 2020. Transformers: State-of-the-art natural language processing. In _Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations_, pages 38–45, Online. Association for Computational Linguistics. 
+*   Xue et al. (2022) Linting Xue, Aditya Barua, Noah Constant, Rami Al-Rfou, Sharan Narang, Mihir Kale, Adam Roberts, and Colin Raffel. 2022. ByT5: Towards a token-free future with pre-trained byte-to-byte models. _Transactions of the Association for Computational Linguistics_, 10:291–306. 
+*   Xue et al. (2021) Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, and Colin Raffel. 2021. mT5: A massively multilingual pre-trained text-to-text transformer. In _Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies_, pages 483–498, Online. Association for Computational Linguistics. 
+*   Zhang et al. (2020) Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q. Weinberger, and Yoav Artzi. 2020. Bertscore: Evaluating text generation with bert. In _Proceedings of the Eighth International Conference on Learning Representations (ICLR’20)_, Addis Abeba, Ethiopia. OpenReview.net. 
+
+Appendix A Model Parametrization and Additional Pretraining Details
+-------------------------------------------------------------------
+
+Models are trained on a TPU v3-8 accelerator on Google Cloud Platform using the JAX framework(Bradbury et al., [2018](https://arxiv.org/html/2203.03759v2#bib.bib2)) and Huggingface Transformers(Wolf et al., [2020](https://arxiv.org/html/2203.03759v2#bib.bib46)). We adopt the T5 v1.1 architecture 12 12 12[text-to-text-transfer-transformer/t511](https://github.com/google-research/text-to-text-transfer-transformer/blob/main/released_checkpoints.md#t511) also used by the mT5 model, improving upon the original T5 by using GeGLU nonlinearities(Shazeer, [2020](https://arxiv.org/html/2203.03759v2#bib.bib40)), scaling model hidden size alongside feedforward layers and pre-training only on unlabeled data, without dropout. All models are pre-trained with a learning rate of 5e-3 and a maximum sequence length of 512 tokens using the Adafactor optimizer Shazeer and Stern ([2018](https://arxiv.org/html/2203.03759v2#bib.bib41)) to reduce the memory footprint of training and are validated on a fixed subset of 15’000 examples. Figure[1](https://arxiv.org/html/2203.03759v2#A1.F1 "Figure 1 ‣ Appendix A Model Parametrization and Additional Pretraining Details ‣ IT5: Text-to-text Pretraining for Italian Language Understanding and Generation") shows the computed loss during the training process for the three standard models (excluding the efficient one). We used the Google Cloud Carbon Footprint tool to estimate the overall amount of CO2 generated by the pre-training process and found it to be approximately equal to 7kgCO2, corresponding approximately to the emissions of a 60km car ride.13 13 13[https://ec.europa.eu/eurostat/cache/metadata/en/sdg_12_30_esmsip2.htm](https://ec.europa.eu/eurostat/cache/metadata/en/sdg_12_30_esmsip2.htm)
+
+![Image 1: Refer to caption](https://arxiv.org/html/2203.03759v2/x1.png)
+
+Figure 1: Loss curves for the masked span prediction task used to pre-train the IT5 models.
+
+Table 3: Full parametrization for IT5 models. Parameters below the line are shared across all configurations.
+
+Table[3](https://arxiv.org/html/2203.03759v2#A1.T3 "Table 3 ‣ Appendix A Model Parametrization and Additional Pretraining Details ‣ IT5: Text-to-text Pretraining for Italian Language Understanding and Generation") shows the full parameter configuration for all IT5 model sizes. The models correspond to the three canonical sizes for T5 models (Small, Base, Large) with T5 v1.1 improvements, and the efficient small version with 32 encoder layers introduced by Tay et al. ([2022](https://arxiv.org/html/2203.03759v2#bib.bib43)) (EL32).
+
+Appendix B Italian Baseline Scores for BERTScore Rescaling
+----------------------------------------------------------
+
+Table [4](https://arxiv.org/html/2203.03759v2#A3.T4 "Table 4 ‣ Appendix C Parametrization for Fine-tuning Experiments ‣ IT5: Text-to-text Pretraining for Italian Language Understanding and Generation") contains the baseline scores computed on the first 1M examples of the Cleaned Italian mC4 Corpus using the same model, which we later use for evaluating generation performances. These should be used alongside the same model and the --rescale_with_baseline option to obtain BERTScore performances directly comparable to the ones reported in this work.
+
+The hash code used for reproducibility by the BERTScore library is dbmdz/bert-base-italian-xxl-uncase 
+
+d_L10_no-idf_version=0.3.11(hug_tr 
+
+ans=4.16.0)-rescaled
+
+Appendix C Parametrization for Fine-tuning Experiments
+------------------------------------------------------
+
+Table[5](https://arxiv.org/html/2203.03759v2#A3.T5 "Table 5 ‣ Appendix C Parametrization for Fine-tuning Experiments ‣ IT5: Text-to-text Pretraining for Italian Language Understanding and Generation") contains task-specific parameters that were used for the fine-tuning experiments. For mT5 Small, IT5 Small and IT5 Base models we use a learning rate of 5e-4 and a batch size of 64 examples, while larger models (mT5 Base and IT5 Large) were fine-tuned with a leaning rate of 5e-5 and a batch size of 32. All models are fine-tuned with linear schedule with no warmup using the AdamW optimizer(Loshchilov and Hutter, [2019](https://arxiv.org/html/2203.03759v2#bib.bib24)).
+
+We highlight that the batch sizes used for fine-tuning are significantly smaller from the canonical batch size of 128 adopted by Raffel et al. ([2020](https://arxiv.org/html/2203.03759v2#bib.bib34)) due to hardware limitations.
+
+Table 4: Baseline scores for using dbmdz/bert-base-italian-xxl-uncased with the BERTScore evaluation framework.
+
+Table 5: Task-specific fine-tuning parameters. SL = Max. source length. TL = Max. target length.
+
+Appendix D Generation Examples using IT5 Base
+---------------------------------------------
+
+Tables[D](https://arxiv.org/html/2203.03759v2#A4 "Appendix D Generation Examples using IT5 Base ‣ IT5: Text-to-text Pretraining for Italian Language Understanding and Generation") and[D](https://arxiv.org/html/2203.03759v2#A4 "Appendix D Generation Examples using IT5 Base ‣ IT5: Text-to-text Pretraining for Italian Language Understanding and Generation") present some generation examples for the IT5 Base model across all the evaluated tasks. We use […] to omit portions of long sources that we judge to be less salient to improve the readability of the examples. Outputs are lowercase because the IT5 Base tokenizer is uncased, while using the EL32 model would produce results with normal casing. Examples shown were randomly sampled among model generations for the respective test sets.
+
+While the quality is generally high, we observe that summarization results, especially for the WITS dataset, tend to contain hallucinated information obtained by combining unrelated portions of the source. For example, "Libro Entertainment" in the first example appears to be a translated version of the actual name of the publishing house, and Paolo Villaggio published an audiobook with the company rather than owning it, as it is stated in the generated summary. This is a well-known problem of abstractive summarization systems(Maynez et al., [2020](https://arxiv.org/html/2203.03759v2#bib.bib25); Ji et al., [2023](https://arxiv.org/html/2203.03759v2#bib.bib17)), which hasn’t been studied extensively for languages other than English.
+
+Table 6: Examples for the summarization and question answering/generation tasks from the respective test sets using the IT5 Base model.
+
+Table 7: Examples of headline style transfer, headline generation and formality style transfer tasks from the respective test sets using the IT5 Base model.