In the last few years, pre-trained neural architectures have provided impressive improvements across several NLP tasks. Still, generative language models are available mainly for English. We develop GePpeTto, the first generative language model for Italian, built using the GPT-2 architecture. We provide a thorough analysis of GePpeTto's quality by means of both an automatic and a human-based evaluation. The automatic assessment consists in (i) calculating perplexity across different genres and (ii) a profiling analysis over GePpeTto's writing characteristics. We find that GePpeTto's production is a sort of bonsai version of human production, with shorter but yet complex sentences. Human evaluation is performed over a sentence completion task, whereGePpeTto's output is judged as natural more often than not, and much closer to the original human texts than to a simpler language model which we take as baseline.
GePpeTto Carves Italian into a Language Model
Dell'Orletta F;
2020
Abstract
In the last few years, pre-trained neural architectures have provided impressive improvements across several NLP tasks. Still, generative language models are available mainly for English. We develop GePpeTto, the first generative language model for Italian, built using the GPT-2 architecture. We provide a thorough analysis of GePpeTto's quality by means of both an automatic and a human-based evaluation. The automatic assessment consists in (i) calculating perplexity across different genres and (ii) a profiling analysis over GePpeTto's writing characteristics. We find that GePpeTto's production is a sort of bonsai version of human production, with shorter but yet complex sentences. Human evaluation is performed over a sentence completion task, whereGePpeTto's output is judged as natural more often than not, and much closer to the original human texts than to a simpler language model which we take as baseline.| Campo DC | Valore | Lingua |
|---|---|---|
| dc.authority.orgunit | Istituto di linguistica computazionale "Antonio Zampolli" - ILC | - |
| dc.authority.people | De Mattei L | it |
| dc.authority.people | Cafagna M | it |
| dc.authority.people | Dell'Orletta F | it |
| dc.authority.people | Nissim M | it |
| dc.authority.people | Guerini M | it |
| dc.collection.id.s | 71c7200a-7c5f-4e83-8d57-d3d2ba88f40d | * |
| dc.collection.name | 04.01 Contributo in Atti di convegno | * |
| dc.contributor.appartenenza | Istituto di linguistica computazionale "Antonio Zampolli" - ILC | * |
| dc.contributor.appartenenza.mi | 918 | * |
| dc.date.accessioned | 2024/02/21 05:17:52 | - |
| dc.date.available | 2024/02/21 05:17:52 | - |
| dc.date.issued | 2020 | - |
| dc.description.abstracteng | In the last few years, pre-trained neural architectures have provided impressive improvements across several NLP tasks. Still, generative language models are available mainly for English. We develop GePpeTto, the first generative language model for Italian, built using the GPT-2 architecture. We provide a thorough analysis of GePpeTto's quality by means of both an automatic and a human-based evaluation. The automatic assessment consists in (i) calculating perplexity across different genres and (ii) a profiling analysis over GePpeTto's writing characteristics. We find that GePpeTto's production is a sort of bonsai version of human production, with shorter but yet complex sentences. Human evaluation is performed over a sentence completion task, whereGePpeTto's output is judged as natural more often than not, and much closer to the original human texts than to a simpler language model which we take as baseline. | - |
| dc.description.affiliations | Department of Computer Science, University of Pisa, Italy; University of Malta, Malta; Istituto di Linguistica Computazionale "Antonio Zampolli", Pisa, Italy; Center for Language and Cognition Groningen, University of Groningen, The Netherlands; Fondazione Bruno Kessler, Trento, Italy; | - |
| dc.description.allpeople | De Mattei, L; Cafagna, M; Dell'Orletta, F; Nissim, M; Guerini, M | - |
| dc.description.allpeopleoriginal | De Mattei L., Cafagna M., Dell'Orletta F., Nissim M., Guerini M. | - |
| dc.description.fulltext | none | en |
| dc.description.numberofauthors | 5 | - |
| dc.identifier.uri | https://hdl.handle.net/20.500.14243/401384 | - |
| dc.language.iso | eng | - |
| dc.relation.conferencedate | 01-03/03/2021 | - |
| dc.relation.conferencename | Seventh Italian Conference on Computational Linguistics (CLiC-it 2020) | - |
| dc.relation.conferenceplace | online | - |
| dc.subject.keywords | natural language generation | - |
| dc.subject.singlekeyword | natural language generation | * |
| dc.title | GePpeTto Carves Italian into a Language Model | en |
| dc.type.driver | info:eu-repo/semantics/conferenceObject | - |
| dc.type.full | 04 Contributo in convegno::04.01 Contributo in Atti di convegno | it |
| dc.type.miur | 273 | - |
| dc.type.referee | Sì, ma tipo non specificato | - |
| dc.ugov.descaux1 | 450797 | - |
| iris.orcid.lastModifiedDate | 2024/04/04 13:06:46 | * |
| iris.orcid.lastModifiedMillisecond | 1712228806705 | * |
| iris.sitodocente.maxattempts | 1 | - |
| Appare nelle tipologie: | 04.01 Contributo in Atti di convegno | |
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
