This paper investigates the ability of XLM language model to transfer linguistic knowledge cross-lingually, verifying if and to which extent syntactic dependency relationships learnt in a language are maintained in other languages. In detail, a structural probe is developed to analyse the cross-lingual syntactic transfer capability of XLM model and comparison of cross-language syntactic transfer among languages belonging to different families from a typological classification, which are characterised by very different syntactic constructions. The probe aims to reconstruct the dependency parse tree of a sentence in order to representing the input sentences with the contextual embeddings from XLM layers. The results of the experimental assessment improved the previous results obtained using mBERT model.

Probing Cross-lingual Transfer of~XLM Multi-language Model

Guarasci, R.
;
Silvestri, S.;Esposito, M.
2024

Abstract

This paper investigates the ability of XLM language model to transfer linguistic knowledge cross-lingually, verifying if and to which extent syntactic dependency relationships learnt in a language are maintained in other languages. In detail, a structural probe is developed to analyse the cross-lingual syntactic transfer capability of XLM model and comparison of cross-language syntactic transfer among languages belonging to different families from a typological classification, which are characterised by very different syntactic constructions. The probe aims to reconstruct the dependency parse tree of a sentence in order to representing the input sentences with the contextual embeddings from XLM layers. The results of the experimental assessment improved the previous results obtained using mBERT model.
2024
Istituto di Calcolo e Reti ad Alte Prestazioni - ICAR
9783031535543
9783031535550
LLMs
XLM
Structural probe
Cross Language
Dependency Tree
File in questo prodotto:
File Dimensione Formato  
Probing_XLM.pdf

non disponibili

Tipologia: Versione Editoriale (PDF)
Licenza: NON PUBBLICO - Accesso privato/ristretto
Dimensione 250.5 kB
Formato Adobe PDF
250.5 kB Adobe PDF   Visualizza/Apri   Richiedi una copia

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/20.500.14243/469846
Citazioni
  • ???jsp.display-item.citation.pmc??? ND
  • Scopus 2
  • ???jsp.display-item.citation.isi??? ND
social impact