In SMT, the instability of MERT, the commonly used optimizer, is an acknowledged problem. This paper presents two methods for smoothing the MERT instability. Both exploit a set of different realizations of the same system obtained by running the optimization stage multiple times. One method averages the sets of different optimal weights; the other combines the translations generated by the various realizations. Experiments conducted on two different sized tasks involving four different language pairs show that both methods are effective in smoothing instability, but also that the average system well competes with the more expensive system combination.

Methods for Smoothing the Optimizer Instability in SMT

Cettolo, Mauro;Bertoldi, Nicola;Federico, Marcello
2011-01-01

Abstract

In SMT, the instability of MERT, the commonly used optimizer, is an acknowledged problem. This paper presents two methods for smoothing the MERT instability. Both exploit a set of different realizations of the same system obtained by running the optimization stage multiple times. One method averages the sets of different optimal weights; the other combines the translations generated by the various realizations. Experiments conducted on two different sized tasks involving four different language pairs show that both methods are effective in smoothing instability, but also that the average system well competes with the more expensive system combination.
File in questo prodotto:
File Dimensione Formato  
paper-FINAL.pdf

accesso aperto

Tipologia: Documento in Pre-print
Licenza: Creative commons
Dimensione 167.12 kB
Formato Adobe PDF
167.12 kB Adobe PDF Visualizza/Apri

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/11582/42180
Citazioni
  • ???jsp.display-item.citation.pmc??? ND
social impact