Page 61 - Fister jr., Iztok, and Andrej Brodnik (eds.). StuCoSReC. Proceedings of the 2016 3rd Student Computer Science Research Conference. Koper: University of Primorska Press, 2016
P. 61
ghts Optimization in Statistical Machine Translation
using Modified Genetic Algorithm
Jani Dugonik
University of Maribor, Faculty of Electrical Engineering and Computer Science
jani.dugonik@um.si
ABSTRACT logical evolution. The GA algorithm repeatedly modifies a
population of individual solutions by relying on bio-inspired
Translations in a statistical machine translations are gen- operators such as mutation, crossover and selection.
erated on the basis of statistical models. These models are
weighted, and different models’ weights gives different trans- The main goal of this paper was to build two SMT sys-
lations. The problem of finding a good translation can be tems for the language pair English-Slovenian, and to im-
regarded as an optimization problem. One way to find good prove translation quality with finding optimal weights using
translations is to find optimal models’ weights. In this paper a modified genetic algorithm. The used translation error
we present the usage of the genetic algorithm with roulette metric was bilingual evaluation understudy (BLEU) [19].
wheel selection to find the optimal models’ weights. Exper-
iments were performed using well-known corpora and the The remainder of the paper is organized as follows. Section 2
most used translation metric in the SMT field. We com- presents related work. In the Section 3 we will describe some
pared the modified genetic algorithm with the basic genetic background of the problem and in Section 4 we will describe
algorithm and the state-of-the-art method. The results show the basic and modified GA algorithms. Experiments are
improvement in the translation quality and are comparable presented in Section 5. We conclude this paper with Section
to the related state-of-the-art method. 6 where we give our conclusion.
Keywords 2. RELATED WORK
Genetic Algorithm, Statistical Machine Translation, Weight The ability to optimize models’ weights according to a trans-
Optimization, Roulette Wheel Selection lation error metric has become a standard assumption in
SMT, due to the wide-spread adoption of Minimum Error
1. INTRODUCTION Rate Training (MERT) [16, 3]. Authors in [16] analyzed
various training criteria which directly optimize translation
Machine translation (MT) can ease the work of a translator. quality. These training criteria make use of an automatic
The most studied and used MT method is the statistical evaluation metrics. They describe a new algorithm for ef-
machine translation (SMT) [2]. SMT is based on statistical ficient training an unsmoothed error count. The results in
methods which were originally used for translating single the paper show that significantly better results can often
words (word-based translation). Now they have progressed be obtained if the final evaluation criterion is taken directly
to the level of translating sequences of words called phrases into account as part of the training procedure.
(phrase-based translation). Currently the most successful
SMT systems are based on phrase-based translation [20]. The problems with MERT can be addressed through the use
Translations in SMT are generated on the basis of statisti- of surrogate loss functions. The Margin Infused Relaxed Al-
cal models. Each model is weighted, and different models’ gorithm (MIRA) [23, 10, 9, 8, 13] employs a structured hinge
weights provide various translations, which can be evaluated loss and is an instance of online learning. In order to im-
by translation error metrics. The problem of finding a good prove generalization, the average of all weights seen during
translation can be regarded as an optimization problem. learning is used on unseen data. D. Chiang in [10] took ad-
The optimization can be done with models’ weights. There vantage of MIRA’s online nature to modify each update to
are various methods for solving this optimization problem better suit SMT. The cost is defined using a pseudo-corpus
and in this paper we used a genetic algorithm (GA) [18]. BLEU which tracks the n-gram statistics of the model-best
It is based on a natural selection process that mimics bio- derivations from the last few updates. This modified cost
matches corpus BLEU better than Add-1 smoothing but
also makes cost time-dependent.
As far as we know, there is only one published method for
tuning SMT based on EAs. In [11] authors used a basic
differential evolution algorithm to see how evolutionary al-
gorithms behave in the field of statistical machine transla-
tion. The results were comparable to the state-of-the-art
StuCoSReC Proceedings of the 2016 3rd Student Computer Science Research Conference 61
Ljubljana, Slovenia, 12 October
using Modified Genetic Algorithm
Jani Dugonik
University of Maribor, Faculty of Electrical Engineering and Computer Science
jani.dugonik@um.si
ABSTRACT logical evolution. The GA algorithm repeatedly modifies a
population of individual solutions by relying on bio-inspired
Translations in a statistical machine translations are gen- operators such as mutation, crossover and selection.
erated on the basis of statistical models. These models are
weighted, and different models’ weights gives different trans- The main goal of this paper was to build two SMT sys-
lations. The problem of finding a good translation can be tems for the language pair English-Slovenian, and to im-
regarded as an optimization problem. One way to find good prove translation quality with finding optimal weights using
translations is to find optimal models’ weights. In this paper a modified genetic algorithm. The used translation error
we present the usage of the genetic algorithm with roulette metric was bilingual evaluation understudy (BLEU) [19].
wheel selection to find the optimal models’ weights. Exper-
iments were performed using well-known corpora and the The remainder of the paper is organized as follows. Section 2
most used translation metric in the SMT field. We com- presents related work. In the Section 3 we will describe some
pared the modified genetic algorithm with the basic genetic background of the problem and in Section 4 we will describe
algorithm and the state-of-the-art method. The results show the basic and modified GA algorithms. Experiments are
improvement in the translation quality and are comparable presented in Section 5. We conclude this paper with Section
to the related state-of-the-art method. 6 where we give our conclusion.
Keywords 2. RELATED WORK
Genetic Algorithm, Statistical Machine Translation, Weight The ability to optimize models’ weights according to a trans-
Optimization, Roulette Wheel Selection lation error metric has become a standard assumption in
SMT, due to the wide-spread adoption of Minimum Error
1. INTRODUCTION Rate Training (MERT) [16, 3]. Authors in [16] analyzed
various training criteria which directly optimize translation
Machine translation (MT) can ease the work of a translator. quality. These training criteria make use of an automatic
The most studied and used MT method is the statistical evaluation metrics. They describe a new algorithm for ef-
machine translation (SMT) [2]. SMT is based on statistical ficient training an unsmoothed error count. The results in
methods which were originally used for translating single the paper show that significantly better results can often
words (word-based translation). Now they have progressed be obtained if the final evaluation criterion is taken directly
to the level of translating sequences of words called phrases into account as part of the training procedure.
(phrase-based translation). Currently the most successful
SMT systems are based on phrase-based translation [20]. The problems with MERT can be addressed through the use
Translations in SMT are generated on the basis of statisti- of surrogate loss functions. The Margin Infused Relaxed Al-
cal models. Each model is weighted, and different models’ gorithm (MIRA) [23, 10, 9, 8, 13] employs a structured hinge
weights provide various translations, which can be evaluated loss and is an instance of online learning. In order to im-
by translation error metrics. The problem of finding a good prove generalization, the average of all weights seen during
translation can be regarded as an optimization problem. learning is used on unseen data. D. Chiang in [10] took ad-
The optimization can be done with models’ weights. There vantage of MIRA’s online nature to modify each update to
are various methods for solving this optimization problem better suit SMT. The cost is defined using a pseudo-corpus
and in this paper we used a genetic algorithm (GA) [18]. BLEU which tracks the n-gram statistics of the model-best
It is based on a natural selection process that mimics bio- derivations from the last few updates. This modified cost
matches corpus BLEU better than Add-1 smoothing but
also makes cost time-dependent.
As far as we know, there is only one published method for
tuning SMT based on EAs. In [11] authors used a basic
differential evolution algorithm to see how evolutionary al-
gorithms behave in the field of statistical machine transla-
tion. The results were comparable to the state-of-the-art
StuCoSReC Proceedings of the 2016 3rd Student Computer Science Research Conference 61
Ljubljana, Slovenia, 12 October