Hopkins and May, EMNLP 2011. Tuning as Ranking
Mark Hopkins and Jonathan May. 2011. Tuning as Ranking. In Proceedings of EMNLP-2011.
This paper presents a simple and scalable method for statistical machine translation parameter tuning based on the pairwise approach to ranking. This pairwise ranking optimization (PRO) method has advantages over MERT Och, 2003 as it is not limited to a handful of parameters and can easily handle systems with thousands of features. In addition, unlike recent approaches built upon the MIRA algorithm of Crammer et al., 2006 (Watanabe et al., 2007), PRO is easy to implement.
Although, MERT is well-understood, easy to implement, and runs quickly, it can behave erratically and would not scale beyond a handful of features. This is a major bottleneck towards working with richer feature representations and structure.
Hence, the authors propose a simpler approach than MIRA, to tuning that similarly scales to high-dimensional feature spaces. Tuning is treated as a ranking problem (Chen et al., 2009), where the explicit goal is to learn to correctly rank candidate translations. The authors describe a pairwise approach to ranking, in which the ranking problem is reduced to the binary classification task of deciding between candidate translation pairs.
The goal of tuning is to learn a weight vector such that assigns a high score to good translations and a low score to bad translations. is given by:
where, the scores for candidate translations are represented in the following form,
Optimization via Pairwise Ranking
MIRA scales well to high-dimensionality candidate spaces. However, its architecture is complex and different to that of MERT. This method assumes a gold scoring function which can be decomposed in the following way:
where, is a local scoring function that scores each candidate translation .
In a pairwise ranking approach, the learning task is framed as the classification of candidate pairs into two categories: correctly ordered and incorrectly ordered. For each pair of translation candidates and , an inequality relation is established:
On simplifying the above, we get
Hence, the optimization problem reduces to that of binary classification. The training data generated is fed directly to any off-the-shelf classification tool that returns a linear classifier, in order to obtain a weight vector that optimizes the above condition. The exact loss function optimized depends on the choice of classifier.
Standard approaches to pairwise ranking enumerate all difference vectors as training data. For tuning, this typically means vectors, where is the cardinality of the largest . This could lead to an exponentially large feature vector space due to high-dimensionality of the data. Hence, a sampler template is used to sample from the space of different vectors in order to make it tractable.
For each source sentence , the sampler generates candidate translation pairs , and accepts each pair with probability . Among the accepted pairs, it keeps the with greatest differential, and adds their difference vectors to the training data. The sampling algorithm is outlined in the figure below:
For each of our systems we identify two feature sets: baseline, which correspond to the typical small feature set reported in current MT literature, and extended, a superset of baseline, which adds hundreds or thousands of features. They use 15 baseline features for phrase-based MT, similar to the baseline features described by Watanabe et al., 2007. They also use 19 baseline features for syntax-based MT, similar to the baseline features described by Chiang et al., 2008.
They used the following feature classes for both syntax-based and phrase-based MT extended scenarios:
- Discount features for rule frequency bins.
- Target word insertion features
Syntax-based MT specific features:
- Rule overlap features
- Node count features
Phrase-based MT specific features:
- Unigram word pair features for the 80 most frequent words in both languages plus tokens for unaligned and all other words.
- Source, target, and joint phrase length features from 1 to 7, e.g. "tgt=4", "src=2", and "src/tgt=2,4".
Experiments and Results
The authors conducted experiments for Urdu-English, Arabic-English and Chinese-English translation tasks and compared the performance of MERT, MIRA and PRO learning algorithms. They evaluated performance on phrase-based and syntax-based MT systems.
The table above shows the size of the data for experiments and results reported in the paper.
The table below shows machine translation performance for the experiments listed in this paper. Scores are case-sensitive IBM BLEU. For every choice of system, language pair, and feature set, PRO performs comparably with the other methods.
 Franz Josef Och, Minimum error rate training in statistical machine translation. 2003. In Proceedings of the 41st Annual Meeting on Association for Computational Linguistics, p.160-167, July 07-12, 2003, Sapporo, Japan
 Taro Watanabe and Jun Suzuki, Hajime Tsukada and Hideki Isozaki. 2007. Online Large-Margin Training for Statistical Machine Translation. In Proceedings of the 2007 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning (EMNLP-CoNLL)
 David Chiang, Yuval Marton, and Philip Resnik. 2008. Online large-margin training of syntactic and structural translation features. In Proceedings of the 2008 Conference on Empirical Methods in Natural Language Processing, pages 224–233, Honolulu, HI, October