System Combination Using Discriminative Cross-Adaptation

Proceedings of the 5th International Joint Conference on Natural Language Processing, pages 667–675, Chiang Mai, Thailand, November 8 – 13, 2011. c⃝2011 AFNLP System Combination Using Discriminative Cross-Adaptation Jacob Devlin and Antti-Veikko I. Rosti and Shankar Ananthakrishnan and Spyros Matsoukas ∗ Raytheon BBN Technologies, 10 Moulton St, Cambridge, MA 02138, USA {jdevlin,arosti,sanantha,smatsouk}@bbn.com Abstract Cross-adaptation (CA) based methods of machine translation (MT) system combi- nation work by adapting the decoding step of a baseline system using information from alternate systems. Generally, the re- quired information is very deep, such as a full decoding forest. In this paper, we de- scribe a method of cross-adaptation based system combination which only requires the final output from each alternate sys- tem. This is achieved by adding a dis- criminatively weighted n-gram confidence feature to our decoder. In order to opti- mize the confidence weight of each sys- tem, we present a novel procedure called non-linear Expected-BLEU optimization that can be used to optimize arbitrary non- linear parameters for any decoding fea- ture. We also describe a method for explic- itly creating an adapted system that is dis- similar from each particular input system, which we have found to be useful in com- bination. Although our new method does not outperform a state-of-the-art confusion network (CN) based combination system on its own, we obtain statistically signif- icant gains of 0.21-0.45 BLEU when the CA output is used as an additional system in CN combination. ∗This work was supported by DARPA/I2O Contract No. HR0011-06-C-0022 under the GALE program (Approved for Public Release, Distribution Unlimited). The views, opin- ions, and/or findings contained in this article/presentation are those of the author/presenter and should not be interpreted as representing the official views or policies, either expressed or implied, of the Defense Advanced Research Projects Agency or the Department of Defense. 1 Introduction In general terms, system combination is the task of using multiple machine translation (MT) sys- tems to produce an output that is better than any input system could produce by itself. The method for doing this is largely shaped by the informa- tion available: some require an actual MT decoder for each system (Li et al., 2009), some require a full forest of derivations (DeNero et al., 2010), and some only require the final output translations from each system (Rosti et al., 2009). Our re- search was done as part of a project where multi- ple sites develop their decoding systems indepen- dently, and then these outputs are combined to- gether to produce a “team” output. Because these systems widely differ in structure, the only infor- mation we have available for system combination is an n-best list of hypotheses. No source-side cor- respondences are available, and some systems are only capable of producing a 1-best list. Additionally, in this procedure, we assume that a strong baseline decoding system is available for each condition where system combination is per- formed. The general principle behind this proce- dure is to run a baseline decoding system with the output adapted …

Content truncated to 500 words