Trait-Based Hypothesis Selection For Machine Translation

2012 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 528–532, Montr´eal, Canada, June 3-8, 2012. c⃝2012 Association for Computational Linguistics Trait-Based Hypothesis Selection For Machine Translation Jacob Devlin and Spyros Matsoukas Raytheon BBN Technologies, 10 Moulton St, Cambridge, MA 02138, USA {jdevlin,smatsouk}@bbn.com Abstract In the area of machine translation (MT) sys- tem combination, previous work on generat- ing input hypotheses has focused on varying a core aspect of the MT system, such as the de- coding algorithm or alignment algorithm. In this paper, we propose a new method for gen- erating diverse hypotheses from a single MT system using traits. These traits are simple properties of the MT output such as “aver- age output length” and “average rule length.” Our method is designed to select hypotheses which vary in trait value but do not signif- icantly degrade in BLEU score. These hy- potheses can be combined using standard sys- tem combination techniques to produce a 1.2- 1.5 BLEU gain on the Arabic-English NIST MT06/MT08 translation task. 1 Introduction In Machine Translation (MT), the output from mul- tiple decoding systems can be used to create a new output which is better than any single input system, using a procedure known as system combination. Normally, the input systems are generated by varying some important aspect of the MT system, such as the alignment algorithm (Xu and Rosti, This work was supported by DARPA/I2O Contract No. HR0011-12-C-0014 under the BOLT program (Approved for Public Release, Distribution Unlimited). The views, opinions, and/or findings contained in this article are those of the au- thor 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 De- fense. 2010) or tokenization algorithm (de Gispert et al., 2009). Unfortunately, creating novel algorithms to perform some important aspect of MT decoding is obviously quite challenging. Thus, it is difficult to increase the number of input systems in a meaning- ful way. In this paper, we show it is possible to create diverse input hypotheses for combination without making any algorithmic changes. Instead, we use traits, which are very simple attributes of the MT output, such as “output length” and “average rule length.” Our basic procedure is to intelligently se- lect hypotheses from our decoding forest which vary in trait value, but have minimal BLEU degradation compared to our baseline. We then combine these to produce a substantial gain. Note that all of the hypotheses are generated from a single decode of a single input system. Additionally, our method is completely compati- ble with multi-system combination, since our proce- dure can be applied to each input system, and then these systems can be combined as normal. Methods for automatically creating diverse hy- potheses from a single system have been explored in speech recognition (Siohan et al., 2005), but we know of no analogous work applied to machine translation. Our procedure does share some surface similarities …

Content truncated to 500 words