Transformation-Based Correction of Rule-Based MT

Transformation-based correction of rule-based MT Jakob Elming Center for Computational Modelling of Language Copenhagen Business School Denmark je.id cbs.dk Abstract We present a pilot study for using transformation-based learning for au- tomatic correction of rule-based machine translation. Correction rules are learned based on a parallel corpus of machine translations from a com- mercial machine translation system and a human-corrected version of these translations. The correction rules exploit information on word forms and part of speech. The experiment results in a relative increase in translation quality of 4.6% measured using the BLEU metric. 1 Introduction The situation in commercial MT today re- veals that most systems are based on rules crafted by language experts. The ideas behind these rule-based systems, however, leave the language experts with bad odds. Not only do they have to give a full descrip- tion of two languages, but they are also ex- pected to specify how to get from one to the other. The experiences from previous at- tempts, e.g. the Eurotra project, has shown that it is in fact a very difficult task, which leaves many uncovered areas. One of the great advantages of rule-based MT as opposed to e.g. statistical MT, is its transparency. Basing the translation on rules makes the output — and thereby the errors — very consistent. If the system makes an error once, it will most likely re- peat this type of error in future translations. This aspect of rule-based MT makes it in- teresting from a machine learning point of view. The ability to predict recurring errors and their corrections is a simple method for improving an MT system; either by means of a post-processing module, or by correct- ing the rules of the system. In the present experiment, transformation-based learning (TBL) utilized as learning algorithm for extractingis rules to correct MT output by means of a post-processing module. The algorithm learns from a parallel corpus of MT output and human-corrected versions of this out- put. The machine translations are provided by a commercial MT system, PaTrans, which is based on Eurotra. In order to add a level of general linguistic knowledge, the data is annotated with part of speech tags. This method results in a substantial improvement of the translation quality of the MT system. For this pilot study, we only look at cor- rections involving substitution. 2 Method The goal of the present experiment is to learn the process of correcting recurring er- rors made by the MT system. This goal is achieved by first word aligning the machine translation to its human-corrected version. This alignment will to a large part consist of links between the same word forms1 at the same sentence positions as exemplified by figure 1. Two sentences only differ where corrections have been made. Secondly, based on a given word align- ment, a relevant correction is identified as a recurring link between differences — either in word form or sentence position — in a 1 ’word form’ here includes any token of a …

Content truncated to 500 words