Learning from Post-Editing: Online Model Adaptation for Statistical Machine Translation

Abstract

Using machine translation output as a starting point for human translation has become an increasingly common application of MT. We propose and evaluate three computationally efficient online methods for updating statistical MT systems in a scenario where post-edited MT output is constantly being returned to the system: (1) adding new rules to the translation model from the post-edited content, (2) updating a Bayesian language model of the target language that is used by the MT system, and (3) updating the MT system's discriminative parameters with a MIRA step. Individually, these techniques can substantially improve MT quality, even over strong baselines. Moreover, we see super-additive improvements when all three techniques are used in tandem.

Online Automatic Post-editing for MT in a Multi-Domain Translation Environment

“…These systems, however, are not able to leverage the feedback of the post-editors in an online translation scenario. The capability to evolve by learning from human feedback has been addressed by several online translation systems but mainly focusing on the MT task (Hardt and Elming, 2010;Bertoldi et al, 2013;Mathur et al, 2013;Simard and Foster, 2013;Ortiz-Martınez and Casacuberta, 2014;Denkowski et al, 2014;Wuebker et al, 2015). From these several online MT systems, we discuss the two that have been used also for the APE task.…”

“…Compared to the suffix arrays used to implement MT dynamic models (Germann, 2014;Denkowski et al, 2014), in which the whole sentence pairs are stored, our technique needs to save more information (all the translation options) but: i) the amount of data in APE is much less that in MT so it can be easily managed by ad hoc solutions, and ii) it allows us to collect global information at translation option level that can result in useful additional features for the model. This last aspect is explored in the next section, in which the reliability of the translation options is measured by looking at the behavior of the post-editors.…”

Response-based Learning for Grounded Machine Translation

“…Since retraining the SMT model after each interaction is too costly, online adaptation after each interaction has become the learning protocol of choice for CAT. Online learning has been applied in generative SMT, e.g., using incremental versions of the EM algorithm (Ortiz-Martínez et al, 2010;Hardt and Elming, 2010), or in discriminative SMT, e.g., using perceptron-type algorithms (Cesa-Bianchi et al, 2008;Martínez-Gómez et al, 2012;Wäschle et al, 2013;Denkowski et al, 2014). In a similar way to deploying human feedback, extrinsic loss functions have been used to provide learning signals for SMT.…”

Online Automatic Post-Editing across Domains

“…In the APE context, the input is a machine-translated segment (optionally with its corresponding source segment), which is processed by the online APE system to fix errors, and then verified by the post-editors. Several online translation systems have been proposed over the years (Hardt and Elming, 2010;Bertoldi et al, 2013;Mathur et al, 2013;Simard and Foster, 2013;Ortiz-Martïnez and Casacuberta, 2014;Denkowski et al, 2014;Wuebker et al, 2015).…”