Invited Speakers

Geza Kovacs, Lilt

Predictive Translation Memory in the Wild - A Study of Interactive Machine Translation Use on Lilt

Predictive Translation Memory is a mixed-initiative approach to translation where an interactive MT system suggests prefix-constrained completions based on the translation the translator has already typed. This technique has been shown to have promising results in small-scale experiments, and can have benefits over post-editing. However, Predictive Translation Memory has yet to be studied at scale on actual commercial translation systems. This talk presents a series of in-the-wild experiments on Lilt, a commercial language service provider and web-based translation platform that implements the Predictive Translation Memory paradigm. We use A/B testing and analyze data logged from translator activity to help us quantify how professional translators interact with Predictive Translation Memory. We will first present a breakdown of what activities translators spend their time on when using Predictive Translation Memory. We will then analyze how much translators make use of our machine translation suggestions, and how much typing effort is saved by Predictive Translation Memory. We then discuss the latency-quality tradeoff inherent in Predictive Translation Memory, and analyze how fast the interactive machine translation system needs to be, in terms of latency, to be useful to translators. Finally, we investigate effects of Predictive Translation Memory on the quality of resulting translations, and the overall time spent by translators.

Rhett Whitaker, Amplexor

The Machine is Blind: Bottom-Up Feedback on the Impact of MT on Human Translation Performance

There is no shortage of research and experimental rigor with regard to cost-effectiveness, accuracy, and other technical considerations in the context of machine translation (MT). What is often absent from these discussions, however, are the unfiltered opinions and experiences of the human translators who work with – and are increasingly at risk of being displaced by – machine translation technology. An understanding of these attitudes remains elusive when gathering hard data and evaluating KPIs, and there are real-world implications for translation production quality resulting from that blind spot. This kind of anecdotal information is also difficult to gather and act upon within an organization in the absence of a dedicated position – specifically tasked with reconciling business requirements with the attitudes and concerns of translators, and ideally with a keen understanding of the motivations of management and vendors alike.

As with any tool, MT is better suited to certain applications than to others. Greater integration with customer terminology and other linguistic or technical resources generally yields better results (i.e. higher quality translations). Likewise, attitudes and outcomes vary on the part of the translators who work with MT on a progressively broader basis; those attitudes and outcomes have potential to aid in decision-making within companies that choose to implement MT. There are ethical as well as financial considerations involved in this arena, and while at times they appear to be at odds with one another at first glance, closer inspection can reveal that they are often aligned. Understanding the short- and longterm effects of MT implementation on the human translator pool may be crucial to continued success in the language services industry, and – along with similar developments in other industries in which automation and digitization are displacing human labor – could have significant societal and economic repercussions on a macro scale.

The concerns and opinions that form the basis of the following discussion are sourced primarily from communications with translators in a live production setting – that is, ad hoc and spontaneous feedback during ongoing translation projects in which MT was utilized and for which the relevant translators were expected to perform MT post-editing services. After a discussion of these acute concerns, we will take a broader view of where things are heading for human translators in the face of steadily improving MT quality and availability, in addition to the potentially hidden effects that translators’ attitudes toward MT may have for language service providers in the future.