1 month ago
December 18, 2020
Why Enterprises Fail to Integrate AI and What to Do About It
Konstantin Savenkov sees the market as past the tipping point for widespread enterprise adoption of machine translation (MT). At SlatorCon Remote December 2020, he offered his advice to businesses: “If you can’t stop it, lead it.”
Savenkov is Co-founder and CEO of Intento, which helps clients identify and deploy the most appropriate and impactful AI services according to their needs. “For now, we’re focusing our efforts on the translation, mainly because here we see the massive gaps between AI capabilities and enterprise adoption,” he said.
According to Savenkov, enterprise integration of AI often fails, but not because the quality is lacking. “The very fact that AI is being treated as software accounts for the vast majority of those failed initiatives,” Savenkov said, adding that worthwhile MT integration requires processes, tools, and expertise different from those used for software. MT, in particular, demands continuous improvement rather than hands-off maintenance.
The good news is that MT has advanced to the point that a business can integrate an enterprise solution with real-time MT and immediately see a return on investment. The problem lies with the traditional pricing model for MT — a flat-rate discount that does not change depending on quality. Set the discount too low and the buyer cannot afford to translate all their content; a discount that is too high hurts post-editing linguists (and, ultimately, buyers as the translation quality drops).
“For now, we’re focusing our efforts on the translation, mainly because here we see the massive gaps between AI capabilities and enterprise adoption”
Popular alternative pricing methods are not foolproof, either. Savenkov advised against using cross metrics such as edit distance to measure effort because costs can become unpredictable and uncorrelated with actual effort. Similarly, running a test project at full rate to estimate a custom discount for a specific client cannot account for unpredictable effort.
Savenkov instead advocated service-level agreements (SLAs). Looking at a client’s content, a language service provider (LSP) estimates what percentage can be handled by MT, and what percentage of MT-translated segments will be perfect. The remaining content will need different editing tasks requiring a range of effort, and the LSP can then estimate the potential translation and post-editing effort and associated discount.
“The very fact that AI is being treated as software accounts for the vast majority of those failed initiatives”
Unlike a project-specific flat discount, the SLA approach allows LSPs to provide the full MT discount without contingency. This becomes important as the share of perfect MT segments reaches up to 70%, as providing such discounts without the SLA imposes too much risk.
“This can be expanded, as the SLA fulfillment depends not only on machine translation quality, but also on source content quality,” Savenkov explained. “A similar approach could then be applied to cover TM [translation memory] and repetition edits.”
Looking ahead, Savenkov said, “some of the largest translation vendors seem to be cautiously optimistic, while smaller LSPs with shorter supply chains and smaller contracts at risk are ready to go.”