So let’s go back to the basics and explain what AI truly is, and then dive into some data to define the actual depth of AI deployments in the language services industry.
Levels of automation
Automation is about using software to define and execute a series of actions that run when instructed to do so or when a trigger or condition dictates (Figure 1).
Rule-based systems deliver the most basic automation. That’s what you use when you process files in a rule-based machine translation (MT) engine that replaces each term with its translation and applies grammar rules to it. Similarly, it’s what you use when a translation management system (TMS) applies a preconfigured project plan for a specific client relying on a predetermined price sheet, translation memory and translator list. When the system encounters an undefined situation, such as a new customer, or new language for a client, the preconfigured workflow stops and calls for human intervention.
Expert systems go a step beyond preloading data for each scenario. Instead, they apply complex rules and conditions to quote rates and turnaround times, choose workflows and select vendors based on project specifications. In this ideally touchless environment, human intervention occurs when the system flags a need for it. For example, the software may discover a shortage of vendor options that can handle the work in the assigned turnaround time.
Generic artificial intelligence is a more advanced form of automation. It refers to technologies that learn from data to perform tasks that would otherwise require human intelligence, but that require instruction and feedback from humans. In these “supervised” systems, humans tell the machine how to learn from the data. For example, an AI-driven system can predict timelines based on actual translator performance for specific types of content. The software flags the odds of a translation passing a preset quality threshold based on analysis of events such as whether the linguist opened a provided glossary.
Machine learning is a further subset of artificial intelligence in which the so-called “unsupervised” systems can learn by analyzing the data without being explicitly told how to. What happens is that you use automation to create further automation.
Deep learning is yet another subfield that uses neural networks that are modeled on brains — for example, today’s neural networks are still most closely modeled on the visual ganglia of insect brains. This is the technology behind neural machine translation.
What constitutes AI is constantly changing. 30 years ago, rules-based systems were considered AI, but today they are not. 20 years ago, expert systems were the rage in AI, but are now considered simply software, not AI. And while statistical MT was central to AI ten years ago, some people already no longer consider it part of AI.
Automation through MT
In a survey of language service providers, CSA Research found that 51% of LSPs in our dataset already use some form of MT software, but fully 80% of these companies have tried neural machine translation, a deep-learning-driven form of machine translation. We applied our LSP Metrix capability-competency model to the use of machine translation and found that MT adoption more or less follows the natural maturity curve of LSPs.
LSP maturity equates to a high level of MT capability. In our analysis, we discovered that 93% of the most mature LSPs are MT-capable, meaning that they have tried out the technology. Those that haven’t experimented with MT are interpreting-centric companies, where automated interpreting software for spoken language is not as market-ready as it is for written language. At the opposite end of the spectrum, when we apply our Metrix model to the least mature LSPs, we find that only about one-third of companies at Stage 0 have some MT capability in-house.
MT capability may not mean LSPs are using it in production. Deployment at the project level remains minimal for most. We found that 57% of respondents using neural MT use it on less than one-tenth of their projects. Most providers don’t trust it sufficiently, feel it hurts more than it helps on jobs, struggle to find a systematic way of integrating the technology with translation processes, or face clients that don’t want it used on their projects.
Systematic users are few and far between. A meager 2% of LSPs with neural MT process almost all of their translation work through that technology. Such companies remain pioneers with business models fully centered around machine translation capabilities and their customers tend to have scalability needs that trigger full adoption.
Automation in TMS
We also examined automation in translation management systems, finding that 80% of our survey respondents declared they’re using a TMS. However, we again found that this high number is no guarantee of deployed automation. Accordingly, we analyzed responses about the level of automation: 1) 54% of respondents say they have “partial” automation capability — that is, the TMS can handle most tasks but requires human project managers for some decisions or approvals; 2) just 28% of our sample claim “full” automation capability, where no humans are involved in this “lights-out” model of project management.
We asked respondents who use some automation what type they used. We found that 86% have built systems based on older rule-based software, another 33% use more advanced but still dated technology, and just 9% claim any artificial intelligence solutions based on machine learning. Thus, most automation on the project management side tends to be old-school, traditional approaches that don’t leverage the data-based knowledge that most LSPs accumulate through their TMS workflows (Figure 2).
The bottom line
Even if neural MT has gained a lot of ground at LSPs, only about half of providers have the capability to use it, but even those do so on only a negligible proportion of their projects when looking at the big picture. However, adoption rates are rising and deployment scenarios will increase over time – even if use cases aren’t in line with the level of talk in the industry.
And on the project management side, providers tend to blur the lines by calling AI things that may support just basic levels of automation. Yet either way, automation is increasing. Technology vendors are bound to provide more AI-driven automation in their commercial products. As a result, it will trigger the democratization of access to the technology and enable even smaller LSPs that don’t have the ability to invest in data scientists to benefit from advanced levels of automation.