If you have ever tried to interact with an online chatbot only to find yourself calling the 1-800 number to reach a live service agent, it can be easy to dismiss the idea that AI is already a developed fixture of everyday life. Nevertheless, with a growing number of smart devices and 5G edge nodes in our daily midst, as well as the relentless push to automate everyday services, AI has indeed managed to become ubiquitous despite its current limitations. On the one hand, this appears to suggest that AI is a surprisingly welcome addition to everyday life, whatever latent concerns may remain about its risks. On the other hand, the question remains of how, if at all, we can expect smart technologies to make the leap to human or near-human capabilities in reasoning and decision making.
Now, reports of recent advances at Google suggest that the next major step for AI may be a surprisingly human one: learning how to tune out distractions. In a formal study, researchers at Google demonstrated that their AttentionAgent software was able to more efficiently navigate a simulated 3D environment from the classic computer game Doom when equipped with a mechanism known as a self-attention bottleneck, which mimics the human tendency not to notice irrelevant details. Far from a unidimensional challenge of improving its recognition on sight, it appears that improving AI’s ability to navigate new visual environments requires endowing it with the ability to refuse to look at certain details, essentially by installing it with an experience bias.
What lessons AttentionAgent holds for language-related AI fields such as NLP (natural language processing) is not immediately clear, but it does demonstrate at least one way in which the work of making AI more powerful can be done by making it more human-like. Moreover, eliminating AI’s eye for distractions in human speech may in fact be the crucial difference between a chatbot that stops chatting for reasons it cannot explain and a chatbot that answers the question you were just about to ask it. For reasons that are of no advantage to a computer, people intrinsically adopt speech patterns and modes of expression that elaborate on their basic intentions in ways that can easily confuse native speakers of the same language, not to mention speakers from other backgrounds. In localization, one key marker of translation quality is the translator’s ability to navigate exactly these challenges, specifically by distilling the essential meaning of a source text and infusing it with expressiveness in the target language. Ultimately, it may be that the insights that separate professional linguists from others are precisely the same capabilities AI will need in order to advance from its current threshold.
As NLP and related capabilities become more central to technical platforms, language and communications services will be vital to ensuring not only the performance of these products but also their availability worldwide. With a global network of expert linguists and subject matter experts in 250+ languages, CSOFT International works closely with providers of enterprise AI solutions to ensure the quality and performance of their platforms in the languages their users demand.