Many apps and programs claim to be able to understand you and are at least capable of engaging in superficial interactions. Spend long enough talking to one of these programs, however, and you’ll no doubt see the hallmarks of imperfectly reproduced natural language. That technology has not yet caught up to the complexities of natural language is no surprise.
According to Gamalon founder Ben Vigoda, a fifteen-word sentence can express something like ten billion ideas. In his talk at ODSC East 2018, Vigoda identifies the necessity of building better, new era of NLP functions to accurately process free-form text: business already receive far more unstructured language-data than they can process, and the amount is only going to keep growing. As Vigoda puts it, “we need an interface between us and that data volume of messages…no way for institutions to handle all their incoming natural language data.”
Beyond Deep Learning
Current training models rely on Pavlovian conditioning, but Vigoda notes that this kind of training is not analogous to human language-learning: “You can’t just teach a system complex ideas by giving it stimulus, showing it a response, and training it to give that response.”
Relying on deep-learning systems that utilize this kind of training results in systems that are capable only of a shallow level of interpretation. Even when the level of interface requires nothing more than the recognition of a few keywords, current models typically rely on time-consuming, labor-intensive training processes, with an end result that is, per Vigoda, never more than 65% accurate or capable of responding to questions outside of a limited amount of prompts. The typical idea-trees used by voice recognition systems are easily confused and are prone to repeat themselves or return to menus when confronted with an unfamiliar idea.
A Model for the New Era of NLP
In creating his product at Gamelon, Vigoda decided to move away from the neural networks that had dominated the sphere of NLP. Instead, he and his team began experimenting with machine learning. Their system relies on Occam’s razor to demystify unstructured data and relies on only a few training models built by their team of data scientists. The results have been revelatory: “One person teaches the model, not thousands; one processor, not a server farm; a handful of training examples per object instead of thousands, hours to train instead of days or months.” Essentially, the system “backsolves” from the unstructured data and fits it to the given model.
Going forward, Vigoda hopes that language processing models will work collaboratively, sharing their idea trees with one another to build more accurate and realistic models for language processing. The billions of possible permutations available to humans are too much for any one company or team to account for in a single model, but massive amounts of unstructured data will make automation of language-processing tasks a necessity.