Next-generation natural language technologies: The deep learning agenda

Dynamic question asking

Ad hoc question asking and answering arguably represents the summit of natural language technologies’ potential, yet is markedly difficult to achieve without applying deep neural networks. According to Cognizer AI CTO Soundar Velu, in much the same way that traditional relational data warehouses require questions upfront for analytics, with conventional natural language technology approaches “you’re pretraining the system with questions, and it cannot answer any questions for you that are outside of this training.” This approach’s limitations are apparent: With unstructured data representing the majority of content available to organizations, it’s difficult to predefine—or even know—all the pertinent questions from all the relevant sources impacting business objectives, such as those for maximizing stock market trading, for example.

Nonetheless, with a confluence of AI approaches, including deep learning, natural language technologies, and machine-reasoning hypergraphs, organizations can succeed in expanding the arsenal of questions answered—and data sources involved—for dynamic, real-time responses of intelligent systems. This synthesis leverages deep learning’s predictions to infer features about language, which becomes the basis for codifying language’s intent into a comprehensive list of frames. “So far, there are about 1,500 frames in English,” Velu said. “All information coming from unstructured information falls into those frames.” By extracting facts from content according to frames and aligning these statements in an intricate hypergraph, users can ask spontaneously conceived questions about products, for example, from customer reviews, internal data sources, customer service interactions, and more. Specific advantages of this method include the following:

♦ Surpassing enterprise search: Instead of inputting keywords and getting a list of sources with which to search through information for actionable business insights, users can simply ask ad hoc questions of those sources and get realtime responses.

♦ Combining structured and unstructured data: Hypergraphs resolve the unstructured data problem by making unstructured data as accessible as structured data is.

♦ Contextual comprehension: Deep learning powers an understanding of what language, words, and phrases mean in context—as opposed to simply extracting predefined concepts from them. This approach comprehends polymorphisms, colloquialisms, and other nuances of language to function “like a real human assistant who’s reading all of your information and is basically right beside you to answer all the questions for you in a human form,” Velu maintained.

Perfecting process automation

Deep learning’s aforementioned impact on ad hoc question-answering hints at the bifurcation of its enterprise benefits for a broadening array of natural language technology use cases. The most immediate gain is the contemporary simplicity in leveraging these technologies because of deep learning’s prediction accuracy. According to Indico CEO Tom Wilde, “Our big moment of inflection as an industry right now is really driven by deep learning’s ability to learn how to solve a problem through examples.” This sentiment is particularly applicable to process automation. By simply giving deep learning models enough annotated training data—the examples Wilde mentioned—they can help complete almost any process, from extracting relevant information from emails to categorizing and redacting information from documents for regulatory compliance. Techniques such as transfer learning decrease the amount of training data necessary to actualize this benefit, making deep learning approaches far easier and quicker to use than other process automation methods—especially for unstructured content.

Besides the drastic reduction in time, labor, and coding skills that process automation allows when employing deep neural networks at scale, the greater advantage of leveraging deep learning for these tasks is the relatively newfound ability to finish them. When given adequate examples of the various steps required of processes and their business relevance, deep learning systems can not only successfully process these tasks, but also issue the decisions for which they’re needed.

Deep learning solutions can identify and extract a variety of information from heterogeneous sources for mortgage applications, for example. Beyond bank forms, common sources might include “tax information from tax authorities,” said Wilde. In addition, he noted, “You might have title and deed information from counties; you might have income verification from the borrower’s bank statements and tax filings.” Deep learning platforms can assemble and process this information to decide whether or not to extend a mortgage to the applicant, and if so, for what amount, as well as the deterministic risks for doing so. This technology empowers mortgage officers in a supervisory capacity to oversee the results and intervene as needed, substantially expanding the scale at which these tasks can be completed. Moreover, deep learning’s proclivity for issuing decisions about automated processes is horizontally applicable.