Artificial intelligence startup nRoad Inc. today announced the launch of its Convus platform, which is focused on helping financial services firms obtain insights from unstructured data.

The company has declared it’s waging “war on documents,” with the aim of helping organizations overcome the difficulty of extracting insights from unstructured data, which includes everything from text-based documents, images, video and audio streams. As nRoad explains, this kind of data makes up around 80% to 90% of all the information that big financial firms collect.

With so much of this “deep, dark data” clogging their servers, financial services firms find it incredibly difficult to tap into and obtain insights that might help them achieve their business goals faster, nRoad said. The problem is that most companies rely on outdated robotic process automation technologies that require extensive human interaction, as well as one-size-fits-all natural language processing algorithms that are quickly overwhelmed by the sheer amount of information.

NRoad reckons those systems simply aren’t viable. It says they’re so slow that organizations collect unstructured data faster than they’re able to process it.

The Convus platform, according to the company, leverages a combination of dedicated neural networks, vision AI and text-based techniques to extract information from any unstructured data source. This information is contextualized and comprehended using Convus’ proprietary natural language understanding algorithms, which leverage knowledge graphs and domain knowledge. Finally, Convus’ consumption engine structures that information into a standard machine readable format where it can be analyzed for insights.

NRoad Chief Executive Aashish Mehta told SiliconANGLE his company has carefully analyzed and identified some of the key challenges in squeezing insights from unstructured data. Those challenges include a lack of domain context and a lack of sophistication, he said.

“Current NLP models, while being robust, fall short on ambiguity resolution,” he said. “Major solution providers overlook the years of institutional and domain knowledge that was built in an enterprise, but the need to incorporate that knowledge is critical. Domain-specific concepts, relationships and business context is required to solve for an enterprise problem statement.”

As for the lack of sophistication, Mehta said existing natural language processing models are heavily template-driven and only support static table and database information formats. “Enterprises are littered with multiple and complex input sources such as long-form documents, multi-dimensional table data, and multilingual documents,” he explained. “Our evolution to a structuring of complex sentences and narratives upon comprehension of complex input sources is highly distinctive.”

It’s distinctive enough that nRoad’s platform has multiple use cases, Mehta said. For example, it can reduce the time to process financial statements by up to 60%. It can also help financial firms to reconcile corporate actions such as the payment of dividends, stock splits and mergers and acquisitions, traditionally a very human capital-intensive task with more than a half-million such actions in the U.S. every year.

Other use cases include mitigating fraud risk during loan underwriting, scanning information from legal and regulatory filings for investment opportunities and other insights, and intelligent contract verification, which includes detecting contract violations.

NRoad makes some big claims, saying Convus has shown itself to be superior to services such as Amazon Web Services Inc.’s Amazon Textract and Google Cloud’s Document AI. Mehta said the problem is that those platforms struggle to extract insights from multidimensional table structures that consist of multiple headers, merged columns or rows.

“While Amazon’s Textract and Google’s Document AI are strong platforms, allowing businesses to solve or automate part of their business processes, we have observed that these platforms are built to address approximately 60% of a specific business problem, leaving the enterprise technology and product teams to solve for the remaining 40%,” Mehta said.

The CEO explained that nRoad has solved these challenges by including thousands of such documents to train its platform, to the extent that it surpasses its cloud rivals in many areas.

Andy Thurai of Constellation Research Inc. agreed, telling SiliconANGLE that making meaning from unstructured data has always been a major challenge for enterprises, with most tools focused on structured information and knowledged.

“The newer sets of unstructured data — such as videos, images, text and the like — need to be properly classified, labeled, categorized, cleansed, and de-biased before the AI systems can create proper models out of it,” Thurai said. “Most existing RPA and OCR solutions fall flat as they are mostly mechanical systems with no true AI behind them. So domain-specific solutions to target specific industries with built-in contextual knowledge can add a lot of value to these mechanical systems.”

“If unstructured data is properly organized, the AI and machine learning models can be more accurate and can create predictable and actionable insights that can help solve enterprise business problems,” he added.

Another key advantage of the Convus platform is it’s easy to implement. Thanks to its microservices-based architecture, Mehta said it can be quickly integrated with existing information technology infrastructures, with minimal training samples required to get it up and running.

Image: nRoad