psdesign1 - Fotolia
When Siri and Alexa fail to understand their user, it may be funny or at worst annoying. In an enterprise setting, a failure of natural language processing application to understand queries could have serious business consequences.
Natural language processing is showing up in an increasing variety of enterprise applications, including analytics tools. But the importance of their workloads means the stakes are high when it comes to making sure applications understand their users. This is the key to unlocking the full benefits of natural language processing.
What is natural language processing?
Natural language processing is a program's ability to understand human language. It used to be that understanding human speech and responding appropriately was the gold standard for artificial intelligence. Today, this capability is available even in low-end Android phones. Virtual assistants can look up information for you on the web, inform you about the weather, remind you of your appointments, send text messages and even tell jokes.
To do all this, they listen to what you say or read what you type and figure out its meaning.
Back in the early days of computers, developers attempted to code grammar rules and vocabulary, but these efforts fell short. In the 1990s and 2000s, developers started applying statistical methods, and the field finally started moving forward. In the last decade, with the use of deep neural networks and other cutting-edge techniques, natural language processing came into its own.
Natural language processing enterprise applications
Natural language processing is found everywhere in the enterprise. Employees use virtual assistants on their phones or type plain English queries into online search engines, and chatbots handle basic customer service requests. But the technology has been slower to move into enterprise applications.
In analytics, natural language processing shows up in two main areas.
The first is in processing data to be analyzed. For example, applications pulling information out of contracts, emails or social media posts will use natural language processing to extract structured meaning from unstructured data.
Natural language processing benefits also include extracting meaning from structured data when the underlying structure is a mess.
For example, with the new European and California privacy laws, companies have to be able to quickly find all the information they've collected about a particular individual, said Jeremy Wortz, senior manager of AI and machine learning at West Monroe Partners. That information can be scattered across multiple systems in a variety of formats.
For a large bank, the challenge of cleaning up that data to make it usable could be an insurmountable task.
Wortz said that his team recently worked on just such a project for a large bank, building a natural language processing capability to find similarities between different systems and tie it all together.
"Now, instead of going through and manually searching various systems, you just type that customer's name into the search bar," he said.
The other enterprise use of natural language processing is in the queries themselves. Employees or customers simply type in their questions instead of writing SQL queries or filling out forms.
Some larger analytics vendors are beginning to include natural language options as part of their platforms.
"Microsoft Power BI has a natural language query to create visualizations," said Alison Thaung Smith, chief scientist and data scientist senior manager at Booz Allen Hamilton. "You can use it for basic queries, like, 'What is the total sales revenue for 2018?'"
Another analytics vendor that's built natural language processing into its platform as a differentiator is ThoughtSpot, Smith said. "They have a search that allows you to ask a general question."
But even with these platforms, users need to learn how to word their queries in order to get useful answers.
"It might be a little less steep of a learning curve than a querying language, but still something you need to worry about," Smith said.
Advanced users who are comfortable with writing SQL queries with different joins, aren't going to use natural language tools. And even with simple queries, users have to be precise with their language, Smith added.
"You have to think about how you word your question so as not to be super open-ended," she said. And the system won't automatically find connections it needs. "If you didn't explicitly create relationships between tables, it won't be captured."
Natural language processing benefits in enterprises
Natural language processing of unstructured files frees human beings from the drudgery of reading documents and manually typing in data, enabling analytics to be performed in domains that were previously impractical to address.
Improved analytics adoption. Appling natural language to query analytics systems can lower the barriers of analytics adoption and make the systems usable in a wider variety of settings, such as on manufacturing floors and with mobile devices.
But natural language processing isn't perfect. Computers are still far from being able to fully decipher what humans are trying to say -- though, humans struggle to understand each other as well.
"It takes time for that type of algorithm and processing capability to improve," said Josh Axelrod, partner/principal in advisory risk cybersecurity practice at EY. "There will be a natural growth curve as we go through adoption."
Faster response to customers. Human language is also notoriously imprecise, and the same word can mean vastly different things in different contexts -- or even when spoken in different tones of voice. As a result, natural language queries work best in narrowly defined domains.
For example, if a utility company is getting a large volume of calls related to power outages, natural language processing benefits the company in reaching customers.
"NLP can get through what customers are saying and improve operations, respond quicker, deploy trucks and get power back on," Axelrod said.
That's why in tightly focused areas such as customer queries in the energy sector, companies are investing half a billion dollars in customer service and information, he said.
"And the self-service part of it has seen dramatic improvements," Axelrod said -- though there's still room for improvement. "Your goal is to take any language and any culture and understand terms that are both commonly used and uncommonly used, and by customers who don't know what they're asking for."
Improved contextual understanding. It also helps to give humans the opportunity to verify that the computer understood them.
For example, instead of saying "The answer is 15," the system can provide more context, such as "Your top sales team increased orders by 15% last month compared to the previous month." Or if the results are in the form of a chart, the labels would be sufficiently descriptive that the user would know immediately if the results are what they're looking for.
Challenges of natural language processing
Outside of customer-facing use cases, there's not much adoption of natural language processing to analyze data in enterprises, said Dave Kuder, principal of cognitive insights and engagement in the U.S. at Deloitte.
Most often, natural language analytics show up when customers are looking for information about their bills or want to find out the status of an order.
Part of the reason for the lack of adoption is that natural language processing requires a lot of training data, and there's a lot more training data to be had in customers asking questions than in a handful of business executives asking questions, Kuder said.
Building natural language processing capabilities into a system is also expensive, and a marginal improvement in productivity for an employee isn't much of a business case, while on the customer service side, natural language processing can help reduce the need for human call center agents answering routine questions.
The one exception is situations where an employee is not able to physically access a keyboard to enter a query, such as when someone needs hands-free access to information, for example, on an assembly line.
"Natural language processing can be quicker and more expedient," Kuder said. "But, ultimately, keyboards are good at being able to do things quickly."
As people get more comfortable with virtual assistants like Alexa, and the technology drops in price, it might become more common in enterprise applications.
"There's potential there," Kuder said.