How NLP is turbocharging business intelligence

Natural language processing (NLP), business intelligence (BI) and analytics have evolved in parallel in recent years. NLP has shown potential to make BI data more accessible. But there is much work ahead to adapt NLP for use in this highly competitive area.  

Integrated NLP-enabled chatbots have become part of many BI-oriented systems along with search and query features. Long-established and upstart BI players alike are in a highly competitive environment, as data science and MLOps technologies pursue similar goals. But the competition has spurred innovation.

Systems such as Domo, Google Looker, Microsoft Power BI, Qlik Insight Advisor Chat, Tableau, SiSense Fusion and ThoughtSpot Everywhere have seen NLP updates. These have made data consumption considerably more convenient as business users retrieve data through natural language queries.

Make room for ChatGPT

There is more innovation in store across a broad product spectrum. As with other technology areas, the field stands to change even more dramatically as large language models like OpenAI’s ChatGPT come online. 

Signs of a ChatGPT boost to NLP efforts appeared last month as Microsoft said Power BI development capabilities based on this model will be available through Azure OpenAI Service. The company followed up this week with generative AI capabilities for Power Virtual Agents.

Also this week, SalesForce announced OpenAI integrations that bring “enterprise ChatGPT” to SalesForce proprietary AI models for a range of tooling, including auto-summarizations that could impact BI workflows.

Up from clunky

“Natural language querying and natural language explanation [are] pretty much routinely found in most every BI analytics product today,” Doug Henschen, analyst at Constellation Research, told VentureBeat. But that road, he said, has at times been rough.

When NLP enhancement originally came to BI systems, “it was kind of clunky,” Henschen said.  Enterprise developers had to work to curate the language that was common within the domain where the users of the data lived. That included identifying synonyms people might use to describe the same thing. Training and behind-the-scenes tools have gotten better at automating setups, he indicated.

“For the most part, BI products have gotten better at handling that,” Henschen said. “Now we’ve got this whole new wave of large language models and generative AI to look at … a whole other level of technology.” 

NLP-enhanced business intelligence

In most BI systems, data is accessed in a traditional way: logging into an application, generating the required report and filtering the insights through dashboards. But this often-lengthy process requires some technical proficiency. That means lower adoption rates.

That’s why companies often resort to hiring data scientists and data analysts to extract insights from their BI systems. But managers also look for wider adoption within the organization. An increasing number of global companies are now adopting NLP-driven business intelligence chatbots that can understand natural language and perform complex tasks related to BI. 

Business intelligence is transforming from reporting the news to predicting and prescribing relevant actions based on real-time data, according to Sarah O’Brien, VP of go-to-market analytics at ServiceNow.

“With the explosion of innovation in natural language processing, these actions can now be constructed in conversational language and pulled from a much wider array of sources,” O’Brien said. “Business intelligence provides the context — and NLP provides the content.”

Today’s chatbots can efficiently abstract data from various sources, such as existing LOB and CRM systems, and integrate with many third-party messaging applications like Skype for Business and Slack, according to Vidya Setlur, director of research at Tableau.

“With NLP-enabled chatbots and question-answering interfaces, visual analytical workflows are no longer tied to the traditional dashboard experience. People can ask questions in Slack to quickly get data insights,” Setlur told VentureBeat. 

That means users can obtain actionable insights through a conversational interface without having to access the BI application every time. Setlur believes this has changed how organizations think of growing their businesses and the types of expertise they hire.

“NLP-driven analytical experiences have democratized how people analyze data and glean insights — without using a sophisticated analytics tool or craft[ing] complex data queries,” added Setlur. 

This convenience plays a significant role in promoting an organization’s analytics culture. By applying NLP to BI tools, even non-technical personnel can independently analyze data rather than rely on IT specialists to generate complex reports.

“Employing NLP enables people who may not have the advanced skillset for sophisticated analysis to ask questions about their data in simple language. As people can get answers to questions from complex databases and large datasets quickly, organizations can make critical data-driven decisions more efficiently,” Setlur explained. 

She added that natural language interfaces (NLIs) that are both voice- and text-based can interpret these questions and provide intelligent answers about the data and insights involved.

Likewise, Ivelize Rocha Bernardo, head of data and applied science at enterprise VR platform Mesmerise, believes that such implementations have made data analytics more transparent, and aided in democratizing organizations’ data. 

“Stakeholders and executives can query the data through questions, and their BI platform could respond by providing relevant graphs. It is the next level of data analysis and unlocking the potential of business intelligence and analytics, where the teams can focus on more detailed follow-up questions and non-straightforward data insights,” Bernardo told VentureBeat. 

Automating your BI workflow with NLP

Organizations can automate many workflow tasks through natural language processing to get the relevant data.

“Search engines can leverage NLP algorithms to recommend relevant results based on previous search history behavior and user intent,” Tableau’s Setlur told VentureBeat. “These search engines have gotten sophisticated [at] answering fact-finding questions like ‘What’s the flight status? or ‘What’s the current score for the Golden State Warriors game?’.”

Predictive text generation and autocompletion have become ubiquitous, from our phones to document and email writing. The algorithms can even recommend words and phrases to suit the tone of the message.

Domains get specific

Collaboration in BI processes is important, according to Mesmerize’s Bernardo. She said that implementing NLP models is a collaboration between teams. It is essential to have the support of a specialist in a domain to refine workflow architectures and work together with the data team.

“There are many successful [use] cases of NLP being used to optimize workflows, and one of them is to analyze social media to identify trends or brand engagement. Another successful case is the chatbots that improve customer service by automating the process of answering frequently asked questions, unblocking employees to focus on tasks that require human interaction,” Bernardo said.

As a seasoned data scientist, Bernardo recommends that the best way to implement such NLP solutions is to work in phases, with small and very objective deliveries, measuring and tracking the results.

“My advice for effectively implementing these solutions is to start by defining the use cases the organization wants to optimize. Then, create long-term and short-term goals. The short-term goals should be associated with deliveries and allocated in a specific project phase. Finally, the team should revisit the long-term plan at the end of each phase to reevaluate and refine it,” Bernardo said. 

She also noted that one of the best practices for implementing NLP solutions is to focus on a specific domain area. “The broader the model’s domain is, the more chances of the NLP model giving not-so-accurate outcomes.”

Current challenges of implementing NLP in BI

One major challenge to implementing NLP in BI is that bias against certain groups or demographics may be found in NLP models. Another is that while NLP systems require vast amounts of data to function, collecting and using this data can raise serious privacy concerns. 

“We should focus on creating models that are fair and unbiased. Before storing any data, organizations need to consider the user benefits, why the data need to be stored, and act according to regulations and best practices to protect user data,” said Bernardo. 

NLP models can also become more complex, and understanding how they arrive at certain decisions can be difficult. Therefore, it is essential to focus on creating explainable models, i.e., making it easier to understand how the model arrived at a particular decision.

“Computer systems would need to be able to parse and interpret the many ways people ask questions about data, including domain-specific terms (e.g., the medical industry). Developing robust and reliable tools that can support BI organizations to analyze and glean insights while maintaining security continue to be issues that the field needs to improve upon further,” added Tableau’s Setlur.

What’s next for NLP in BI? 

While NLP has advanced, and can help solve a range of problems, language itself is still complicated and ambiguous.

According to Yashar Behzadi, CEO and founder of synthetic data platform Synthesis AI, generative AI approaches to NLP are still new, and a limited number of developers understand how to properly build and fine-tune the models. 

“Naive utilization of these approaches may lead to bias and inaccurate summarization. However, there are startups and more established companies creating enterprise versions of these systems to streamline the development of fine-tuned models, which should alleviate some of the current challenges,” said Behzadi. 

Behzadi predicts that in the coming years, enterprise-grade turnkey solutions will enable companies to fine-tune large language models on their data. He also said that model monitoring and feedback solutions will become commonplace to help assess in-the-wild performance and continually refine the underlying models.

“Traditional BI should be complemented [by] and not replaced with new NLP approaches for the next few years. The technology is maturing quickly, but core business-driven decisions should rely on tried-and-true BI approaches until confidence is established with new approaches,” added Behzadi.

For his part, Yaniv Makover, CEO and co-founder of AI copywriting platform Anyword, said that his company is observing an increasing need for “copy intelligence,” a BI approach to managing communications with the market across channels. Makover says that we might see BI integrations with generative AI in the near future. 

“With the emergence of LLMs, NLP algorithms can summarize much more accurately and understand the meaning of user-generated content without extracting an endless stream of examples, copied word for word. This will make query summarization much more powerful,” said Makover.

Understanding end users’ preferences and needs is a continuing imperative for NLP and business intelligence, as is the need to programmatically sort through masses of data.

It is important to note that LLMs like ChatGPT can also help address developer-side bottlenecks for BI.

Such generative AI can help out with software programming languages, not just the language of business, noted Doug Henschen.

“As the next generation of natural language, generative AI also generates code,” he said. “That’s huge.”

But he cites a caveat, which he calls “the human in the loop caution.”

“There have been so many stories and examples of someone trying something with the model, and it delivered gibberish. So, the more context that software makers can build in, the more reliable the result will be.”

Henschen said enterprises will continue to need human supervision and oversight. Still, he said, models like ChatGPT “promise to save a huge amount of time, and to get you started on generating language-generating code that is very close to what’s needed.” 

“But you have to make sure that it’s right.”