Tech Talk: How Emerging NLP Models Will Transform the Enterprise


CF Su, VP machine learning, Hyperscience, joins Neha Pradhan Kulkarni to outline the implications of natural language processing in enterprise. Su discusses why sentient AI models are becoming a phenomenon today.

In this edition of Tech Talk, Su explains how natural language processing will transform back-office processes and make it easier for machines and humans to communicate. He also shares his two cents on how the AI community can make progress toward securing an unbiased regulatory framework.

Key Takeaways on How Emerging NLP Models Will Transform the Enterprise:

  • Usher in a future where machines and humans can communicate via images.
  • Help machines further extract information from human documents and with little error.
  • Be able to extract data from video, audio, and even photographs.

Here are the edited excerpts from our exclusive interview with CF Su, VP machine learning, Hyperscience:

SWNI: The AI industry was abuzz after a Google engineer’s claim that Google’s LaMDA language model has acquired a soul. So, can AI become sentient?

CF Su: Despite the allurement of AI becoming human-like, or coming alive, sentient AI is more of a fantasy than fact. For decades, society has mused about the possibility of humanoid robots and technology overcoming humans. While AI has mastered complex games and various applications of intelligence—like pokerOpens a new window or AlphaGoOpens a new window —this has ultimately been achieved at the behest of humans. 

Google’s language model happened to be excellent at mimicking human speech but mimicking human speech does not make something living. 

“The LaMDA language model is exceptional, but the full, unedited transcript of its communications is still less impressive knowing it often trails off or responds randomly or inconsequentially to questions. This showcases a current limitation in the language model itself.”

Fundamentally, the large language model represents a distillation and compressed representation of a vast corpora of human knowledge in the training set. The model is built based on statistical correlation derived from the training data. It is very far from being sentient.

SWNI: Why are sentient AI models becoming a phenomenon today? How would you describe this belief to be ‘more problematic or solution providing’ for the enterprise?

CF Su: There is significant progress in the development of the Large Language Model (LLM) in recent years, including the GPT-3 from Open AI and the largest open-source LLM  BLOOMOpens a new window .  These LLMs became so good that they could explain jokes and answer simple questions with a natural language response that appeared to be coherent and logical. Their performance on some natural language tasks were unthinkable just a couple of years ago. Naturally the progress generates lots of excitement in the public and raises misguided questions whether the statistical models underlying LLM are sentient because of the impressive text outputs. While sentient AI isn’t a current reality, there’s no doubt that rapid advancements in AI are changing the world — and the enterprise. Despite these improvements, enterprises in our current age are struggling to take full advantage of AI for digital transformation efforts and in mission critical processes. What this tells us is that there’s a disconnect between AI research and advancement, and how AI is being implemented in practice for the enterprise world. 

“For IT decision makers, it can be all too easy to get lost in the noise and lack understanding of how AI can operationalize an organization and impact the bottom line.”

While those in the industry should celebrate and keep a pulse on AI achievements, it’s critical to think rationally about how the technology is being used in practice. Otherwise, one might simply scoff at the idea of AI having a soul, and fail to recognize that soul or not, AI can transform a business and the lives of employees and consumers for the better.

See More: How Spotify’s Voice Recognition Device Is Making In-Car Listening Better

SWNI: Can you explain how advanced AI will transform back-office processes and make it easier for machines and humans to communicate?

CF Su: Natural Language Processing, or NLP, refers to the area of computer science technologies that allow computers to process and understand text and spoken words in a similar manner as a human being. In the past, the technologies were rooted in computational linguistics, statistics, and data science, however thanks to recent breakthroughs, like Google’s LaMDA, things have begun to evolve from Deep Learning, or Neural networks. 

“The Large Language Model, for example, powers a long list of NLP tasks, such as text classification, speech recognition and machine translation, delivering outstanding results that were inconceivable a couple of years ago. That generated lots of excitement throughout the industry because it shows how we are just scratching the surface of what’s possible.”

Not only is there a whole segment of Intelligent Document Processing (IDP) which leverages NLP to process human-readable documents encountered in transactions, but most business transactions are in fact powered by human-readable content of various types. However, most are currently processed by human office workers—making NLP a great tool to further productivity, quality and speed, while reducing the cost because of its automated capabilities. 

See More: AI Trends for 2022: 10 Experts on How Artificial Intelligence Will Evolve Next Year

SWNI: Things like sarcasm or code-switching don’t work well with today’s NLP engines. In your opinion, will sentient AI overcome these NLP barriers and have a real place in the enterprise?

CF Su: As discussed, sentient AI is fiction rather than fact. However, more advanced NLP and LLM will usher in a future where machines and humans can communicate via a variety of modalities, like images for example. 

“Machine learning (ML) could generate texts to describe a photo and help classify and index the images included in insurance claims or other business transactions. Beyond NLP-powered data extraction and processing, other applications focused on the downstream stages of business processes, such as ML-powered decision making, could soon become a reality.”

This would not only cause an explosion of machine-readable data in the enterprise world—requiring data science and analytical technologies to process that data—but we could then use NLP text generation to convert raw data insights into a human-readable format more easily consumed by human workers. It is already a reality that the ML system could help news agencies generate sports news for sport matches or earning stories for US public companies.

See More: “AI in Messaging Is Evolving To Assist Conversations, Not Take Their Place”: CEO, Holler

SWNI: There is an ongoing debate over crucial aspects of AI performance, including bias, ethics, and overall control. How can the AI community make progress toward securing an unbiased regulatory framework?

CF Su: Those in the industry must make a commitment to ethical AI and adopt solutions, like explainable AI, that can determine why AI arrived at a certain decision. 

“By using solutions that showcase why AI decided, it’s much easier to gauge whether the decision was made with biased data.”

From a regulatory standpoint, we’ll rightfully see an increase in legislation that demands ethical considerations when using AI in hiring software, loan approvals, and in any processes that involve protected groups and access to service and economic opportunities. 

See More: AI Energy Storage To Make Electric Grids More Responsive: Q&A With Fluence, CTO

SWNI: With the rising adoption of conversational AI models in enterprises, which trends are most likely to directly shape the future of AI?

CF Su: Our society is already accustomed to using conversational IVR (Interactive Voice Response) and online chat bots, but we’ll see more adoption as the technology becomes more advanced and can take on historically human tasks. 

“On the back end, advanced NLP will help machines further extract information from human documents effortlessly and with little error. This will be beneficial to further digitize back-office processes and allow organizations to extract actionable insights needed for decision making, marketing, sales and the list goes on.”

With our current IDP capabilities, AI models can extract data from unstructured documents like emails and PDFS, but there are still complexities that require looping in a knowledge worker to correct or complete a process. The human-in-the-loop approach of human and machine collaboration will reign supreme no matter the level of sophistication of IDP.  

Finally, with more advanced language models, machines will be able to extract data from video, audio, and even photographs, which will allow for a newfound collection of data with the potential to change how we work, and even how we communicate. 

About CF SuOpens a new window

As vice president of machine learning, Ching-Fong (CF) is responsible for supporting the teams who deliver short-term ML projects that create value for our customers as well as longer-term bets that push the technical frontier in our mission to become the world’s leading automation company. CF has more than 15 years of R&D experience in the tech industries, leading engineering teams in fast-paced start-ups as well as big Internet giants. His expertise includes areas of search ranking, content classification, online advertisement, and data analytics. Prior to joining Hyperscience, CF was the head of machine learning at Quora, where his teams developed ML applications of recommendation systems, content understanding, and text classification models. 

About Hyperscience

Hyperscience is transforming the future of work to elevate human potential. The company’s human-centered approach to automation enables a new era of human and machine collaboration that delivers dramatically improved organizational agility, without the legacy cost and burden of change management. By combining data, people, and processes into digital assembly lines, the Hyperscience Platform turns complex processes into simple, configurable workflows. Their industry-leading machine learning technology continuously learns and evolves, to involve humans only when needed.

About Tech TalkOpens a new window

Tech Talk is an interview series that features notable CTOs and senior technology executives from around the world. Join us as we talk to these technology and IT leaders who share their insights and research on data, analytics, and emerging technologies. If you are a tech expert and wish to share your thoughts, write to [email protected]Opens a new window .

How is natural language processing shaping the future in your enterprise? Share your findings with us on LinkedInOpens a new window , Facebook,Opens a new window TwitterOpens a new window .