Executive Q&A: Cognitive Search and Extraction Pave the Way to New Breakthroughs
New generations of cognitive search and text extraction are allowing data workers to focus more of their time on meaningful work. We spoke with Ryan Welsh, CEO of Kyndi, an artificial intelligence company serving critical government and commercial institutions, to learn more.
- By James E. Powell
- October 25, 2021
Artificial intelligence (AI) and machine learning (ML) are subject to some well-known limitations, not the least of which is producing or acquiring training data for models. CEO Ryan Welsh of Kyndi, an artificial intelligence company serving critical government and commercial institutions, explains how innovations in cognitive search are addressing these roadblocks and streamlining processes. He also explains about a new direction in AI -- neuro-symbolic AI -- and how it helps organizations deal with unstructured text and maximize the value of AI investments.
Upside: What are the characteristics of the next generation of cognitive search and extraction?
Upside Extra:
Hear from our Speaker
Natural language processing has been getting a lot of attention lately given the proliferation of AI and ML data. Where do you see the future (or market) going?
Ryan Welsh: It’s really an evolution of representation, from the simplest taxonomies and indexes to these richer representations of relatedness of words. It first started with semantic technologies, then evolved to machine learning word embeddings -- approaches that were essentially two different ways of attempting to achieve the same thing: a conceptual model with stronger semantics.
The next evolution combines those two worlds, which is critically important. Although it’s great that semantic technologies are able to establish a strong semantic representation, you need humans to implement it. It’s great that machine learning can show the relatedness of words without human intervention, but it can’t yet get to a representation that is as semantically strong as symbolic AI representation. When you combine these two techniques in a nonsuperficial way, you get a system that can start to comprehend natural language.
What broad trends are driving the need for cognitive search and extraction?
There are macro and micro trends. At the macro level, we’re all familiar with the exponential growth in computational power, storage capacity, and data transmission. What we’re not as familiar with is that the utility of software hasn’t kept up with that exponential growth of hardware. We are basically getting the same results out of software that we did 30 years ago. We can store more data and move more information around faster than we ever could, but when consuming the information, we are still reading with the same speed as our parents and grandparents. As humans, we are the bottleneck in the modern production process.
Talking to large global enterprises, we’ve heard multiple times that to meet their 10-year growth goals, at a minimum they need to double their employee count. There are not enough employees in the world to handle that level of hiring across the Global 2000. They also believe that their employees are not as productive as they could be if they had better software.
This giant macro trend is on the demand and the supply side. The demand side comes from enterprises not having optimal productivity from employees and needing twice the workforce to support their growth goals over the next 10 years. On the supply side, the technologies that people have been working on for the past 75 years are starting to be quite productive. These two things coming together are driving the need for cognitive search and extraction.
For micro-level trends, it all boils down to our penchant for communicating online. Written texts and emails have replaced the phone as our primary communication method. When we do call, it’s transcribed into text. These digital communications are valuable to organizations as they often hold answers that organizations can extract information from to drive business value.
How did previous cognitive search work? Why weren’t these approaches sufficient?
Previous cognitive search used knowledge engineers to build the knowledge representation. This was the wave of “semantic technologies” that people brought to market in the late 90s and 2000s. The benefit was that you’d get a semantically rich representation. The downside was that you needed to manually build the knowledge representation.
Currently, there are a lot of machine learning-based cognitive search products. The benefit is that you can learn from data instead of having to manually build the knowledge representation. However, the problem is that you cannot get as semantically rich a representation, so the tradeoffs are manually build and manage a knowledge representation or use ML to learn an inferior representation without the knowledge engineers. Neuro-symbolic AI gives you the best of both worlds.
Cognitive search is just one aspect of AI. Why do organizations have difficulty implementing AI solutions?
For supervised ML, the main limitation is that these systems that require tons of labeled training data. This technology works really well in situations in which you have hordes of labeled training data. For example, the reason why the tech giants are good at doing image analysis is because they have hundreds of millions of users globally hashtagging images, effectively labeling them. Unless you have large volumes of labeled data, supervised machine learning doesn’t work very well, to the degree that supervised machine learning is prohibitive in most enterprise use cases.
People also run into problems with AI solutions because they’re not built for enterprise AI. They’re built for being in a lab, in an academic setting, with no specific design for solving business problems. Enterprises don’t have enough labeled training data to train these systems; they have to spend all the time, money, and effort to collect and label data before they can train a model to prove a use case.
The next challenge is how technical you need to be to get AI and ML to work. Once I collect all this labeled training data, I need to understand which specific technique to use to train a system to build a model, then use the model to search over underlying text data. In this instance, the machine learning model may be searching over a corpus of documents, for example, and searching for some specific answer to a business problem such as "What month did my sales peak in 2019?” or anything else that’s relevant to the user.
Enterprises that want mature AI capabilities will, on average, need to hire 600 full-time AI employees and pay them $120,000 annual salaries. Over the next three-to-five years, the Global 2000 need to hire 1.2 million full-time AI employees and pay them $240 billion per year in total salaries.
Processing, storing, and sharing unstructured text is also a problem because it’s all about representation. Text data just isn’t represented in a computable way.
Explain how natural language processing (NLP) helps organizations handle the unstructured data problem.
Enterprises want to do two things with their unstructured text data. They want to make it searchable and they want to extract information from it for analytics. NLP allows them to do that. When searching unstructured text today, there are technologies that comprehend language better than previous technologies did. Instead of just matching a keyword, it comprehends your question and the text that it’s reading, and steers you to the answer. It’s the same with extracting information from text.
Robotic process automation bots are good at extracting information from applications if that information is in the same spot, but when it’s in different locations, the system needs to read more thoroughly to find what you’re looking for. RPA bots can’t do that. NLP, because it comprehends text better than other technologies, understands that this number in this paragraph is what should be extracted.
What is neuro-symbolic AI and how does it help organizations deal with unstructured text and maximize the value of AI investments?
There’s a long history of neuro-symbolic AI where people thought about combining these the two main AI fields of AI: symbolic and connectionist approaches. Symbolic approaches are a nod to symbolic reasoning. This is the knowledge side of AI in which there are terms, confirmed definitions, and hierarchies of terms that intelligent systems use to reason with, partly based on rules. Symbolic AI has strong reasoning capabilities, but it has no learning capability and its ability to perceive is weak. The lack of learning is why I said the representation needs to be manually built.
Connectionist approaches are simply machine learning. ML approaches learn from data (their representation is a vector) but have no contextual capability and minimal reasoning ability. The merits of each approach overcome the other’s deficits. Supervised machine learning requires lots of data to be trained. Representations in symbolic AI allow you to be data efficient because they can be reused in multiple tasks. With neuro-symbolic AI you don’t need all the labeled training data for each new task, which solves a major hurdle to becoming successful with supervised learning systems.
The second hurdle is explainability for end users and developers. Neuro-symbolic AI is more explainable because of its language-like representation, and that representation is more amenable to human understanding. The system can link back to underlying data sources and communicate in a natural language way versus numbers, which don’t mean anything to most people.
From a developer perspective, explainability allows you to tune the system in the direction you want it to go. When a system is able to explain to the developer “these are the words I don’t understand,” that’s a targeted approach to teaching the computer those specific words.
What benefits can companies expect from neuro-symbolic AI? What drawbacks and obstacles remain?
The benefit for end users is a productivity increase from finding answers and extracting information from text. They’ll do that quicker than they ever did before. The benefit to the data science team, software developers, and IT-at-large organizations is the time to value in building AI solutions. The success of this work making it to deployment significantly increases the impact on end users and drives organizational value. This technology unlocks the value of the human being and human intelligence.
The remaining drawbacks and obstacles for neuro-symbolic approaches contrast with specific ML approaches. Neuro-symbolic methods are more amenable to returning snippets of text versus the precise answer itself. If you ask, “when was JFK assassinated,” ML systems will return “1963.” Neuro-symbolic approaches as they’re built today will bring back the snippet of text that has the answer in it. In the underlying document, it’ll bring back the sentence that says, ‘John F. Kennedy was assassinated in 1963’. The answer from ML systems -- 1963 -- is arguably more accurate. However, getting that answer could take a very long time and require thousands of labeled training examples.
Where is all this technology headed?
My hope is that we’ll be able to spend our valuable time doing our most meaningful work. So much of what we do today isn’t the meaningful work we want to do; it’s preparation for it. These modern and evolving technologies allow individuals to focus their time and effort on being creative, resourceful, and leveraging all those attributes that make us human. I also hope for a return to the 40-hour work week. Because of the above-mentioned inefficiencies, knowledge workers are working 12-14 hours days, seven days a week. Using cognitive search, we can all get back to delivering our most meaningful work daily -- in eight hours or less.