Artificial Intelligence (AI) and machine learning applications, particularly those performing Natural Language Understanding (NLU) are becoming a part of our everyday lives. As consumers we see them in products such as digital assistants from Amazon, Google, and Apple, sales and customer service chatbots, and of course our mobile phones.
So, it stands to reason that the rest of us in the public and private sectors would start looking for ways to leverage this technology for our applications. After all, 80% of the data in the world is unstructured, a good portion of it in social media posts, and we know that there are valuable insights to be had in understanding that data.
We at Gamalon meet a lot of data analysts and scientists working for consumer products companies, automobile manufacturers, and similar businesses that have an immediate goal to be more accessible, responsive, and fair to their customers and constituents. They are looking for a natural language understanding solution that is easy to implement and practical for a business person to use.
But natural language understanding technology has evolved significantly from the natural language processing we have been using the past 20 years to do things like route emails based on keywords. This can make it a challenge to keep up to date and to go in a direction that won’t be rendered obsolete in a few years. There are two primary categories of solutions:
- Supervised learning solutions from big names like IBM, Microsoft and Google, which give organizations off the shelf solutions, but would need data scientists to customize
- Unsupervised learning technologies for document classification and exploration
Open source technologies can be leveraged within either approach, but require skilled developers and data science teams to create custom solutions.
While each approach has its advantages and challenges, the big question is which path will be most effective at delivering a system that adapts to changes in the business and continues to learn so that it can be functional well into the future? For many companies, it’s not practical to continually engage data scientists and developers.
At Gamalon, we believe the key to building successful ML programs is to get business domain experts more involved in the process from beginning to end. When subject matter experts close to the business can guide the system in its learning, see how it is making its decisions, provide feedback to the system, and inject new information into the system when the business changes, the system will learn faster, be more accurate, and will have a significantly longer useful life.
The Practicality of Supervised Learning, Unsupervised Learning
Let’s take a closer look at the supervised and unsupervised learning techniques.
Supervised learning is the more commonly used approach. There are neural networks and deep learning, in which humans train the natural language understanding system by providing large amounts of labeled data, or keywords or phrases it wants the system to recognize. For example, all the different forms in which a “date” might be represented, or all the words that might represent dissatisfaction or buying intent. Another supervised learning technique is rule-based, for which humans define the rules of behavior of the model.
While there are advantages of supervised systems in terms of how the models are trained and the utility over time, users are finding five categories of challenges:
- Data – Requires large quantities of labeled data
- Manpower – Requires large teams to continually label data and technical teams to update the model or retrain the system
- Long time to setup – It can take significant time and effort to get the system properly trained
- Opaque decision making – With neural network, they never really know how the system is making its decisions
- Lack of accuracy – Systems are only as accurate as their labeled data sets and human-defined models. While these can seem promising in laboratory conditions, when it comes to implementation, real world accuracies are often as low as 60%. Rules based systems in particular can be very difficult to improve.
Unsupervised learning is where the system parses large amounts of text on its own, such as documents, and then classifies them into categories based on the common words it finds between them. While these natural language understanding systems can be quick to initially implement, and avoid the manpower and data requirements of supervised learning, there are three categories of challenges users are finding:
- Order and context – While these systems are good at recognizing words, some systems lack an understanding of the context and order of words which can change their meaning
- Inability to assess decision making – Without labeled data to which to compare the results, there’s limited ability to measure the success of a particular model on its decision making
- Poor results – These systems often produce categories and other results that are far afield from what a human would have done, let alone can use
Neither Approach Works for Regulated Industries
It is important to note that the lack of transparency in decision making in both of these natural language understanding approaches, is a real issue for regulated industries, as discussed in the CIO article “Risky AI business: Navigating regulatory and dangers to come”. While rules-based approaches are often favored among regulated industries because of the decision making transparency, the trade off with accuracy makes this approach problematic as well. Two common-sense examples where transparency is especially critical are:
- Financial Services – You need to be able to tell clients your rationale behind recommendations such as investment decisions
- Medical Services – You need to be able to describe to a patient and their insurance company how you reached a diagnosis and why you recommend a specific course of treatment
A Pragmatic Approach to AI and Machine Learning for Natural Language Understanding
A new generation of machine learning for natural language understanding is emerging that combines the quick time to market of unsupervised learning with the human knowledge of supervised learning. This is called “Idea Learning”. The approach is essentially next-generation, human-guided unsupervised learning.
Idea Learning = Human + Machine Learning
In this approach, the machine already knows that words are made up of letters, and phrases are made up of words. It understands language by understanding words and phrase-level synonyms within the context, order, and logic within a hierarchy. As a result, it can learn “ideas” based on the context of words, not just the individual words. For example, that:
- “Logging in to” and “getting into” are synonyms, when used in conjunction with “Problem” and “My account”
- “Foundation” can mean one thing when in the context of concrete, but something else when in the context of beauty products
- “Help me please, my account shut off,” is very different than, “Please help me shut off my account.
With Idea Learning, the natural language understanding system:
- Learns on its own and is able to identify “ideas” in an unsupervised fashion
- Graphically displays the decision sequence it took
- Asks for confirmation from a human on a decision it is making when unsure of itself
Benefits of Human-guided Machine Learning
We won’t go into the technical details here, but this new generation leverages a combination of AI approaches, and makes it usable by a subject matter expert, not just a data scientist. Idea Learning combines the best of the advances that have been accomplished in machine learning and natural language understanding.
- Requires less human interaction up front – Idea Learning is unsupervised learning that injects a human into the process to assist the learning
- Transparent decision making – Idea Learning is transparent in how it makes decisions and when it’s uncertain, it surfaces that to subject-matter-experts for guidance, so it gets smarter faster
- Accepts business-specific information or “expertise” from a business person or subject matter expert – A human can guide the learning from within an intuitive UI and insert domain-specific information into the model
- Works well even for small, short messages – Does not require lengthy documents, but can manage on short-form text like chat transcripts or social media posts
- Is more accurate – Tests show it is more accurate, more quickly than supervised learning systems
- Identifies the all of the “Ideas” or “stories” behind the words – Very important for sales, customer service, and survey responses, it doesn’t categorize a message as a single topic, but can identify all of the ideas contained in that one message
- Is constantly learning – Is able to continually discover new insights, as well as continually measure and monitor specific topics
- Applicable to a broader audience – Able to be easily implemented by businesses without data science teams or where transparency is tantamount, such as regulated industries
Learn More about the Next Generation of AI Machine Learning for NLU
We at Gamalon, believe that humans and machines need to work together to make each other smarter. The next generation of machine learning for natural language understanding is about borrowing and synthesizing the best ideas from across the communities and unifying them in a mathematical way that puts them in the hands of a wider audience, is quick and easy to implement, and significantly outperforms the other approaches.
To learn more, download our white paper “Idea Learning Defined”, that describes this fundamentally new approach to AI and machine learning for natural language understanding.