Artificial Intelligence is Only as Good as Data Labeling

Data Labeling with SynerScope

Recent events in my home country inspired me to write this blog. Every day we hear stories about businesses and government organizations struggling to sufficiently understand individual files or cases. Knowledge gaps and lack of access to good information hurts individual-and-organizational well-being. Sometimes, the prosperity of society itself is affected. For example, with large-scale financial crime and remediation cases in banking, insurance, government, and pandemics.

We simply have little understanding of the data which means AI and analytics are set up to fail. In addition, it’s difficult to see what data we can or may collect to run human-computer processes of extracting relevant information to solve those issues.

Unlimited Data with No Application

The COVID19 pandemic shows not only how difficult it is to generate the right data, but also how difficult it is to use existing data. Therefore, data-driven decision-making often shows gaps in understanding data.

Banks spend billions on technology and people in KYC, AML, and customer remediation processes. Yet, they’re still not fully meeting desired regulatory goals.

Governments also show signs of having difficulties with data. For example, recent scandals in the Dutch tax office, such as the Toeslagenaffaire, show how difficult it is to handle tens of thousands of cases in need of remediation. And the Dutch Ministry of Economic Affairs is struggling to determine individual compensation in Groningen, where earthquakes caused by gas extraction have damaged homes.

Today, the world is digitized to an unbelievable extent. So, society, from citizens to the press to politicians and the legal system, overestimate the capabilities of organizations to get the right information from the data which is so plenty available.

After all, those organizations, their data scientists, IT teams, cloud vendors, and scholars have promised a world of well-being and benevolence based on data and AI. Yet, their failure to deliver on those promises is certainly not a sign that conspiracy theories are true. Rather, it shows the limits of AI in a world where organizations understand less than half of the data they have when it is not in a machine processing ready state. After all, if you don’t know what you have, you can’t tell what data you’re missing.

Half of All Data is Dark Data

Gartner coined the term “Dark Data” to refer to that half of all data that we know nothing about. And, if Dark Matter influences so much in our universe, could Dark Data not have a similar impact on our ability to extract information and knowledge from the data?

We have come to believe in the dream of AI too much, because what if dark data behaves as dark matter? By overestimating what is possible with data-driven decision making, people may believe that the powers that be are manipulating this data.

SynerScope’s driving concept is based on our technology to assess Dark Data within organizations. By better understanding our dark data, we can better understand our world, get better results from human and computer intelligence (AI) combined.

Algorithms Rely on Labeled Datasets

Today’s AI, DL (Deep Learning, and ML (Machine learning) need data to learn – and lots of it. Data bias is a real problem for that process. The better training data is, the better the model performs. So, the quality and quantity of training data has as much impact on the success of an AI project as the algorithms themselves.

Unfortunately, unstructured data and even some well-structured data, is not labeled in a way that makes it suitable as a training set for models. For example, sentiment analysis requires slang and sarcasm labels. Chatbots require entity extraction and careful syntactic analysis, not just raw language. An AI designed for autonomous driving requires street images labeled with pedestrians, cyclists, street signs, etc.

Great models require solid data as a strong foundation. But how do we label the data that could help us improve that foundation. For chatbots, for self-driving vehicles, and for the mechanisms behind customer remediation, fraud prevention, government support programs, pandemics, and accounting under IFRS?

Regulation and pandemics appear in the same sentence because, from a data perspective, they’re similar. They both represent a sudden or undetected arrival that requires us to extract new information from existing data. Extracting that new information is only manageable for AI if training data has been labeled with that goal in mind.

Let me explain with an easy example of self-driving vehicles. Today, training data is labelled for pedestrians, bicycles, cars, trucks, road signs, prams, etc. What if, tomorrow, we decide that the AI also must adapt to the higher speed of electric bikes? You will need a massive operation of collecting new data and re-training of that data, as the current models would be unlikely to perform well for this new demand.

Companies using software systems with pre-existing meta data models or business glossaries have the same boundaries. They work by selecting and applying labels without deriving any label from the content – otherwise they must label by hand, which is labor and time intensive – and often too much so to allow for doing this under the pressure of large-scale scandals and crises.

Automatic Data Labeling and SynerScope

The need to adapt data for sudden crises does not allow for manual labeling. Instead, automatic labeling is a better choice. But, as we know from failures by organizations and by government, AI alone is not accurate enough to take individual content into account.

For SynerScope, content itself should always drive descriptive labeling. Labeling methodology should always evolve with the content. That’s why we use a combination of algorithm automation and human supervision, to bring the best of both worlds together – for fast and efficient data labeling.

If you want to learn more about how our labelling works, feel free to contact us at info@synerscope.com