Have you ever spoken to Alexa, used Google Translate or an auto-correct tool like Grammarly? Many of these tools which make our lives easier are powered by Natural Language Processing, or NLP, and we interact with them on a daily basis.
NLP lies at the intersection of linguistics and artificial intelligence and enables computers to understand natural human language.
Understanding language as a human is, for the most part, an easy task. Most people have no problem reading a blog post, having a conversation with a friend or writing a story.
However, this isn’t the case for a computer. Computers love structure which is why they struggle with the fluid nature of language. In fact, a computer understanding language is such a difficult task that Alan Turing made it the centrepiece for his test for computer intelligence in the Turing test.
Although NLP has been around since the 1950s, the vast quantity of text data which is generated on the internet every day has driven progress in the field. There have been major advances in the last couple of years driven by some of the heavy-hitters of machine learning, notably: Google, Facebook, Amazon and OpenAI.
Computers learn how to understand language in a way similar to a child: through examples. Whether you want to translate your favourite book into German, summarise a blog post in 100 words or work out if someone is saying ‘Alexa tell me a joke’ you will need to train your model using examples.
You also can’t just throw examples at your model without telling your model what to do with it. If you wanted to train Alexa for example, you would need to show it something like this:
Once Alexa has seen enough examples of people asking it to tell a joke it will be able to identify similar phrases which it has never seen before, such as ‘Alexa, pull my leg’.
So how many examples are enough? That depends on the complexity of your task and the number of examples available. For a simple model, the bare minimum would usually be a few thousand examples, though the more examples the better.
To create a state-of-the-art model you need as much data as you can get your hands on. The largest NLP model made to date, GPT-3, was trained on a dataset of half a trillion words. This model was trained on a supercomputer but if you wanted to train it yourself using a top-of-the-range cloud computer it would take 355 years and cost you over £3.3m in computing costs!
Past getting Alexa to tell us jokes, there are countless practical applications for NLP that can improve our world.
One area which has been historically lacking in technological advancements - but would greatly benefit from the latest research in AI - is debt recovery & management.
Dealing with debt can be incredibly difficult. This is only exacerbated by (as well as caused by) other issues, whether physical or mental. Some examples include alcoholism, mental health struggles, a physical disability or a recent bereavement. In fact, StepChange has estimated that 1 in 5 people who find themselves in debt are struggling with something else on top of their financial situation.
When consumers are in vulnerable circumstances, it may affect the way they engage with financial services. Vulnerable consumers may be significantly less able to represent their own interests, they may have different needs and may have more behavioural biases that negatively impact their decision making - The FCA
Debt collectors therefore have a huge responsibility in ensuring that everybody receives the right care and the right outcome, with a particular focus on vulnerable customers.
Often, the problem that traditional debt collectors struggle with is not supporting the individual once they are aware of their problem, but detecting it in the first place. Personal struggles can be incredibly difficult to open up about and so debt agencies need to be constantly on the ball to ensure that nobody falls through the cracks.
At Ophelos we have developed OLIVE, an NLP model that is capable of predicting a customer’s vulnerability and identifying the possible causes.
OLIVE works by scanning text that customers send us, via email or live chat, and flagging up potential vulnerability to our customer support agents in real time. Not only can the model give our agents a predicted vulnerability score from 0 to 1 (0 being no vulnerability, 1 being highly vulnerable) but it can also suggest the driving force behind the vulnerability. For example, if a customer writes “lockdown feels so tough this time”, OLIVE would indicate a vulnerability score and offer ‘Coronavirus’ and ‘Mental health’ as possible causes.
Using NLP to detect vulnerability ensures that no customer is left behind and everybody receives the right help for their personal situation. Our customer support agents are of course highly trained and are great at listening to each customer’s situation, though OLIVE provides that extra safety net for vulnerable customers by allowing our agents to always be on their A-game.
The data that we gather using OLIVE can also be used to confront the debt crisis at its source. Rather than just catering for customers who fall into debt, we can prevent people from falling into it in the first place by analysing the demographic of those struggling with debt and the reasons for which they are struggling.
By relaying this information back to our clients, we can help companies to build better products that will not be detrimental to their customer’s financial health. For example, if we can spot that a lot of 45 to 65-year-old men are struggling with gambling addictions, our clients could use this information to implement a gambling block on their product.
If you're interested in finding out more about the ways tech is improving money management, why not check out this blog post?