When I started university in mainland China my dream was to be a physicist. Then I became interested in Big Data and Artificial Intelligence (AI) and everything changed.
It was amazing to find that a machine can ‘think’ like a human and know that ‘apple’ and ‘iPhone’ could mean the same thing; that ‘Cappuccino’ and ‘Latte’ were different kinds of coffee. Or that when you typed ‘Pandas’ it could tell whether you meant bears or the dataframe used in Python.
I knew at that time, AI would change the world, and I fell in love with it. I found myself an internship in Microsoft Research in Beijing and said goodbye to physics. After that I worked with Baidu, Yahoo and Tencent in China’s Silicon Valley in Beijing.
Then I made another big change. I decided to leave the Big Tech world and join HSBC as a senior data scientist for the bank’s PayMe app that’s used by two million people in Hong Kong with a 70 per cent market share by P2P value.
I made the change because I saw my chance to be part of the fast-growing adoption of AI in the financial sector.
After starting at HSBC I found that although tech companies and the bank have totally different kinds of products, their core values of being customer-focused and business value-driven are the same.
As technologists, we use data to inform decisions that can bring the biggest impact to the bank and its customers, instead of just trying to come up with the fanciest machine learning (ML) models.
It was with this mindset that I found many opportunities to use my knowledge of Natural Language Processing (NLP) to find insights in the unstructured text data that could help provide new and better services for customers.
For example, we applied a state-of-the-art (SOTA) NLP deep-learning model to identify key common themes in customer conversations. That led to the development of chatbots to improve the efficiency and user experience of customers when they contact the bank.
Before employing this model, we needed to manually label the data and spend lots of time trying to understand the conversations embedded in the data. This was particularly confusing and complex because of the many languages used in Hong Kong.
The creative part for this chatbot solution is transfer learning. We use open source models that have been trained on billions of data points and clusters of graphic processing units (GPU) for days or weeks, modify their model structures and do further training with our own conversational banking data in our own secured environment.
By using transfer learning we are quickly able to build a highly accurate chatbot without much human effort on labeling. It also saves on computing resource and makes the whole thing very scalable and operational.
There’s so much that AI and ML tools can do to help analyse customer sentiment so we can get a better understanding of their experiences. Knowing where their pain points are in our journeys allows us to react quickly with solutions and improvements.
Pingping explains more about how SOTA NLP and deep learning can enable innovation in digital payments, conversational banking and personal banking at the Women in Data Science (WiDS) talks in Shanghai, part of the Global WiDS Conference held at Stanford University.