By 2020, the total amount of digital data created worldwide is expected to reach 44 zettabytes thanks to the advent of connected devices and software applications, according to IDC. This is exciting, but it also presents new challenges. How can business users make sense of all of this data efficiently? More importantly, how can business users quickly access the right data without being an expert in data mining?
At Salesforce Research we spend a lot of time thinking about data and how we can make it easier for sales, service, marketing leaders to get real-time access to the data they need to be competitive—without having to be experts in advanced database queries. Today, we’re excited to announce a new AI model that addresses this exact challenge by allowing business users to interact with data more intuitively using natural language.
Introducing Seq2SQL, a new state-of-the-art system that allows business users to communicate with relational databases and get the answers they need by simply typing a question in a search box or speaking to a device—without having to “speak” database languages such as SQL. The Seq2SQL deep learning model uses reinforcement learning—meaning the model learns best actions based on rewards—on real-time database search queries using the SQL structure. As a result, the model achieves significant gains in performance when compared to other deep learning methods.
Another contribution of Salesforce Research to the field of database question answering is a large training dataset. Datasets are vitally important to training powerful AI models. That’s why we’re excited to introduce WikiSQL, an open source dataset that is orders of magnitude larger than existing datasets. Additionally, WikiSQL will help further advance research in natural language interfaces. Together, the model and dataset create user interfaces that are more intuitive and increase the speed at which users can access data. Check out our academic paper for more on how our models work.
The ability to quickly and easily communicate with relational databases allows business users to achieve new levels of productivity and a better understanding of their customers. For example, with Seq2SQL, we take a first step towards a world in which service leaders can quickly access their most important key performance indicators (KPIs) in plain English versus manually selecting columns or inputting conditions. By dictating or typing, “Show me all accounts with the lowest customer satisfaction score” or “Show me all Tier 1 accounts that are likely to attrite,” service leaders can quickly and easily search historical trends, performance metrics and more. This allows them to better prioritize their day and make more impactful decisions-- all without prior knowledge of database schemas or object and field nomenclature. In turn, they can provide more personalized and proactive services to resolve cases faster and improve customer satisfaction.
The impact of this research does not stop at service, and can also benefit sales and marketing by unlocking valuable insights about incoming leads, opportunities and overall pipeline health. This work, led by Research Scientist Victor Zhong, is an important breakthrough as we continue to build on our mission to bring cutting edge AI to Salesforce customers. For more on Salesforce Research and our recent projects, check out Einstein.com.