Deep learning has shown that it is able to achieve incredible performance on certain tasks. Lately, surfing on the hype wave surrounding chatbots, natural language understanding has been a huge focus for businesses investing in machine learning. One of our clients asked us to help automate their document processing pipeline by creating a deep learning module that would extract relevant information from contracts automatically. By training a deep learning model naïvely, a surprisingly good performance was already reached.
However, it was far from enough. After investigating, many mistakes made by the model could be explained by the ambiguity surrounding the documents it had to deal with.
Tables were difficult to find, words would be laid out physically on the sheet in a way that was difficult to translate into a sequence of words. These documents were created to be read by humans, not computers. It is often argued that the goal of artificial intelligence is to reach human-level cognition. But it doesn’t need to be such a one-sided effort. Humans can help bridge that gap by rethinking the way they design, by taking machines into account. For instance, road signs could be designed to be easily recognized by computer vision systems, fonts could be modified to improve optical character recognition, or products and packaging could be shaped in ways that robotic arms can more easily handle them. If artificial intelligence will help us in the future, we need to help it today.