Teaching Computers to Solve Word Problems
For many people, the word problem was the dread of their mathematical career. Converting language to algebraic equations is just not something everyone is comfortable doing, and computers have an even harder time than people. Researchers at MIT though have successfully created a new computer system that actually can read and solve word problems on its own.
Computers may be far superior to humans when it comes to performing mathematical operations, but computers lack the ability to understand and interpret language well enough to parse paragraphs. A person can read a word problem and correctly isolate the important pieces of information, because we understand the words being used. Even with natural language processing, a computer can struggle to do this, especially as semantic parsing has typically focused on single sentences. This new work however had the system actually examining multiple sentences at a time, and creating a graph from it. That graph is then mapped to a template for the algebra system the software uses. Using machine learning, the software created the templates itself, though the researchers did supply it with examples to study.
The researchers used two different methods to teach the system. One had the researchers feed it four hundred example word problems with their algebraic translations, and the other only gave it a few examples with their translations, with more examples only including their final solutions. As you can expect, the former method improved the system's performance as it solved 70% of its problems, while the latter method left it at 46% solved. That may be a failing grade for a student, but for a computer system that is quite impressive and demonstrates that this approach could be generalized for more complex problems.