Mathematical expressions are an important means of scientific communication, used not only for numerical calculation and theorem proving but also for clarifying concept definitions and ensuring formal operations are not ambiguous. In this paper, we address the problem of understanding the meaning of mathematical expressions. In our approach, we formulate the problem as a task of translation from Presentation MathML to Content MathML expressions. We propose a machine learning-based approach that combines automatic extraction of fragment rules and statistical machine translation. Experimental results showed the potential of our approach for understanding the meaning of mathematical expressions.