Mathematical Expression Extraction from Unstructured Plain Text

Mathematical expressions are often found embedded inline with unstructured plain text in the web and documents. They can vary from numbers and variable names to average-level mathematical expressions. Traditional rule-based techniques for mathematical expression extraction do not scale well across a wide range of expression types, and are less robust for expressions with slight typos and lexical ambiguities. This research employs sequential, as well as deep learning classifiers to identify mathematical expressions in a given unstructured text. We compare CRF, LSTM, Bi-LSTM with word embeddings, and Bi-LSTM with word and character embeddings. These were trained with a dataset containing 102K tokens and 9K mathematical expressions. Given the relatively small dataset, the CRF model out-performed RNN models.