Dowsing for Math Answers
Yin Ki Ng, Dallas J. Fraser, Besat Kassaie, Frank Wm. Tompa · Lecture notes in computer science · 2021
Mathematical Information Retrieval (MathIR) focuses on using mathematical formulas and terminology to search and retrieve documents that include mathematical content. To index mathematical documents, we convert each formula into a token list that is compatible with natural language text. Then, given a natural language query that includes formulas, we select key terms and formulas from the query, again convert the query formulas into token lists, and finally search and rank results using standard search engine techniques. In this paper, we describe our approach in detail for a Community Question Answering task and evaluate the weight to be given to formula tokens versus text tokens. We also evaluate a regression-based approach to re-ranking based on metadata associated with the documents returned from the search.