Syntactic role identification of mathematical expressions

This paper presents a prediction algorithm to infer the syntactic role (SR) of mathematical expressions (ME), or SRme, in ME-plaintext mixed sentences. SRme is a predicted syntax label of ME, which could be integrated into any constituent parser to improve their accuracy in sentence parsing. SRME is based upon three features of ME placement in a sentence: properness of Sentence structure (feature F3), properties of ME (feature F2), and PoS of the Local neighbor plain text (feature F1). An inside-outside inspired algorithm is proposed for SRME by maximizing the probability of a relaxed parsing tree. Features in F2 was found to fit into both exponential and Poisson distributions, which could fuse with other features to re-weight the prediction rule that improves the prediction precision for SRme as a noun phrase (noun modifier) by 3.6% (18.7%). F1, F2, and F3 were found to complement each other. Significant discriminative patterns on the part-of-speech (PoS) of the neighbor plaintext are adopted to build a Naïve Bayesian classifier, which is fused with the F3 baseline that improved the precision of the prediction of SRme as a sentence by 10%. The overall error rate of the SRME prediction algorithm was found to be 15.1% based on an experiment using a public ME-plaintext mixed parsing tree data set provided by Elsevier.