The modifications of the LRV algorithm in a new method of word segmentation ESA

This article proposes two modifications of a new unsupervised method of word segmentation consisting of three phases: Evaluation, Selection, and Adjustment (ESA), which was presented in our early paper. Lowest Relative Value (LRV) is the core algorithm in ESA The whole method has only one parameter (the exponent in LRV) that can be approximately predicted by the empirical formulae. In this article, we further evaluate one of the empirical formulae on SIGHAN Bakeoff-3 dataset. And we correct the formula for the better prediction. Additionally, we modify another part of the LRV and demonstrate how the part alleviates the sparse data problem. Meanwhile, the results indicate that the modified ESA can produce better results than the original ESA and other methods.