Optimisation of the Largest Annotated Tibetan Corpus Combining Rule-based, Memory-based, and Deep-learning Methods

Abstract

This paper presents the new and improved version of the Annotated Corpus of Classical Tibetan (ACTib). These segmented and POS-tagged versions of all available texts in the Buddhist Digital Resource Center (BDRC) were annotated automatically using a memory-based tagger (see Meelen and Hill 2017). While this method had certain clear advantages - large amounts of data could quickly be split into meaningful words and grammatical markers, provided with highly detailed morpho-syntactic labels - the accuracy of these initial results can be improved in various ways. In this paper, we present a thorough error analysis and focus on correcting and improving these results using a combination of optimised memory-based, neural networks and rule-based methods.