Fast Construction of a Word-Number Index for Large Data

The paper presents a work still in progress, but with promising results. We offer a new method of construction of word to number and number to word indices for very large corpus data (tens of billions of tokens), which is up to an order of magnitude faster than the current approach. We use HAT-trie for sorting the data and Daciuk’s algorithm for building a minimal deterministic finite state automaton from sorted data. The latter we reimplemented and our new implementation is roughly three times faster and with smaller memory footprint than the one of Daciuk. This is useful not only for building word↔number indices, but also for many other applications, e.g. building data for morphological analysers.