A Generative Deep Learning for Generating Korean Abbreviations

摘要

An abbreviation is a short form of a sequence of words or phrases. Abbreviations have been widely used as an efficient way of communicating within a human community, and nowadays they are used more widely and more often because electronic communications such as world wide web or twitter get available. One critical issue about abbreviations is that they are continuously generated whenever a new material such as a new novel or a TV drama is made. Therefore, a method to understand generation and detection of abbreviations is required for further processing of the abbreviations. The simple and well-known method for abbreviation generation is to use the rules that are well designed by human experts, but such rule-based methods are not appropriate for Korean abbreviations. This is due to two major reasons. The first is that Korean abbreviations are much irregularly generated compared to English ones, and thus the rules become too complex for managing all irregularities. The other is that many Korean abbreviations contain characters or syllables that do not appear at the original sequence of words due to a pronunciation issue. As a result, a great number of rules to generate new characters or syllables should be made, which makes the rule-based methods impractical. As a solution to this problem, this paper proposes a generative deep learning architecture to generate Korean abbreviations. The proposed architecture consists of two Long Short Term Memory (LSTM) networks, in which one LSTM encodes a variable-length source sequence into a fixed-length vector and the other LSTM decodes the vector into a variable-length target shorter sequence. According to our experiments on the Korean abbreviations set from National Institute of Korean Language, the proposed method achieves 21.4% of accuracy, which is 420% improved accuracy over a simple rule-based method. This result proves that the proposed method is effective in generating Korean abbreviations.