Korean Morphological Analysis with Tied Sequence-to-Sequence Multi-Task Model

Abstract

Korean morphological analysis has been considered as a sequence of morpheme processing and POS tagging. Thus, a pipeline model of the tasks has been adopted widely by previous studies. However, the model has a problem that it cannot utilize interactions among the tasks. This paper formulates Korean morphological analysis as a combination of the tasks and presents a tied sequence-tosequence multi-task model for training the two tasks simultaneously without any explicit regularization. The experiments prove the proposed model achieves the state-of-the-art performance.

Transformer‐based reranking for improving Korean morphological analysis systems

“…For our experiments, we utilized the Sejong corpus (used in the literature [18–24, 29, 30]), UCorpus [37], and Everyone's Corpus [39]. In line with previous studies for comparison purposes, the Sejong corpus underwent training using a single model without separation.…”

“…In recent years, Korean morphological analyses have witnessed a diverse range of methodologies [3–31]. The agglutinative nature of the Korean language poses challenges that have inspired researchers to devise innovative solutions, laying the foundation for future investigations.…”

“…To address the drawbacks of dictionary dependence, syllable-by-syllable morphological analysis has emerged as an alternative [8,9,11,13,14,17,19,[25][26][27][29][30][31]. This approach involves either tagging each syllable and then applying a base-form restoration dictionary [14,26] or tagging the syllable with the base form already restored [31].…”

Multi-Task Learning in Natural Language Processing: An Overview

“…Targeting community question answering, [139] uses the result of question category prediction to enhance document representations. [110] feeds the result of morphological tagging to a POS tagging model and the two models are further tied by skip connections. To enable asynchronous training of multi-task models, [152] initializes input from other tasks by a uniform distribution across labels.…”

“…For example, [93] converts the parsing of Alexa meaning representation language into three independent tagging tasks for intents, types, and properties, respectively. [110] transforms the pipeline relation between POS tagging and morphological tagging into a parallel relation and further builds a joint MTL model.…”