This paper proposes probabilistic models for Korean morphological analysis and part-of-speech tagging, which are the most fundamental problems in natural language processing.
Contrary to the previous morphological analyzers that depend on manually con...
This paper proposes probabilistic models for Korean morphological analysis and part-of-speech tagging, which are the most fundamental problems in natural language processing.
Contrary to the previous morphological analyzers that depend on manually constructed linguistic knowledge, the proposed one can fully automatically acquire the linguistic knowledge from annotated corpora (e.g. part-of-speech tagged corpora).
Besides, without any modification of the system, it can be easily applied to other corpora having different tagsets and annotation guidelines.
This paper presents four probabilistic models for morphological analysis that are based on Korean linguistic units (such as Eojeol, morpheme, and syllable) and substring.
The Eojeol-unit model and the syllable-unit model compensate the previous methods that only depend on morpheme-unit analysis.
The Eojeol-unit model is very simple and highly efficient, and also contributes the improvement of precision.
The syllable-unit model is robust to the unknown-word problem and shows high precision.
This paper describes the models and presents the experimental results on three corpora with a wide range of conditions.
Through the experiments, the proposed system achieved comparable and improved performances with the previous systems.
The proposed part-of-speech models are based on the Markov model and the maximum entropy model, which are the most representative probabilistic models.
They consist of the morphological analysis model and the Eojeol transition model.
By the morphological analysis model, the system can consider lexical forms.
By the Eojeol transition model, the system can consider Eojeol-unit context.
This paper proposes various tagging models according to the contexts used in the Eojeol transition model.
Through the experiments, the second-order Markov model shows the best performance among the other models including the previous one.