Unsupervised chunking based on graph propagation from bilingual corpus.

Area of Science:

• Natural Language Processing
• Computational Linguistics
• Machine Learning

Background:

• Unsupervised shallow parsing models for Chinese text often require annotated data or Chinese linguistic information.
• Existing methods for bilingual knowledge transfer may not be optimal for unsupervised parsing tasks.

Purpose of the Study:

• To develop a novel unsupervised shallow parsing model for Chinese text.
• To achieve high accuracy without relying on annotated Chinese data or Chinese linguistic features.
• To explore the effectiveness of graph-based label propagation for cross-lingual knowledge transfer in parsing.

Main Methods:

• The proposed approach trains a shallow parsing model on unannotated Chinese text from a parallel Chinese-English corpus.
• It employs graph-based label propagation for bilingual knowledge transfer from English to Chinese.
• Projected labels derived from the English side are utilized as features within the unsupervised Chinese model.

Main Results:

• The novel approach demonstrates improved performance in unsupervised shallow parsing of Chinese.
• Experimental comparisons show significantly higher accuracy (F-score) compared to state-of-the-art algorithms.
• The method effectively transfers knowledge across languages without direct use of Chinese annotations.

Conclusions:

• The proposed unsupervised shallow parsing method is effective for Chinese text.
• Graph-based label propagation is a viable technique for bilingual knowledge transfer in this context.
• This approach offers a promising direction for low-resource or unannotated language parsing.