TY  - THES
T1  - Learning Chinese language structures with multiple views
A1  - Sun,Weiwei
Y1  - 2012/05/25
N2  - Motivated by the inadequacy of single view approaches in many areas in NLP, we study multi-view Chinese language processing, including word segmentation, part-of-speech (POS) tagging, syntactic parsing and semantic role labeling (SRL), in this thesis.  We consider three situations of multiple views in statistical NLP: (1) Heterogeneous computational models have been designed for a given problem; (2) Heterogeneous annotation data is available to train systems; (3) Supervised and unsupervised machine learning techniques are applicable.

First, we comparatively analyze successful single view approaches for Chinese lexical, syntactic and semantic processing.  Our analysis highlights the diversity between heterogenous systems built on different views, and motivates us to improve the state-of-the-art by combining or integrating heterogeneous approaches.  Second, we study the annotation ensemble problem, i.e. learning from multiple data sets under different annotation standards.  We propose a series of generalized stacking models to effectively utilize heterogeneous labeled data to reduce approximation errors for word segmentation and parsing.  Finally, we are concerned with bridging the gap between unsupervised and supervised learning paradigms.  We introduce feature induction solutions that harvest useful linguistic knowledge from large-scale unlabeled data and effectively use them as new features to enhance discriminative learning based systems.

For word segmentation, we present a comparative study of word-based and character-based approaches.  Inspired by the diversity of the two views, we design a novel stacked sub-word tagging model for joint word segmentation and POS tagging, which is robust to integrate different models, even models trained on heterogeneous annotations.  To benefit from unsupervised word segmentation, we derive expressive string knowledge from unlabeled data which significantly enhances a strong supervised segmenter.

For POS tagging, we introduce two linguistically motivated improvements: (1) combining syntax-free sequential tagging and syntax-based chart parsing results to better capture syntagmatic lexical relations and (2) integrating word clusters acquired from unlabeled data to better capture paradigmatic lexical relations.

For syntactic parsing, we present a comparative analysis for generative PCFG-LA constituency parsing and discriminative graph-based dependency parsing.  To benefit from the diversity of parsing in different formalisms, we implement a previously introduced stacking method and propose a novel Bagging model to combine complementary strengths of grammar-free and grammar-based models.  In addition to the study on the syntactic formalism, we also propose a reranking model to explore heterogenous treebanks that are labeled under different annotation scheme.  Finally, we continue our efforts on combining strengths of supervised and unsupervised learning, and evaluate the impact of word clustering on different syntactic processing tasks.

Our work on SRL focus on improving the full parsing method with linguistically rich features and a chunking strategy.  Furthermore, we developed a partial parsing based semantic chunking method, which has complementary strengths to the full parsing based method.  Based on our work, Zhuang and Zong (2010) successfully improve the state-of-the-art by combining full and partial parsing based SRL systems.

KW  - Chinesisch
KW  - Sprachverarbeitung
KW  - Syntaktische Analyse
KW  - Semantische Analyse
KW  - Lexikalische Analyse
KW  - Maschinelles Lernen
CY  - Saarbrücken
PB  - Universitäts- und Landesbibliothek
AD  - Postfach 151141, 66041 Saarbrücken
UR  - http://scidok.sulb.uni-saarland.de/volltexte/2012/4863
ER  -