Automatic Multi-Lingual Information Extraction

Information Extraction(IE) is a burgeoning technique because of the explosion of internet. So far, most of the IE systems are focusing on English text; and most of them are in the supervised learning framework, which requires large amount of human labor; and most of them can only work in narrow domain, which is domain dependent. These systems are difficult to be ported to other languages, other domains because of these inherent shortcomings. Currently, besides western languages like English, there are many other Asian languages which are much different from English. In English, words are delimited by white-spaces so computer can easily tokenize the input text string. In many languages like Chinese, Japanese, Thai and Korea, they do not have word boundaries between words. This poses a difficult problem for the information extraction for those languages. In this thesis, we intend to implement a self-contained, language independent automatic IE system. The system is automatic because we are using a unsupervised learning framework in which no labeled data is required for training or a semi-supervised learning framework in which small amount of labeled data and large amount of unlabeled data are used. Specificly, we deal with Chinese and English languages name entity recognition and entity relation extraction, but the system can be easily extended to any other languages and other tasks. We implement an unsupervised Chinese word segmenter, a Chinese POS tagger, and we extend maximum entropy models to incorporate unlabeled data for general information extraction.