BUPT Systems in the SIGHAN Bakeoff 2007

Chinese Word Segmentation(WS), Name Entity Recognition(NER) and Part-OfSpeech(POS) are three important Chinese Corpus annotation tasks. With the great improvement in these annotations on some corpus, now, the robustness, a capability of keeping good performances for a system by automatically fitting the different corpus and standards, become a focal problem. This paper introduces the work on robustness of WS and POS annotation systems from Beijing University of Posts and Telecommunications(BUPT), and two NER systems. The WS system combines a basic WS tagger with an adaptor used to fit a specific standard given. POS taggers are built for different standards under a two step frame, both steps use ME but with incremental features. A multiple knowledge source system and a less knowledge Conditional Random Field (CRF) based systems are used for NER. Experiments show that our WS and POS systems are robust.