Pronunciation-translated names (P-Names) bring more ambiguities to Chinese word segmentation and generic named entity recognition. As there are few annotated resources that can be used to develop a good P-Name extraction system, this paper presents a bootstrapping algorithm, called PN-Finder, to tackle this problem. Starting from a small set of P-Name characters and context cue-words, the algorithm iteratively locates more P-Names from the Internet. The algorithm uses a combination of P-Name and context word probabilities to identify new P-Names. Experiments show that our PN-Finder is able to locate a large number of P-Names (over 100,000) from the Internet with a high recognition accuracy of over 85%. Further tests on the MET-2 test set show that our PN-Finder can achieve a performance of over 90% in F1 value in locating P-Names. The results demonstrate that our PN-Finder is effective.
展开▼
