Research on recognition of semantic chunk boundary in Tibetan

Semantic chunk is able to well describe the sentence semantic framework. It plays a very important role in Natural Language Processing applications, such as machine translation, QA system and so on. At present, the Tibetan chunk researches are mainly based on rule-methods. In this paper, according to the distinctive language characteristics of Tibetan, we firstly put forward the descriptive definition of the Tibetan semantic chunk and its labeling scheme and then we propose a feature selection algorithm to select the suitable ones automatically from the candidate feature-templates. Through the experiment conducted on the two different kinds of Tibetan corpus, namely corpus-sentence and corpus-discourse, the F-Measure achieves 95.84%, 94.95% and 91.97%, 88.82% by using of Conditional Random Fields (CRF) model and Maximum Entropy (ME) model respectively. The positive results show that the definition of Tibetan semantic chunk in this paper is reasonable and operable. Furthermore, its boundary recognition is feasible and effective via statistical techniques in small scale corpus.