Research On Chinese Prepositional Phrase Identification Based On Multi-layer Conditional Random Fields

Prepositional phrases, as a class of important phrases, account for a rather large proportion in Chinese. Therefore, prepositional phrase identification has significant meaning which simplifies the structure of sentence, reduces the number of candidate main verbs and makes the parsing easily. In this paper, we present a system of prepositional phrase identification based on Conditional Random Fields (CRFs). Moreover, a transformation-based error-driven learning approach is adopted to revise the prepositional phrase identification results of CRFs model.This paper coverts the task of prepositional phrase identification into sequence labeling, and adopt CRFs model as our identification model. Through analyzing the structural characteristic of prepositional phrases, six features are extracted as our feature set and an effective feature template is selected based on incremental learning method. For the situation of more than one prepositional phrase existing in a sentence, in order to reduce the complexity of phrases and improve the accuracy of prepositional phrase identification, a multi-layer method, which identifies prepositional phrase from right to left based on CRFs and replace the identified preposition phrases, is proposed in this paper. For further improve the identification results, a transformation-based error-driven learning approach is adopted to revise the identification results based on CRFs.Experiment shows that, the multi-layer identification method based on CRFs is effective. Experiments carried out on the corpus of the People’s Daily2000containing more than7,000prepositional phrases, the precision, recall and F-value can achieve91.45%,91.39%and91.42%respectively. With the help of transformation-based error-driven learning, the performances of CRFs based prepositional phrase identification are improved to91.98%,91.92%and91.96%.Our research on prepositional phrase identification achieves better performance, which can apply to the fields of parsing, machine translation and so on.