Identification Of English Prepositional Phrases’ Syntactic Functions And Its Application In Translation

The number of English prepositions is small, but the prepositional phrases (PPs) they form are statistically frequent, functionally versatile and positionally flexible, which is likely to cause ambiguity in a sentence. In machine translation (MT), identification of English PPs’ syntactic functions is the precondition of shallow parsing, which can contribute to improving the quality of MT. This study aims to explore the identification of PPs’syntactic functions within business domain for the purpose of MT under the guidance of systemic functional grammar (SFG) and apply this recognition to PP translation.This thesis brought Halliday’s systemic functional grammar (SFG) into the recognition of PPs’syntactic functions in MT. Researchers in MT only focused on the methods used in the automatic identification system, and seldom paid attention to PPs and their syntactic functions. In MT, the binary PP attachment, that is, whether a PP is attached to the verb or the noun was explored, which was not delicate enough for the purpose of MT. Under the guidance of SFG, we refined PPs’syntactic functions into four types:adjunct (AD), complement (C), post-modifier (PM) and part of process (POP). This study combined the knowledge of part of speech (POS) to identify PPs’syntactic functions. The corpus we adopted was a small English-Chinese bilingual corpus in business domain. The POS tag set we used was Perm Treebank tag set with four adjustments being made for the purpose of PP recognition: distinguishing subordinating conjunctions from prepositions, defining particles in broad sense, defining functional words and distinguishing different syntactic functions of the word to. The construction of the training corpus was automatically completed with a manual correction. The identification system employed conditional random fields (CRFs), suitable for the corpus of a small scale. A comparable experiment was carried out to test the effectiveness of the adjustments made in the POS tag set. The corpus in Experiment One was annotated based on revised Perm Treebank tag set, while the corpus in Experiment Two was tagged using Perm Treebank tag set. In the two experiments, the same five-fold cross-validation tests were performed in order to have reliable results. Experimental results showed that the adjustments made in the POS tag set gave rise to5.06%,5.27%and5.14%increase in the precision, recall and F-score respectively. The accuracy achieved in our system was88.45%, higher than other reported approaches. Moreover, this research applied identification of PPs’ syntactic functions into PP translation. Translation of the top two PPs, OF-PPs and IN-PPs, was explored. Results proved that if the syntactic functions of PPs were different, their translation was different.