| Syntactic parsing is one of the key technologies in the field of Natural Language Processing (NLP), which plays an important role in NLP. Syntactic parsing is widely used in the area of machine translation, information retrieval, speech recognition research and so on. In some applications of machine translation, we only need syntactic phrase boundary information. Mainstream syntactic analysis methods not only generate syntactic phrase boundaries but also generate phrase labels, and further, in the process of analysis these methods need to use the phrase label information. The aim of this paper is only to get phrase boundaries of a syntactic tree, no phrase labels, and therefore we want to explore the feasibility of directly predicting phrase boundaries based on word and part-of-speech (POS) information without using phrase label information. The problem of boundary parsing based only on word and POS information is equal to the problem of syntactic parsing based on Treebank without phrase labels.In syntactic trees, there are two kinds of phrases, which are unary phrases and non-unary phrases. Because unary phrases account for a small proportion and separating them will decrease ambiguity, in this paper we first predict non-unary phrase boundaries and then predict unary phrase boundaries based on the results of non-unary phrase boundaries.In this paper, we present two possible ways of boundaries analysis, one is based on the level analysis method, and the other is based on the shift-reduce method. Boundary analysis based on hierarchical analysis can be considered as a multi-level chunking problem, in the order of bottom to up. This method achieves close to 85% of F-measure on the Penn English Treebank standard test set. Boundary analysis based on shift-reduce method is divided into two phases, firstly boundary analysis of the basic phrases, and then boundary analysis of nested phrase analysis. The boundary identification of the basic phrase uses chunking method, while nested phrase boundary analysis uses shift-reduce style method. We also test this method on Penn English Treebank standard test set and the F-measure is close to 88%. These two methods did not use phrase label information in the process of boundary analysis, and the results did not produce the phrase labels.At the end of end this paper, we do unary phrase boundary prediction on the basis of the results of above boundary analysis based on shift-reduce method. We convert this problem to a number prediction problem and predict it in the order of bottom to up by level. The results show that the method can effectively predict a unary phrase boundary.At last we get a complete boundary parser whose performance is comparable to Stanford parser. All the work in this paper shows that directly syntactic boundaries prediction based on word and POS information without phrase labels is feasible. |
