Research On Application Of Semantic Parsing And Language Generation In Machine Translation

Semantic parsing,language generation and machine translation are important research topics in natural language processing.Semantic parsing aims to transform natural language into semantic expression,while language generation in turn aims to transform semantic expression back to natural language.The main content of this paper includes:Focusing on semantic parsing,this paper explores the application of hierarchical phrase-based statistical machine translation and neural machine translation in semantic parsing.On one hand,considering the characteristics of semantic parsing tasks,this paper improves the hierarchical phrase-based translation framework from three aspects:(1)it introduces structure informed non-terminals,better guiding the parsing in favor of well formed structure,instead of using a uninformed non-terminal in synchronous context-free grammar(SCFG);(2)it explores different word alignments,and summarizes the proper word alignment for semantic parsing;and(3)it proposes a method for generating translation rules for unknown words via their synonyms.On the other hand,this paper explores multi-language neural semantic parsing.Taking advantage of multilingual inputs,this paper updates the traditional encode-decode model to a double-encode and decode model.Experimental results show that multilingual inputs can effectively improve the performance of semantic parsing.Focusing on language generation,this paper also explores the application of hierarchical phrase-based statistical machine translation and neural machine translation in language generation.On one hand,considering the characteristics of language generation tasks,this paper improves hierarchical phrase-based translation framework from two aspects:(1)it explores several different word alignments,and summarizes the proper word alignment for language generation;and(2)given n-best translation results,it proposes a filtering method by predicting the question answer types of the input and all translation results.On the other hand,this paper also explores the application of neural machine translation in multilingual language generation.Based on the traditional encode-decode framework,this paper proposes a double-decode model.The experimental results showthat the model can effectively improve the performance of language generation.Finally,this paper combines semantic parsing and language generation,and explores their application in machine translation.To this aim,it first translates a sentence of source language into a semantic expression and then converts the semantic expression into a sentence of target language.In order to alleviate the error propagation,this paper employs the n-best semantic expressions of semantic parsing to improve the performance of translation.