| Lexical substitution aims to find suitable substitutes for target words in sentences.Word substitution is applied to many downstream tasks in natural language processing,such as word simplification,word sense disambiguation and assisted writing.The latest class of word substitution methods have significantly improved the performance of word substitution,but they still have limitations.The current vocabulary replacement method is based on the language pre-training model BERT.The method based on BERT loses the semantic information of the target word due to the partial mask of the target word,and pays more attention to the words associated with the context information,and may eventually get sentences with completely opposite semantics;In addition,such methods can only generate words composed of one tag,and cannot generate words composed of two or more tags as candidate substitutes.In order to better deal with the above problems,this paper proposes two word-replacement methods based on the retelling model and an English-assisted writing system.The research contents of this paper are as follows:(1)In order to solve the problem that the word substitution method based on BERT cannot generate substitute words composed of two or more tokens,a word suvstitution method based on the retelling model is proposed This method proposes a decoding strategy that only focuses on the change of target words,which is used to generate substitute words from the retelling model.Compared with the existing word substitution methods,the method proposed in this paper has been significantly improved by evaluating on two widely used data sets LS07 and CoInCo.(2)In order to solve the problem that the word substitution method based on BERT relies on the information about the sequence structure when generating substitute words.We propose a word substitution method based on BART,and use BART model as the retelling model.Because BART is based on the standard sequence-to-sequence model,it is more suitable for generating tasks than the BERT-based method,and it also adds two-way context information.In addition,the substitute words are put into the original sentence,and the semantic relevance of the original sentence and the substituted sentence is evaluated with the evaluation indicators BERTScore and BLERT generated by the text.(3)An English writing assistant system is designed to help users better write in English.The system is divided into system configuration module,view module and function module,which are respectively responsible for the configuration of language model and WEB server;The page design and layout of the program,and the corresponding HTML file;Realize the functions of word substitution,sentence repetition and prompt generation. |
