Study On Invariance Of Text Representation In The Dynamic Context

This paper mainly investigates a variety of text representation methods,especially word embedding,the most popular word vector method.It was found that most of the papers focus on the training process of word representation and improve the final evaluation score.Therefore,this paper focused on the invariance of semantic representation of the word embedding.That is,though training words representation in different corpus,the vectors still held invariance of representation.This paper mainly includes the following aspects:1,Studied on the quantification representation of the space mapping relations of word vectors,and calculated the mapping matrix between two semantic spaces by several methods.Used the accuracy metrics to compare the differences between these methods and the reasons caused the differences.2,verified and analyzed on invariance of the vector differences and angles of word vectors base on the classical evaluation set.By comparing the different types of words,this paper found the reasons for the existence of these invariant characteristics,and analyzed what kind of words or word pairs have stronger representation invariance.3,Detected and analyzed semantic changes of words by the mapping relationship of word vectors and the invariance of word representation.The semantic change and semantic change direction of a word are analyzed,and the relationship between words and similar words is analyzed by clustering method.