Text Representation Method Combining Multi- level Semantic Features

Abstract

The text vector representation transforms text from unstructured to structured, from high dimensional to low dimensional, and from sparse to dense, which is the basic task of text analysis. The senLDA model obtains the multinomial distribution of topics on the document based on the sentence, but due to the lack of semantic information for words, there is incomplete coverage of the high-value information and thus affects the effect of text representation. Aiming at this problem, a method that combines senLDA with Word2Vec's word-level features is proposed, which fuses three-level semantic features of words, sentences and documents to realize the text representation. F1 value of three datasets were increased by 11.41%, 17.88%, 17.63% respectively compared to the senLDA method, and increased by 4.65%, 7.73%, 8.62% respectively compared to Word2Vec.