Composition scoring method based on attention mechanism

Abstract

The invention provides a composition scoring method based on an attention mechanism; the method comprises: using a neural network attention frame of word-sentence-document trilayer structure in a composition scoring system, using artificially extracted features to perform fusing in sentence and document levels of the frame, and setting attention weights of the sentence and document levels. The influences of factors, such as local features and global features of a language, completeness of sentences, accuracy of word usage, diversity of expressions, coherency of sentences and digressing or zero digressing from the subject, upon a scoring task are comprehensively considered herein, and composition scoring effect is maximally improved.

Description

technical field [0001] The invention relates to a composition scoring method, in particular to a composition scoring method based on an attention mechanism with a multi-hop structure. Background technique [0002] Compared with the manual scoring system, AES (Automated Essay Scoring) has the advantages of being more objective, efficient and low-cost. The composition scoring system constructed by AES has been successfully applied to GMAT, TOEFL, GRE and other examination systems in the United States. With the advancement of technology, the machine scoring system has gradually become a trend. Traditional automatic scoring systems rely on machine learning techniques such as natural language processing to model and analyze text through shallow text semantic features. However, text features need to be manually designed by experts in related fields, which is too expensive and wastes a lot of manpower. With the development of deep learning in recent years, technology based on dee...

AI Technical Summary

Problems solved by technology

Compared with natural language processing tasks such as sentiment analysis and text classification, the automatic grading of articles has a stronger priori. Traditional automatic grading systems for composition generally use manual extraction of features and statistical methods of multi-regression regression for analysis and evaluation. The number of words in a composition usually exceeds 500 words. Recurrent neural network models such as LSTM are difficult to capture long-distance information. Language modeling that relies solely on recurrent neural networks is not enough to capture complex language structures, and the existing automatic scoring technology is capable of analyzing the model. Not enough, the basis for grading the composition is not sufficient

Images