| Automated Essay Scoring(AES)is a typical application of Natural Language Processing(NLP)in the field of education.With the popularization of language learning,in order to examine language proficiency of language learners,the need for essay scoring has been further expanded.However,the traditional manual scoring method is time-consuming and laborious,and it is difficult to meet the corresponding needs,so the automatic essay scoring technology came into being.The existing automatic essay scoring technology mainly focuses on two types of scenarios: the in-prompt essay scoring(the essay to be graded and the existing scored essay come from the same essay prompt),and the cross-prompt essay scoring(the essay to be graded and the existing scored essay From different essay prompts).In the two types of scenarios,it is relatively easy to indicate the score in same prompt essay scoring,and the model's training data and test data have a consistent distribution of data.At present,good progress has been made in the in-prompt scenario.Although the cross-prompt essay scoring scenario is relatively practical,the model's training data and test data may differ in essay topics(writing content),test difficulty level,score range,writing genre,etc.,which gives essay The design of the scoring model brings great challenges.At present,the study of this scenario is not deep enough.Here we mainly research on the cross-prompt automatic essay scoring scenario from two aspects: aiming at the characteristics of the difference between the model training data and the test data on the essay theme,we study the prompt-aware automatic scoring model that is sensitive to essay prompts;Aiming at the characteristics of the difference in test difficulty level and score range between model training data and test data,a cross-prompt automatic scoring model based on small sample learning is studied.Specifically,the main work here is as follows:1 This paper presents a prompt-aware cross-prompt automatic essay scoring model.Existing cross-prompt automatic essay scoring method usually focus on evaluating the quality of essay and ignore the prompt-aware features.This is harmful for cross-prompt essay scoring since the score of essay may be different when the same essay is scored under different prompts.In order to solve this problem,we propose a prompt-aware neural network(PANN)model which emphasizes prompt adherence for cross-prompt automated essay scoring.Specifically,we design two sub-networks to capture the information of essay from the aspects of both prompt adherence and quality of writing.In addition,we design three weakly-supervised pre-training tasks to specialize the capabilities of the two sub-networks.Experiments on two public datasets demonstrate the effectiveness of PANN.2 This paper presents a cross-prompt automatic essay scoring model based on few-shot learning.Existing cross-prompt automatic essay scoring method usually only use the essay set with the same difficulty level and score range to conduct model training for the scoring scenarios,but this kind of scenario will make it impossible to automatically score some essays without the same difficulty level samples for training in the essay exams.In order to deal with this problem,this paper proposes a cross prompt automatic essay scoring model TGBN.The main idea is to make full use of the manifold structure between a small number of marked samples and a large number of unmarked samples to achieve the score alignment in the target essay set by using the graph based direct push learning method.Specifically,the model includes three main components: the essay encoder,the graph convolution module and the ordered classifier,which are respectively used to extract text features,manifold structure features and classification considering category ordering.In addition,in order to make full use of the historical essay data set,the model adopts the”segment” training strategy in meta-learning to train the model from the essay sets with different difficulty levels and score ranges in history,so as to achieve rapid adaptation on the target essay sets.Finally,the effectiveness of TGBN is verified by experiments on two public data sets. |
PDF Full Text Request