Research On Text Analysis Method Based On Pivots And Meta Learning

When Natural Language Processing(NLP)technology is applied to text analysis tasks,it has problems such as large differences in data characteristics and small amount of data in specific fields compared with general fields,which makes it difficult to train effective deep learning models.This article takes the Global View in Engineering Practice of Engineering College Students as an example to conduct research on text analysis methods in the context of sustainable development.Sustainable Development View is an important part of Engineering education.It contains moral values and social responsibility.At present,the quantitative and objective evaluation of such non-technical literacy is not perfect.With the help of artificial intelligence technology,the tester’s sustainable development view can be evaluated quantitatively to some extent,mainly through deep learning and natural language processing,to assist in judging what positive sustainable development ideas and attitudes the tester holds,and to assess the degree of recognition of the concept of sustainable development itself.When semantic understanding and categorization tasks in the field of natural language processing are applied to the above scenarios,there are several problems:(1)Pre-training language models have good generalization performance in the field of general knowledge and poor generalization performance in specific areas such as sustainable development,resulting in poor performance of the model in achieving task goals in this field.(2)The traditional classification method cannot meet the multi-label classification task with few samples.(3)The concepts and connotations of sustainable development have different official reference texts(UNESCO,CSSD,etc.).The single reference answer text similarity comparison method does not perform well in evaluating the cognitive level of the concept of sustainable development of Engineering students.In order to solve the above problems,this dissertation takes BERT as the basic network framework,and the main work is as follows:1.To solve the problem that the pre-training language model has good generalization performance in general domain and poor generalization performance in specific domain,this dissertation uses unsupervised domain adaptation method based on keyword guidance.This method integrates unsupervised data in the field of sustainable development,uses Chinese Whole word pre-training model,integrates pivots through word frequency statistics,improves the way of covering language model,and finally generates a fine-tuned model after training,which can achieve better results in the tasks downstream.2.In the evaluation of the view of sustainable development for Engineering College students,it is necessary to classify the descriptive texts answered by students in the questionnaire with multiple tags(multiple perspectives of sustainable development).Since such studies do not expose datasets,they are multi-label classification problems with few samples.This dissertation improves the Meta-Bert(Meta-Learning with BERT)model to make it suitable for multi-label classification tasks and achieve good classification results.3.In view of College Students’ cognitive problems(semantic understanding)of sustainable development concepts,this dissertation uses the semantic similarity analysis method combined with comparative learning to compare the text description of the understanding of sustainable development concepts with a variety of reference answer anchor samples(UNESCO and CSSD)set in the students’ answers,and comprehensively determines the similarity between students’ descriptions and reference answers,and maps them into score intervals.This is used to assess students’ awareness of the concept of sustainable development.The experimental results show that the methods and models presented in this dissertation are effective and have a certain degree of improvement compared with the traditional methods.