Research And Application Of Intelligent Labeling Of Knowledge Points For Mathematical Problems Based On BiLSTM Model

Since the development plan of "deep integration of information technology and education" was proposed,there have been many excellent intelligent educational products in the field of mathematics,all of which need to be based on the accurate marking of knowledge points.However,the existing knowledge point annotation is basically done manually,which makes the annotation process time-consuming and laborious,and there are also problems such as missing and mis-marking.The use of intelligent annotation of mathematical questions to improve the efficiency of annotation and thus promote the development of social productivity is still relatively little researched.Therefore,the main problem of this thesis is how to implement an accurate and convenient intelligent labelling system for mathematical questions based on the characteristics of the mathematical domain.This thesis takes junior high school mathematics test text as the research object,aiming to optimize the intelligent labeling model for mathematics question knowledge points.Two models are proposed: the MFE-BiLSTM model combining mathematical feature extraction and the MFEHS-BiLSTM model for semantic enhancement of tag relevance.Finally,an intelligent tagging system for mathematical question knowledge points is designed and implemented based on the improved MFEHS-BiLSTM model.The specific work is as follows:(1)The deep semantic and temporal information in the text of mathematics test questions can seriously affect the accuracy of knowledge point labeling,which can be effectively solved by using a bi-directional long and short-term memory network(BiLSTM)as an encoder.However,BiLSTM lacks the ability to capture inferential information features in mathematical questions,so this thesis proposes a new encoder structure that can extract inferential information features,and together with the LSTM decoder structure based on the attention mechanism,we construct a coder-decoder model MFE-BiLSTM that incorporates numerical features extraction.The encoder adds a mathematical feature extraction part to the original BiLSTM encoder,uses the test text semantics to calculate the correlation with the a priori knowledge semantics,and performs a normalisation operation on multiple correlations to use them as inference information semantic weights,and then adopts a spliced feature fusion to incorporate them into the test text semantics,thus solving the problem of inference information extraction in mathematics test text.Compared with the best Att-BiLSTM encoder model in the baseline model,the MFE-BiLSTM model performs better on the multi-label mathematics test question dataset with a 4.3% reduction in Hamming loss,while the Micro-F1 improves by 5.7%.(2)As the decoder part of the MFE-BiLSTM model cannot effectively use the correlation between labels and the problems that the errors in label classification will continue to accumulate,this thesis proposes a two-channel decoder structure based on the hierarchical label structure model HMTC with LSTM.However,the HMTC model is not a good solution to the problems that many mathematical questions test multiple knowledge points that tend to appear in the same chapter,as well as the semantic similarity of some of the knowledge point labels.This thesis further optimises the HMTC model by proposing a new hierarchical tag structure model,HOSRHS,which incorporates the semantic relevance of horizontally oriented knowledge point tags,and solves the problem of semantic intersection of knowledge points by integrating the structural and semantic double embedding of tags.Further,it uses the sibling nodes in the hierarchical structure to influence the label probability of the current label node,thus solving the problem of knowledge point contribution.Finally,this new decoder structure is combined with the encoder structure combining mathematical feature extraction proposed in Chapter 3 to jointly construct the coder-decoder model MFEHS-BiLSTM combining mathematical feature extraction and semantic enhancement of the hierarchical label structure.experimental results show that,compared with the MFE-BiLSTM model,the MFE-BiLSTM model on the multi-label mathematics test data set has been further improved with a 3.5% reduction in Hamming loss and a 2.3% improvement in Micro-F1.(3)Based on the research results described above,this thesis further designs and implements an intelligent tagging system for mathematical questions through flask and Py Torch framework technology.Firstly,the overall system requirements are analysed,followed by a detailed description of the design methodology.Finally,this thesis uses the Py Torch framework to train the improved model and save the parameters,which are accessed according to the interface provided by flask,in order to implement an intelligent labelling system for maths question tags.