Solving Mathematical Application Problems In Primary And Secondary Schools Based On UniLM And Word Vector Adjustment Algorithm

As the informatization of education has risen to a national strategy,machine solutions to mathematics are the most important and difficult research problem in the informatization of education,which has great research value.At present,most of the cutting-edge math problem machine solutions are based on the deep learning pretraining language model.Among them,the Bert model is changed to the deep learning pre-training language model UniLM in the form of Seq2Seq.However,the deep learning pre-training language models introduced by technology companies such as Google and Microsoft are all trained using general knowledge,and are better at solving general tasks.However,the professional tasks such as solving mathematical application problems in primary and secondary schools are all inadequate in the two stages of pretraining and fine-tuning.First of all,in the pre-training stage of the deep learning pretraining language model,the model uses general large corpus training to train word vectors.The lack of professional math problem corpus training makes the model’s understanding of the concept and terminology of math problems inaccurate,leading to the model cannot pay attention to the most important details when seeking attention.If the task is determined,pre-training with professional domain corpus will lead to low corpus utilization and insufficient word vector training.In response to this problem,this paper proposes an algorithm for adjusting word vectors based on a knowledge tree,which uses corpus of professional fields to efficiently retune word vectors,and alleviates the inaccurate understanding of the concepts and terminology of the deep learning pre-training language model UniLM.Second,during the fine-tuning stage of the deep learning pre-training language model,the model is prone to over-learning,leading to phenomena such as over-fitting.In response to this problem,this paper proposes an improved method for the deep learning pre-training language model UniLM.In the primary and middle school mathematics application problem solving task,it reduces the difficulty of decoding the UniLM model,reduces the training parameters of the UniLM model,and achieves alleviation of over-fitting.In order to reduce the use of rules in the machine answering system,so that the machine answering system can get better and more intelligent answers faster,this paper aims at the insufficiency of the machine answering of math problems for primary and middle school students based on the deep learning language model,and makes the following innovative work:(1)The UniLM model is used as a control group to set up experiments.The accuracy of UniLM single-mode in the large public data set Ape210K is 4.61%higher than the deep learning model proposed by the domestic education giant Ape Guidance,which verifies the advanced nature of the UniLM model.A new data structure-knowledge tree is proposed.Use the word vector adjustment algorithm based on the knowledge tree to fine-tune the word vector again,and return the fine-tuned Embedding table to UniLM.This can alleviate the inaccurate understanding of the concepts and terminology of math problems in the UniLM model.And set up experiments to verify that the improved UniLM’s accuracy rate in the large public data set Ape210K is improved by 0.72%compared with UniLM single-mode,which verifies the effectiveness of adjusting the word vector.(2)It is observed that the mathematical expressions generated in the data set are composed of input and fixed mathematical symbols.According to the idea of Pointer Networks,a dynamic vocabulary composed of input and several fixed symbols is used to replace the search field of the UniLM model,and the search field is changed from 21128 It is reduced to about 20,innovatively changing the structure of the model,thereby reducing the difficulty of the decoding task.And set up experiments,in the large public data set Ape210K compared to UniLM single-mode accuracy rate increased by 0.8%,verifying the effectiveness of narrowing the search domain.The setup experiment verifies that the improved UniLM is compatible with the algorithm for adjusting the word vector based on the knowledge tree,and the use of both at the same time increases by 1.02%compared with the single-mode UniLM.