Automatic Verification Of LSTM-RNN Based Language-model In K12 English Exercises

At present,there are a large number of students and a limited number of teachers in China.With the development of education and teaching,adaptive and efficient learning has become the demand of more and more teachers and students.In this environment,students need to be suitable for their ability level of the topic,the teacher needs to reduce the burden.The research and application of the adaptive testing system will be possible to effectively solve the appeal of teachers and students.This paper focuses on how to optimize the traditional title recommendation scheme that relies too much on the semaphore and exposure in the CAT title recommendation strategy,so as to increase its complexity,improve accuracy and improve the way of automatic verification.How to design an automated test scheme to reduce the burden on teachers and professionals in building question bank,provide them with a parameter reference for building question bank,and promote the development of automatic test;The debugging of the language model and the overall engineering application,how to complete from the word vector training to the final output of the language model topic sentence probability and explore how to use these probabilities to carry out the actual auxiliary automatic verification;Research methods of this article is mainly the engineering application practice,after establishing the model language to achieve auxiliary driving strategies to improve and question bank construction tasks,the actual used include high-performance crawler,large data sets clean language,word vector and neural network model technology,completed a project,and achieved the actual output.The conclusion of the final output shows that different sentences do have actual probability differences in the verification of the language model,and the language model can definitely be closer to the learned corpus and closer to common semantics by giving a higher average probability.Sentences that are less closely related to common semantics and learned corpora are negated by giving a lower average probability.Through these judgments,in the question bank construction stage and the topic recommendation screening stage,combined with other calculation information or feedback to the user a parameter information,so as to provide auxiliary decision-making.