Language reading understanding system based on artificial intelligence

Abstract

The invention belongs to the technical field of natural language processing, and discloses a language reading understanding system based on artificial intelligence, and the system comprises a data collection module which is used for collecting language reading text data and reading understanding question data; the data processing module is used for preprocessing the collected language reading text data and reading understanding question data to obtain text feature data and understanding question feature data; answering the reading understanding question based on the text feature data to obtain a reading understanding question score; analyzing the influence of the understanding question feature data on answering the reading understanding question to obtain influence feature data; fusing the understanding question feature data, the influence feature data and the text feature data to obtain a comprehensive feature data set; the model training module is used for constructing an understanding question answer error prediction model, inputting the comprehensive feature data set into the understanding question answer error prediction model, and predicting to obtain a question solving answer error; and the learner is helped to quickly identify own defects so as to adjust a learning strategy.

Description

Technical Field

[0001] The present invention relates to the technical field of natural language processing, and more specifically, to a language reading comprehension system based on artificial intelligence. Background Art

[0002] With the development of artificial intelligence technology, especially the progress of natural language processing, intelligent education systems have played an increasingly important role in personalized teaching and automated assessment; however, existing language reading comprehension systems still have shortcomings in accuracy and flexibility, and it is difficult to effectively adapt to the needs of students of different age groups and reading levels.

[0003] The patent with the patent publication number CN111326040A discloses a language reading comprehension intelligent test and intelligent tutoring system and method. Among them, the system is the application of emerging technologies such as the Internet and artificial intelligence in the field of language reading comprehension teaching and evaluation, including an intelligent topic selection subsystem based on a language reading comprehension ability model, an automatic evaluation subsystem for language reading comprehension based on artificial intelligence, a personalized and hierarchical tutoring subsystem for language reading comprehension, and a language reading comprehension teaching auxiliary subsystem based on data mining. The system can intelligently select topics based on the learners' reading ability, and can automatically score and evaluate language reading comprehension questions, including subjective questions; and on the basis of accurately grasping the specific learning situation of each learner's reading comprehension, further targeted and individualized intelligent tutoring is implemented, so as to realize the teaching of language reading comprehension in accordance with students' aptitude. In addition, the system can also provide assistance for offline teaching through the generated data.

[0004] Existing language reading comprehension systems rely only on a single feature to evaluate answers, which may ignore other important information, resulting in incomplete and inaccurate evaluation results. Ignoring the comprehensive consideration of multilingual features may make it impossible for the evaluation system to accurately capture the deep meaning and logical relationship in the answer. Sorting candidate answers based on a single feature alone will lead to unreasonable sorting results, or even mislead users or the system to choose the wrong answer. It does not take into account that each student has their own unique learning habits and strengths and weaknesses, and cannot provide students with targeted learning suggestions and strategies, which may cause students to take detours in learning and waste time and energy. It does not use the problem-solving order impact model to guide students to adjust the problem-solving order. Students may solve problems in an unreasonable order, resulting in spending too much time on more difficult questions and rushing to answer easier questions, thus affecting overall learning efficiency.

[0005] In the existing technology, the collection and understanding of various types of incorrect answer text data such as deviations, errors in the main idea, errors in understanding details, and errors in reasoning are not considered. Without fine-grained error analysis, teachers may only have a general understanding of students' errors, but cannot accurately identify the specific types and causes of errors. This will lead to insufficient diagnostic accuracy and difficulty in formulating effective improvement measures. Without targeted improvement measures, students may not be able to effectively solve their own problems, resulting in a waste of time and energy on content that has already been mastered, and a failure to make substantial progress on content that has not been mastered.

[0006] In view of this, the present invention proposes a language reading comprehension system based on artificial intelligence to solve the above problems. Summary of the invention

[0007] In order to overcome the above-mentioned defects of the prior art and to achieve the above-mentioned purpose, the present invention provides the following technical solution: a language reading comprehension system based on artificial intelligence, comprising:

[0008] A data collection module, used to collect language reading text data and reading comprehension question data;

[0009] A data processing module is used to pre-process the collected language reading text data and reading comprehension question data to obtain text feature data and comprehension question feature data; answer the reading comprehension questions based on the text feature data to obtain the reading comprehension question scores;

[0010] Analyze the influence of comprehension question feature data on answering reading comprehension questions to obtain influence feature data; fuse comprehension question feature data, influence feature data and text feature data to obtain a comprehensive feature data set;

[0011] The model training module is used to build a comprehension question answer error prediction model, input the comprehensive feature data set into the comprehension question answer error prediction model, and predict the error of the answer to the question;

[0012] The error assessment module is used to compare the predicted answer error with the preset answer error threshold to determine whether the student's reading comprehension score is qualified;

[0013] The learning feedback module provides further learning suggestions if the student's reading comprehension score is qualified; if the student's reading comprehension score is unqualified, targeted feedback will be provided to point out the student's shortcomings and provide improvement measures; each module is connected by wired and / or wireless means.

[0014] Furthermore, the language reading text data includes lexical data, syntactic data, semantic data and text structure data; the reading comprehension question data includes multiple-choice questions, true-or-false questions, fill-in-the-blank questions and short-answer questions.

[0015] Furthermore, the method of preprocessing the collected language reading text data and reading comprehension question data to obtain text feature data and comprehension question feature data includes:

[0016] The density clustering algorithm is used to detect and eliminate abnormal data in the language reading text data and reading comprehension question data to obtain processed text feature data and comprehension question feature data. The text feature data and comprehension question feature data are normalized and converted into a standard normal distribution to obtain normalized text feature data and comprehension question feature data.

[0017] Furthermore, the method of answering reading comprehension questions based on text feature data and obtaining reading comprehension question scores includes:

[0018] Extract features from text feature data, use the part-of-speech tagging tool in the natural language processing library NLTK to extract part-of-speech tagging and lemmatization information, extract stem information through the stem extraction tool Porter Stemmer, create a mapping from part-of-speech tags to integer indexes through a hash table, mark each word in the text at the corresponding position in the feature vector according to its part-of-speech tag, lemmatization result, and stem extraction result, quantify the part-of-speech tag, lemmatization result, and stem extraction result through a hash function, and construct a lexical feature vector;

[0019] Segment the sentence into words or morphemes; assign part-of-speech tags to each word, including nouns, verbs, adjectives, adverbs, prepositions, conjunctions, and pronouns; use the dependency parser in the NLTK natural language processing library, input the preprocessed sentence into the parser, and obtain the sentence's dependency syntactic tree;

[0020] Dependency syntactic analysis is used to extract the syntactic structure of the sentence and form a syntactic feature vector:

[0021] S41. Traverse the dependency syntax tree, starting from the core verb of the sentence, and traverse the entire tree;

[0022] S42, identifying the subject-verb-object relationship, finding the node that points to the verb and is marked as the subject relationship, and finding the node that is pointed to by the verb and is marked as the object;

[0023] S43, identify the modifying relationship: identify the relationship between the attributive and the central word and the relationship between the adverbial and the central word;

[0024] S44, construct syntactic feature vector and define feature space: according to task requirements, define a set of syntactic features including subject type, predicate type, object type and main modification relationship type;

[0025] S45, using one-hot encoding to represent each relationship type; for a relationship type involving a specific word or part of speech, incorporating the relationship type into a syntactic feature vector;

[0026] S46, combining the extracted features of all relationship types into a complete syntactic feature vector;

[0027] Obtain word embeddings through the pre-trained BERT model, and use semantic role labeling SRL to extract the predicate-argument structure, and combine all relationship types to build a semantic feature vector; determine the feature dimensions that need to be extracted based on the specific content and structure of the text, extract the corresponding feature values ​​for each feature dimension, and arrange the extracted feature values ​​in a certain order to build a text structure feature vector;

[0028] The candidate answers and the preset standard answers are composed of text feature data and comprehension question feature data respectively; the similarity between the candidate answers and the preset standard answers is calculated, and they are sorted according to the similarity scores. The reading comprehension question score is calculated through the sorting and similarity scores of the candidate answers.

[0029] Furthermore, the method of calculating the similarity between the candidate answers and the preset standard answers, sorting them according to the similarity scores, and calculating the reading comprehension question score through the sorting of the candidate answers and the similarity scores includes:

[0030] A comprehensive similarity function is defined to measure the similarity between the candidate answer and the preset standard answer; the comprehensive similarity function is:

[0031]

[0032] Among them, Osim(A,A*) is the similarity between the candidate answer and the preset standard answer; is the similarity between the lexical feature vector and the preset standard lexical feature vector; is the similarity between the syntactic feature vector and the preset standard syntactic feature vector; is the similarity between the semantic feature vector and the preset standard semantic feature vector; is the similarity between the text structure feature vector and the preset standard text structure feature vector; α1, α2, α3 and α4 are weight factors of the similarity of each vector;

[0033] The candidate answers are evaluated using the comprehensive similarity function and sorted according to the similarity scores to obtain a sorted answer candidate list {A1, A2, ..., A n}; Calculate the reading comprehension score based on the ranking and similarity scores of the candidate answers;

[0034] The calculation formula for the reading comprehension question score is: Among them, A i is the i-th answer in the sorted candidate answer list; A * is the preset standard answer; i is the weight assigned according to the ranking position of the candidate answers; n is the total number of candidate answers; i is an index from 1 to n, used to traverse all candidate answers.

[0035] Furthermore, the method of analyzing the influence of comprehension question feature data on answering reading comprehension questions and fusing text feature data to obtain a comprehensive feature data set includes:

[0036] Collect students' problem-solving order O, problem-solving time T, and final answer data when answering reading comprehension questions; calculate the average problem-solving time for students to answer each reading comprehension question: Where, is the average problem-solving time; is the time taken by the jth student to solve the i′th problem; N is the total number of students;

[0037] Define a problem-solving order influence model to calculate the influence of the problem-solving order on the correct rate of the i′th question; the problem-solving order influence model is: Odr = f(Odr i′ ,Ur i′ );Odr i′ is the order of solving the i′th problem; Ur i′ is the correct rate of the i′th question;

[0038] The correct rate of the i′th question is: Among them, Ur i′,j is the answer of the jth student to the i′th question, 1 for correct answer and 0 for wrong answer;

[0039] Construct a multivariate regression model to quantify the combined effects of problem-solving order and problem-solving time on the accuracy of answering reading comprehension questions, and obtain the influencing characteristic data;

[0040] The multivariate regression model is: Uri′,j=δ+ε1·Oi′,j+ε2·Ti′,j+∈i′,j; where O i′,j The order in which the jth student answers the i′th question; T i′,j is the time it takes for the jth student to answer the i′th question; δ is a constant term; ε1 is the coefficient of the order of solving questions on the correct rate of answering questions; ε2 is the coefficient of the time of solving questions on the correct rate of answering questions; ∈i ′ ,j is the error term of the multivariate regression model;

[0041] According to the results of the multivariate regression model analysis, optimize students' problem-solving strategies; if ε1 is a negative value, it is recommended that students put more difficult or more important questions first; if ε2 is a positive value, it is recommended that students allocate time reasonably and spend enough time thinking on more difficult questions; if ε1 is a positive value, it is recommended that students quickly browse the overall difficulty of the questions and give priority to answering easier questions; if ε2 is a negative value, it is recommended that students pay attention to time management and avoid spending too much time on a certain question;

[0042] The comprehension question feature data, impact feature data and text feature data are weighted and fused to obtain a comprehensive feature data set.

[0043] Furthermore, the training method of the comprehension question answer error prediction model includes:

[0044] Divide the dataset into training set, validation set and test set, train the model and evaluate the model performance; build the error prediction model of the answer to the comprehension question, including the input layer, GRU layer, fully connected layer and output layer;

[0045] The input layer of the comprehension question answer error prediction model is a historical comprehensive feature dataset, and the number of neurons in the input layer should match the number of features in the historical comprehensive feature dataset. The GRU layer is used to process the historical comprehensive feature dataset, and the number of GRU layers and neurons are adjusted according to the complexity of the task. The fully connected layer is used to provide additional nonlinear transformations.

[0046] The output layer of the model is the reading comprehension question answer error. The number of neurons in the output layer corresponds to the number of prediction targets. The prediction value is output through one neuron, and the identity function is used as the activation function. The comprehension question answer error prediction model is a gated recurrent unit model.

[0047] The mean square error is used as the loss function to measure the error between the model's predicted value and the actual value; the mean square error loss function is: Where n is the number of data sets; Y i is the actual value of data point i in the data set; is the predicted value of data point i in the data set;

[0048] Use the training set data to train the model, minimize the loss function through the Adam optimizer, and tune the model's hyperparameters; evaluate the model through the test set, stop testing when the model's performance in the prediction task reaches the preset performance threshold, and obtain the comprehension question answer error prediction model; input the real-time comprehensive feature data set into the comprehension question answer error prediction model to obtain the real-time problem-solving answer error.

[0049] Furthermore, the method for tuning the hyperparameters of the model includes:

[0050] For the comprehension question answer error prediction model, a grid space is set, wherein the grid space includes at least two groups of hyperparameter combinations, and each group of hyperparameters varies within a preset range;

[0051] For each set of hyperparameter combinations in the grid space, use these hyperparameters to configure the training model and use accuracy to evaluate the performance of the model; compare the model performance corresponding to each set of hyperparameters, select the set of hyperparameters with the highest accuracy, and set them as the optimal hyperparameters for the comprehension question answer error prediction model.

[0052] Furthermore, the method of comparing the predicted answer error with a preset answer error threshold to determine whether the student's reading comprehension score is qualified includes:

[0053] If the predicted answer error is less than the preset answer error threshold, the student's reading comprehension score is considered qualified;

[0054] If the predicted answer error is greater than or equal to the preset answer error threshold, the student's reading comprehension score is judged to be unqualified.

[0055] Furthermore, if a student fails to pass the reading comprehension test, targeted feedback will be provided to point out the student's shortcomings and provide improvement measures, including:

[0056] If the student's reading comprehension score is not satisfactory, the student's wrong answer text data is collected, and the wrong answer text data includes comprehension deviation data, main idea error data, detail comprehension error data and reasoning error; the wrong answer text data is preprocessed and converted into a feature sequence;

[0057] Assume that X is the text feature sequence of students’ wrong answers, and Y is the corresponding error type label sequence. The sequence probability labeling model is used to maximize the probability of the corresponding error type label sequence Y given the text feature sequence X of students’ wrong answers.

[0058] The sequence probability labeling model is: Among them, Y is the error type label sequence; T′ is the length of the sequence; X is the text feature sequence of the student’s wrong answer; y t is the error type label at time step t; y t-1 is the error type label at time step t-1; P(y t |y t-1 ,X) is given the previous label y t-1 and the student's incorrect answer to the text feature sequence X, the tth label is y t The conditional probability of

[0059] The conditional probability is defined by a feature transfer model, which is:

[0060] Among them, f k (y t-1 ,X) is the feature function defined on the given error type label sequence Y and student error answer text feature sequence X; k is the index of the feature function; λ k is the characteristic function f k The associated weight parameter; Z(y t-1 ,X) is the partition function, which makes the sum of the probabilities of all labels being wrong type labels yt equal to 1;

[0061] Construct a weighted optimization model and preset the weight parameter λ k The maximum threshold is λ max , weight parameter λ k The minimum threshold is λ min ; Combine the weighted optimization model and the weight parameter threshold to adjust the weight parameter λ k to limit the scope of

[0062] The weighted optimization model is: Where P(Y|X) is the probability of a given student’s wrong answer text feature sequence X corresponding to an error type label sequence Y; N′ is the number of error type label sequences Y;

[0063] Calculate the weight parameter λ through the weighted optimization model k , select multiple groups of weight parameters with the minimum threshold λ min and the maximum threshold value of the weight parameter λ max The intermediate weight parameter is substituted into the feature transfer model to calculate the given previous label y t-1 and the student's incorrect answer to the text feature sequence X, the tth label is y t Multiple conditional probabilities P′;

[0064] From multiple conditional probabilities P′, select the highest probability P corresponding to the error type label sequence Y under the given student's wrong answer text feature sequence X max , the weight parameter λ at this time k is the optimal weight parameter; through the feature transfer model, the teacher automatically identifies the error types in the students' wrong answer text data, and provides targeted feedback and improvement measures accordingly.

[0065] The technical effects and advantages of the language reading comprehension system based on artificial intelligence of the present invention are as follows:

[0066] The present invention captures the similarities and differences between candidate answers and standard answers more comprehensively by comprehensively considering multiple language features such as lexical, syntactic, semantic and text structure; more accurately evaluates the quality of candidate answers and reduces the deviation caused by single feature evaluation; sorts candidate answers according to similarity scores and calculates the final score based on the sorting position and the total number of candidate answers, taking into account the accuracy of the answer and the confidence of the answer; helps to select the most likely correct answer from multiple candidate answers and give a reasonable score;

[0067] By collecting data on students' problem-solving order, problem-solving time, and final answers when answering reading comprehension questions, we can analyze each student's personalized learning habits and strengths and weaknesses. Based on these data, we can provide customized learning suggestions for each student to help them optimize their problem-solving strategies. The problem-solving order impact model can reveal the impact of different question orders on students' answer accuracy. By adjusting the problem-solving order, students can achieve a higher accuracy rate in a shorter time, thereby improving learning efficiency.

[0068] By collecting and understanding various types of incorrect answer text data, such as deviations, errors in the main idea, errors in understanding details, and errors in reasoning, students' errors can be analyzed in a fine-grained manner; this helps to diagnose students' deficiencies more accurately; using the sequence probability labeling model and the feature transfer model, teachers can automatically identify the types of errors in students' incorrect answer text data; automated identification not only improves efficiency, but also reduces the subjectivity and errors of human judgment; through targeted feedback and improvement measures, students can focus more on improving their own deficiencies and avoid wasting time and energy on content they have already mastered; after clarifying their own deficiencies and obtaining specific improvement measures, students can study and practice more specifically, thereby accelerating their own progress. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 A schematic diagram of the structure of a language reading comprehension system based on artificial intelligence of the present invention;

[0070] Figure 2 The present invention is a flow chart of a language reading comprehension method based on artificial intelligence. DETAILED DESCRIPTION

[0071] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0072] Embodiment 1

[0073] See also Figure 1 As shown, the language reading comprehension system based on artificial intelligence described in this embodiment includes:

[0074] A data collection module, used to collect language reading text data and reading comprehension question data;

[0075] A data processing module is used to pre-process the collected language reading text data and reading comprehension question data to obtain text feature data and comprehension question feature data; answer the reading comprehension questions based on the text feature data to obtain the reading comprehension question scores;

[0076] Analyze the influence of comprehension question feature data on answering reading comprehension questions to obtain influence feature data; fuse comprehension question feature data, influence feature data and text feature data to obtain a comprehensive feature data set;

[0077] The model training module is used to build a comprehension question answer error prediction model, input the comprehensive feature data set into the comprehension question answer error prediction model, and predict the error of the answer to the question;

[0078] The error assessment module is used to compare the predicted answer error with the preset answer error threshold to determine whether the student's reading comprehension score is qualified;

[0079] The learning feedback module provides further learning suggestions if the student's reading comprehension score is qualified; if the student's reading comprehension score is unqualified, targeted feedback will be provided to point out the student's shortcomings and provide improvement measures; each module is connected by wired and / or wireless means.

[0080] The language reading text data includes lexical data, syntactic data, semantic data and text structure data; the reading comprehension question data includes multiple-choice questions, true-or-false questions, fill-in-the-blank questions and short-answer questions.

[0081] The method of preprocessing the collected language reading text data and reading comprehension question data to obtain text feature data and comprehension question feature data comprises:

[0082] The density clustering algorithm is used to detect and eliminate abnormal data in the language reading text data and reading comprehension question data to obtain processed text feature data and comprehension question feature data. The text feature data and comprehension question feature data are normalized and converted into a standard normal distribution to obtain normalized text feature data and comprehension question feature data.

[0083] The method for answering reading comprehension questions based on text feature data and obtaining reading comprehension question scores includes:

[0084] Extract features from text feature data, use the part-of-speech tagging tool in the natural language processing library NLTK to extract part-of-speech tagging and lemmatization information, extract stem information through the stem extraction tool Porter Stemmer, create a mapping from part-of-speech tags to integer indexes through a hash table, mark each word in the text at the corresponding position in the feature vector according to its part-of-speech tag, lemmatization result, and stem extraction result, quantify the part-of-speech tag, lemmatization result, and stem extraction result through a hash function, and construct a lexical feature vector;

[0085] For example, if the text is "I like watching movies", the possible results after part-of-speech tagging are: ["I / r", "like / v", "watch / v", "movie / n"], where "r" represents a pronoun, "v" represents a verb, and "n" represents a noun; for the word "watch i ng", the basic form that may be obtained after word form restoration is "watch";

[0086] Suppose there is a simple text "I am watching a movie", and it has been segmented, POS tagged, lemmatized and stemmed; word segmentation: ["I","am","watch i ng","a","movie"]; POS tagging: ["I / PRP","am / VBP","watch i ng / VBG","a / DT","movie / NN"]; lemmatization: ["I","be","watch","a","movie"]; stemming: ["i","am","watch","a","movi"].

[0087] Segment the sentence into words or morphemes; assign part-of-speech tags to each word, including nouns, verbs, adjectives, adverbs, prepositions, conjunctions, and pronouns; use the dependency parser in the NLTK natural language processing library, input the preprocessed sentence into the parser, and obtain the sentence's dependency syntactic tree;

[0088] Dependency syntactic analysis is used to extract the syntactic structure of the sentence and form a syntactic feature vector:

[0089] S41. Traverse the dependency syntax tree, starting from the core verb of the sentence, and traverse the entire tree;

[0090] S42, identifying the subject-verb-object relationship, finding the node that points to the verb and is marked as the subject relationship, and finding the node that is pointed to by the verb and is marked as the object;

[0091] S43, identify the modifying relationship: identify the relationship between the attributive and the central word and the relationship between the adverbial and the central word;

[0092] S44, construct syntactic feature vector and define feature space: according to task requirements, define a set of syntactic features including subject type, predicate type, object type and main modification relationship type;

[0093] S45, using one-hot encoding to represent each relationship type; for a relationship type involving a specific word or part of speech, incorporating the relationship type into a syntactic feature vector;

[0094] S46, combining the extracted features of all relationship types into a complete syntactic feature vector;

[0095] For example, suppose the sentence is "Xiao Ming likes watching movies". After dependency syntactic analysis, the following relations may be obtained: SBV (like, Xiao Ming); VOB (like, watching movies); VOB (watching, movies); the actual analysis results may vary depending on the tool, and "watching movies" may be identified as a verb phrase, directly forming a VOB relationship with "like", rather than being further decomposed into VOB (watching, movies); among them, SBV is used to indicate the relationship between the subject and the predicate, and VOB represents the verb-object relationship.

[0096] Obtain word embeddings through the pre-trained BERT model, and use semantic role labeling SRL to extract the predicate-argument structure, and combine all relationship types to build a semantic feature vector; determine the feature dimensions that need to be extracted based on the specific content and structure of the text, extract the corresponding feature values ​​for each feature dimension, and arrange the extracted feature values ​​in a certain order to build a text structure feature vector;

[0097] For example, take the sentence "Xiao Ming eats apples" as an example, where "eat" is the predicate, "Xiao Ming" and "apple" are arguments; the predicate-argument structure of this sentence can be expressed as: predicate: "eat"; argument 1 (agent): "Xiao Ming" (subject); argument 2 (patient): "apple" (object);

[0098] This structure shows that the action of "eating" is initiated by "Xiao Ming" and acts on "apple".

[0099] The candidate answers and the preset standard answers are composed of text feature data and comprehension question feature data respectively; the similarity between the candidate answers and the preset standard answers is calculated, and they are sorted according to the similarity scores. The reading comprehension question score is calculated through the sorting and similarity scores of the candidate answers.

[0100] The method of calculating the similarity between the candidate answers and the preset standard answers, sorting them according to the similarity scores, and calculating the reading comprehension question score through the sorting and similarity scores of the candidate answers includes:

[0101] A comprehensive similarity function is defined to measure the similarity between the candidate answer and the preset standard answer; the comprehensive similarity function is:

[0102]

[0103] Among them, Osim(A,A*) is the similarity between the candidate answer and the preset standard answer; is the similarity between the lexical feature vector and the preset standard lexical feature vector; is the similarity between the syntactic feature vector and the preset standard syntactic feature vector; is the similarity between the semantic feature vector and the preset standard semantic feature vector; is the similarity between the text structure feature vector and the preset standard text structure feature vector; α1, α2, α3 and α4 are weight factors of the similarity of each vector;

[0104] The candidate answers are evaluated using the comprehensive similarity function and sorted according to the similarity scores to obtain a sorted answer candidate list {A1, A2, ..., A n}; Calculate the reading comprehension score based on the ranking and similarity scores of the candidate answers;

[0105] The calculation formula for the reading comprehension question score is: Among them, A i is the i-th answer in the sorted candidate answer list; A * is the preset standard answer; i The weights assigned according to the ranking of candidate answers, usually ω1>ω2>…>ω n ; n is the total number of candidate answers; i is an index from 1 to n, used to traverse all candidate answers.

[0106] The method of analyzing the influence of comprehension question feature data on answering reading comprehension questions and fusing text feature data to obtain a comprehensive feature data set includes:

[0107] Collect students' problem-solving order O, problem-solving time T, and final answer data when answering reading comprehension questions; calculate the average problem-solving time for students to answer each reading comprehension question: Where, is the average problem-solving time; is the time taken by the jth student to solve the i′th problem; N is the total number of students;

[0108] Define a problem-solving order influence model to calculate the influence of the problem-solving order on the correct rate of the i′th question; the problem-solving order influence model is: Odr = f(Odr i′ ,Ur i′ );Odr i′ is the order of solving the i′th problem; Ur i′is the correct rate of the i′th question;

[0109] The correct rate of the i′th question is: Among them, Ur i′,j is the answer of the jth student to the i′th question, 1 for correct answer and 0 for wrong answer;

[0110] Construct a multivariate regression model to quantify the combined effects of problem-solving order and problem-solving time on the accuracy of answering reading comprehension questions, and obtain the influencing characteristic data;

[0111] The multivariate regression model is: Uri′,j=δ+ε1·Oi′,j+ε2·Ti′,j+∈i′,j; where O i′,j The order in which the jth student answers the i′th question; T i′,j is the time it takes for the jth student to answer the i′th question; δ is a constant term; ε1 is the coefficient of the order of solving questions on the correct rate of answering questions; ε2 is the coefficient of the time of solving questions on the correct rate of answering questions; ∈ i′,j is the error term of the multivariate regression model;

[0112] According to the results of the multivariate regression model analysis, the students' problem-solving strategies are optimized; if ε1 is a positive value, it means that as the problem-solving order increases, that is, the questions are placed later, the correct answer rate increases, which means that the students gradually get into the state in the process of solving the previous questions, and their thinking becomes more active and focused, thus performing better in the later questions; if ε1 is a negative value, it means that as the problem-solving order increases, that is, the questions are placed later, the correct answer rate decreases, which may be due to student fatigue or decreased attention;

[0113] If ε2 is a positive value, it means that the longer the time to solve the problem, the higher the correct answer rate, which may be because the students put more thought into the problem; if ε2 is a negative value, it means that the longer the time to solve the problem, the lower the correct answer rate; it may be that some students may spend too much time on some questions, but did not find the correct solution, but wasted time, resulting in a decrease in the correct rate;

[0114] If ε1 is a negative value, it is recommended that students answer more difficult or more important questions first; if ε2 is a positive value, it is recommended that students allocate time reasonably and spend enough time thinking on more difficult questions; if ε1 is a positive value, it is recommended that students quickly browse the overall difficulty of the questions and answer easier questions first; if ε2 is a negative value, it is recommended that students pay attention to time management and avoid spending too much time on a certain question;

[0115] The comprehension question feature data, impact feature data and text feature data are weighted and fused to obtain a comprehensive feature data set.

[0116] The training method of the comprehension question answer error prediction model includes:

[0117] Divide the dataset into training set, validation set and test set, train the model and evaluate the model performance; build the error prediction model of the answer to the comprehension question, including the input layer, GRU layer, fully connected layer and output layer;

[0118] The input layer of the comprehension question answer error prediction model is a historical comprehensive feature dataset, and the number of neurons in the input layer should match the number of features in the historical comprehensive feature dataset. The GRU layer is used to process the historical comprehensive feature dataset, and the number of GRU layers and neurons are adjusted according to the complexity of the task. The fully connected layer is used to provide additional nonlinear transformations.

[0119] The output layer of the model is the reading comprehension question answer error. The number of neurons in the output layer corresponds to the number of prediction targets. The prediction value is output through one neuron, and the identity function is used as the activation function. The comprehension question answer error prediction model is a gated recurrent unit model.

[0120] The mean square error is used as the loss function to measure the error between the model's predicted value and the actual value; the mean square error loss function is: Where n is the number of data sets; Y i is the actual value of data point i in the data set; is the predicted value of data point i in the data set;

[0121] Use the training set data to train the model, minimize the loss function through the Adam optimizer, and tune the model's hyperparameters; evaluate the model through the test set, stop testing when the model's performance in the prediction task reaches the preset performance threshold, and obtain the comprehension question answer error prediction model; input the real-time comprehensive feature data set into the comprehension question answer error prediction model to obtain the real-time problem-solving answer error.

[0122] The method for tuning the hyperparameters of the model includes:

[0123] For the comprehension question answer error prediction model, a grid space is set, wherein the grid space includes at least two groups of hyperparameter combinations, and each group of hyperparameters varies within a preset range;

[0124] For each set of hyperparameter combinations in the grid space, use these hyperparameters to configure the training model and use accuracy to evaluate the performance of the model; compare the model performance corresponding to each set of hyperparameters, select the set of hyperparameters with the highest accuracy, and set them as the optimal hyperparameters for the comprehension question answer error prediction model.

[0125] The method of comparing the predicted answer error with a preset answer error threshold to determine whether the student's reading comprehension score is qualified includes:

[0126] If the predicted answer error is less than the preset answer error threshold, the student's reading comprehension score is considered qualified;

[0127] If the predicted answer error is greater than or equal to the preset answer error threshold, the student's reading comprehension score is judged to be unqualified.

[0128] If a student fails the reading comprehension test, targeted feedback will be provided to point out the student's shortcomings and provide improvement measures, including:

[0129] If the student's reading comprehension score is not satisfactory, the student's wrong answer text data is collected, and the wrong answer text data includes comprehension deviation data, main idea error data, detail comprehension error data and reasoning error; the wrong answer text data is preprocessed and converted into a feature sequence;

[0130] Assume that X is the text feature sequence of students’ wrong answers, and Y is the corresponding error type label sequence. The sequence probability labeling model is used to maximize the probability of the corresponding error type label sequence Y given the text feature sequence X of students’ wrong answers.

[0131] The sequence probability labeling model is: Among them, Y is the error type label sequence; T′ is the length of the sequence; X is the text feature sequence of the student’s wrong answer; y t is the error type label at time step t; y t-1 is the error type label at time step t-1; P(y t |y t-1 ,X) is given the previous label y t-1 and the student's incorrect answer to the text feature sequence X, the tth label is y t The conditional probability of

[0132] The conditional probability is defined by a feature transfer model, which is:

[0133] Among them, f k (y t-1 ,X) is the feature function defined on the given error type label sequence Y and student error answer text feature sequence X; k is the index of the feature function; λ k is the characteristic function f k The associated weight parameter; Z(y t-1 ,X) is the partition function, which makes the sum of the probabilities of all labels being wrong type labels yt equal to 1;

[0134] Construct a weighted optimization model and preset the weight parameter λ k The maximum threshold is λ max , weight parameter λk The minimum threshold is λ min ; Combine the weighted optimization model and the weight parameter threshold to adjust the weight parameter λ k to limit the scope of

[0135] The weighted optimization model is: Where P(Y|X) is the probability of a given student’s wrong answer text feature sequence X corresponding to an error type label sequence Y; N′ is the number of error type label sequences Y;

[0136] Calculate the weight parameter λ through the weighted optimization model k , select multiple groups of weight parameters with the minimum threshold λ min and the maximum threshold value of the weight parameter λ max The intermediate weight parameter is substituted into the feature transfer model to calculate the given previous label y t-1 and the student's incorrect answer to the text feature sequence X, the tth label is y t Multiple conditional probabilities P′;

[0137] From multiple conditional probabilities P′, select the highest probability P corresponding to the error type label sequence Y under the given student's wrong answer text feature sequence X max , the weight parameter λ at this time k is the optimal weight parameter; through the feature transfer model, the teacher automatically identifies the error types in the students' wrong answer text data, and provides targeted feedback and improvement measures accordingly.

[0138] The preset answer error threshold is set by the staff. Different answer errors are collected through the language reading comprehension management terminal, and the average of multiple answer errors is taken as the preset distance threshold. Similarly, the preset performance threshold, weight parameter minimum threshold and weight parameter maximum threshold are set.

[0139] This embodiment, by comprehensively considering multiple language features such as lexical, syntactic, semantic and text structure, more comprehensively captures the similarities and differences between candidate answers and standard answers; more accurately evaluates the quality of candidate answers and reduces the deviation caused by single feature evaluation; sorts candidate answers according to similarity scores, and calculates the final score in combination with the sorting position and the total number of candidate answers, taking into account the accuracy of the answer and the confidence of the answer; helps to select the most likely correct answer from multiple candidate answers and give a reasonable score; by collecting the order, time and final answer data of students when answering reading comprehension questions, the personalized learning habits and strengths and weaknesses of each student can be analyzed; based on these data, customized learning suggestions are provided for each student to help students optimize their problem-solving strategies; the problem-solving order influence model can reveal the impact of different question orders on the accuracy of students' answers; by adjusting the problem-solving order, students can obtain a higher accuracy rate in a shorter time, thereby improving learning efficiency;

[0140] By collecting and understanding various types of incorrect answer text data, such as deviations, errors in the main idea, errors in understanding details, and errors in reasoning, students' errors can be analyzed in a fine-grained manner; this helps to diagnose students' deficiencies more accurately; using the sequence probability labeling model and the feature transfer model, teachers can automatically identify the types of errors in students' incorrect answer text data; automated identification not only improves efficiency, but also reduces the subjectivity and errors of human judgment; through targeted feedback and improvement measures, students can focus more on improving their own deficiencies and avoid wasting time and energy on content they have already mastered; after clarifying their own deficiencies and obtaining specific improvement measures, students can study and practice more specifically, thereby accelerating their own progress.

[0141] Embodiment 2

[0142] See also Figure 2 As shown, the part not described in detail in this embodiment is described in Example 1, which provides a language reading comprehension method based on artificial intelligence, including:

[0143] S1, collect language reading text data and reading comprehension question data;

[0144] S2. Preprocessing the collected language reading text data and reading comprehension question data to obtain text feature data and comprehension question feature data; answering the reading comprehension questions based on the text feature data to obtain reading comprehension question scores;

[0145] Analyze the influence of comprehension question feature data on answering reading comprehension questions to obtain influence feature data; fuse comprehension question feature data, influence feature data and text feature data to obtain a comprehensive feature data set;

[0146] S3, constructing a comprehension question answer error prediction model, inputting the comprehensive feature data set into the comprehension question answer error prediction model, and predicting the error of the answer to the question;

[0147] S4, comparing the predicted answer error with a preset answer error threshold to determine whether the student's reading comprehension score is qualified;

[0148] S5. If the student's reading comprehension score is qualified, further learning suggestions will be given; if the student's reading comprehension score is unqualified, targeted feedback will be provided to point out the student's shortcomings and provide improvement measures.

[0149] Since the electronic device introduced in this embodiment is an electronic device used to implement a language reading comprehension system based on artificial intelligence in the embodiment of the present application, based on the language reading comprehension system based on artificial intelligence introduced in the embodiment of the present application, a person skilled in the art can understand the specific implementation of the electronic device of the present embodiment and its various variations, so how the electronic device implements the method in the embodiment of the present application is not described in detail here. As long as a person skilled in the art implements an electronic device used by a language reading comprehension system based on artificial intelligence in the embodiment of the present application, it belongs to the scope of protection of this application.

[0150] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters and thresholds in the formula are set by technicians in this field according to actual conditions.

[0151] The above is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technical users in this technical field, some improvements and modifications without departing from the principle of the present invention should also be regarded as the protection scope of the present invention.

Claims

1. A language reading comprehension system based on artificial intelligence, characterized in that: include: A data collection module, used to collect language reading text data and reading comprehension question data; A data processing module, used for preprocessing the collected language reading text data and reading comprehension question data to obtain text feature data and comprehension question feature data; Answer reading comprehension questions based on text feature data and obtain reading comprehension question scores; Analyze the influence of comprehension question feature data on answering reading comprehension questions to obtain influence feature data; fuse comprehension question feature data, influence feature data and text feature data to obtain a comprehensive feature data set; The model training module is used to build a comprehension question answer error prediction model, input the comprehensive feature data set into the comprehension question answer error prediction model, and predict the error of the answer to the question; The error assessment module is used to compare the predicted answer error with the preset answer error threshold to determine whether the student's reading comprehension score is qualified; The learning feedback module provides further learning suggestions if the student's reading comprehension score is qualified; if the student's reading comprehension score is unqualified, targeted feedback will be provided to point out the student's shortcomings and provide improvement measures; each module is connected by wired and / or wireless means.

2. The artificial intelligence-based language reading comprehension system according to claim 1, characterized in that: The language reading text data includes lexical data, syntactic data, semantic data and text structure data; the reading comprehension question data includes multiple-choice questions, true-or-false questions, fill-in-the-blank questions and short-answer questions.

3. The artificial intelligence-based language reading comprehension system according to claim 2, characterized in that: The method of preprocessing the collected language reading text data and reading comprehension question data to obtain text feature data and comprehension question feature data comprises: The density clustering algorithm is used to detect and eliminate abnormal data in the language reading text data and reading comprehension question data to obtain processed text feature data and comprehension question feature data. The text feature data and comprehension question feature data are normalized and converted into a standard normal distribution to obtain normalized text feature data and comprehension question feature data.

4. The artificial intelligence-based language reading comprehension system according to claim 3, characterized in that: The method for answering reading comprehension questions based on text feature data and obtaining reading comprehension question scores includes: Extract features from text feature data, use the part-of-speech tagging tool in the natural language processing library NLTK to extract part-of-speech tagging and lemmatization information, extract stem information through the stem extraction tool Porter Stemmer, create a mapping from part-of-speech tags to integer indexes through a hash table, mark each word in the text at the corresponding position in the feature vector according to its part-of-speech tag, lemmatization result, and stem extraction result, quantify the part-of-speech tag, lemmatization result, and stem extraction result through a hash function, and construct a lexical feature vector; Segment the sentence into words or morphemes; assign part-of-speech tags to each word, including nouns, verbs, adjectives, adverbs, prepositions, conjunctions, and pronouns; use the dependency parser in the NLTK natural language processing library, input the preprocessed sentence into the parser, and obtain the sentence's dependency syntactic tree; Dependency syntactic analysis is used to extract the syntactic structure of the sentence and form a syntactic feature vector: S41. Traverse the dependency syntax tree, starting from the core verb of the sentence, and traverse the entire tree; S42, identifying the subject-verb-object relationship, finding the node pointing to the verb and marked as the subject relationship, and finding the node pointed to by the verb and marked as the object; S43, identify the modifying relationship: identify the relationship between the attributive and the central word and the relationship between the adverbial and the central word; S44, construct syntactic feature vector and define feature space: according to task requirements, define a set of syntactic features including subject type, predicate type, object type and main modification relationship type; S45, using one-hot encoding to represent each relationship type; for a relationship type involving a specific word or part of speech, incorporating the relationship type into a syntactic feature vector; S46, combining the extracted features of all relationship types into a complete syntactic feature vector; Obtain word embeddings through the pre-trained BERT model, and use semantic role labeling SRL to extract the predicate-argument structure, and combine all relationship types to build a semantic feature vector; determine the feature dimensions that need to be extracted based on the specific content and structure of the text, extract the corresponding feature values ​​for each feature dimension, and arrange the extracted feature values ​​in a certain order to build a text structure feature vector; The candidate answers and the preset standard answers are composed of text feature data and comprehension question feature data respectively; the similarity between the candidate answers and the preset standard answers is calculated, and they are sorted according to the similarity scores. The reading comprehension question score is calculated through the sorting and similarity scores of the candidate answers.

5. The artificial intelligence-based language reading comprehension system according to claim 4, characterized in that: The method of calculating the similarity between the candidate answers and the preset standard answers, sorting them according to the similarity scores, and calculating the reading comprehension question score through the sorting and similarity scores of the candidate answers includes: A comprehensive similarity function is defined to measure the similarity between the candidate answer and the preset standard answer; the comprehensive similarity function is: Among them, Osim(A,A*) is the similarity between the candidate answer and the preset standard answer; is the similarity between the lexical feature vector and the preset standard lexical feature vector; is the similarity between the syntactic feature vector and the preset standard syntactic feature vector; is the similarity between the semantic feature vector and the preset standard semantic feature vector; is the similarity between the text structure feature vector and the preset standard text structure feature vector; α1, α2, α3 and α4 are weight factors of the similarity of each vector; The candidate answers are evaluated using the comprehensive similarity function and sorted according to the similarity scores to obtain a sorted answer candidate list {A1, A2, ..., A n }; Calculate the reading comprehension score based on the ranking and similarity scores of the candidate answers; The calculation formula for the reading comprehension question score is: Among them, A i is the i-th answer in the sorted candidate answer list; A * is the preset standard answer; i is the weight assigned according to the ranking position of the candidate answers; n is the total number of candidate answers; i is an index from 1 to n.

6. The artificial intelligence-based language reading comprehension system according to claim 5, characterized in that: The method of analyzing the influence of comprehension question feature data on answering reading comprehension questions and fusing text feature data to obtain a comprehensive feature data set includes: Collect students' problem-solving order O, problem-solving time T, and final answer data when answering reading comprehension questions; calculate the average problem-solving time for students to answer each reading comprehension question: Where, is the average problem-solving time; is the time taken by the jth student to solve the i′th problem; N is the total number of students; Define a problem-solving order influence model to calculate the influence of the problem-solving order on the correct rate of the i′th question; the problem-solving order influence model is: Odr = f(Odr i′ ,Ur i′ );Odr i′ is the order of solving the i′th problem; Ur i′ is the correct rate of the i′th question; The correct rate of the i′th question is: Among them, Ur i′,j is the answer of the jth student to the i′th question, 1 for correct answer and 0 for wrong answer; Construct a multivariate regression model to quantify the combined effects of problem-solving order and problem-solving time on the accuracy of answering reading comprehension questions, and obtain the influencing characteristic data; The multivariate regression model is: Uri′,j=δ+ε1·Oi′,j+ε2·Ti′,j+∈i′,j; where O i′,j The order in which the jth student answers the i′th question; T i′,j is the time it takes for the jth student to answer the i′th question; δ is a constant term; ε1 is the coefficient of the order of solving questions on the correct rate of answering questions; ε2 is the coefficient of the time of solving questions on the correct rate of answering questions; ∈ i′,j is the error term of the multivariate regression model; According to the results of the multivariate regression model analysis, optimize students' problem-solving strategies; if ε1 is a negative value, remind students to put more difficult or more important questions first; if ε2 is a positive value, send students a reminder to allocate time reasonably and spend enough time thinking on more difficult questions; if ε1 is a positive value, remind students to quickly browse the overall difficulty of the questions and give priority to answering easier questions; if ε2 is a negative value, remind students to pay attention to time management; The comprehension question feature data, impact feature data and text feature data are weighted and fused to obtain a comprehensive feature data set.

7. The artificial intelligence-based language reading comprehension system according to claim 6, characterized in that: The training method of the comprehension question answer error prediction model includes: Divide the dataset into training set, validation set and test set, train the model and evaluate the model performance; build the error prediction model of the answer to the comprehension question, including the input layer, GRU layer, fully connected layer and output layer; The input layer of the comprehension question answer error prediction model is a historical comprehensive feature dataset, and the number of neurons in the input layer should match the number of features in the historical comprehensive feature dataset. The GRU layer is used to process the historical comprehensive feature dataset, and the number of GRU layers and neurons are adjusted according to the complexity of the task. The fully connected layer is used to provide additional nonlinear transformations. The output layer of the model is the reading comprehension question answer error. The number of neurons in the output layer corresponds to the number of prediction targets. The prediction value is output through one neuron, and the identity function is used as the activation function. The comprehension question answer error prediction model is a gated recurrent unit model. The mean square error is used as the loss function to measure the error between the model's predicted value and the actual value; the mean square error loss function is: Where n is the number of data sets; Y i is the actual value of data point i in the data set; is the predicted value of data point i in the data set; Use the training set data to train the model, minimize the loss function through the Adam optimizer, and tune the model's hyperparameters; evaluate the model through the test set, stop testing when the model's performance in the prediction task reaches the preset performance threshold, and obtain the comprehension question answer error prediction model; input the real-time comprehensive feature data set into the comprehension question answer error prediction model to obtain the real-time problem-solving answer error.

8. The artificial intelligence-based language reading comprehension system according to claim 7, characterized in that: The method for tuning the hyperparameters of the model includes: For the comprehension question answer error prediction model, a grid space is set, wherein the grid space includes at least two groups of hyperparameter combinations, and each group of hyperparameters varies within a preset range; For each set of hyperparameter combinations in the grid space, use these hyperparameters to configure the training model and use accuracy to evaluate the performance of the model; compare the model performance corresponding to each set of hyperparameters, select the set of hyperparameters with the highest accuracy, and set them as the optimal hyperparameters for the comprehension question answer error prediction model.

9. The artificial intelligence-based language reading comprehension system according to claim 8, characterized in that: The method of comparing the predicted answer error with a preset answer error threshold to determine whether the student's reading comprehension score is qualified includes: If the predicted answer error is less than the preset answer error threshold, the student's reading comprehension score is considered qualified; If the predicted answer error is greater than or equal to the preset answer error threshold, the student's reading comprehension score is judged to be unqualified.

10. The artificial intelligence-based language reading comprehension system according to claim 9, characterized in that: If a student fails the reading comprehension test, targeted feedback will be provided to point out the student's shortcomings and provide improvement measures, including: If the student's reading comprehension score is not satisfactory, the student's wrong answer text data is collected, and the wrong answer text data includes comprehension deviation data, main idea error data, detail comprehension error data and reasoning error; the wrong answer text data is preprocessed and converted into a feature sequence; Assume that X is the text feature sequence of students’ wrong answers, and Y is the corresponding error type label sequence. The sequence probability labeling model is used to maximize the probability of the corresponding error type label sequence Y given the text feature sequence X of students’ wrong answers. The sequence probability labeling model is: Among them, Y is the error type label sequence; T′ is the length of the sequence; X is the text feature sequence of the student’s wrong answer; y t is the error type label at time step t; y t-1 is the error type label at time step t-1; P(y t |y t-1 ,X) is given the previous label y t-1 and the student's incorrect answer to the text feature sequence X, the tth label is y t The conditional probability of The conditional probability is defined by a feature transfer model, which is: Among them, f k (y t-1 ,X) is the feature function defined on the given error type label sequence Y and the student’s wrong answer text feature sequence X; k is the index of the feature function; λ k is the characteristic function f k The associated weight parameter; Z(y t-1 ,X) is the partition function, which makes the sum of the probabilities of all labels being wrong type labels yt equal to 1; Construct a weighted optimization model and preset the weight parameter λ k The maximum threshold is λ max , weight parameter λ k The minimum threshold is λ min ; Combine the weight optimization model and the weight parameter threshold to adjust the weight parameter λ k to limit the scope of The weighted optimization model is: Where P(Y|X) is the probability of a given student’s wrong answer text feature sequence X corresponding to an error type label sequence Y; N′ is the number of error type label sequences Y; Calculate the weight parameter λ through the weighted optimization model k , select multiple groups of weight parameters with the minimum threshold λ min and the maximum threshold value of the weight parameter λ max The intermediate weight parameters are substituted into the feature transfer model to calculate the given previous label y t-1 and the student's incorrect answer to the text feature sequence X, the tth label is y t Multiple conditional probabilities P′; From multiple conditional probabilities P′, select the highest probability P corresponding to the error type label sequence Y under the given student's wrong answer text feature sequence X max , the weight parameter λ at this time k is the optimal weight parameter; through the feature transfer model, the teacher automatically identifies the error types in the students' wrong answer text data, and provides targeted feedback and improvement measures accordingly.

Images