Question and answer evaluation method and system for autism intervention

Abstract

The invention discloses a question and answer evaluation method and system for autism intervention, and relates to the field of medical electronic systems.The method comprises the steps that historical data are extracted, noise reduction processing is carried out to obtain noise reduction data after noise reduction processing, and a database is built according to the noise reduction data, a word frequency-inverse document frequency algorithm and a BERT model; extracting question and answer information of a plurality of different autism stages in the database, and constructing a judgment model according to the question and answer information and the database; obtaining current question and answer data, and obtaining a judgment result of the child according to the current question and answer data and the judgment model; the teacher resume in the database is extracted, multiple recommendation teachers are determined according to the teacher resume and a judgment result, the two are combined to form a multi-level semantic feature matrix, local word frequency information is reserved, global context understanding is fused, and therefore the problem that feature extraction of a single model is one-sided is solved; accurate recognition capability of autism core symptoms (such as social obstacles and communication defects) is remarkably improved, and a more reliable behavior characteristic basis is provided for an intervention scheme.

Description

Technical Field

[0001] The present invention relates to the field of medical electronic systems, and in particular to a question-and-answer evaluation method and system for autism intervention. Background Art

[0002] Autism is also a pervasive developmental disorder. Common types of the disease include apathetic type, active but weird type, and passive type. The interaction between genetic and environmental factors is the main cause of the disease. Studies have shown that the heritability can be as high as 80%-90%. In addition, abnormal brain structure and function, biological mechanisms, and adverse factors during maternal pregnancy can also affect the onset of autism. Most patients begin to develop social interaction disorders, communication disorders, and obvious interests, stereotypes, and repetitive behaviors in childhood. Autism is an increasingly common neurodevelopmental disorder. The prevalence of autism is mainly in male children and is increasing year by year.

[0003] Traditional single models (such as those relying solely on word frequency-inverse document frequency algorithms or basic bag-of-words models) often have significant limitations in evaluating autism intervention question-answering due to a lack of deep understanding of contextual semantics. For example, they can only count word frequencies but cannot parse the behavioral associations implicit in expressions such as "eye avoidance" and "social withdrawal," or have difficulty distinguishing the semantic differences of "repetitive actions" in different contexts (such as anxiety relief vs. stereotyped behaviors). Single models also have weak generalization capabilities for synonyms (such as "talking to oneself" and "monologue") and metaphorical expressions (such as "living in one's own world"), resulting in one-sided feature extraction and, in turn, affecting the accurate identification of core autism symptoms (such as social impairment and communication deficits). Therefore, the present invention proposes a question-answering evaluation method and system for autism intervention. Summary of the Invention

[0004] To solve the above technical problems, a question-answering evaluation method and system for autism intervention are provided, which solves the problem that the above-mentioned traditional single model (such as relying solely on the word frequency-inverse document frequency algorithm or the basic bag-of-words model) often has significant limitations in autism intervention question-answering evaluation due to the lack of deep understanding of contextual semantics. For example, it can only count word frequencies but cannot analyze the behavioral correlations implied in expressions such as "eye avoidance" and "social withdrawal", or it is difficult to distinguish the semantic differences of "repetitive actions" in different contexts (such as anxiety relief vs. stereotyped behaviors). The single model has weak generalization ability for synonyms (such as "talking to oneself" and "monologue") and metaphorical expressions (such as "living in one's own world"), resulting in one-sided feature extraction, which in turn affects the accurate identification of core symptoms of autism (such as social disorders and communication defects).

[0005] In order to achieve the above objects, the technical solution adopted by the present invention is: A question-and-answer assessment method for autism intervention, comprising: Extract historical data and perform noise reduction processing to obtain noise-reduced data, and build a database based on the noise-reduced data, the word frequency-inverse document frequency algorithm, and the BERT model; Extract question-and-answer information from the database for multiple different stages of autism, and build a judgment model based on the question-and-answer information and the database; Obtain the current question and answer data, and obtain the child's judgment result based on the current question and answer data and the judgment model; Extract teacher resumes from the database and determine multiple recommended teachers based on the teacher resumes and judgment results.

[0006] Preferably, extracting historical data and performing noise reduction processing to obtain noise-reduced data after noise reduction processing, and constructing a database based on the noise-reduced data, the term frequency-inverse document frequency algorithm and the BERT model include the following steps: Extract historical data and perform noise reduction processing on the historical data to obtain noise-reduced data after noise reduction processing; Construct a feature matrix based on denoised data, the term frequency-inverse document frequency algorithm, and the BERT model; Extract the child ID of each child in the denoised data; A database is constructed based on each child's ID and feature matrix.

[0007] Preferably, extracting historical data and performing noise reduction processing on the historical data to obtain noise-reduced data after noise reduction processing includes the following steps: Extract historical data and scan the historical data to identify abnormal data and missing data; Identify and filter abnormal data based on the recurrent neural network algorithm; Fill in missing data based on interpolation; The historical data is cleaned based on regular expressions and stop word filtering to obtain the denoised data.

[0008] Preferably, the constructing of a feature matrix based on the noise reduction data, the term frequency-inverse document frequency algorithm and the BERT model comprises the following steps: Scanning the denoised data to obtain each word or sentence in the denoised data; Generate a feature vector set of denoised data based on the term frequency-inverse document frequency algorithm and the BERT model; Determine a feature vector of the child ID based on the child ID; Construct a feature matrix based on the eigenvectors.

[0009] Preferably, extracting question-and-answer information of multiple different autism stages from the database and constructing a judgment model based on the question-and-answer information and the database includes the following steps: Extract the IDs of children at different stages of autism from the database; Extract the corresponding feature matrix based on multiple children's IDs; Determine multiple question and answer information based on the characteristic matrices of different autism stages; Extract keywords from multiple Q&A messages and label them according to different autism stages; Construct the judgment model of the corresponding stage through the marked keywords.

[0010] Preferably, the construction of the judgment model for the corresponding stage by using the marked keywords includes the following steps: Extract the frequency of keywords in each stage; Construct threshold ranges for different keywords based on frequency; Construct a preliminary judgment model based on the random forest algorithm and threshold range; Divide the denoised data into training set and validation set; The preliminary judgment model is modified based on the training set and the validation set to obtain the judgment model.

[0011] Preferably, the obtaining of the current question and answer data includes the following steps based on the current question and answer data and the judgment model: Get current question and answer data; Scan the current question and answer data to obtain the frequency of occurrence of feature words in the question and answer data; Compare the feature words with the keywords in the judgment model and filter out the same words; The input set is constructed by the same words and the frequency of occurrence of the same words; Input the input set into the judgment model to obtain the judgment result of the child.

[0012] Preferably, extracting teacher resumes from the database and determining multiple recommended teachers based on the teacher resumes and the judgment results includes the following steps: Extract the teacher's resume and the judgment result of the child from the database; Determine the type of teacher to be selected based on the judgment result, filter the teacher resumes based on the teacher type, and obtain multiple teachers to be recommended; Get the work schedule and working hours of the teacher to be recommended; Based on their work schedules and working hours, multiple teachers to be recommended are screened a second time to obtain recommended teachers.

[0013] Preferably, a question-answering evaluation system for autism intervention is proposed, which is used to implement the above-mentioned question-answering evaluation method for autism intervention, including: Control module: The control module is used to control data transmission within the system; Data storage module: The data storage module is used to store the data of each child and teacher, and build a database based on the child's data; Data analysis module: The data analysis module is used to build a judgment model, determine the current stage of the child, and recommend suitable teachers; Interaction module: The interaction module is used for users to input question and answer data; Display module: The display module is used to display the user's question and answer process and the teacher's resume.

[0014] Compared with the existing technology, the advantages of the present invention are: the present invention effectively addresses the semantic understanding deficiencies of single models (such as the word frequency-inverse document frequency algorithm or bag-of-words model) in autism intervention question-answering evaluation through a fusion feature extraction architecture of the BERT model and the word frequency-inverse document frequency algorithm. BERT, based on a bidirectional Transformer architecture, captures the contextual dependencies of words and sentences in question-answering texts, can parse the potential behavioral relevance of expressions such as "eye avoidance" and "social withdrawal", and distinguish the semantic differences of "repetitive actions" in the context of anxiety relief and stereotyped behavior. At the same time, it uses a self-attention mechanism to identify the deep semantic commonalities of synonyms (such as "talking to oneself" and "monologue") and metaphorical expressions (such as "living in one's own world"). The word frequency-inverse document frequency algorithm strengthens the statistical significance of key terms. The combination of the two forms a multi-level semantic feature matrix that not only retains local word frequency information but also integrates global contextual understanding, thereby overcoming the one-sided feature extraction problem of a single model, significantly improving the ability to accurately identify core symptoms of autism (such as social disorders and communication deficits), and providing more reliable behavioral feature basis for intervention plans. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 Schematic diagram of the process of steps S100-S400 in a question-answer assessment method and system for autism intervention proposed by the present invention; Figure 2 Schematic diagram of the flow of steps S101-S104 in a question-and-answer assessment method and system for autism intervention proposed by the present invention; Figure 3 Schematic diagram of the process of steps S1011-S1014 in a question-and-answer assessment method and system for autism intervention proposed by the present invention; Figure 4 Schematic diagram of the flow of steps S1021-S1024 in a question-and-answer assessment method and system for autism intervention proposed by the present invention; Figure 5 Schematic diagram of the flow of steps S201-S205 in a question-and-answer assessment method and system for autism intervention proposed by the present invention; Figure 6Schematic diagram of the process of steps S2051-S2055 in a question-and-answer assessment method and system for autism intervention proposed by the present invention; Figure 7 Schematic diagram of the flow of steps S301-S305 in a question-and-answer assessment method and system for autism intervention proposed by the present invention; Figure 8 Schematic diagram of the flow of steps S401-S401 in a question-answer assessment method and system for autism intervention proposed by the present invention; Figure 9 This is a structural block diagram of a question-and-answer evaluation method and system for autism intervention proposed by the present invention. DETAILED DESCRIPTION

[0016] The following description is intended to disclose the present invention so that those skilled in the art can implement the present invention. The preferred embodiments described below are merely examples, and those skilled in the art may conceive of other obvious variations.

[0017] Reference Figure 1-9 As shown, a question-and-answer assessment method for autism intervention includes: S100, extracting historical data and performing noise reduction processing to obtain noise-reduced data after noise reduction processing, and constructing a database based on the noise-reduced data, the word frequency-inverse document frequency algorithm and the BERT model; S200, extracting question-and-answer information of multiple different autism stages from the database, and building a judgment model based on the question-and-answer information and the database; S300: Obtain current question and answer data, and obtain the judgment result of the child based on the current question and answer data and the judgment model; S400, extracting teacher resumes from the database, and determining multiple recommended teachers based on the teacher resumes and the judgment results; It is understandable to those skilled in the art that it is necessary to establish a standardized, high-quality autism intervention knowledge base, eliminate outliers (such as logically contradictory question and answer records), fill in missing values ​​(such as unrecorded developmental milestones), ensure data integrity, convert text into structured feature vectors through the word frequency-inverse document frequency algorithm and the BERT model, quantify children's behavioral performance, use the child ID as the index, associate the feature matrix with the stage label, support efficient retrieval and model training, extract question and answer features of different stages (mild / moderate / severe) from the database (such as the keyword "imitation ability" only appears frequently in the mild stage), build a multi-stage classifier based on the random forest algorithm, and During the keyword frequency and threshold range determination stage, the question and answer data input by parents are scanned, keywords are extracted and the frequency is calculated (for example, "repetitive stereotyped behavior" appears 5 times), the input features are matched with the judgment model, and the stage judgment results (such as "moderate autism tendency") and dimension scores (social ability 60 points / 100 points) are output. Teachers with corresponding qualifications are recommended according to the child's stage (for example, "ABA certification + more than 5 years of experience" is required for the severe stage). Teachers who can provide immediate service are screened based on the teacher's geographical location and class schedule. Teachers are re-matched according to the child's progress rate (such as quarterly evaluation) (for example, upgrading from a "basic social training" teacher to an "advanced language development" teacher).

[0018] like Figure 2 As shown, extracting historical data and performing noise reduction processing to obtain noise-reduced data after noise reduction processing, and building a database based on the noise reduction data, the word frequency-inverse document frequency algorithm and the BERT model include the following steps: S101, extracting historical data and performing noise reduction processing on the historical data to obtain noise-reduced data after noise reduction processing; S102. Construct a feature matrix based on the noise reduction data, the term frequency-inverse document frequency algorithm, and the BERT model; S103, extracting the child ID of each child in the noise reduction data; S104, constructing a database based on each child's ID and feature matrix; Those skilled in the art will understand that historical data denoising cleans noise and redundant information from the original data, improves data quality, and provides a reliable basis for subsequent analysis; feature matrix construction converts unstructured text data into structured feature vectors, quantifies children's behavioral performance, and supports machine learning model training; child ID extraction establishes a unique association between data and individual children, supports personalized analysis and tracking, extracts unique identifiers (such as "CHILD_001") from denoised data, binds child IDs to feature matrices, and forms individualized data archives; database construction integrates child IDs, feature matrices, and stage labels to build a queryable and scalable autism intervention knowledge base; child IDs, feature vectors, and stage labels (mild / moderate / severe) are stored in a relational database (such as MySQL); indexes are established for child IDs to support fast retrieval (such as millisecond-level response); data update time and source are recorded, and data traceability and auditing are supported.

[0019] like Figure 3 As shown, extracting historical data and performing noise reduction processing on the historical data to obtain noise-reduced data after noise reduction processing includes the following steps: S1011. Extract historical data and scan the historical data to identify abnormal data and missing data in the historical data; S1012. Identify and filter abnormal data based on a recurrent neural network algorithm; S1013. Fill in missing data using interpolation method; S1014. Clean the historical data based on regular expressions and stop word filtering to obtain denoised data. It is understandable to those skilled in the art that the RNN model can scan the logical contradictions in the historical data (such as the mismatch between the age of the child and the developmental stage, such as "a 2-year-old child can recite a complete ancient poem"), detect data format errors (such as confusing date formats, values ​​outside a reasonable range); accurately identify and eliminate noise data through deep learning models, improve data credibility, mark unrecorded key fields (such as missing language ability assessment results, social interaction frequency), use labeled data (such as confirmed abnormal medical records) to train the RNN model, learn the logical patterns in the data, input new data into the RNN model, calculate the anomaly score (such as a score > 0.8 is considered an anomaly), filter high-scoring data (such as deleting contradictory questions and answers), and accurately identify and eliminate noise data through deep learning models. records); complete incomplete data, maintain data continuity, support statistical analysis, use linear / spline interpolation methods for longitudinal data (such as multiple follow-up records), estimate missing values ​​(such as interpolating the middle value based on the "language vocabulary" of two previous and subsequent records), and use the mean / median of the same group to fill in transverse data (such as missing fields in a single assessment) (such as using the mean language ability of children of the same age group to fill missing values); purify text data, remove irrelevant information, improve feature extraction efficiency, delete sensitive information (such as name, address, contact information), unify the format (such as standardizing the date format to "YYYY-MM-DD"), and remove meaningless words (such as modal particles such as "um", "ah", and "that").

[0020] like Figure 4 As shown in the figure, constructing a feature matrix based on denoised data, the term frequency-inverse document frequency algorithm, and the BERT model includes the following steps: S1021, scanning the noise reduction data to obtain each word or sentence in the noise reduction data; S1022. Generate a feature vector set of noise reduction data based on the term frequency-inverse document frequency algorithm and the BERT model; S1023. Determine a feature vector of the child ID based on the child ID; S1024. Construct a feature matrix based on the feature vector; Those skilled in the art will understand that, using a domain-adapted word segmentation tool (such as a Chinese word segmenter optimized for the field of autism) to segment the noise-reduced data into sentences and words, extracting individual words and phrases (such as "little eye contact" and "repeatedly opening and closing the door"), unifying synonyms (such as "language delay" and "late speech" are normalized to the same label), breaking down the text data into the smallest semantic units, providing basic input for feature extraction, and calculating the weight of words and phrases in the document based on the word frequency-inverse document frequency algorithm (i.e., the word frequency-inverse document frequency algorithm) (such as "stereotyped behavior" in a child's record has a word frequency-inverse document frequency algorithm = 0.85, indicating that the word and phrase is highly important to the child), generating a sparse feature vector (the dimension is the vocabulary size, such as 10,000 dimensions), and inputting the words and phrases into the pre-trained BERT model to generate a 768-dimensional semantic vector (such as "little eye contact" The vector of "social avoidance" is close to the vector of "social avoidance" in the semantic space), capturing the deep semantics of words and sentences (such as the potential association between "repetitive behavior" and "compulsive action"), concatenating the word frequency-inverse document frequency algorithm vector with the BERT vector to form a high-dimensional feature vector (such as 10,768 dimensions), taking into account both statistical significance and semantic relevance, taking the mean / weighted average of all word and sentence feature vectors under the same child ID to generate the child's overall feature vector, dynamically adjusting the vector weight according to the frequency of word and sentence occurrence (such as the word "social disorder" has a higher weight), constructing a dense matrix with the child ID as the row index and the feature vector as the column value, and performing Min-Max normalization on the feature vector (such as scaling the vector value to the interval [0,1]) to eliminate dimensionality effects. The sparse vectors of the word frequency-inverse document frequency algorithm are stored in CSR format to save storage space.

[0021] like Figure 5 As shown, extracting question-and-answer information of multiple different autism stages from the database and building a judgment model based on the question-and-answer information and the database includes the following steps: S201, extracting IDs of children at different stages of autism from a database; S202, extracting corresponding feature matrices based on multiple children's IDs; S203, determining a plurality of question-answer information based on characteristic matrices of different autism stages; S204, extracting keywords from the multiple question-and-answer messages, and marking the keywords according to different autism stages; S205, constructing a judgment model for the corresponding stage through the marked keywords; Those skilled in the art will appreciate that, based on the stage labels (such as mild, moderate, and severe autism) or diagnostic criteria (such as ADOS-2 scores) in the database, a list of child IDs corresponding to the stage is extracted to ensure balanced sample sizes in each stage (such as adjusting data distribution through oversampling / undersampling) to avoid model bias, and feature vectors of corresponding rows are extracted from the global feature matrix based on the child ID. The feature vectors of all children in the same stage are merged to form a submatrix of [number of stage samples × feature dimension]. Based on the child ID and the stage label, corresponding question and answer records are extracted from the database, and the question and answer records are associated with the feature vectors. High-frequency keywords are extracted from the question and answer records using a word frequency-inverse document frequency algorithm or a TextRank algorithm. Based on the stage labels of the question and answer records, stage labels are assigned to the keywords. Irrelevant words are filtered out in combination with autism field dictionaries (such as the DSM-5 Autism Diagnostic Criteria), and the feature matrix and keyword tags are input to output the stage prediction results.

[0022] like Figure 6 As shown, building a judgment model for the corresponding stage through the marked keywords includes the following steps: S2051. Extract the frequency of keywords in each stage; S2052. Construct threshold ranges for different keywords based on frequency; S2053. Construct a preliminary judgment model based on the random forest algorithm and the threshold range; S2054. Divide the denoised data into a training set and a validation set; S2055. Modify the preliminary judgment model based on the training set and the validation set to obtain a judgment model; Those skilled in the art will understand that the question and answer records of each stage are segmented and keywords are extracted (such as "eye avoidance" and "repetitive actions"), the number of occurrences of each keyword in the stage is counted, and the relative frequency of the keyword is calculated (such as the frequency of "eye avoidance" in the mild stage = the number of occurrences of the word / the total number of words in the stage), the influence of sample size differences is eliminated, and for each keyword, its frequency distribution in different stages is calculated (such as the mean frequency of "eye avoidance" in the mild stage = 0.3, standard deviation = 0.1), and the threshold is corrected in combination with domain knowledge (such as the DSM-5 diagnostic criteria) (such as the threshold of "no language ability" in the severe stage is forced to be ≥0.9), and the keyword frequency is used as an input feature (such as "eye avoidance frequency" and "repetitive action frequency"). Combined with the threshold range, binary features were generated (e.g., "eye avoidance > 0.4" = 1, otherwise = 0). A random forest algorithm was used to input features and stage labels and output stage predictions (e.g., mild / moderate / severe). The threshold rules were embedded in the model as hard constraints (e.g., "frequency of non-verbal ability ≥ 0.9" was directly classified as severe). The denoised data was randomly divided into a training set (80%) and a validation set (20%) in an 8:2 ratio to ensure consistent distribution (e.g., the same sample ratio at each stage). Random forest parameters (e.g., number of trees, maximum depth) were adjusted on the training set. Performance (e.g., precision and recall) was evaluated on the validation set. The threshold range was modified based on feedback from the validation set (e.g., adjusting the mild threshold for "repetitive actions" from 0.3 to 0.35 to reduce false positives).

[0023] like Figure 7 As shown, obtaining the current question and answer data, based on the current question and answer data and the judgment model, includes the following steps: S301, obtaining current question and answer data; S302: Scan the current question and answer data to obtain the frequency of occurrence of feature words in the question and answer data; S303, comparing the feature words with the keywords in the judgment model to filter out the same words; S304, constructing an input set based on the same words and the frequencies of occurrence of the same words; S305: Input the input set into the judgment model to obtain the judgment result of the child; Those skilled in the art will appreciate that children's question and answer records (such as "Do you avoid eye contact?" and "Do you have repetitive movements?") are obtained through clinical systems, mobile applications, or parent questionnaires. Feature words (such as "eye avoidance," "repetitive movements," and "language delay") are extracted from the question and answer records using a domain dictionary (such as an autism behavior dictionary). The frequency of occurrence of each feature word is calculated (such as "eye avoidance" appears 2 times / total number of words 10 times = 0.2). The feature words in the current question and answer data are compared with the keyword library of the judgment model (such as "eye avoidance" and "inability to speak") to filter out identical words, organize the identical words and their frequencies into a dictionary or vector (such as {"eye avoidance": 0.2, "repetitive movements": 0.4}), fill in default values ​​(such as frequency = 0) or ignore words that do not appear in the model keywords, input the input set into the judgment model (such as a random forest classifier), output the stage prediction result (such as mild / moderate / severe), and interpret the result by feature importance (such as "eye avoidance" has the highest weight) or keyword threshold (such as "repetitive movement frequency > 0.3" triggers a moderate judgment).

[0024] like Figure 8 As shown, extracting teacher resumes from the database and determining multiple recommended teachers based on the teacher resumes and the judgment results includes the following steps: S401, extracting the teacher's resume and the judgment result of the child from the database; S402: Determine the teacher type to be selected based on the judgment result, filter the teacher resumes based on the teacher type, and obtain multiple teachers to be recommended; S403, obtaining the work schedule and working hours of the teacher to be recommended; S404: Perform secondary screening of multiple teachers to be recommended based on their work schedules and work hours to obtain recommended teachers; Those skilled in the art will appreciate that resume data (such as professional background, work experience, and areas of expertise) can be extracted from the teacher information database, and the child's current autism stage (such as mild / moderate / severe) and behavioral characteristics (such as "eye avoidance" and "language delay") can be obtained from the diagnosis result database. The teacher information and the child's diagnosis results are associated with unique identifiers (such as child ID and teacher ID) to form a teacher-child matching data set. A mapping rule between teacher type and autism stage is preset (such as "mild → ABA teacher" and "severe → speech therapist"), and the type is refined in combination with the child's behavioral characteristics (such as "language delay" is matched with speech therapists first). According to the teacher type, matching fields (such as professional qualifications and years of intervention experience) are extracted from the resume, and the matching fields are filtered using a rule engine (Drools) or conditional filtering (P Python Pandas) is used to screen qualified teachers, obtain teachers' working hours (9:00-17:00, Monday to Friday) from the teacher scheduling system (such as Google Calendar API) or internal schedule, parse the worksheet into computable time intervals (such as converting "9:00-17:00" into a timestamp range), compare the children's intervention demand time (10:00-11:00 every Wednesday) with the teacher worksheet, eliminate conflicting teachers, use time interval algorithms (such as Allan's Interval Algorithm) to efficiently detect overlapping time periods, sort the remaining teachers according to the matching degree (professional qualification weight 60%, time matching 40%), and output the final recommendation list (for example, recommending 3 teachers, sorted by priority).

[0025] like Figure 9 As shown, a question-answering evaluation method and system for autism intervention are proposed, which are used to implement the above-mentioned question-answering evaluation method for autism intervention, including: Control module: The control module is used to control data transmission within the system; Data storage module: The data storage module is used to store the data of each child and teacher, and build a database based on the child's data; Data analysis module: The data analysis module is used to build a judgment model, determine the current stage of the child, and recommend suitable teachers; Interaction module: The interaction module is used for users to input question and answer data; Display module: The display module is used to display the user's question and answer process and the teacher's resume.

[0026] In summary, the advantages of the present invention are: through the fusion feature extraction architecture of the BERT model and the word frequency-inverse document frequency algorithm, it effectively addresses the semantic understanding deficiencies of single models (such as the word frequency-inverse document frequency algorithm or the bag-of-words model) in the evaluation of autism intervention questions and answers. BERT, based on the bidirectional Transformer architecture, captures the contextual dependencies of words and sentences in the question and answer text, can parse the potential behavioral relevance of expressions such as "eye avoidance" and "social withdrawal", and distinguish the semantic differences of "repetitive actions" in the scenarios of anxiety relief and stereotyped behavior. At the same time, it uses the self-attention mechanism to identify the deep semantic commonalities of synonyms (such as "talking to oneself" and "monologue") and metaphorical expressions (such as "living in one's own world"). The word frequency-inverse document frequency algorithm strengthens the statistical significance of key terms. The combination of the two forms a multi-level semantic feature matrix, which not only retains local word frequency information but also integrates global contextual understanding, thereby overcoming the one-sided feature extraction problem of a single model, significantly improving the ability to accurately identify core symptoms of autism (such as social disorders and communication deficits), and providing more reliable behavioral feature basis for intervention plans.

[0027] The above shows and describes the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions merely illustrate the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of protection claimed by the present invention is defined by the appended claims and their equivalents.

Claims

1. A question-answering evaluation method for autism intervention, characterized in that: include: Extract historical data and perform noise reduction processing to obtain noise-reduced data, and build a database based on the noise-reduced data, the word frequency-inverse document frequency algorithm, and the BERT model; Extract question-and-answer information from the database for multiple different stages of autism, and build a judgment model based on the question-and-answer information and the database; Obtain the current question and answer data, and obtain the child's judgment result based on the current question and answer data and the judgment model; Extract teacher resumes from the database and determine multiple recommended teachers based on the teacher resumes and judgment results.

2. A question-and-answer assessment method for autism intervention according to claim 1, characterized in that: The extracting of historical data and performing noise reduction processing to obtain noise-reduced data after noise reduction processing, and constructing a database based on the noise-reduced data, the term frequency-inverse document frequency algorithm and the BERT model include the following steps: Extract historical data and perform noise reduction processing on the historical data to obtain noise-reduced data after noise reduction processing; Construct a feature matrix based on denoised data, the term frequency-inverse document frequency algorithm, and the BERT model; Extract the child ID of each child in the denoised data; A database is constructed based on each child's ID and feature matrix.

3. The question-and-answer assessment method for autism intervention according to claim 2, characterized in that: The extracting of historical data and performing noise reduction processing on the historical data to obtain noise-reduced data after noise reduction processing includes the following steps: Extract historical data and scan the historical data to identify abnormal data and missing data; Identify and filter abnormal data based on the recurrent neural network algorithm; Fill in missing data based on interpolation; The historical data is cleaned based on regular expressions and stop word filtering to obtain the denoised data.

4. The question-and-answer assessment method for autism intervention according to claim 3, characterized in that: The construction of a feature matrix based on the noise reduction data, the term frequency-inverse document frequency algorithm and the BERT model includes the following steps: Scanning the denoised data to obtain each word or sentence in the denoised data; Generate a feature vector set of denoised data based on the term frequency-inverse document frequency algorithm and the BERT model; Determine a feature vector of the child ID based on the child ID; Construct a feature matrix based on the eigenvectors.

5. The question-and-answer assessment method for autism intervention according to claim 4, characterized in that: The step of extracting question-and-answer information of multiple different autism stages from the database and constructing a judgment model based on the question-and-answer information and the database includes the following steps: Extract the IDs of children at different stages of autism from the database; Extract the corresponding feature matrix based on multiple children's IDs; Determine multiple question and answer information based on the characteristic matrices of different autism stages; Extract keywords from multiple Q&A messages and label them according to different autism stages; Construct the judgment model of the corresponding stage through the marked keywords.

6. The question-and-answer assessment method for autism intervention according to claim 5, characterized in that: The construction of the judgment model for the corresponding stage by using the marked keywords includes the following steps: Extract the frequency of keywords in each stage; Construct threshold ranges for different keywords based on frequency; Construct a preliminary judgment model based on the random forest algorithm and threshold range; Divide the denoised data into training set and validation set; The preliminary judgment model is modified based on the training set and the validation set to obtain the judgment model.

7. The question-and-answer assessment method for autism intervention according to claim 6, characterized in that: The acquisition of current question and answer data, based on the current question and answer data and the judgment model, includes the following steps: Get current question and answer data; Scan the current question and answer data to obtain the frequency of occurrence of feature words in the question and answer data; Compare the feature words with the keywords in the judgment model and filter out the same words; The input set is constructed by the same words and the frequency of occurrence of the same words; Input the input set into the judgment model to obtain the judgment result of the child.

8. The question-and-answer assessment method for autism intervention according to claim 6, characterized in that: Extracting teacher resumes from the database and determining multiple recommended teachers based on the teacher resumes and the judgment results includes the following steps: Extract the teacher's resume and the judgment result of the child from the database; Determine the type of teacher to be selected based on the judgment result, filter the teacher resumes based on the teacher type, and obtain multiple teachers to be recommended; Get the work schedule and working hours of the teacher to be recommended; Based on their work schedules and working hours, multiple teachers to be recommended are screened a second time to obtain recommended teachers.

9. A question-answer evaluation and diagnosis system for autism intervention, used to implement a question-answer evaluation and diagnosis method for autism intervention as described in claims 1-8, characterized in that: include: Control module: The control module is used to control data transmission within the system; Data storage module: The data storage module is used to store the data of each child and teacher, and build a database based on the child's data; Data analysis module: The data analysis module is used to build a judgment model, determine the current stage of the child, and recommend suitable teachers; Interaction module: The interaction module is used for users to input question and answer data; Display module: The display module is used to display the user's question and answer process and the teacher's resume.

Images