Intelligent test paper judging method and system for adaptive multi-modal subjective questions based on nursing field

Abstract

The invention discloses a self-adaption multi-modal subjective item intelligent paper judgment method and system based on the nursing field, relates to the field of intelligent pension, online education and artificial intelligence application, and aims to solve the problems that in nursing training, manual paper judgment of subjective items is low in efficiency, high in subjectivity, non-uniform in standard and high in efficiency. And a general AI paper judging system is difficult to adapt to non-standard language expression in the nursing field and lacks effective teaching feedback. Performing cleaning and standardized conversion on oral, dialected and non-standard term answers of students through deep preprocessing of nursing field data, extracting semantic features and identifying key score points; performing multi-dimensional scoring based on content accuracy, logic coherence and term normalization to generate a preliminary score; dynamic score correction is carried out in combination with trainee portraits, and step-by-step elastic score giving is carried out on reasonable innovation operation; and finally, multi-modal feedback including text interpretation, 3D / AR visual demonstration, a knowledge point thermodynamic diagram and intelligent make-up examination suggestions is generated.

Description

Technical Field

[0001] The present invention belongs to the fields of intelligent elderly care, online education technology and artificial intelligence applications, and in particular relates to an adaptive multi-modal subjective question intelligent grading method and system based on the nursing field. Background Art

[0002] As the nursing industry develops, the professional skills of nurses are directly linked to the quality of nursing services. Scientific and effective training and assessment are crucial for improving nursing skills. Subjective questions, crucial for assessing nurses' comprehensive understanding, analytical application, and practical operational thinking, play a crucial role in nursing training and assessment.

[0003] Currently, in the field of vocational skills training such as nursing, online examinations have become an important means of assessing students' mastery. For objective question types such as multiple-choice questions and true-or-false questions, automated computer grading technology is quite mature and widely used. However, for subjective questions that require assessing students' comprehensive understanding, analytical application, and practical operational thinking (such as short-answer questions, essay questions, and case analysis questions), especially in the assessment of specific practical skills (such as nursing operations), manual grading is still the main method. This situation is particularly prominent in nursing training, where practitioners with relatively limited educational background account for a large proportion.

[0004] Manual grading has many inherent drawbacks: First, it consumes a lot of manpower and material resources, is inefficient and costly, and this problem is even more pronounced in large-scale training programs. Second, different graders have different understandings and grasps of the grading criteria, which are easily influenced by subjective factors such as personal experience and fatigue. This makes it difficult to ensure the objectivity and consistency of the grading results, which in turn affects the fairness of the assessment. Third, manual grading can usually only provide a final score, making it difficult to achieve instant feedback. In addition, the feedback is mostly limited to text and the content is relatively simple, making it impossible to provide students with detailed error analysis and targeted improvement suggestions.

[0005] Furthermore, when answering subjective questions, nursing practitioners often lack a grasp of language habits, local dialects, and specialized terminology, resulting in non-standard and colloquial responses, often with typos and incoherent sentences. This further increases the difficulty and complexity of manual grading. Furthermore, the same nursing procedure or concept may be expressed in a variety of different dialects or idioms across different regions, institutions, and even individuals, creating challenges for consistent and accurate assessment.

[0006] In addition, the existing system has obvious deficiencies and low efficiency in teaching feedback: on the one hand, traditional grading methods and most AI systems can only give scores, and cannot deeply analyze students' weaknesses and generate specific, actionable and personalized improvement suggestions; on the other hand, the feedback form is not intuitive enough. For students with low cultural level and who rely more on visual thinking, simple text feedback is difficult to be effectively understood and absorbed. They need more intuitive and visual demonstrations to understand the causes of errors and correct operations.

[0007] In summary, the current review of subjective questions in the nursing field has many problems in terms of efficiency, objectivity, accuracy and teaching feedback, which cannot meet the urgent needs of nursing training for precise guidance and skill improvement. Therefore, it is of great practical significance to develop an intelligent marking system that can adapt to the characteristics of the nursing field, understand non-standard expressions, and provide effective feedback. Summary of the Invention

[0008] This paper proposes an intelligent grading and feedback method for nursing subjective questions based on the integration of domain knowledge and artificial intelligence algorithms. The core of this solution lies in the construction of an intelligent processing process tailored to the specific characteristics of the nursing field. First, a specially designed "Nursing Field Data Deep Preprocessing Submodule" cleans and standardizes students' original answers, which are full of colloquialisms, dialects, and non-standard terminology. Deep learning models are then used to extract deep semantic features. These features are then fed into a "Multi-Dimensional Intelligent Scoring Module," which outputs a preliminary objective score based on multiple pre-set dimensions, such as content accuracy, logical coherence, and terminology standardization. This module combines the dynamic scoring corrections tailored to student profiles and the flexible scoring mechanism for innovative operations in the "Nursing Field Adaptive and Flexible Adjustment Module" to create a flexible scoring mechanism for innovative operations. Finally, the scoring results and detailed error analysis are linked to the "Multimodal Teaching Feedback and Intelligent Linkage Module" to generate comprehensive feedback, including text explanations, visual and interactive operation demonstrations, personalized ability heat maps, and intelligent re-examination exercises. The entire system aims to accurately assess students' mastery of nursing knowledge and skills through automated and intelligent means, and provide easy-to-understand, targeted teaching and guidance, thereby improving training effectiveness.

[0009] The present invention provides a multi-modal intelligent grading method for subjective questions based on self-adaptation in the nursing field, comprising:

[0010] Receive the answer text submitted by the students;

[0011] Perform format check and formatting on the received answer text;

[0012] Cleaning and standardizing the formatted answer text to obtain standardized text;

[0013] Perform in-depth analysis on standardized text and convert it into semantic feature vectors;

[0014] Identify key intentions and scoring points based on semantic feature vectors;

[0015] Conduct a comprehensive assessment and produce a preliminary score;

[0016] Adaptive and flexible adjustments to preliminary scores;

[0017] The system calculates the final scoring result.

[0018] In one embodiment of the present invention, the adaptive and flexible adjustment of the preliminary score includes:

[0019] Obtain student information from the student portrait database and decide whether to activate the "dialect tolerance mode" to correct the preliminary score;

[0020] Evaluate whether there are reasonable but non-standard innovative operations in the answer text, and apply the "operation integrity coefficient α" to give flexible scores.

[0021] In one embodiment of the present invention, it further includes:

[0022] Generate multimodal feedback based on the scoring results, including text explanations, 3D operation demonstrations, personal ability heat maps, and intelligent re-examination suggestions;

[0023] Present the final scoring results and multimodal feedback information to students or teachers.

[0024] In one embodiment of the present invention, cleaning and standardizing the formatted answer text to obtain a standardized text includes:

[0025] From the input text S input Starting from the current position, search for the longest non-standard term s that can match in the dictionary non-std ;

[0026] For the fragments that are not directly hit, calculate their similarity with the non-standard entries in the dictionary.

[0027]

[0028] Where |s| represents the length of string s. When Sim(s candidate ,s dict_entry )>τ sim When , the match is considered successful;

[0029] Convert suspected pinyin input text segments into possible Chinese character candidates, and then verify them based on context and fuzzy matching;

[0030] When a non-standard expressionnon-std Mapping to multiple standard terms L std When s non-std The surrounding context C is associated with each candidate standard term The combined conditional probability Select the one with the highest probability, or filter the most relevant standard terms based on the nursing module to which the question belongs.

[0031] In one embodiment of the present invention, performing in-depth analysis on the standardized text and converting it into a semantic feature vector includes:

[0032] The BERT model, fine-tuned on nursing text, processes the token sequence of the input standardized text, captures the bidirectional semantic information of each token in the deep context, and outputs context-aware word embedding vectors.

[0033] The word embedding sequence output by the BERT layer is input into the BiLSTM network to further capture the long-range temporal dependencies of the text and enhance the representation of sequence information.

[0034] Generate a fixed-dimensional dense vector representing the semantics of the answer, namely the semantic feature vector E A .

[0035] In one embodiment of the present invention, identifying key intentions and scoring points based on semantic feature vectors includes:

[0036] Assume that the semantic embedding of the student's answer is

[0037] Predefined key scoring points / operation steps, k∈K points , whose embedding vector is

[0038] Calculate the similarity as a score,

[0039]

[0040] Where sim is the dot product or cosine similarity, a k Indicates the degree of attention of the answer to the score point k, that is, the similarity.

[0041] In one embodiment of the present invention, performing a comprehensive assessment and obtaining a preliminary score includes:

[0042] Calculate content accuracy scores,

[0043]

[0044] Among them, cosine similarity is:

[0045]

[0046] w content is the weight of content accuracy;

[0047] Obtain the sequence of operation steps through intention recognition,

[0048] Seq A =[step1,step2,…,step n ],

[0049] Seq A With the predefined standard operating procedure template Seq Std contrast,

[0050]

[0051] Train a classifier M logic JudgeSeq A The logical score p logic ∈[0,1],

[0052]

[0053] in is the step sequence Seq A The embedding representation of

[0054] Get the final logic score,

[0055] S logic =w logic ×(λS logic,rule +(1-λ)S logic,model ),

[0056] Where λ is the fusion coefficient;

[0057] After cleaning the non-standard answers, let the set of standard nursing terms used correctly in the answers be T correct The expected set of standard terms related to the topic is T expected ,

[0058] The term normativeness score is,

[0059] Calculate the total score S total =clip(S content +S logic +S term ,S min ,S max );

[0060] where w content ,w logic ,w term is the preset weight of each dimension, Smin ,S max are the minimum and maximum values ​​of the total score.

[0061] In one embodiment of the present invention, evaluating whether there are reasonable but non-standard innovative operations in the answer and applying the "operation integrity coefficient α" for flexible scoring includes:

[0062] Assume that the standard operation includes N core steps {KP1,…,KP N}, M core steps are identified in the students’ answers, and K operation steps are identified as “reasonable but non-standard” or “beneficial and innovative” by the “nursing knowledge base / expert rules” {IP1,…,IP K};

[0063] IP at every innovation step j Assign an innovation weight based on its rationality and innovation

[0064] The operational integrity coefficient α is defined as:

[0065]

[0066] The final score is calculated as follows:

[0067] S'=S×α';

[0068] where α'=clip(α,α min ,α max ), is the operational integrity coefficient after constraints.

[0069] The present invention also provides a multi-modal intelligent grading system for subjective questions based on self-adaptation in the field of nursing, comprising:

[0070] The user interaction and presentation module is configured to provide a student-side interface for receiving text input of answers to subjective questions and presenting the grading results and multimodal feedback to students or teachers;

[0071] The data access and preprocessing module is configured to receive and format answers, clean non-standard answers, enhance semantics and extract features, and identify key intents;

[0072] A multi-dimensional intelligent scoring module is configured to comprehensively score based on three dimensions: content accuracy, logical coherence, and terminology standardization. The comprehensive scoring unit calculates the preliminary score; and

[0073] The nursing field adaptive and flexible adjustment module is configured to perform adaptive and flexible adjustments based on the preliminary scores.

[0074] In one embodiment of the present invention, it further includes:

[0075] The multimodal teaching feedback and intelligent linkage module is configured to generate multimodal feedback, update knowledge point heat maps, and push intelligent re-examination suggestions;

[0076] The core resource and knowledge base module is configured to provide nursing terminology conversion, standard answers, 3D / AR materials and student portrait data support.

[0077] The present invention has the following beneficial effects:

[0078] (1) Aiming at the common problems of colloquial expressions, non-standard terminology, and dialect differences among nursing students (especially those from specific age and cultural background groups), the system improves its ability to recognize, understand, and correctly evaluate non-standard answers by building a term conversion library and semantic understanding model dedicated to the nursing field.

[0079] (2) Not only does it give scores, but it also generates detailed error cause analysis, knowledge point mastery heat maps, targeted improvement suggestions, and combines 3D nursing operation demonstrations, AR marking and other visual means to make feedback more intuitive, easy to understand and effective.

[0080] (3) By introducing a dynamic scoring correction mechanism (such as dialect error tolerance mode) and designing a visual feedback form that conforms to its cognitive characteristics, the user-friendliness and teaching effectiveness of the system can be improved.

[0081] (4) Through a step-by-step scoring mechanism and an operational integrity coefficient, reasonable evaluations are given to students’ partially correct or innovative answers that have practical significance, thus avoiding a rigid “one-size-fits-all” scoring approach.

[0082] (5) Through a systematic intelligent processing process, the burden of manual marking is reduced, the quality of evaluation is improved, and an effective teaching-learning-evaluation closed loop is formed.

[0083] The core purpose of this proposed technology, based on adaptive multimodal intelligent grading of subjective questions in the nursing field, is to address practical technical issues in nursing training, such as low efficiency, strong subjectivity, inconsistent standards, and poor adaptability of general systems. These processing and control processes fully adhere to the natural laws of information transmission, feature extraction, and logical reasoning, ultimately achieving the technical effect of improving grading accuracy and scoring consistency, and generating intuitive and effective multimodal teaching feedback, significantly optimizing the technical process of nursing training assessment. BRIEF DESCRIPTION OF THE DRAWINGS

[0084] Figure 1 A flow chart of a method for intelligently grading multimodal subjective questions based on nursing field adaptation in one embodiment of the present invention is shown;

[0085] Figure 2 The figure shows a block diagram of a multi-modal intelligent scoring system for subjective questions based on self-adaptation in the nursing field in one embodiment of the present invention;

[0086] Figure 3 A schematic diagram showing the data structure of an "error-material mapping library" in one embodiment of the present invention is shown; and

[0087] Figure 4 An embodiment of the present invention is shown for executing Figure 3 Pseudocode example diagram of the "triggering and calling logic" of a mapping rule. DETAILED DESCRIPTION

[0088] In the following description, the present invention is described with reference to various embodiments. However, those skilled in the art will recognize that the embodiments may be implemented without one or more of the specific details or with other alternative and / or additional methods, materials, or components. In other cases, well-known structures, materials, or operations are not shown or described in detail to avoid obscuring the inventive aspects of the present invention. Similarly, for purposes of explanation, specific numbers, materials, and configurations are set forth to provide a comprehensive understanding of the embodiments of the present invention. However, the present invention is not limited to these specific details.

[0089] In the present invention, each embodiment is only intended to illustrate the aspects of the present invention and should not be construed as limiting.

[0090] In this specification, reference to "one embodiment" or "the embodiment" means that a particular feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment of the present invention. The appearances of the phrase "in one embodiment" in various places in this specification are not necessarily all referring to the same embodiment.

[0091] In addition, the numbering of the steps of the methods of the present invention does not limit the order in which the steps are to be performed. Unless otherwise specified, the steps of the methods may be performed in different orders.

[0092] The present invention will be further described below with reference to the accompanying drawings in conjunction with specific embodiments.

[0093] Figure 1 The flowchart of the method for intelligent grading of multimodal subjective questions based on nursing field adaptation in one embodiment of the present invention is shown.

[0094] like Figure 1 As shown, in one embodiment of the present invention, the process of the multi-modal subjective question intelligent grading method based on nursing field self-adaptation is as follows:

[0095] S202 Student answer input: the system first receives the subjective question answer text submitted by the student through the student-side interface 101 of the user interaction and presentation module 100.

[0096] S203 Answer reception and formatting: The received data is subjected to preliminary format checking and standardization processing by the answer reception and formatting unit 201 in the data access and pre-processing module 200 to provide raw data for subsequent analysis.

[0097] S204 Deep preprocessing of nursing data:

[0098] S204a: Non-standard answer cleaning. After preliminary processing, the answer text enters the nursing domain data deep preprocessing submodule 202. At this stage, the non-standard answer cleaning engine 202a is activated, calling the nursing term conversion library 601 to identify, clean, and standardize dialects, colloquial expressions, and typos in the text to reduce interference with subsequent semantic understanding.

[0099] S204b: Semantic enhancement and feature extraction: The cleaned and standardized text is then input into the semantic enhancement and feature extraction engine 202b. This engine uses a deep learning model (such as a combination of BERT and BiLSTM) to perform in-depth analysis of the text, capture its inherent semantic information, and convert it into high-dimensional semantic feature vectors. These feature vectors lay the foundation for subsequent scoring and intent recognition.

[0100] S204c identifies key intent and scoring points. The extracted semantic feature vectors are fed into the key intent and scoring point identification network 202c. This network uses algorithms such as the attention mechanism to match the semantics of the student's answer with the core knowledge points preset in the standard answer and scoring point library 602, identifying whether the student has answered the key points and which key points they have answered.

[0101] S205 Multi-dimensional Intelligent Scoring: Based on key intent identification, the system enters the scoring phase. The multi-dimensional intelligent scoring module 300 is activated to comprehensively evaluate the answer based on three dimensions: content accuracy, logical coherence, and terminology standardization. The comprehensive score calculation unit 304 then generates a preliminary score.

[0102] S206 Adaptation and flexible adjustment: The nursing field adaptation and flexible adjustment module 400 decides whether to activate the "dialect tolerance mode" to correct the score based on the student information (such as age and region) obtained from the student portrait database 604; at the same time, it will also evaluate whether there are reasonable but non-standard innovative operations in the answer, and apply the "operation integrity coefficient α" to give flexible scores.

[0103] S207 generates the final score. After multi-dimensional scoring and adaptive adjustment, the system calculates the final, objective scoring result.

[0104] S208 Multimodal Teaching Feedback Generation and Linkage: After scoring is completed, the multimodal teaching feedback and intelligent linkage module 500 is activated. Based on the scoring results and error analysis, the system calls corresponding visualization resources from the 3D / AR nursing operation material library 603 to generate multimodal feedback including text explanations, 3D operation demonstrations, personal ability heat maps, and intelligent re-examination suggestions.

[0105] S209 Result feedback and presentation: The final score and rich teaching feedback information are clearly presented to students or teachers through the user interaction and presentation module 100 so that they can make subsequent learning and teaching improvements.

[0106] Figure 2 The figure shows a block diagram of a multi-modal intelligent scoring system for subjective questions based on nursing field adaptation in one embodiment of the present invention.

[0107] like Figure 2 As shown, in one embodiment of the present invention, a multi-modal intelligent scoring system for subjective questions in the field of nursing that is adaptive includes:

[0108] The user interaction and presentation module 100 includes:

[0109] The student-side interface 101 receives the text of the answers to the subjective questions submitted by the students, and displays the final scoring results, multimodal teaching feedback (such as text explanations, 3D operation demonstrations, AR markings, etc.), knowledge point heat maps and intelligent re-examination suggestions.

[0110] The interface 102 on the teacher side displays data such as the overall grading status, grading distribution, and common error analysis of students to the teacher, and supports the teacher in viewing and managing the system grading results.

[0111] The data access and preprocessing module 200 includes:

[0112] The answer receiving and formatting unit 201 receives the answers submitted by the students, performs preliminary format checking (such as removing invalid symbols and unifying the text format) and standardization processing, and provides standardized original data for subsequent analysis.

[0113] Nursing field data deep preprocessing module 202:

[0114] The non-standard answer cleaning engine 202a calls the nursing term conversion library 601 to identify and clean non-standard content such as dialects, colloquial expressions, typos, etc. in the answers, and convert them into standard nursing terms (such as converting "turn the patient over and pat him" to "turn the patient over and tap his back").

[0115] The semantic enhancement and feature extraction engine 202b uses deep learning models such as BERT and BiLSTM to perform in-depth analysis on the cleaned text, extract deep semantic features, and generate high-dimensional semantic feature vectors.

[0116] The key intention and scoring point recognition network 203c matches the semantic features of the student's answer with the core knowledge points in the standard answer and scoring point library 602 through algorithms such as the attention mechanism, and identifies whether the student's answer covers the key scoring points and specific content.

[0117] Multi-dimensional intelligent scoring module 300:

[0118] The content accuracy evaluation unit 301 calculates the semantic similarity between the student's answer and the standard answer, and evaluates the degree of matching between the answer content and the core knowledge points.

[0119] The logical consistency evaluation unit 302 compares the consistency of the operation step sequence in the student's answer with the standard process, and evaluates the logical rationality through rules (such as the longest common subsequence) and model fusion.

[0120] The terminology standardization evaluation unit 303 counts the proportion of correctly used standard nursing terms and expected terms in the answers to evaluate the degree of standardization of terminology use.

[0121] The comprehensive score calculation unit 304 integrates the score results of the above three dimensions and calculates a preliminary total score.

[0122] Nursing Adaptive and Flexible Adjustment Module 400:

[0123] The student portrait analysis unit 401 extracts information such as student age and region from the student portrait database 604 to provide a basis for subsequent score adjustments.

[0124] Dynamic score correction causes 402. The dialect tolerance mode is activated based on the student's profile (such as age exceeding the threshold or coming from a dialect-prone area), and the initial score is corrected to improve tolerance.

[0125] The distributed and innovative scoring engine 403 identifies reasonable but non-standard innovative operation steps in the answer, and flexibly scores them based on the operation integrity coefficient α, giving reasonable scores to partially correct or innovative points.

[0126] Multimodal teaching feedback and intelligent linkage module 500:

[0127] Personalized feedback generation engine 501:

[0128] The error cause location and text feedback unit 501a analyzes the errors in the students' answers and generates detailed text explanations (such as "missing the disinfection step") and targeted improvement suggestions.

[0129] The visual feedback matching and calling unit 501b calls corresponding visual resources (such as 3D animation, AR markers) from the 3D / AR nursing operation material library 603 according to the error type to intuitively display the correct operation.

[0130] Trainee Capacity Diagnosis and Improvement Engine 502:

[0131] The knowledge point mastery heat map generating unit 502a generates a heat map of the student's knowledge point mastery based on the LDA topic model or the knowledge tracking algorithm, and intuitively presents the strengths and weaknesses.

[0132] The intelligent make-up exam and personalized practice push unit 502b pushes targeted make-up exam questions and personalized practice from the intelligent question bank and knowledge point bank 605 based on error analysis and heat map.

[0133] Core Resources and Knowledge Base Module 600:

[0134] Nursing terminology conversion library 601 uses efficient structures such as hash tables to store the mapping relationship between dialects, colloquial expressions and standard terms (such as "erythema" corresponds to "pressure ulcer stage I"), and supports the cleaning of non-standard answers.

[0135] The standard answer and scoring point library 602 stores the standard answer, core scoring points and operation steps of each subjective question, providing a basis for key intention identification and scoring.

[0136] The 3D / AR nursing operation material library 603 stores visualization resources such as 3D animations and AR models related to nursing operations, which correspond one-to-one with error types (e.g., "wrong step sequence" corresponds to a 3D demonstration of the correct process).

[0137] The student portrait database 604 records information such as the student's age, region, historical answer data, and knowledge point mastery, and supports student portrait analysis and dynamic score correction.

[0138] The intelligent question bank and knowledge point bank 605 stores exercises, supplementary exam questions and corresponding knowledge point tags in the nursing field, providing resources for intelligent supplementary exams and personalized exercise push.

[0139] In one embodiment of the present invention, the non-standard answer cleaning and terminology conversion algorithm applied to the non-standard answer cleaning engine 202a is as follows:

[0140] Core data structure: Nursing term conversion library 601, stored in an efficient search format (such as hash table, Trie tree). Keys are non-standard expressions such as dialects / colloquialisms / typos non-std , the value is the corresponding standard nursing term list (May contain multiple candidates, with context clues or frequency of use).

[0141] For example, "turn the patient over and pat his back" = "turn the patient over and pat his back", "erythema" = "pressure ulcer stage I", "redness of the skin".

[0142] Match and replace strategy:

[0143] Maximal Match Algorithm: From the input text S input Starting from the current position, search for the longest non-standard term s that can match in dictionary D non-std .

[0144] Fuzzy matching algorithm: For fragments that are not directly hit, the edit distance (Levenshtein

[0145] Distance)d(s1,s2) calculates its similarity with the non-standard entries in the dictionary. If the normalized similarity Sim(s1,s2) is greater than the preset threshold τ sim , then replace it.

[0146]

[0147] Where |s| represents the length of string s. When Sim(s candidate ,s dict_entry )>τ sim , the match is considered successful.

[0148] Pinyin matching and error correction: Text fragments suspected of being pinyin input are converted into possible Chinese character candidates, and then verified by combining context and fuzzy matching.

[0149] Contextual disambiguation: When a non-standard expression s non-std Mapping to multiple standard terms L std The following strategies can be adopted:

[0150] Based on N-gram language model: calculate s non-std The surrounding context C is associated with each candidate standard term The combined conditional probability Choose the one with the highest probability.

[0151] Rule-based: For example, the most relevant standard terms are selected based on the nursing module to which the topic belongs (such as "basic nursing care" and "emergency nursing care").

[0152] In one embodiment of the present invention, the semantic enhancement and feature extraction algorithm applied to the semantic enhancement and feature extraction engine 202b is as follows:

[0153] The present invention uses a deep learning model to extract semantic features from the cleaned answer text. A preferred embodiment is to use a combination of a pre-trained language model and a sequence model, including but not limited to:

[0154] Combination of BERT (Bidirectional Encoder Representations from Transformers) + BiLSTM (Bidirectional Long Short-Term Memory):

[0155] BERT layer: This layer uses a BERT model fine-tuned on nursing text to process the input token sequence, capture the bidirectional semantic information of each token in the deep context, and output a context-aware word embedding vector.

[0156] BiLSTM layer: The word embedding sequence output by the BERT layer is input into the BiLSTM network to further capture the long-distance temporal dependencies of the text and enhance the representation of sequence information.

[0157] Other alternative implementations include using other advanced Transformer architecture models (such as RoBERTa, ALBERT, etc.) or other sequence models (such as GRU). The ultimate goal is to generate a fixed-dimensional dense vector E that can accurately represent the semantics of the answer. A .

[0158] BERT (Bidirectional Encoder Representations from Transformers):

[0159] Input: After cleaning and Tokenization (segmentation, adding special tags [CLS], [SEP]) of the student's answer Token sequence T = [tok0, tok1, ..., tok N ]Where tok0=[CLS].

[0160] BERT processing: H BERT =BERT(T)=[h0,h1,…,h N ],in is the context-aware word embedding vector of the i-th Token.

[0161] Sentence-level embedding E sent In a preferred embodiment of the present invention, the output vector h0 corresponding to the [CLS] tag can be used as the sentence-level embedding. Other methods include pooling the output vectors of all tokens (such as Mean Pooling or Max Pooling).

[0162] BiLSTM(Bidirectional Long Short-Term Memory):

[0163] Input and dimension matching: embed the Token output by the BERT layer into the sequence H BERT As input. To ensure dimensionality consistency, if necessary, a linear transformation layer can be added between the BERT output and the BiLSTM input.

[0164] Forward LSTM: where h i For H BERT input.

[0165] Backward LSTM: where h i For H BERT input.

[0166] BiLSTM output: is the concatenation of the forward and backward hidden states.

[0167] Final feature vector generation:

[0168] The ultimate goal is to generate a fixed dimension d that can accurately represent the semantics of the answer final The dense vector Serves as input for subsequent scoring and intent recognition.

[0169] Generation method: The output sequence H of all time steps of BiLSTM ″ =[h0 ″ ,…,h N ″ ] Information is aggregated through a pooling layer (e.g., attention pooling or max pooling), and then transformed through a fully connected layer (Fully ConnectedLayer), and finally a fixed-dimensional answer semantic embedding E is obtained. A .

[0170] In one embodiment of the present invention, the key intent and scoring point identification algorithm applied to the key intent and scoring point identification network 202c is as follows:

[0171] Attention-based matching:

[0172] Assume that the semantic embedding of the student's answer is (or its Token sequence is embedded in H A =[h A,1 ,…,hA,M ]).

[0173] Predefined “key scoring points / operation steps” k∈K points , whose embedding vector is

[0174] Attention score (e.g., using scaled dot product attention):

[0175]

[0176] In this scenario, E A (or h A,i ) as the query Q, as key K (and value V), or vice versa. More simply, the similarity can be directly calculated as attention:

[0177]

[0178] Where sim can be the dot product or cosine similarity. k Indicates the degree of attention the answer pays to score point k.

[0179] Sequence labeling models (such as CRF on top of BiLSTM / BERT):

[0180] The task is regarded as a sequence of tokens T = [tok1,…,tok N ]'s every tok i Label a label y i ∈Y tags (such as B-operation step, I-operation step, O-other).

[0181] The model typically contains an embedding layer (such as BERT output), a sequence encoding layer (such as BiLSTM), and a CRF layer.

[0182] The goal of the CRF layer is to find the label sequence that maximizes the conditional probability P(Y|T)

[0183]

[0184] The training data needs to be labeled with a lot of scoring points (tok i →y i ) of nursing answer sample.

[0185] In one embodiment of the present invention, the multi-dimensional intelligent scoring algorithm applied to the multi-dimensional intelligent scoring module 300 is as follows:

[0186] Assume that the final semantic embedding of the student’s answer is E A, the embedding of the standard answer / reference point j is

[0187] Content Accuracy Score (S content ):

[0188]

[0189] Or use average similarity:

[0190]

[0191] Among them, cosine similarity is:

[0192]

[0193] w content is the weight of content accuracy.

[0194] Logical coherence score (S logic ):

[0195] If the sequence of operation steps is obtained through intention recognition

[0196] Seq A =[step1,step2,…,step n ].

[0197] Rule-based: Seq A With the predefined standard operating procedure template Seq Std For example, to calculate Seq A and Seq Std The normalized longest common subsequence (LCS) length or the inverse of the normalized edit distance between them.

[0198]

[0199] Model-based: train a classifier M logic (such as sequence classifier based on RNN or Transformer) to judge Seq A The logical score p logic ∈[0,1].

[0200]

[0201] in is the step sequence Seq A Embedding representation of .

[0202] Final logic score:

[0203] S logic =w logic ×(λSlogic,rule +(1-λ)S logic,model ).

[0204] Where λ is the fusion coefficient.

[0205] Term normative score (S term ):

[0206] After cleaning the non-standard answers, let the set of standard nursing terms used correctly in the answers be T correct , the expected standard term set related to the topic is T expected .

[0207]

[0208] Or based on the quality of the match between the terms identified in the answer and the term base.

[0209] Total score (S total ):

[0210] S total =clip(S content +S logic +S term ,S min ,S max ).

[0211] where w content ,w logic ,w term is the preset weight of each dimension (here it is assumed that it has been integrated into the scoring of each sub-item), S min ,S max is the minimum and maximum value of the total score.

[0212] In one embodiment of the present invention, the dynamic score modification and step-by-step scoring algorithms used in the dynamic score modification engine 402 and the step-by-step and innovative scoring engine 403 are as follows:

[0213] Dialect fault tolerance trigger:

[0214] According to the τ in the student portrait age Age threshold, R dialect A collection of high-incidence areas of dialects to determine whether to trigger dialect tolerance.

[0215] Operational integrity coefficient α and innovation bonus:

[0216] Assume that the standard operation includes N core steps {KP1,…,KP N}. M core steps were identified in the students’ answers.

[0217] Identify K operational steps {IP1,…,IP K}.

[0218] IP at every innovation step j Assign an innovation weight based on its rationality and innovation (e.g., evaluated by an expert system or a small classifier).

[0219] The operational integrity factor α is defined as:

[0220]

[0221] α is usually constrained to a reasonable range, such as [α min ,α max ](such as [0.5,1.3]).

[0222] α'=clip(α,α min ,α max ), is the operational integrity coefficient after constraints.

[0223] Final score adjustments (e.g., content only or total score):

[0224] S' total =S total ×α' or S' content =S content ×α'.

[0225] The total score is then recalculated.

[0226] In one embodiment of the present invention, the knowledge mastery degree heat map generation algorithm applied to the knowledge point mastery degree heat map generation unit 502a is as follows:

[0227] Based on the LDA (Latent Dirichlet Allocation) topic model:

[0228] The LDA model assumes that the document generation process:

[0229] For each document d, from the Dirichlet distribution Dir(α LDA ) extracts the topic distribution θ d .

[0230] For each word w in document d d,n :

[0231] a. From the topic multinomial distribution Multinomial(θ d ) to extract a topic z d,n .

[0232] b. From the selected topic d,n Corresponding word multinomial distribution Extract word w d,n .

[0233] where α LDA and β LDA (Dirichlet prior parameter, φ t ~Dir(β LDA ) is a hyperparameter.

[0234] The goal of model learning is to infer the latent variables: topic distribution Θ = {θ d} and topic-word distribution

[0235] Φ={φ t}.

[0236] A student maps to a specific nursing knowledge point KP j Mastery of the subject It can be approximated by the average probability (or weighted average) of all its answers on topic t:

[0237]

[0238] Based on knowledge point tracking:

[0239] The BKT model estimates the student's mastery of knowledge point k P(L k ), the model usually has four parameters:

[0240] P(L0): Initial mastery probability.

[0241] P(T): The transition probability from unknown to mastered (learning rate).

[0242] P(G): Probability of guessing correctly (Guess).

[0243] P(S): Probability of answering incorrectly (Slip, mastering the question but answering incorrectly).

[0244] After the student answers a question related to knowledge point k, P(L k ∣evidence).

[0245] In one embodiment of the present invention, the 3D / AR visual feedback matching and calling logic applied to the visual feedback matching and calling unit 501b is as follows:

[0246] ERROR - MaterialMappingLibrary: Data structure can be either a nested dictionary or a database table.

[0247] Figure 3A schematic diagram showing the data structure of an "error-material mapping library" in one embodiment of the present invention is shown; and

[0248] Figure 4 An embodiment of the present invention is shown for executing Figure 3 Pseudocode example diagram of the "triggering and calling logic" of a mapping rule.

[0249] like Figure 3 , as shown in Figure 4, the triggering and calling logic is as follows:

[0250] Input: List of detected error codes L errors =[err1,err2,…].

[0251] For L errors Each err in i :

[0252] 1. Find err in error_to_asset_map i .

[0253] 2. If found, get feedback type F type and material ID A id and parameter P params .

[0254] 3. According to F type Call the corresponding rendering engine:

[0255] IF F type =="3D_Animation"THEN Call_3D_Engine(A id ,P params );

[0256] IF F type =="AR_Overlay"THEN Call_AR_Engine(A id ,P params ).

[0257] Although various embodiments of the present invention have been described above, it should be understood that they are presented by way of example only and not limitation. It will be apparent to those skilled in the relevant art that various combinations, modifications, and variations may be made thereto without departing from the spirit and scope of the present invention. Therefore, the breadth and scope of the present invention disclosed herein should not be limited by the exemplary embodiments disclosed above, but should be defined solely in accordance with the appended claims and their equivalents.

Claims

1. A multi-modal intelligent scoring method for subjective questions based on self-adaptation in the field of nursing, characterized by: include: Receive the answer text submitted by the students; Perform format check and formatting on the received answer text; Cleaning and standardizing the formatted answer text to obtain standardized text; Perform in-depth analysis on standardized text and convert it into semantic feature vectors; Identify key intentions and scoring points based on semantic feature vectors; Conduct a comprehensive assessment and produce a preliminary score; Adaptive and flexible adjustments to preliminary scores; The system calculates the final scoring result.

2. The method for intelligent grading of multimodal subjective questions based on nursing field adaptation according to claim 1 is characterized in that: The adaptive and flexible adjustment of the preliminary score includes: Obtain student information from the student portrait database and decide whether to enable "dialect tolerance mode" to correct the initial score; Evaluate whether there are reasonable but non-standard innovative operations in the answer text, and apply the "operation integrity coefficient α" to give flexible scores.

3. The method for intelligent grading of multimodal subjective questions based on nursing field adaptation according to claim 1 is characterized in that: Also includes: Generate multimodal feedback based on the scoring results, including text explanations, 3D operation demonstrations, personal ability heat maps, and intelligent re-examination suggestions; Present the final scoring results and multimodal feedback information to students or teachers.

4. The method for intelligent grading of multimodal subjective questions based on nursing field adaptation according to claim 1 is characterized in that: The formatted answer text is cleaned and standardized to obtain standardized text, including: From the input text S input Starting from the current position, search for the longest non-standard term s that can match in the dictionary non-std ; For the fragments that are not directly hit, calculate their similarity with the non-standard entries in the dictionary. Where |s| represents the length of string s. When Sim(s candidate ,s dict_entry )>τ sim When , the match is considered successful; Convert suspected pinyin input text segments into possible Chinese character candidates, and then verify them based on context and fuzzy matching; When a non-standard expression non-std Mapping to multiple standard terms L std When s non-std The surrounding context C is associated with each candidate standard term The combined conditional probability Select the one with the highest probability, or filter the most relevant standard terms based on the nursing module to which the question belongs.

5. The method for intelligent grading of multimodal subjective questions based on nursing field adaptation according to claim 1 is characterized in that: In-depth analysis of standardized text and conversion of it into semantic feature vectors includes: The BERT model, fine-tuned on nursing text, processes the token sequence of the input standardized text, captures the bidirectional semantic information of each token in the deep context, and outputs context-aware word embedding vectors. The word embedding sequence output by the BERT layer is input into the BiLSTM network to further capture the long-range temporal dependencies of the text and enhance the representation of sequence information. Generate a fixed-dimensional dense vector representing the semantics of the answer, namely the semantic feature vector E A .

6. The method for intelligent grading of multimodal subjective questions based on self-adaptation in the field of nursing according to claim 1 is characterized in that: Identifying key intents and scoring points based on semantic feature vectors includes: Assume that the semantic embedding of the student's answer is Predefined key scoring points / operation steps, k∈K points , whose embedding vector is Calculate the similarity as a score, Where sim is the dot product or cosine similarity, a k Indicates the degree of attention of the answer to the score point k, that is, the similarity.

7. The method for intelligent grading of multimodal subjective questions based on self-adaptation in the field of nursing according to claim 1 is characterized in that: A comprehensive assessment is conducted and a preliminary score is generated, including: Calculate content accuracy scores, Among them, cosine similarity is: w content is the weight of content accuracy; Obtain the sequence of operation steps through intention recognition, Seq A =[step1,step2,…,step n ], Seq A With the predefined standard operating procedure template Seq Std contrast, Train a classifier M logic JudgeSeq A The logical score p logic ∈[0,1], in is the step sequence Seq A The embedding representation of Get the final logic score, Where λ is the fusion coefficient; After cleaning the non-standard answers, let the set of standard nursing terms used correctly in the answers be T correct The expected set of standard terms related to the topic is T expected , The term normativeness score is, Calculate the total score S total =clip(S content +S logic +S term ,S min ,S max ); where w content ,w logic ,w term is the preset weight of each dimension, S min ,S max are the minimum and maximum values ​​of the total score.

8. The method for intelligent grading of multimodal subjective questions based on nursing field adaptation according to claim 2 is characterized in that: Evaluate whether there are reasonable but non-standard innovative operations in the answer, and apply the "operational integrity coefficient α" for flexible scoring, including: Assume that the standard operation includes N core steps {KP1,…,KP N }, M core steps are identified in the students’ answers, and K operation steps are identified as “reasonable but non-standard” or “beneficially innovative” by the “nursing knowledge base / expert rules” {IP1,…,IP K }; IP for each innovation step j Assign an innovation weight iw according to its rationality and innovation j ∈[0,1]; The operational integrity coefficient α is defined as: The final score is calculated as follows: S'=S×α'; where α ' =clip(α,α min ,α max ), is the operational integrity coefficient after constraints.

9. A multi-modal intelligent scoring system for subjective questions based on self-adaptation in the field of nursing, characterized by: include: The user interaction and presentation module is configured to provide a student-side interface for receiving text input of answers to subjective questions and presenting the grading results and multimodal feedback to students or teachers; The data access and preprocessing module is configured to receive and format answers, clean non-standard answers, enhance semantics and extract features, and identify key intents; A multi-dimensional intelligent scoring module is configured to comprehensively score based on three dimensions: content accuracy, logical coherence, and terminology standardization. The comprehensive scoring unit calculates the preliminary score; and The nursing field adaptive and flexible adjustment module is configured to perform adaptive and flexible adjustments based on the preliminary scores.

10. The multi-modal intelligent scoring system for subjective questions based on nursing field adaptation according to claim 9 is characterized in that: Also includes: The multimodal teaching feedback and intelligent linkage module is configured to generate multimodal feedback, update knowledge point heat maps, and push intelligent re-examination suggestions; The core resource and knowledge base module is configured to provide nursing terminology conversion, standard answers, 3D / AR materials and student portrait data support.

Images