A health data screening method and device based on AI
By generating text feature vector sequences, combining semantic association and structural information, the problems of low efficiency and poor accuracy of health survey text data processing in the prior art are solved, and the in-depth understanding and authenticity evaluation of health data are achieved, and the data quality is improved. .
Patent Information
- Application Number
- CN202411359739.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-27
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2044-09-27
AI Technical Summary
The prior art is difficult to accurately capture the deep semantic information in health survey text data, and traditional methods ignore the logical relationships in text structure, resulting in inefficient processing efficiency and high error rates, making it difficult to meet the needs of large-scale data processing.
By obtaining semantic coded vector sequences and structural coded vector sequences, generating text feature vector sequences, combining semantic associations and structural information, performing multi-level understanding and evaluation, and using the authenticity evaluation model for detailed screening.
It significantly improves the authenticity and reliability of health data, realizes the accurate removal of false information, and improves the quality of health survey data.
Smart Images

Figure CN118861269B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an AI-based health data screening method and device. Background Art
[0002] With the rapid development of information technology, health data has become an important resource in the fields of medicine, public health, and health management. As an important part of health data, health survey text data contains a large amount of information such as patient symptom descriptions, living habits, and medical history, which is of great significance for the prevention, diagnosis, and treatment of diseases. However, since health survey text data usually comes from respondents with different backgrounds and different education levels, the authenticity, accuracy, and consistency of its content are often difficult to guarantee. In addition, health survey text data is highly complex and diverse. Traditional manual screening methods are not only inefficient, but also prone to errors, and it is difficult to meet the needs of large-scale data processing.
[0003] In order to improve the processing efficiency and accuracy of health survey text data, relevant technologies have begun to explore the use of artificial intelligence technology (AI) for automated screening. Among the relevant technologies, they mainly rely on natural language processing technology (NLP) and machine learning algorithms. These methods usually perform shallow processing and analysis of text content through text classification, sentiment analysis, keyword extraction and other means. However, these methods often have the following problems when processing complex and multi-level health survey text data: First, it is difficult to accurately capture the deep semantic information in the text, resulting in insufficient understanding of the text content; second, there is a lack of consideration of the text structure, ignoring the logical relationship between the various parts of the text. Summary of the invention
[0004] In order to at least overcome the above-mentioned deficiencies in the prior art, the purpose of the embodiments of the present application is to provide an AI-based health data screening method and device.
[0005] According to one aspect of the present application, a health data screening method based on AI is provided, the method comprising:
[0006] Obtaining a semantic coding vector sequence and a structural coding vector sequence of the health survey text data to be screened; the semantic coding vector sequence includes a keyword node vector of the health survey text data to be screened at a basic semantic understanding depth and a keyword node vector of the health survey text data to be screened at a derived semantic understanding depth;
[0007] Generate a text feature vector sequence of the health survey text data to be screened based on the semantic coding vector sequence and the structural coding vector sequence; the text feature vector sequence includes a keyword node vector of the health survey text data to be screened at a basic semantic understanding depth and a reinforced semantic understanding vector of the health survey text data to be screened at X semantic understanding depths, and the reinforced semantic understanding vector of the health survey text data to be screened at any one semantic understanding depth is obtained based on the keyword node vector of the health survey text data to be screened at a derived semantic understanding depth and the structural coding vector sequence;
[0008] The authenticity of the text feature vector sequence is evaluated through an authenticity evaluation model to generate an authenticity evaluation result of the health survey text data to be screened, and the health survey text data to be screened is screened according to the authenticity evaluation result; wherein the authenticity evaluation result includes the authenticity evaluation result of each text feedback unit in the health survey text data to be screened.
[0009] In a possible implementation of the first aspect, generating a text feature vector sequence of the health survey text data to be screened according to the semantic coding vector sequence and the structural coding vector sequence includes:
[0010] Strengthening and blending the keyword node vectors of the health survey text data to be screened at the derived semantic understanding depth and the structural coding vector sequence to generate a strengthened structural coding vector sequence;
[0011] The keyword node vector of the health survey text data to be screened at the derived semantic understanding depth is merged with the enhanced structure coding vector sequence to generate the enhanced semantic understanding vector of the health survey text data to be screened at the derived semantic understanding depth;
[0012] A text feature vector sequence of the health survey text data to be screened is generated based on the keyword node vectors of the health survey text data to be screened at the basic semantic understanding depth and the enhanced semantic understanding vectors of the health survey text data to be screened at the derived semantic understanding depth.
[0013] In a possible implementation of the first aspect, the method further includes:
[0014] Iteratively extract semantic features of the enhanced semantic understanding vector of the health survey text data to be screened at the derived semantic understanding depth by Y semantic feature extraction units to generate the enhanced semantic understanding vector of the health survey text data to be screened at Y semantic understanding depths, where Y is a positive integer;
[0015] The method of generating a text feature vector sequence of the health survey text data to be screened based on the keyword node vector of the health survey text data to be screened at the basic semantic understanding depth and the enhanced semantic understanding vector of the health survey text data to be screened at the derived semantic understanding depth includes:
[0016] The keyword node vectors of the health survey text data to be screened at the basic semantic understanding depth, the enhanced semantic understanding vectors of the health survey text data to be screened at the derived semantic understanding depth, and the enhanced semantic understanding vectors of the health survey text data to be screened at Y semantic understanding depths are loaded into a sequence to generate a text feature vector sequence of the health survey text data to be screened.
[0017] In a possible implementation of the first aspect, the step of reinforcing and blending the keyword node vectors of the health survey text data to be screened at the derived semantic understanding depth and the structure coding vector sequence to generate an enhanced structure coding vector sequence includes:
[0018] Performing feature combination on the keyword node vector of the health survey text data to be screened at the derived semantic understanding depth and the structural coding vector sequence according to the semantic association focusing strategy to generate a first feature combination result;
[0019] Performing feature combination on the keyword node vector of the health survey text data to be screened at the derived semantic understanding depth and the structure coding vector sequence according to the logical structure focusing strategy to generate a second feature combination result;
[0020] Perform feature combination on the combination result of the first feature combination result and the second feature combination result according to the recurrent neural network unit to generate a reinforced structure encoding vector sequence.
[0021] In a possible implementation of the first aspect, the step of fusing the keyword node vector of the health survey text data to be screened at the derived semantic understanding depth with the enhanced structure encoding vector sequence to generate the enhanced semantic understanding vector of the health survey text data to be screened at the derived semantic understanding depth includes:
[0022] Recursively operating the enhanced structure coding vector sequence through a target recursive model to generate a recursive result of the enhanced structure coding vector sequence;
[0023] The recursive result of the enhanced structural encoding vector sequence is accumulated on an element-by-element basis with the keyword node vector of the health survey text data to be screened at the derived semantic understanding depth to generate an enhanced semantic understanding vector of the health survey text data to be screened at the derived semantic understanding depth.
[0024] In a possible implementation of the first aspect, the method further includes:
[0025] Semantically decompose the keyword node vectors of the health survey text data to be screened at the basic semantic understanding depth and the enhanced semantic understanding vectors of the health survey text data to be screened at X semantic understanding depths to generate X+1 core component vector sets and X+1 auxiliary component vector sets;
[0026] According to the adaptive focusing strategy, the X+1 core component vector sets and the X+1 auxiliary component vector sets are respectively feature enhanced to generate X+1 enhanced core component vector sets and X+1 enhanced auxiliary component vector sets as component vector sets at the multi-level semantic understanding level;
[0027] Using the component vector set of the multi-level semantic understanding level to enhance the text feature vectors in the text feature vector sequence, to generate an enhanced text feature vector sequence;
[0028] The process of performing semantic decomposition on the keyword node vectors of the health survey text data to be screened at the basic semantic understanding depth and the enhanced semantic understanding vectors of the health survey text data to be screened at X semantic understanding depths to generate X+1 core component vector sets and X+1 auxiliary component vector sets includes:
[0029] Recursively operate the keyword node vectors of the health survey text data to be screened at the basic semantic understanding depth through the first context knowledge space and the second context knowledge space respectively to generate the core recursive vector and the auxiliary recursive vector of the health survey text data to be screened at the basic semantic understanding depth;
[0030] Performing semantic compression processing on the core recursive vector and the auxiliary recursive vector of the health survey text data to be screened at the basic semantic understanding depth, respectively, to generate a first core component vector set and a first auxiliary component vector set;
[0031] Recursively operate the enhanced semantic understanding vector of the health survey text data to be screened at the a+1th semantic understanding depth according to the first contextual knowledge space and the second contextual knowledge space respectively, to generate a core recursive vector and an auxiliary recursive vector of the health survey text data to be screened at the a+1th semantic understanding depth, where a is a positive integer not greater than X;
[0032] The core recursive vectors and the auxiliary recursive vectors of the health survey text data to be screened at the a+1th semantic understanding depth are semantically compressed to generate an a+1th core component vector set and an a+1th auxiliary component vector set.
[0033] In a possible implementation of the first aspect, the adaptive focusing strategy includes a recursive unit, a semantic association focusing unit, and a logical structure focusing unit; the step of performing feature enhancement on the X+1 core component vector sets and the X+1 auxiliary component vector sets according to the adaptive focusing strategy to generate X+1 enhanced core component vector sets and X+1 enhanced auxiliary component vector sets includes:
[0034] According to the recursive unit, feature combination is performed on the b-th core component vector set and the b-th auxiliary component vector set respectively to generate a feature combination result of the b-th core component vector set and a feature combination result of the b-th auxiliary component vector set, where b is a positive integer not greater than X+1;
[0035] According to the feature combination result of the b-th core component vector set and the feature combination result of the b-th auxiliary component vector set by the logical structure focusing unit, the logical structure dimension is strengthened to generate the b-th core reference feature and the b-th auxiliary reference feature;
[0036] According to the semantic association focusing unit, the b-th core reference feature and the b-th auxiliary reference feature are respectively enhanced in terms of semantic association dimensions to generate a b-th enhanced core component vector set and a b-th enhanced auxiliary component vector set.
[0037] In a possible implementation of the first aspect, the component vector set of the multi-level semantic understanding level includes X+1 enhanced core component vector sets and X+1 enhanced auxiliary component vector sets; and the step of using the component vector set of the multi-level semantic understanding level to enhance the text feature vectors in the text feature vector sequence to generate an enhanced text feature vector sequence includes:
[0038] Merging the X+1 enhanced core component vector sets to generate a core semantic enhancement map, and merging the X+1 enhanced auxiliary component vector sets to generate an auxiliary semantic enhancement map;
[0039] The core semantic enhancement graph and the keyword node vector of the health survey text data to be screened are enhanced and integrated at the basic semantic understanding depth through a recursive model to generate a first enhanced integration result;
[0040] According to the recursive model, the auxiliary semantic enhancement map and the enhanced semantic understanding vector of the health survey text data to be screened at the Xth semantic understanding depth are enhanced and blended to generate a second enhanced fusion result;
[0041] Among them, the process of merging the X + 1 enhanced core component vector sets to generate the core semantic enhancement graph includes:
[0042] Using a semantic merging network model to merge the feature vector data loaded at the c-th time and the (X - c + 1)-th enhanced core component vector set to generate the c-th merging result;
[0043] Among them, c is a positive integer not greater than X; when c = 1, the feature vector data loaded at the c-th time is the (X + 1)-th enhanced core component vector set; when 1 < c < X + 1, the feature vector data loaded at the c-th time is the (c - 1)-th merging result;
[0044] Among them, the semantic merging network model includes a context-based mapping model; the process of using the semantic merging network model to merge the feature vector data loaded at the c-th time and the (X - c + 1)-th enhanced core component vector set to generate the c-th merging result includes:
[0045] Calibrating the feature vector data loaded at the c-th time with the semantic compression result according to the context-based mapping model to generate a calibration result; the semantic compression result is obtained by performing semantic compression processing on the (X - c + 1)-th enhanced core component vector set;
[0046] Adjusting the position adjustment coefficient and the weight adjustment coefficient according to the calibration result to generate an adjusted position adjustment coefficient and an adjusted weight adjustment coefficient;
[0047] Calculating the dot product of the (X - c + 1)-th enhanced core component vector set and the first recursively extended result to generate a reference semantic vector sequence; the first recursively extended result is obtained by recursively extending the adjusted weight adjustment coefficient;
[0048] Accumulating the reference semantic vector sequence and the second recursively extended result at the element level to generate the c-th merging result; the second recursively extended result is obtained by recursively extending the adjusted position adjustment coefficient.
[0049] In a possible implementation manner of the first aspect, the obtaining of the semantic encoding vector sequence and the structural encoding vector sequence of the health survey text data to be screened includes:
[0050] Performing feature extraction on the health survey text data to be screened through a first semantic feature extraction unit to generate a keyword node vector of the health survey text data to be screened at the basic semantic understanding depth;
[0051] Performing feature extraction on the keyword node vector of the health survey text data to be screened at the basic semantic understanding depth by a second semantic feature extraction unit to generate the keyword node vector of the health survey text data to be screened at the derived semantic understanding depth;
[0052] The keyword node vector of the health survey text data to be screened at the basic semantic understanding depth and the keyword node vector of the health survey text data to be screened at the derived semantic understanding depth are used as the semantic encoding vector sequence of the health survey text data to be screened;
[0053] Performing dependency syntactic analysis on the health survey text data to be screened according to the logical structure model to generate dependency syntactic analysis results of the health survey text data to be screened;
[0054] Performing structured mapping processing on the dependency syntax analysis results of the health survey text data to be screened according to the structured mapping table to generate a structured encoding vector sequence of the health survey text data to be screened;
[0055] The step of performing dependency syntactic analysis on the health survey text data to be screened according to the logical structure model to generate the dependency syntactic analysis result of the health survey text data to be screened includes:
[0056] The health survey text data to be screened is divided into independent words, and after marking the part of speech for each word, each word is restored to its basic form;
[0057] According to the part of speech marked for each word, the core words in the health survey text data to be screened are identified as the central nodes of the dependency relationship;
[0058] According to a preset dependency relationship rule, an initial dependency edge is established for each word in the health survey text data to be screened, wherein the dependency edge connects the word with the word on which it depends;
[0059] Extracting feature data of each word and position information and context information of each word in the sentence, wherein the feature data includes part-of-speech feature, morphological feature, and semantic role feature;
[0060] Inputting the feature data of each word and the position information and context information of each word in the sentence into the trained logical structure model for reasoning to generate an initial dependency graph;
[0061] According to the initial dependency graph, a dependency syntax tree is constructed as the dependency syntax analysis result, wherein each node in the dependency syntax tree represents a word, and the edges represent the dependency relationship between words.
[0062] According to one aspect of an embodiment of the present application, an AI-based health data screening device is provided, wherein the AI-based health data screening device comprises a processor and a machine-readable storage medium, wherein the machine-readable storage medium stores machine-executable instructions, and the machine-executable instructions are loaded and executed by the processor to implement the AI-based health data screening method in any one of the aforementioned possible implementations.
[0063] In the technical solutions provided in some embodiments of the present application, the embodiments of the present application significantly improve the authenticity and reliability of health data by deeply analyzing the semantic and structural information of the health survey text data to be screened. First, a semantic coding vector sequence and a structural coding vector sequence are constructed, which not only captures the basic semantic features of the text, but also deeply mines the information of the derived semantic level, and combines the structural features of the text to generate a comprehensive text feature vector sequence. In this process, the integrity and accuracy of the text information are ensured through enhanced processing of multiple semantic understanding depths. Subsequently, the authenticity evaluation model is used to conduct a detailed evaluation of the feature vector sequence, which not only obtains the authenticity judgment of the entire text, but also the authenticity evaluation results of each text feedback unit, realizing the accurate elimination of false or misleading information, which can greatly improve the quality of health survey data. BRIEF DESCRIPTION OF THE DRAWINGS
[0064] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required to be used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can be extracted in combination with these drawings without creative work.
[0065] Figure 1 A flowchart of an AI-based health data screening method provided in an embodiment of the present application;
[0066] Figure 2 This is a schematic block diagram of the architecture of an AI-based health data screening device provided in an embodiment of the present application. DETAILED DESCRIPTION
[0067] The following description is intended to enable one of ordinary skill in the art to implement and combine the present application, and the description is provided in the context of a specific application scenario and its requirements. It will be apparent to one of ordinary skill in the art that various changes may be made to the disclosed embodiments, and that the general principles defined in the present application may be applied to other embodiments and application scenarios without departing from the principles and scope of the present application. Therefore, the present application is not limited to the described embodiments, but should be given the broadest scope consistent with the claims.
[0068] Figure 1 This is a flow chart of an AI-based health data screening method provided by an embodiment of the present application. The AI-based health data screening method is introduced in detail below.
[0069] Step S110, obtaining a semantic coding vector sequence and a structural coding vector sequence of the health survey text data to be screened. The semantic coding vector sequence includes a keyword node vector of the health survey text data to be screened at a basic semantic understanding depth and a keyword node vector of the health survey text data to be screened at a derived semantic understanding depth.
[0070] In detail, the semantic coding vector sequence refers to a sequence of key semantic information in text data converted into numerical vectors through semantic analysis technology. In this semantic coding vector sequence, each semantic coding vector represents the coded representation of a certain keyword or key phrase in the text at a specific semantic understanding depth. These semantic coding vectors capture the semantic meaning of words, enabling machines to understand the text content.
[0071] The basic semantic understanding depth refers to the most basic semantic analysis of the text, identifying the key words, phrases and their basic meanings in the text. At this depth, the focus is on the literal meaning of words and the direct contextual relationship. The derived semantic understanding depth is based on the basic semantic understanding, and further explores the deep connections between words, implicit meanings, and derived meanings in the context. At this depth, attempts are made to understand more complex semantic phenomena such as indirect relationships, metaphors, and symbols between words.
[0072] The keyword node vector refers to the conversion of keywords or key phrases in the text into numerical vectors through some algorithms (such as word embedding, TF-IDF, etc.). These vectors can represent the position of keywords in a specific semantic space, so as to be used for subsequent semantic analysis and processing. The structural encoding vector sequence refers to the sequence of converting the syntactic structure, paragraph structure, etc. of text data into numerical vectors. These vectors capture the organizational structure information of the text, allowing the machine to understand the relationship and hierarchy between the various parts of the text.
[0073] For example, in this embodiment, it is assumed that the server receives a text data from a health questionnaire, the content of which is "I have been feeling a little tired recently, my sleep quality is not very good, I often wake up at night, and my appetite has decreased. I don't have any cold symptoms, but I occasionally have headaches. I suspect that the work pressure is too great." The server needs to process this health survey text data to be screened to obtain its semantic encoding vector sequence and structural encoding vector sequence.
[0074] Specifically, the first semantic feature extraction unit starts working, analyzes the input text, and identifies words with key semantic information. For example, in this text, words such as "tired", "poor sleep quality", "loss of appetite", "headache", and "high work pressure" are key information describing health conditions. After being processed by the feature extraction unit, these words will be converted into vector form. Assume that the word "tired" is converted into a specific vector [v1] = [0.1, 0.2, -0.1] (this is just a simple representation of the vector for the sake of example, the actual vector dimension may be higher), and "poor sleep quality" is converted into [v2] = [0.3, -0.1, 0.2], etc. These vectors are combined to form a keyword node vector at the basic semantic understanding depth.
[0075] Then, the second semantic feature extraction unit further mines semantics based on the first semantic feature extraction unit. For the words in the keyword node vector under the basic semantics obtained above, it will consider the derivative relationship between the words. For example, there may be a causal derivative relationship between "fatigue" and "high work pressure". This relationship will be captured by the second semantic feature extraction unit and feature extracted. Assume that "fatigue" has a vector representation of [v1']=[0.2, 0.3, -0.2] (the vector value has changed to reflect the derived semantics) under the derived semantic understanding depth due to its relationship with "high work pressure". Similarly, "poor sleep quality" has a vector representation of [v2']=[0.4, -0.2, 0.3] under the derived semantics due to its possible association with "headache". These new vectors are combined into the keyword node vector of the health survey text data to be screened under the derived semantic understanding depth.
[0076] Next, the keyword node vectors obtained above at the basic semantic understanding depth (e.g., composed of [v1], [v2], etc.) and the keyword node vectors at the derived semantic understanding depth (e.g., composed of [v1'], [v2'], etc.) are combined to form a semantic encoding vector sequence. This sequence contains the encoding of the semantic information of the text data from different semantic understanding depths.
[0077] Finally, the logical structure model is called to process the text data.
[0078] First, the health survey text data to be screened are divided into independent words, such as "I", "recently", "feeling", "body", "some", "tired", etc., and the part of speech is marked for each word. For example, "I" is a pronoun, "recently" is a time adverb, "feeling" is a verb, etc. Then each word is restored to its basic form (for example, "feeling" does not need to be restored, and "I" does not need special restoration, etc.).
[0079] According to the part of speech marked for each word, the core words in the health survey text data to be screened are identified as the central nodes of the dependency relationship. In this example, "feeling" may be a core word because it is a key verb that expresses the subject's feelings.
[0080] According to the preset dependency rules, an initial dependency edge is established for each word in the health survey text data to be screened. For example, "fatigue" depends on "feeling", "sleep quality" depends on "feeling", etc.
[0081] Extract the feature data of each word and the position information and context information of each word in the sentence. For the word "tired", its part of speech feature is an adjective, its morphological feature may not have special changes, and its semantic role feature is to describe the physical state; its position in the sentence is after "feeling", and its context is to describe the physical feeling. Input this information (such as various features, position information, context information, etc. of "tiredness") into the trained logical structure model for reasoning to generate the initial dependency graph.
[0082] According to the initial dependency graph, a dependency syntax tree is constructed as the result of dependency syntax analysis. In the dependency syntax tree, the node "feeling" may be in the center, connected to nodes such as "tired" and "poor sleep quality", indicating the dependency relationship.
[0083] Furthermore, the present embodiment is also provided with a structured mapping table, which stores the mapping relationship between the dependency syntactic structure and the structural coding vector. For the dependency syntactic tree generated previously, it is converted into a structural coding vector sequence according to this mapping table. For example, if the dependency relationship between "feeling" and "tiredness" in the dependency syntactic tree corresponds to a specific vector [s1]=[0.5, -0.1, 0.3] in the mapping table, the dependency relationship between "feeling" and "poor sleep quality" corresponds to [s2]=[0.4, 0.2, -0.1], etc., these vectors are combined to form a structural coding vector sequence.
[0084] Step S120, generating a text feature vector sequence of the health survey text data to be screened based on the semantic coding vector sequence and the structural coding vector sequence. The text feature vector sequence includes a keyword node vector of the health survey text data to be screened at a basic semantic understanding depth and a reinforced semantic understanding vector of the health survey text data to be screened at X semantic understanding depths, and the reinforced semantic understanding vector of the health survey text data to be screened at any one semantic understanding depth is obtained based on the keyword node vector of the health survey text data to be screened at a derived semantic understanding depth and the structural coding vector sequence.
[0085] In detail, the text feature vector sequence refers to the integration of various semantic and structural features of the text to form a vector sequence that can fully represent the text information. This sequence includes the keyword node vector of the text at the basic semantic understanding depth, the enhanced semantic understanding vector at multiple derived semantic understanding depths, etc., which is used for subsequent tasks such as text classification and authenticity assessment.
[0086] The enhanced semantic understanding vector is a vector obtained by enhancing the keyword node vector by combining semantic association and structural information at the derived semantic understanding depth. These vectors not only contain the basic semantic information of the keyword, but also incorporate the overall structure and contextual relationship of the text, so that the machine can understand the meaning of the text more deeply.
[0087] For example, in this embodiment, after obtaining the previous semantic coding vector sequence and structural coding vector sequence, a text feature vector sequence of the health survey text data to be screened is generated based on these information.
[0088] In detail, the keyword node vector and the structure coding vector sequence of the health survey text data to be screened at the derived semantic understanding depth can be feature combined according to the semantic association focusing strategy to generate a first feature combination result.
[0089] For example, the "fatigue" (vector [v1']=[0.2, 0.3, -0.2]) in the keyword node vector under the derived semantic understanding depth and the dependency vector related to "fatigue" in the structural encoding vector sequence (such as [s1]=[0.5, -0.1, 0.3]) may be multiplied and added to obtain a new vector [r1]=[0.20.5+0.3(-0.1)+(-0.2)0.3]=[0.1-0.03-0.06]=[0.01] (the calculation is simplified here, and the actual operation is more complicated), which is part of the first feature combination result. The same operation is performed on other vectors in the keyword node vector under the derived semantic understanding depth and related vectors in the structural encoding vector sequence to obtain the first feature combination result.
[0090] On this basis, the logical structure focusing strategy may combine from the perspective of the logical relationship of the sentence structure. For example, a special combination is made for the word vectors related to the causal relationship in the sentence. If "work pressure is high" (vector [v3'] under the derived semantics) and the structural vector representing the causal relationship (assuming it is [s3]), according to the logical structure focusing strategy, a logical structure-oriented combination will be performed to obtain a new vector [r2], which is part of the second feature combination result. The second feature combination result is obtained by combining all relevant vectors.
[0091] Assuming that the vector set obtained by the first feature combination result is [R1]={[r1], [r2], ...}, and the vector set obtained by the second feature combination result is [R2]={[r3], [r4], ...}, the recurrent neural network unit will input the vectors in [R1] and [R2] in a certain order and perform a cyclic calculation. For example, for [r1] and [r3], after the calculation of the recurrent neural network unit, a new vector [s1'] is obtained. After performing such calculations on all vectors, a sequence of reinforced structural encoding vectors is obtained.
[0092] The target recursive model is used to recursively operate the enhanced structure encoding vector sequence to generate a recursive result of the enhanced structure encoding vector sequence.
[0093] Assume that the sequence of reinforced structure encoding vectors is {[s1'], [s2'], ...}, and the target recursive model will process these vectors according to a certain recursive formula. For example, for [s1']=[0.1, 0.2, -0.1], after one recursive operation, it becomes [s1'']=[0.2, 0.3, -0.2] (the recursive operation here is performed according to the algorithm of the target recursive model, just a simple example), and the recursive result is obtained after the entire sequence of reinforced structure encoding vectors is recursively operated.
[0094] Assuming that one of the vectors in the keyword node vector under the derived semantic understanding depth is [v1'] = [0.2, 0.3, -0.2], and the corresponding vector in the recursive result of the enhanced structure encoding vector sequence is [s1''] = [0.2, 0.3, -0.2], then the new vector obtained after element-level accumulation is [v1''] = [0.4, 0.6, -0.4]. After performing such operations on all vectors, the enhanced semantic understanding vector of the health survey text data to be screened under the derived semantic understanding depth is obtained.
[0095] Assume that the server needs to generate enhanced semantic understanding vectors at 3 semantic understanding depths (i.e., X=3). Y semantic feature extraction units are used to iteratively extract the enhanced semantic understanding vectors of the health survey text data to be screened at the derived semantic understanding depths to generate enhanced semantic understanding vectors of the health survey text data to be screened at Y semantic understanding depths, where Y is a positive integer.
[0096] For example, for the previously obtained enhanced semantic understanding vector at the derived semantic understanding depth, the semantic feature extraction unit will mine the semantic information again and iterate. Assume that after the first iteration, a new enhanced semantic understanding vector is obtained, in which one vector changes from [v1'']=[0.4, 0.6, -0.4] to [v1''']=[0.5, 0.7, -0.5]. After the second iteration, another new enhanced semantic understanding vector is obtained, and so on.
[0097] Assuming that the keyword node vector at the basic semantic understanding depth is {[v1], [v2], ...}, the enhanced semantic understanding vector at the derived semantic understanding depth is {[v1''], [v2''], ...}, and the enhanced semantic understanding vector at the semantic understanding depth after iteration is {[v1'''], [v2'''], ...}, these vectors are combined in sequence to form a text feature vector sequence.
[0098] Step S130, conduct an authenticity assessment on the text feature vector sequence through an authenticity assessment model, generate an authenticity assessment result of the health survey text data to be screened, and screen the health survey text data to be screened according to the authenticity assessment result. The authenticity assessment result includes the authenticity assessment result of each text feedback unit in the health survey text data to be screened.
[0099] In detail, the authenticity assessment model is a model built using machine learning or deep learning technology to assess the authenticity of text data. The authenticity assessment model determines whether the text content is authentic and reliable by analyzing the feature vector sequence of the text, thereby helping to filter out false information or misleading content.
[0100] The text feedback unit refers to the part of the text data that can be evaluated for authenticity separately. For example, in a health survey text, each symptom description, each medical experience, etc. can be regarded as a text feedback unit. The authenticity evaluation model will evaluate these units one by one and generate the authenticity evaluation results for each unit.
[0101] For example, this embodiment has obtained a text feature vector sequence of the health survey text data to be screened, and now uses the authenticity assessment model to judge the authenticity of this text data and performs screening based on the results.
[0102] First, the authenticity assessment model has been pre-trained with a large amount of health survey text data, which has labeled authenticity information (for example, true health status reports are marked as 1 and false ones are marked as 0).
[0103] The text feature vector sequence of the health survey text data to be screened can be input into the authenticity assessment model. For example, the vectors {[v1], [v2], ...} in the text feature vector sequence, the authenticity assessment model will analyze these vectors according to its internal algorithm structure. Assume that there are some neural network layers in the model, such as a fully connected layer and an activation function layer. For the input vector, the fully connected layer calculates the product of the vector and the weight and the sum of the bias, and then performs a nonlinear transformation through an activation function (such as a ReLU function). After multiple layers of such calculations, the model will output a probability value, indicating the probability that this text data is true. Assume that this probability value is 0.8. Since this probability value is high, it means that the model believes that this text data has a greater probability of being true. At the same time, the model can also evaluate each text feedback unit in the text data. For example, for the part describing "fatigue", the probability of its authenticity is evaluated to be 0.9, and for the part describing "poor sleep quality", the probability of its authenticity is evaluated to be 0.85, etc. These individual evaluation results constitute the authenticity evaluation results of the health survey text data to be screened.
[0104] If the overall authenticity assessment result (for example, the probability value of 0.8 obtained above) exceeds the threshold preset by the server (assuming it is 0.7), then the health survey text data to be screened will be judged as credible data and can be stored in the database of real health survey data for subsequent health analysis and other operations. For example, if the authenticity assessment result of a text feedback unit is lower than a certain sub-threshold (for example, the authenticity probability of the "headache" part is 0.6, which is lower than the sub-threshold 0.7), then the server may mark this text feedback unit for subsequent manual review or further analysis. If the overall authenticity assessment result is lower than the preset threshold, then this data may be marked as suspicious data and stored in a separate database for further review.
[0105] Based on the above steps, the embodiment of the present application significantly improves the authenticity and reliability of health data by deeply analyzing the semantic and structural information of the health survey text data to be screened. First, a semantic coding vector sequence and a structural coding vector sequence are constructed, which not only captures the basic semantic features of the text, but also deeply mines the information of the derived semantic level, and combines the structural features of the text to generate a comprehensive text feature vector sequence. In this process, the integrity and accuracy of the text information are ensured through enhanced processing of multiple semantic understanding depths. Subsequently, the authenticity evaluation model is used to conduct a detailed evaluation of the feature vector sequence, which not only derives the authenticity judgment of the entire text, but also the authenticity evaluation results of each text feedback unit, realizing the accurate elimination of false or misleading information, which can greatly improve the quality of health survey data.
[0106] In a possible implementation, step S120 includes:
[0107] Step S121, strengthening and blending the keyword node vectors of the health survey text data to be screened at the derived semantic understanding depth and the structural coding vector sequence to generate a strengthened structural coding vector sequence.
[0108] Step S122, fusing the keyword node vector of the health survey text data to be screened at the derived semantic understanding depth with the enhanced structure coding vector sequence to generate the enhanced semantic understanding vector of the health survey text data to be screened at the derived semantic understanding depth.
[0109] Step S123, generating a text feature vector sequence of the health survey text data to be screened based on the keyword node vectors of the health survey text data to be screened at the basic semantic understanding depth and the enhanced semantic understanding vectors of the health survey text data to be screened at the derived semantic understanding depth.
[0110] In this embodiment, when processing health survey text data, the keyword node vector and structure encoding vector sequence under the derived semantic understanding depth are operated according to the semantic association focusing strategy. Take the previously mentioned health survey text "I feel a little tired recently, my sleep quality is not very good, I often wake up at night, and my appetite has also decreased. I don't have cold symptoms, but I occasionally have headaches. I suspect that the work pressure is too high." as an example.
[0111] The vector corresponding to "tiredness" in the keyword node vector under the derived semantic understanding depth is assumed to be [v_{tiredness}']=[0.2, 0.3, -0.2], and the dependency vector related to "tiredness" in the structural encoding vector sequence (such as the structural relationship vector related to expressing physical state) is assumed to be [s_{tiredness}]=[0.5, -0.1, 0.3]. According to the semantic association focus strategy, the association value of the two is calculated. One possible calculation method is to first multiply the corresponding elements to obtain [0.2×0.5, 0.3×(-0.1), -0.2×0.3] or [0.1, -0.03, -0.06], and then add these results to obtain a combined value, which is the combination result of the keyword "tiredness" and the related structural vector. The same operation is performed on other keywords in the keyword node vector under the derived semantic understanding depth, such as "poor sleep quality" and "headache" and their corresponding structural encoding vectors, and finally the first feature combination result is obtained.
[0112] Based on the logical structure focus strategy, the server will pay attention to the logical relationship in the sentence. For example, there is a causal logical relationship between "high work pressure" and "fatigue" in the text. The vector of "high work pressure" under the derived semantics is assumed to be [v_{high work pressure}'], and the corresponding structural vector representing the causal relationship is assumed to be [s_{causal}]. The server will perform special combination calculations based on the logical structure focus strategy. For example, according to a specific logical structure weight formula, a new vector is calculated as the combination result of this pair of vectors. This operation is performed on all keyword vectors and structural vectors with logical relationships in the text to obtain the second feature combination result.
[0113] Next, the recurrent neural network unit starts working and combines the first feature combination result and the second feature combination result obtained previously. Assume that the vector set obtained from the first feature combination result is [R_1]={[r_1], [r_2], ...}, and the vector set obtained from the second feature combination result is [R_2]={[r_3], [r_4], ...). The recurrent neural network unit inputs the vectors in [R_1] and [R_2] in sequence according to its internal loop calculation rules. For example, first input [r_1] and [r_3], and after calculation within the unit, a new vector [s_1'] is obtained, and then the next pair of vectors are input for calculation, and finally the enhanced structure encoding vector sequence {[s_1'], [s_2'], ...} is obtained.
[0114] The server calls the target recursive model to process the sequence of enhanced structural coding vectors. Taking the vector [s_1']=[0.1, 0.2, -0.1] in the sequence of enhanced structural coding vectors as an example, the target recursive model processes [s_1'] according to its internal recursive algorithm (such as a recursive calculation method based on time steps). Assume that after one recursive operation, it becomes [s_1'']=[0.2, 0.3, -0.2]. This recursive operation is performed on each vector in the sequence of enhanced structural coding vectors to obtain the recursive result of the entire sequence of enhanced structural coding vectors.
[0115] For the vectors in the keyword node vector under the derived semantic understanding depth, such as [v_{tired}']=[0.2, 0.3, -0.2] corresponding to "fatigue", and the corresponding vector [s_1'']=[0.2, 0.3, -0.2] in the recursive result of the enhanced structural encoding vector sequence, element-wise accumulation is performed. The new vector [v_{tired}'']=[0.4, 0.6, -0.4] is obtained. The other vectors in the keyword node vector under the derived semantic understanding depth are operated in the same way, and finally the enhanced semantic understanding vector of the health survey text data to be screened under the derived semantic understanding depth is obtained.
[0116] Assume that the server needs to consider the enhanced semantic understanding vectors at three semantic understanding depths (i.e., Y=3 mentioned earlier). Based on the enhanced semantic understanding vectors at the derived semantic understanding depth obtained previously, iterative semantic feature extraction is performed through the semantic feature extraction unit. For example, for the enhanced semantic understanding vector [v_{tired}'']=[0.4, 0.6, -0.4] corresponding to "tired", after the first iteration, according to the algorithm of the semantic feature extraction unit (which may involve further mining of surrounding semantic relationships), it becomes [v_{tired}''']=[0.5, 0.7, -0.5]. Similar iterative operations are performed on the enhanced semantic understanding vectors corresponding to other keywords to obtain enhanced semantic understanding vectors at different semantic understanding depths.
[0117] The server loads the keyword node vectors of the health survey text data to be screened at the basic semantic understanding depth (such as {[v_{basic 1}], [v_{basic 2}], ...}), the enhanced semantic understanding vectors at the derived semantic understanding depth (such as {[v_{tired}''], [v_{other}''], ...}), and the enhanced semantic understanding vectors at the iterative semantic understanding depth (such as {[v_{tired}'''], [v_{other}'''], ...}) into the sequence in order. This sequence constitutes the text feature vector sequence of the health survey text data to be screened.
[0118] In a possible implementation, the method further includes: performing iterative semantic feature extraction on the enhanced semantic understanding vector of the health survey text data to be screened at a derived semantic understanding depth through Y semantic feature extraction units to generate an enhanced semantic understanding vector of the health survey text data to be screened at Y semantic understanding depths, where Y is a positive integer.
[0119] Step S123 includes: loading the keyword node vectors of the health survey text data to be screened at the basic semantic understanding depth, the enhanced semantic understanding vectors of the health survey text data to be screened at the derived semantic understanding depth, and the enhanced semantic understanding vectors of the health survey text data to be screened at Y semantic understanding depths into a sequence to generate a text feature vector sequence of the health survey text data to be screened.
[0120] In this embodiment, the enhanced semantic understanding vector of the health survey text data to be screened under the derived semantic understanding depth has been obtained before. Take the previous health survey text "I feel a little tired recently, my sleep quality is not very good, I often wake up at night, and my appetite has also decreased. I don't have cold symptoms, but I occasionally have headaches. I suspect that the work pressure is too high." as an example, the enhanced semantic understanding vector of the keyword "tired" under the derived semantic understanding depth is assumed to be [v_{tired}''] = [0.4, 0.6, -0.4].
[0121] The server starts to iterate using the semantic feature extraction unit. For the first semantic feature extraction unit, it analyzes the previous enhanced semantic understanding vector. Taking the vector [v_{tired}''] corresponding to "fatigue" as an example, this semantic feature extraction unit will deeply explore the semantic relationship in the text. For example, it will consider the connection between "fatigue" and other related symptoms (such as "poor sleep quality") at a broader semantic level. Through a predefined semantic analysis algorithm, this algorithm may be based on the statistical information of a semantic knowledge base or a large-scale corpus. For example, if it is found in the semantic knowledge base that "fatigue" has a strong association with long-term "poor sleep quality", and this association is not fully reflected in the current text, then [v_{tired}''] will be adjusted. Assume that after the first iteration, the vector corresponding to "fatigue" becomes [v_{tired}''']=[0.5, 0.7, -0.5]. The same operation is performed on the enhanced semantic understanding vectors corresponding to other keywords (such as "poor sleep quality" and "headache", etc.), so as to obtain the enhanced semantic understanding vector at the first semantic understanding depth.
[0122] If Y is greater than 1, for example, Y=3, the server will continue to iterate using subsequent semantic feature extraction units. For the second semantic feature extraction unit, it will perform semantic mining again based on the results of the previous iteration. Taking the vector [v_{tired}'''] corresponding to "fatigue" as an example, this unit may further consider the relationship with "high work pressure" in a more complex semantic network. Assume that there is a multi-factor association model in the semantic network, taking into account the comprehensive impact of multiple factors such as work pressure and living habits on fatigue. According to this model and the information in the current text, [v_{tired}'''] is adjusted again to obtain [v_{tired}'''']. In the same way, the enhanced semantic understanding vectors corresponding to all keywords are processed to obtain the enhanced semantic understanding vectors at the second semantic understanding depth. If there is a third semantic feature extraction unit, similar operations will continue to be performed, and finally the enhanced semantic understanding vectors at Y semantic understanding depths will be obtained.
[0123] The server has obtained the keyword node vectors of the health survey text data to be screened at the basic semantic understanding depth, assuming that they are {[v_{basic 1}], [v_{basic 2}], ...}. For example, the vector [v_{basic-fatigue}] corresponding to "fatigue" at the basic semantics is an element in this set. There are also enhanced semantic understanding vectors at the derived semantic understanding depth, such as {[v_{fatigue}''], [v_{other}''], ...}, and enhanced semantic understanding vectors after Y semantic understanding depth iterations (assuming Y=3), such as {[v_{fatigue}''''], [v_{other}''''], ...}.
[0124] The server loads these vectors into a sequence in a specific order. First, the elements in the keyword node vector at the basic semantic understanding depth are put into the sequence in sequence, such as [v_{basic-fatigue}] first. Then the elements in the enhanced semantic understanding vector at the derived semantic understanding depth are put into the sequence, such as [v_{fatigue}''], etc. Finally, the elements in the enhanced semantic understanding vector at Y semantic understanding depths are put into the sequence, such as [v_{fatigue}''''], etc. The final generated sequence is the text feature vector sequence of the health survey text data to be screened. This sequence completely contains information from basic semantics to derived semantics at different depths, providing a comprehensive data foundation for subsequent authenticity assessment and other operations.
[0125] In a possible implementation, step S121 includes:
[0126] Step S1211, based on the semantic association focusing strategy, feature combination is performed on the keyword node vector of the health survey text data to be screened at the derived semantic understanding depth and the structural coding vector sequence to generate a first feature combination result.
[0127] Step S1212, based on the logical structure focusing strategy, feature combination is performed on the keyword node vector of the health survey text data to be screened at the derived semantic understanding depth and the structure coding vector sequence to generate a second feature combination result.
[0128] Step S1213: Perform feature combination on the combination result of the first feature combination result and the second feature combination result according to the recurrent neural network unit to generate an enhanced structure coding vector sequence.
[0129] In this embodiment, the server obtains the keyword node vector and structure encoding vector sequence of the health survey text data to be screened at the derived semantic understanding depth. Take the health survey text "I feel a little tired recently, my sleep quality is not very good, I often wake up at night, and my appetite has also decreased. I don't have cold symptoms, but I occasionally have headaches. I wonder if the work pressure is too high." as an example. The keyword node vector corresponding to "tired" at the derived semantic understanding depth is assumed to be [v_{tired}']=[0.2, 0.3, -0.2], and the dependency relationship vector related to "tired" in the structure encoding vector sequence (for example, a vector representing the relationship between physical status) is assumed to be [s_{tired}]=[0.5, -0.1, 0.3].
[0130] According to the semantic association focusing strategy, the server will perform feature combination on the keyword node vector and the corresponding vector in the structure encoding vector sequence. For the keyword "tired", one possible combination method is to calculate the inner product of the vector or multiply it element by element and then sum it. First, multiply [v_{tired}'] and [s_{tired}] element by element to get [0.2×0.5, 0.3×(-0.1), -0.2×0.3]=[0.1, -0.03, -0.06], and then add these results to get a combined value, which is the combination result of the keyword "tired" and the related structure vector. The server will operate on other keywords in the keyword node vector under the derived semantic understanding depth, such as [v_{sleep}'] corresponding to "poor sleep quality" and the related vector [s_{sleep}] in the structure encoding vector sequence according to the same semantic association focusing strategy, and finally get the first feature combination result.
[0131] Furthermore, the server will identify logical relationships in the health survey text data to be screened. In the above text, "high work pressure" and "fatigue" have a causal logical relationship. The keyword node vector of "high work pressure" under the derived semantics is assumed to be [v_{high work pressure}'], and the corresponding structural vector representing the causal relationship is assumed to be [s_{causal}].
[0132] According to the logical structure focusing strategy, the server performs special combination calculations. For example, according to the pre-set logical structure weight formula, for [v_{work pressure is high}'] and [s_{cause and effect}], the elements in [v_{work pressure is high}'] may be weighted and adjusted first (the weight is determined according to factors such as the importance of the cause and effect relationship), and then multiplied and summed element by element with [s_{cause and effect}] to obtain a new vector as the combination result of this pair of vectors. This operation is performed on all keyword vectors and structural vectors with logical relationships in the text. For example, "no cold symptoms" and the structural vector representing a negative logical relationship are also combined according to a similar logical structure focusing strategy to obtain the second feature combination result.
[0133] Finally, the server combines the first feature combination result and the second feature combination result obtained previously. Assume that the vector set obtained by the first feature combination result is [R_1]={[r_1], [r_2], ...}, and the vector set obtained by the second feature combination result is [R_2]={[r_3], [r_4], ...}. Combining these two vector sets together, a new vector set [R]={[r_1], [r_2], [r_3], [r_4], ...} is obtained.
[0134] Specifically, the recurrent neural network unit starts working and inputs the vectors in [R] in order according to its internal loop calculation rules. For example, first input [r_1] and [r_2], and the recurrent neural network unit calculates according to its internal neuron connection weights and activation function (for example, using the ReLU activation function) to obtain a new vector [s_1']. Then continue to input the next pair of vectors for calculation, such as inputting [r_3] and [r_4], and then calculate [s_2'] again in the unit. All vectors are processed in this way, and finally the enhanced structure encoding vector sequence {[s_1'], [s_2'], ...} is obtained.
[0135] In a possible implementation, step S122 includes:
[0136] Step S1221, recursively operate the enhanced structure coding vector sequence through the target recursive model to generate a recursive result of the enhanced structure coding vector sequence.
[0137] Step S1222, performing element-level accumulation on the recursive result of the enhanced structural coding vector sequence and the keyword node vector of the health survey text data to be screened at the derived semantic understanding depth, to generate an enhanced semantic understanding vector of the health survey text data to be screened at the derived semantic understanding depth.
[0138] In this embodiment, the server obtains the previously generated enhanced structure encoding vector sequence. Taking the previous health survey text "I feel a little tired recently, my sleep quality is not very good, I often wake up at night, and my appetite has also decreased. I don't have cold symptoms, but I occasionally have headaches. I suspect that the work pressure is too high." as an example, assume that a vector in the enhanced structure encoding vector sequence is [s_1']=[0.1, 0.2, -0.1].
[0139] The server calls the target recursive model to perform recursive operations on the vectors in the sequence of reinforced structure encoding vectors. The target recursive model has its own specific recursive algorithm. For example, it may be calculated based on the concept of time steps. For the vector [s_1'], in the first recursive step, it will calculate each element of the vector according to the parameters and algorithms inside the model. Assume that the calculation is based on the following rules: the first element (0.1) becomes (0.1+0.05times(0.1)) (here (0.05) is a coefficient learned by the model according to the data characteristics), and (0.105) is obtained; the second element (0.2) becomes (0.2+0.05times(0.2)), and (0.21) is obtained; the third element (-0.1) becomes (-0.1+0.05times(-0.1)), and (-0.105) is obtained. After this recursive operation, [s_1'] becomes [0.105, 0.21, -0.105].
[0140] The server also performs the same recursive operation on other vectors in the enhanced structure encoding vector sequence, such as [s_2'], according to the algorithm of the target recursive model, thereby obtaining the recursive result of the entire enhanced structure encoding vector sequence.
[0141] Next, the server obtains the keyword node vector at the derived semantic understanding depth. For the previous health survey text, the keyword node vector of "fatigue" at the derived semantic understanding depth is assumed to be [v_{fatigue}'] = [0.2, 0.3, -0.2]. At the same time, the server has obtained the recursive result of the enhanced structure encoding vector sequence, assuming that the vector corresponding to "fatigue" is [s_{fatigue}''] = [0.105, 0.21, -0.105] after the recursive operation.
[0142] The server accumulates the recursive results and the derived semantic keyword node vectors on an element-by-element basis. For the vector corresponding to the keyword "exhausted", the corresponding elements of [v_{exhausted}'] and [s_{exhausted}''] are added to obtain a new vector [v_{exhausted}'']. The specific calculation is: the first element (0.2+0.105=0.305); the second element (0.3+0.21=0.51); the third element (-0.2+(-0.105)=-0.305), so [v_{exhausted}'']=[0.305, 0.51, -0.305].
[0143] In the same way, the server performs element-level accumulation operations on other vectors in the keyword node vector at the derived semantic understanding depth and the corresponding vectors in the recursive results of the enhanced structural encoding vector sequence, and finally obtains the enhanced semantic understanding vector of the health survey text data to be screened at the derived semantic understanding depth.
[0144] In a possible implementation, the method further includes:
[0145] Step A110, semantically decompose the keyword node vectors of the health survey text data to be screened at the basic semantic understanding depth and the enhanced semantic understanding vectors of the health survey text data to be screened at X semantic understanding depths, and generate X+1 core component vector sets and X+1 auxiliary component vector sets.
[0146] Step A120, based on the adaptive focusing strategy, feature enhancement is performed on the X+1 core component vector sets and the X+1 auxiliary component vector sets respectively, to generate X+1 enhanced core component vector sets and X+1 enhanced auxiliary component vector sets as component vector sets of the multi-level semantic understanding level.
[0147] Step A130, using the component vector set of the multi-level semantic understanding level to enhance the text feature vectors in the text feature vector sequence, to generate an enhanced text feature vector sequence.
[0148] In this embodiment, we continue to take the previous health survey text "I have been feeling a little tired recently, my sleep quality is not very good, I often wake up at night, and my appetite has also decreased. I don't have any cold symptoms, but I occasionally have headaches. I wonder if the work pressure is too high." as an example, assuming that the keyword node vector corresponding to "tired" at the basic semantic understanding depth is [v_{tired}]=[0.3, 0.2, -0.1].
[0149] The server recursively operates on [v_{tired}] through the first context knowledge space and the second context knowledge space. The first context knowledge space may contain semantic information about the description of the physical state, and the second context knowledge space may contain semantic information related to emotions and living habits. In the recursive operation under the first context knowledge space, it is assumed that according to the semantic rules therein, the elements in [v_{tired}] are adjusted to obtain the core recursive vector [v_{tired-core1}]=[0.35, 0.22, -0.12] and the auxiliary recursive vector [v_{tired-auxiliary1}]=[0.05, -0.02, 0.02].
[0150] Then, semantic compression is performed on these two recursive vectors. Semantic compression may be weighted summation of elements according to certain weights. Assume that after semantic compression, a vector [c_{fatigue-core}]=[0.4] is obtained in the first core component vector set, and a vector [a_{fatigue-auxiliary}]=[0.03] is obtained in the first auxiliary component vector set. The server also performs the same operation on other vectors in the keyword node vector under the basic semantic understanding depth (such as the vector corresponding to "poor sleep quality", etc.), thereby generating the first core component vector set and the first auxiliary component vector set.
[0151] Assuming X=2, for the enhanced semantic understanding vector corresponding to “fatigue” at the first semantic understanding depth, [v_{fatigue-enhancement1}]=[0.4, 0.3, -0.2].
[0152] Recursive operations are performed based on the first context knowledge space and the second context knowledge space. In the first context knowledge space, the core recursive vector [v_{fatigue-core2}]=[0.42, 0.33, -0.22] and the auxiliary recursive vector [v_{fatigue-auxiliary2}]=[0.02, -0.03, 0.02] are obtained.
[0153] After semantic compression, we get a vector [c_{fatigue-core2}] in the second core component vector set and a vector [a_{fatigue-auxiliary2}] in the second auxiliary component vector set. We process the enhanced semantic understanding vectors at the second semantic understanding depth and the enhanced semantic understanding vectors corresponding to other keywords in the same way, and finally generate X+1 (i.e. 3) core component vector sets and X+1 auxiliary component vector sets.
[0154] Next, the server operates on the bth core component vector set (taking b=1 as an example, i.e., the first core component vector set). Assume that the vectors in the first core component vector set are [c_{fatigue-core}]=[0.4], etc.
[0155] The recursive unit is used to combine features. The recursive unit may be adjusted according to the previous semantic decomposition results. For example, if the previous semantic decomposition found that "fatigue" has some correlation with other symptoms, the recursive unit will adjust [c_{fatigue-core}] according to this correlation and obtain the feature combination result [c_{fatigue-core-combination}] = [0.45].
[0156] Then the logical structure focusing unit strengthens the logical structure dimension of [c_{fatigue-core-combination}]. Assuming that according to the logical relationship between "fatigue" and other symptoms in the text (such as causal relationship, etc.), [c_{fatigue-core-combination}] is weighted and other operations are performed to obtain the first core reference feature [c_{fatigue-core-reference}]=[0.5].
[0157] Finally, the semantic association focusing unit strengthens the semantic association dimension of [c_{fatigue-core-reference}]. For example, according to the semantic association between concepts such as "fatigue" and "high work pressure" in the semantic knowledge base, [c_{fatigue-core-reference}] is adjusted to obtain a vector [c_{fatigue-core-strengthening}]=[0.55] in the first strengthened core component vector set.
[0158] The server performs the same operation on other core component vector sets (b=2, 3, etc.) to obtain X+1 enhanced core component vector sets.
[0159] For the b-th auxiliary component vector set (taking b=1 as an example), assume that the vector is [a_{fatigue-auxiliary}]=[0.03].
[0160] Following steps similar to the core component vector set, the recursive unit performs feature combination to obtain [a_{exhaustion-auxiliary-combination}], the logical structure focus unit performs logical structure dimension enhancement processing to obtain [a_{exhaustion-auxiliary-reference}], and the semantic association focus unit performs semantic association dimension enhancement processing to obtain a vector [a_{exhaustion-auxiliary-enhancement}] in the first enhanced auxiliary component vector set. The same operation is performed on other auxiliary component vector sets to obtain X+1 enhanced auxiliary component vector sets.
[0161] Then, the server merges the X+1 enhanced core component vector sets to generate a core semantic enhancement map. For example, the enhanced core component vectors corresponding to "fatigue" are merged, which may be connected in a certain order or based on a certain semantic relationship.
[0162] At the same time, X+1 sets of enhanced auxiliary component vectors are merged to generate an auxiliary semantic enhancement graph.
[0163] Next, the server uses a recursive model to enhance and blend the core semantic enhancement graph with the keyword node vector of the health survey text data to be screened at the basic semantic understanding depth. For example, the information in the core semantic enhancement graph is blended with the keyword node vector [v_{fatigue}] of "fatigue" under the basic semantics, which may be obtained through weighted summation of vector elements, element replacement and other operations to obtain the first enhanced fusion result.
[0164] According to the recursive model, the auxiliary semantic enhancement map and the enhanced semantic understanding vector of the health survey text data to be screened at the Xth semantic understanding depth are enhanced and fused to obtain the second enhanced fusion result.
[0165] Finally, these fusion results are used to replace or adjust the text feature vectors in the text feature vector sequence, so as to generate an enhanced text feature vector sequence.
[0166] In a possible implementation, step A110 includes:
[0167] Step A111, recursively operate the keyword node vectors of the health survey text data to be screened at the basic semantic understanding depth through the first context knowledge space and the second context knowledge space respectively, to generate the core recursive vector and auxiliary recursive vector of the health survey text data to be screened at the basic semantic understanding depth.
[0168] Step A112, semantic compression processing is performed on the core recursive vectors and the auxiliary recursive vectors of the health survey text data to be screened at the basic semantic understanding depth, to generate a first core component vector set and a first auxiliary component vector set.
[0169] Step A113, recursively operate the enhanced semantic understanding vector of the health survey text data to be screened at the a+1th semantic understanding depth according to the first contextual knowledge space and the second contextual knowledge space, to generate the core recursive vector and auxiliary recursive vector of the health survey text data to be screened at the a+1th semantic understanding depth, where a is a positive integer not greater than X.
[0170] Step A114, semantic compression processing is performed on the core recursive vector and the auxiliary recursive vector of the health survey text data to be screened at the a+1th semantic understanding depth, to generate the a+1th core component vector set and the a+1th auxiliary component vector set.
[0171] When processing health survey text data in this embodiment, operations are performed on the keyword node vector at the basic semantic understanding depth. Continuing with the previous health survey text "I feel a little tired recently, my sleep quality is not very good, I often wake up at night, and my appetite has also decreased. I don't have cold symptoms, but I occasionally have headaches. I suspect that the work pressure is too high." As an example, assume that the keyword node vector of "tired" at the basic semantic understanding depth is [v_{tired}=[0.3, 0.2, -0.1]].
[0172] The first context knowledge space of the server contains semantic knowledge related to the physiological state of the body, and the second context knowledge space contains semantic knowledge related to the psychological state and lifestyle. When recursively operating [v_{tired}] through the first context knowledge space, for example, if there is a rule in the first context knowledge space indicating that physical fatigue may be related to energy level, then the elements of [v_{tired}] are adjusted according to this rule. Assume that according to a certain calculation method (such as based on the predefined relationship between energy level and vector elements), the core recursive vector of [v_{tired}] in the first context knowledge space is obtained [v_{tired-core1}=[0.32, 0.21, -0.11]].
[0173] At the same time, recursive operation is performed through the second context knowledge space. If there is a rule in the second context knowledge space indicating that fatigue may be related to psychological stress, [v_{fatigue}] is adjusted according to this rule to obtain the auxiliary recursive vector [v_{fatigue-auxiliary1}=[0.02, -0.01, 0.01]].
[0174] The server also recursively operates other keyword node vectors at the basic semantic understanding depth (such as the vector corresponding to "poor sleep quality") through the first context knowledge space and the second context knowledge space in a similar manner to obtain their respective core recursive vectors and auxiliary recursive vectors.
[0175] The "tired" core recursive vector [v_{tired-core1}] and the auxiliary recursive vector [v_{tired-auxiliary1}] are semantically compressed. The semantic compression process may be a weighted summation of the vector elements according to a predefined weight. For example, for the core recursive vector [v_{tired-core1}], assuming the weight is [0.4, 0.3, 0.3], then [0.4*0.32+0.3*0.21+0.3*(-0.11)=0.128+0.063-0.033=0.158] is calculated, and this result is used as an element in the first core component vector set.
[0176] For the auxiliary recursive vector [v_{fatigue-auxiliary1}], assuming the weights are [0.5, 0.3, 0.2], calculate [0.5*0.02+0.3*(-0.01)+0.2*0.01=0.01-0.003+0.002=0.009], and this result is used as an element in the first auxiliary component vector set.
[0177] In the same way, the server performs semantic compression processing on the core recursive vectors and auxiliary recursive vectors obtained by recursive operations on other keyword node vectors at the basic semantic understanding depth, thereby generating a first core component vector set and a first auxiliary component vector set.
[0178] Assume X=2, when a=0 (i.e. the first semantic understanding depth), the reinforced semantic understanding vector for “fatigue” at this semantic understanding depth is [v_{fatigue-reinforcement1}=[0.4, 0.3, -0.2]].
[0179] Recursive operation is performed based on the first context knowledge space. If there is a rule in the first context knowledge space indicating that fatigue is related to the ability of physical function to recover as the health condition is further understood, [v_{fatigue-enhancement1}] is adjusted according to this rule to obtain [v_{fatigue-enhancement1-core2}=[0.42, 0.32, -0.22]] as the core recursive vector.
[0180] Recursive operation is performed based on the second context knowledge space. If there is a rule in the second context knowledge space indicating that fatigue is related to changes in the rhythm of life, [v_{fatigue-enhancement 1}] is adjusted according to this rule to obtain [v_{fatigue-enhancement 1-auxiliary 2}=[0.03, -0.02, 0.02]] as the auxiliary recursive vector.
[0181] When a=1 (i.e., the second semantic understanding depth), the corresponding enhanced semantic understanding vector is also recursively operated in the same way based on the first context knowledge space and the second context knowledge space to obtain their respective core recursive vectors and auxiliary recursive vectors.
[0182] The core recursive vector [v_{fatigue-reinforcement1-core2}] of "fatigue" when a=0 is semantically compressed. Assuming the weights are [0.3, 0.4, 0.3], calculate [0.3*0.42+0.4*0.32+0.3*(-0.22)=0.126+0.128-0.066=0.188], and this result is used as an element in the second core component vector set.
[0183] The auxiliary recursive vector [v_{fatigue-reinforcement1-auxiliary2}] is semantically compressed. Assuming the weights are [0.4, 0.3, 0.3], calculate [0.4*0.03+0.3*(-0.02)+0.3*0.02=0.012-0.006+0.006=0.012], and this result is used as an element in the second auxiliary component vector set.
[0184] In the same way, the core recursive vector and the auxiliary recursive vector obtained by recursive operation of the enhanced semantic understanding vector under different a values (not greater than X) are semantically compressed to generate the a+1th core component vector set and the a+1th auxiliary component vector set.
[0185] In a possible implementation, the adaptive focusing strategy includes a recursive unit, a semantic association focusing unit and a logical structure focusing unit. Step A120 includes:
[0186] Step A121, performing feature combination on the b-th core component vector set and the b-th auxiliary component vector set according to the recursive unit, to generate feature combination results of the b-th core component vector set and feature combination results of the b-th auxiliary component vector set, where b is a positive integer not greater than X+1.
[0187] Step A122, based on the feature combination results of the b-th core component vector set and the feature combination results of the b-th auxiliary component vector set by the logical structure focusing unit, respectively perform logical structure dimension enhancement processing to generate the b-th core reference feature and the b-th auxiliary reference feature.
[0188] Step A123, based on the semantic association focusing unit, the bth core reference feature and the bth auxiliary reference feature are respectively enhanced in the semantic association dimension to generate the bth enhanced core component vector set and the bth enhanced auxiliary component vector set.
[0189] Take the previously generated core component vector set as an example. Assume X = 2, when b = 1 (i.e. the first core component vector set), assume that a vector element in this core component vector set is [c_{fatigue-core1}=0.4] related to "fatigue" (here it is simplified to a scalar element representing an element in a vector set).
[0190] The recursive unit combines features based on the information from the previous semantic decomposition process. For example, in the previous semantic decomposition, it was found that "fatigue" is associated with "poor sleep quality", and there is a corresponding association weight in the semantic network. The recursive unit adjusts [c_{fatigue-core1}] according to this association weight. Assuming the association weight is 0.2 (indicating the weight of the impact of "poor sleep quality" on "fatigue"), after calculation, [c_{fatigue-core1-combination}=0.4*(1+0.2)=0.48] is obtained, which is an element in the feature combination result of the first core component vector set.
[0191] For other elements in the core component vector set (if there are vectors with multiple elements), feature combination operations are performed through recursive units in a similar manner based on the association information between elements in the previous semantic decomposition process.
[0192] When b=1, for the auxiliary component vector set, assume that one of the vector elements representing "fatigue" is [a_{fatigue-auxiliary1}=0.05].
[0193] The recursive unit also operates based on the information from the previous semantic decomposition process. For example, if it is found that "fatigue" and "high work pressure" are related to each other at the auxiliary semantic level, and the association weight is 0.3, after calculation, [a_{fatigue-auxiliary1-combination}=0.05*(1+0.3)=0.065] is obtained, which is an element in the feature combination result of the first auxiliary component vector set. Similarly, similar operations are performed on other elements in the auxiliary component vector set.
[0194] Next, the server obtains the feature combination result of the b-th core component vector set (taking b=1 as an example), assuming that the element related to "fatigue" is [c_{fatigue-core1-combination}=0.48].
[0195] The logical structure focus unit analyzes the logical structure relationship in the text. In the health survey text "I feel a little tired recently, my sleep quality is not very good, I often wake up at night, and my appetite has decreased. I don't have cold symptoms, but I have headaches occasionally. I suspect that it is because of too much work pressure." "Fatigue" and "decreased appetite" may have a logical parallel relationship, and a causal relationship with "poor sleep quality".
[0196] According to these logical relationships, the logical structure focusing unit strengthens [c_{fatigue-core1-combination}]. Assuming that the logical structure strengthening weight set for causality is 0.1 (indicating the importance weight of causality in the logical structure), after calculation, [c_{fatigue-core1-reference}=0.48*(1+0.1)=0.528] is obtained, which is an element in the first core reference feature.
[0197] For other elements in the feature combination result of the core component vector set, similar operations are performed according to the logical structure relationship in the text and the corresponding logical structure reinforcement weights.
[0198] The feature combination result of the b-th auxiliary component vector set is obtained (taking b=1 as an example). Assume that the element related to "fatigue" is [a_{fatigue-auxiliary1-combination}=0.065].
[0199] The logical structure focusing unit processes the logical structure relationship in the text. For example, for the logical relationship between "fatigue" and "high work pressure" at the auxiliary semantic level, the corresponding logical structure reinforcement weight is set to 0.2, and after calculation, [a_{fatigue-auxiliary1-reference}=0.065*(1+0.2)=0.078] is obtained, which is an element in the first auxiliary reference feature. Similarly, similar operations are performed on other elements in the feature combination result of the auxiliary component vector set.
[0200] Then, the server obtains the b-th core reference feature (taking b=1 as an example), assuming that the element related to "fatigue" is [c_{fatigue-core1-reference}=0.528].
[0201] The semantic association focusing unit operates with reference to the information in the semantic knowledge base. In the semantic knowledge base, "fatigue" and "physical function decline" have a semantic association, and this association has a certain importance in the overall semantic understanding of the current text. Assume that the semantic association reinforcement weight corresponding to this importance is 0.15.
[0202] After calculation, we get [c_{fatigue-core1-strengthening}=0.528*(1+0.15)=0.6072], which is an element in the first strengthened core component vector set. Similar operations are performed on other elements in the core reference feature according to the semantic association information in the semantic knowledge base and the corresponding semantic association reinforcement weights.
[0203] The b-th auxiliary reference feature is obtained (taking b=1 as an example). Assume that the element related to "fatigue" is [a_{fatigue-auxiliary1-reference}=0.078].
[0204] The semantic association focusing unit operates according to the information in the semantic knowledge base. For example, "tired" and "depressed" are semantically related, and the corresponding semantic association reinforcement weight is set to 0.2.
[0205] After calculation, we get [a_{fatigue-auxiliary1-reinforcement}=0.078*(1+0.2)=0.0936], which is an element in the auxiliary component vector set after the first reinforcement. Similarly, similar operations are performed on other elements in the auxiliary reference feature.
[0206] According to the above method, operations are performed on different b values (not greater than X+1) respectively, so as to generate X+1 enhanced core component vector sets and X+1 enhanced auxiliary component vector sets.
[0207] In a possible implementation, the component vector set of the multi-level semantic understanding level includes X+1 enhanced core component vector sets and X+1 enhanced auxiliary component vector sets. Step A130 includes:
[0208] Step A131, merging the X+1 enhanced core component vector sets to generate a core semantic enhancement map, and merging the X+1 enhanced auxiliary component vector sets to generate an auxiliary semantic enhancement map.
[0209] Step A132, using a recursive model, the core semantic enhancement graph and the keyword node vector of the health survey text data to be screened are enhanced and integrated at the basic semantic understanding depth to generate a first enhanced integration result.
[0210] Step A133, based on the recursive model, the auxiliary semantic enhancement map and the enhanced semantic understanding vector of the health survey text data to be screened at the Xth semantic understanding depth are enhanced and blended to generate a second enhanced fusion result.
[0211] In this embodiment, the operation is performed by taking the previously generated X+1 enhanced core component vector sets as an example. Assuming X=2, there are 3 enhanced core component vector sets.
[0212] Taking the health survey text "I have been feeling a little tired recently, my sleep quality is not very good, I often wake up at night, and my appetite has also decreased. I don't have any cold symptoms, but I have headaches occasionally. I wonder if it's because of too much work pressure." as an example, the elements related to "fatigue" in these three sets are [c_{fatigue-core1-reinforcement}=0.6072] (from the first reinforced core component vector set), [c_{fatigue-core2-reinforcement}=0.65] (from the second reinforced core component vector set), and [c_{fatigue-core3-reinforcement}=0.7] (from the third reinforced core component vector set).
[0213] The way the server merges these elements may be to construct a graph structure. For example, starting from the first element, the elements are connected according to the semantic relationship between them (such as the progressive relationship between elements as the depth of semantic understanding increases). [c_{fatigue-core1-reinforcement}] is used as the basic node, and then connected to [c_{fatigue-core2-reinforcement}], indicating the development of the core semantics related to "fatigue" as the depth of semantic understanding increases, and then connected to [c_{fatigue-core3-reinforcement}] to form a directed graph structure. The elements corresponding to other keywords are also connected in a similar way to generate a core semantic reinforcement graph.
[0214] For the X+1 enhanced auxiliary component vector sets, we also assume that X = 2. Taking the "fatigue" related elements as an example, in the three sets, they are [a_{fatigue-auxiliary 1-enhancement}=0.0936], [a_{fatigue-auxiliary 2-enhancement}=0.1], and [a_{fatigue-auxiliary 3-enhancement}=0.12].
[0215] The server constructs the auxiliary semantic reinforcement graph in a similar way. Take [a_{fatigue-auxiliary 1-reinforcement}] as the starting node and connect it to [a_{fatigue-auxiliary 2-reinforcement}] and [a_{fatigue-auxiliary 3-reinforcement}] according to the semantic relationship. The same operation is performed for the elements corresponding to other keywords to form an auxiliary semantic reinforcement graph.
[0216] Next, the server obtains the core semantic enhancement graph and the keyword node vector of the health survey text data to be screened at the basic semantic understanding depth. Taking "fatigue" as an example, the keyword node vector at the basic semantic understanding depth is [v_{fatigue}=[0.3, 0.2, -0.1]], as well as the nodes and connection relationships related to "fatigue" in the core semantic enhancement graph generated earlier.
[0217] Then, the enhanced blending operation is performed through the recursive model. The recursive model may perform interactive calculations based on the node relationships in the graph and the elements in the keyword node vector.
[0218] For example, for the first element 0.3 in [v_{fatigue}], the recursive model calculates according to the first node [c_{fatigue-core1-reinforcement}=0.6072] related to "fatigue" in the core semantic reinforcement graph, according to certain calculation rules (such as weighted summation, the weight is determined according to the previous semantic analysis, assumed to be 0.5), and obtains (0.3*0.5+0.6072*0.5=0.4536).
[0219] For the second element 0.2 in [v_{fatigue}], according to the second node [c_{fatigue-core2-reinforcement}=0.65] related to "fatigue" in the core semantic reinforcement graph, the calculation is performed according to the same or different calculation rules (assuming the weight is 0.4), and the result is (0.2*0.4+0.65*0.6=0.43).
[0220] For the third element -0.1, according to the third node related to "fatigue" in the core semantic reinforcement graph [c_{fatigue-core3-reinforcement}=0.7], the calculation is performed according to the calculation rules (assuming the weight is 0.3), and the result is (-0.1*0.3+0.7*0.7=0.46).
[0221] Combining these calculation results, we get [v_{fatigue-fusion 1}=[0.4536, 0.43, 0.46]], which is part of the first enhanced fusion result corresponding to "fatigue". The elements corresponding to other keywords in the keyword node vector under the basic semantic understanding depth are also fused and calculated in a similar way through the recursive model and the core semantic enhancement map, and finally the first enhanced fusion result is obtained.
[0222] Assuming X=2, the server obtains the auxiliary semantic reinforcement map and the reinforcement semantic understanding vector of the health survey text data to be screened at the second semantic understanding depth. Taking "fatigue" as an example, the reinforcement semantic understanding vector at the second semantic understanding depth is [v_{fatigue-reinforcement2}=[0.45, 0.35, -0.25]], as well as the nodes and connection relationships related to "fatigue" in the auxiliary semantic reinforcement map generated previously.
[0223] Recursive models perform reinforcement blending operations according to specific rules.
[0224] For the first element 0.45 in [v_{fatigue-reinforcement 2}], according to the first node related to "fatigue" in the auxiliary semantic reinforcement graph [a_{fatigue-auxiliary 1-reinforcement}=0.0936], according to certain calculation rules (such as weighted summation, assuming the weight is 0.3), we get (0.45*0.7+0.0936*0.3=0.33708).
[0225] For the second element 0.35, according to the second node related to "tiredness" in the auxiliary semantic enhancement graph [a_{tiredness - auxiliary2 - enhancement}=0.1], calculate according to the calculation rule (assuming the weight is 0.4), and get (0.35 * 0.6 + 0.1 * 0.4 = 0.25).
[0226] For the third element -0.25, according to the third node related to "tiredness" in the auxiliary semantic enhancement graph [a_{tiredness - auxiliary3 - enhancement}=0.12], calculate according to the calculation rule (assuming the weight is 0.3), and get (-0.25 * 0.7 + 0.12 * 0.3 = -0.149).
[0227] Combine these calculation results to get [v_{tiredness - blend2}=[0.33708, 0.25, -0.149]], which is a part of the second enhanced blending result corresponding to "tiredness". For the elements corresponding to other keywords in the enhanced semantic understanding vector at the Xth semantic understanding depth, also perform blending calculations with the auxiliary semantic enhancement graph through a recursive model in a similar manner, and finally obtain the second enhanced blending result.
[0228] These first enhanced blending results and second enhanced blending results will be used to perform operations such as replacement or adjustment on the text feature vectors in the text feature vector sequence, so as to generate an enhanced text feature vector sequence.
[0229] In a possible implementation manner, step A131 includes: using a semantic merging network model to merge the cth loaded feature vector data and the (X - c + 1)th enhanced core component vector set to generate the cth merging result.
[0230] Among them, c is a positive integer not greater than X. When c = 1, the cth loaded feature vector data is the (X + 1)th enhanced core component vector set. When 1 < c < X + 1, the cth loaded feature vector data is the (c - 1)th merging result.
[0231] In this embodiment, the previously assumed (X=2) is used as an example for operation. At this time, there are (X+1=3) sets of core component vectors after strengthening. For the "fatigue" related element in the health survey text "I feel a little tired recently, my sleep quality is not very good, I often wake up at night, and my appetite has also decreased. I don't have cold symptoms, but I occasionally have headaches. I wonder if the work pressure is too high.", it is assumed that [c_{fatigue-core3-strengthening}=0.7] in the (3rd) set of core component vectors after strengthening (here it is simplified to a scalar representing an element in the vector set), [c_{fatigue-core2-strengthening}=0.65] in the (2nd) set of core component vectors after strengthening, and [c_{fatigue-core1-strengthening}=0.6072] in the (1st) set of core component vectors after strengthening.
[0232] When (c=1), the (c)th loading feature vector data is the (X+1)th enhanced core component vector set, that is, the element [c_{fatigue-core3-enhancement}] in the (3)th enhanced core component vector set. It is to be merged with the (X-c+1)th enhanced core component vector set. At this time, (X-c+1=2-1+1=2), that is, the (2)th enhanced core component vector set.
[0233] The semantic merging network model of the server includes operations such as context-based mapping models. First, the set of core component vectors after the (2)th enhancement is semantically compressed, and it is assumed that the semantic compression result [s_{fatigue-core2-compression}=0.63] is obtained (the semantic compression result here is obtained according to the semantic compression algorithm inside the model).
[0234] According to the context-based mapping model, the element [c_{fatigue-core3-enhancement}] in the (3)th enhanced core component vector set is calibrated with the semantic compression result [s_{fatigue-core2-compression}]. The calibration process may be to calculate the correlation between the two based on the predefined contextual relationship. Assume that the calibration result [r_{fatigue-calibration}=0.68] is obtained (this calibration result is calculated according to the algorithm of the model based on the relationship between the two).
[0235] The position adjustment coefficient and weight adjustment coefficient are adjusted according to the calibration results. Assuming that the initial position adjustment coefficient is (p_0=0.5) and the weight adjustment coefficient is (w_0=0.6), according to the calibration result [r_{fatigue-calibration}], after the adjustment rules within the model (for example, adjustment based on the comparison between the calibration result and the predefined threshold), the adjusted position adjustment coefficient (p_1=0.55) and the adjusted weight adjustment coefficient (w_1=0.63) are obtained.
[0236] Calculate the dot product of the (2)th enhanced core component vector set and the first recursive expansion result. The first recursive expansion result is obtained by recursively expanding the adjusted weight adjustment coefficient (w_1). Assume that the result after recursive expansion is [w_{1-expanded}=[0.63, 0.63]] (here simply represented as two vectors of the same elements, and different vectors may be obtained according to the recursive algorithm). The (2)th enhanced core component vector set is assumed to be [c_{exhausted-core2-enhanced-vector}=[0.65, 0.65]] (also simplified as two vectors of the same elements). Calculate the dot product to obtain the reference semantic vector sequence [v_{exhausted-reference}=0.65*0.63+0.65*0.63=0.819] (here the calculation is simplified, and it may actually be the dot product calculation of the vectors).
[0237] The reference semantic vector sequence [v_{exhaustion-reference}] and the second recursive expansion result are accumulated on an element-by-element basis. The second recursive expansion result is obtained by recursively expanding the adjusted position adjustment coefficient (p_1). Assuming that the result after recursive expansion is [p_{1-expansion}=[0.55, 0.55]], the element-by-element accumulation is [v_{exhaustion-merge 1}=0.819+0.55+0.55=1.919], which is the (1)th merge result.
[0238] When (c=2), the (c)th loading feature vector data is the (c-1)th merged result, that is, [v_{fatigue-merge1}=1.919] obtained above. It needs to be merged with the (X-c+1)th enhanced core component vector set. At this time, (X-c+1=2-2+1=1), that is, the (1)th enhanced core component vector set.
[0239] Perform semantic compression on the set of core component vectors after the (1)th enhancement, and assume that the semantic compression result is [s_{fatigue-core1-compression}=0.60].
[0240] According to the context-based mapping model, the feature vector data [v_{exhaustion-merge 1}] loaded for the second time is calibrated with the semantic compression result [s_{exhaustion-core 1-compression}], and it is assumed that the calibration result [r_{exhaustion-calibration 2}=0.7] is obtained.
[0241] The position adjustment coefficient and weight adjustment coefficient are adjusted according to the calibration results. Assuming that the initial position adjustment coefficient is (p_0'=0.4) and the weight adjustment coefficient is (w_0'=0.5), according to the calibration result [r_{fatigue-calibration2}], the adjusted position adjustment coefficient (p_2=0.45) and the adjusted weight adjustment coefficient (w_2=0.53) are obtained through the adjustment rules within the model.
[0242] Calculate the dot product of the (1)th enhanced core component vector set and the first recursive expansion result. The first recursive expansion result is obtained by recursively expanding the adjusted weight adjustment coefficient (w_2). Assume that the result after recursive expansion is [w_{2-expansion}=[0.53, 0.53]], and the (1)th enhanced core component vector set is assumed to be [c_{exhaustion-core1-enhancement-vector}=[0.6072, 0.6072]], and calculate the dot product to obtain the reference semantic vector sequence [v_{exhaustion-reference2}=0.6072*0.53+0.6072*0.53=0.643728].
[0243] The reference semantic vector sequence [v_{exhaustion-reference 2}] and the second recursive expansion result are accumulated on an element-by-element basis. The second recursive expansion result is obtained by recursively expanding the adjusted position adjustment coefficient (p_2). Assuming that the result after recursive expansion is [p_{2-expansion}=[0.45, 0.45]], the element-by-element accumulation is [v_{exhaustion-merge 2}=0.643728+0.45+0.45=1.543728], which is the (2)th merge result.
[0244] In this way, when (X=2), the process of merging (X+1) enhanced core component vector sets to generate a core semantic enhancement map is completed through the operations of (c=1) and (c=2). If (X) takes other values, the same steps can be followed.
[0245] In a possible implementation, the semantic merging network model includes a context-based mapping model. The semantic merging network model is used to merge the c-th loaded feature vector data and the X-c+1-th enhanced core component vector set to generate the c-th merged result, including:
[0246] Step A1311, calibrating the c-th loaded feature vector data with the semantic compression result according to the context-based mapping model to generate a calibration result. The semantic compression result is obtained by semantically compressing the X-c+1-th enhanced core component vector set.
[0247] Step A1312, adjusting the position adjustment coefficient and the weight adjustment coefficient according to the calibration result to generate an adjusted position adjustment coefficient and an adjusted weight adjustment coefficient.
[0248] Step A1313, calculating the dot product of the X-c+1th enhanced core component vector set and the first recursive expansion result to generate a reference semantic vector sequence. The first recursive expansion result is obtained by recursively expanding the adjusted weight adjustment coefficient.
[0249] Step A1314: perform element-level accumulation on the reference semantic vector sequence and the second recursive expansion result to generate a cth merging result. The second recursive expansion result is obtained by recursively expanding the adjusted position adjustment coefficient.
[0250] In this embodiment, it is assumed that the "fatigue" related vector in the health survey text data is being processed, and (X=3), (c=2). The (c)th loading feature vector data is assumed to be [v_{fatigue-loading}=[0.5, 0.4, 0.3]] (this is the result of a previous operation). The (X-c+1=3-2+1=2)th enhanced core component vector set, it is assumed that the "fatigue" related vector is [c_{fatigue-core2-enhancement}=[0.6, 0.5, 0.4]].
[0251] First, the set of vectors of the core component after the (2)th reinforcement [c_{tired-core2-strengthened}] is semantically compressed. The semantic compression process may be based on some predefined rules, such as weighted summation of vector elements. Assuming the weight is ([0.4, 0.3, 0.3]), the semantic compression result [s_{tired-core2-compressed}] is (0.4*0.6+0.3*0.5+0.3*0.4=0.45).
[0252] The context-based mapping model starts working by first analyzing the contextual relationship between the (c)th loading feature vector data [v_{fatigue-load}] and the semantic compression result [s_{fatigue-core2-compression}]. This relationship may be based on the previous semantic analysis of the text data. For example, if the previous semantic analysis found that "fatigue" is associated with other symptoms (such as poor sleep quality) in a specific context, this degree of association will affect the calibration process.
[0253] Assume that the context-based mapping model calculates the calibration relationship between the two based on this information. It may obtain the calibration result [r_{fatigue-calibration}] by calculating some distance metric between the vectors (such as Euclidean distance or cosine similarity, etc., and according to the predefined transformation rules]. For example, if cosine similarity is used for calculation, the calibration result may be a value between (0) and (1), assuming (0.7). This calibration result reflects the matching degree or association strength between [v_{fatigue-load}] and [s_{fatigue-core2-compression}] in the current context.
[0254] This embodiment presets the initial values of the predefined position adjustment coefficient and weight adjustment coefficient. Assume that the initial position adjustment coefficient (p_0=0.3) and the initial weight adjustment coefficient (w_0=0.4).
[0255] The position adjustment coefficient and the weight adjustment coefficient are adjusted according to the calibration result [r_{fatigue-calibration}=0.7]. The adjustment rule may be based on a predefined function or a lookup table. For example, if the calibration result is greater than (0.5), the coefficient is increased in a certain proportion. For the weight adjustment coefficient (w_0), it is assumed that the adjustment is made according to the rule of (w_0*(1+0.1*(r_{fatigue-calibration}-0.5))), and the adjusted weight adjustment coefficient is calculated to be (w_1=0.4*(1+0.1*(0.7-0.5))=0.408).
[0256] For the position adjustment coefficient (p_0), assuming that it is adjusted according to the rule of (p_0*(1+0.2*(r_{fatigue-calibration}-0.5))), the adjusted position adjustment coefficient is obtained (p_1=0.3*(1+0.2*(0.7-0.5))=0.312).
[0257] The adjusted weight adjustment coefficient (w_1=0.408) is recursively expanded. The recursive expansion may be based on a recursive algorithm predefined in the server, for example, the coefficient is multiplied by a fixed factor (assuming (1.1)) each time, and recursive expansion is performed (3) times (the number of times here can be set according to specific needs or models). The first recursive expansion result [w_{1-expanded}=[0.408, 0.408*1.1, 0.408*1.1*1.1]=[0.408, 0.4488, 0.49368]] is obtained.
[0258] Calculate the dot product of the (X-c+1=2)th enhanced core component vector set [c_{fatigue-core2-enhancement}=[0.6, 0.5, 0.4]] and the first recursive expansion result [w_{1-expansion}]. According to the calculation rules of the dot product, the reference semantic vector sequence [v_{fatigue-reference}] is calculated.
[0259] [v_{fatigue-reference}[0]=0.6*0.408=0.2448]
[0260] [v_{fatigue-reference}[1]=0.5*0.4488=0.2244]
[0261] [v_{fatigue-reference}[2]=0.4*0.49368=0.197472]
[0262] So the reference semantic vector sequence [v_{fatigue-reference}=[0.2448, 0.2244, 0.197472]].
[0263] The adjusted position adjustment coefficient (p_1=0.312) is recursively expanded. According to a similar recursive algorithm, assuming that (3) recursive expansions are performed, the second recursive expansion result [p_{1-expansion}=[0.312, 0.312*1.1, 0.312*1.1*1.1]=[0.312, 0.3432, 0.37752]] is obtained.
[0264] The reference semantic vector sequence [v_{fatigue-reference}=[0.2448, 0.2244, 0.197472]] and the second recursive expansion result [p_{1-expansion}=[0.312, 0.3432, 0.37752]] are accumulated element-wise.
[0265] [v_{fatigue-merge}[0]=0.2448+0.312=0.5568]
[0266] [v_{fatigue-merge}[1]=0.2244+0.3432=0.5676]
[0267] [v_{fatigue-merge}[2]=0.197472+0.37752=0.574992]
[0268] So the (c=2)th merge result is [v_{exhaustion-merge}=[0.5568, 0.5676, 0.574992]].
[0269] Following these steps, for different (X) and (c) values, the server can use the semantic merging network model to merge the corresponding vectors and finally generate a core semantic enhancement graph.
[0270] In a possible implementation, step S110 includes:
[0271] Step S111, performing feature extraction on the health survey text data to be screened by a first semantic feature extraction unit, and generating a keyword node vector of the health survey text data to be screened at a basic semantic understanding depth.
[0272] Step S112, performing feature extraction on the keyword node vector of the health survey text data to be screened at the basic semantic understanding depth by a second semantic feature extraction unit, and generating the keyword node vector of the health survey text data to be screened at the derived semantic understanding depth.
[0273] Step S113, taking the keyword node vector of the health survey text data to be screened at the basic semantic understanding depth and the keyword node vector of the health survey text data to be screened at the derived semantic understanding depth as the semantic encoding vector sequence of the health survey text data to be screened.
[0274] Step S114, performing dependency syntactic analysis on the health survey text data to be screened according to the logical structure model, and generating dependency syntactic analysis results of the health survey text data to be screened.
[0275] Step S115, performing structured mapping processing on the dependency syntax analysis results of the health survey text data to be screened according to the structured mapping table, and generating a structured coding vector sequence of the health survey text data to be screened.
[0276] In a possible implementation, step S114 includes:
[0277] Step S1141, dividing the health survey text data to be screened into independent words, marking the part of speech for each word, and then restoring each word to its basic form.
[0278] Step S1142, based on the part of speech marked for each word, identify the core words in the health survey text data to be screened as the central node of the dependency relationship.
[0279] Step S1143, according to the preset dependency relationship rules, an initial dependency edge is established for each word in the health survey text data to be screened, and the dependency edge connects the word with the word on which it depends.
[0280] Step S1144, extracting feature data of each word and position information and context information of each word in the sentence, wherein the feature data includes part-of-speech feature, morphological feature, and semantic role feature.
[0281] Step S1145, input the feature data of each word and the position information and context information of each word in the sentence into the trained logical structure model for reasoning to generate an initial dependency graph.
[0282] Step S1146: construct a dependency syntax tree according to the initial dependency graph as the dependency syntax analysis result, wherein each node in the dependency syntax tree represents a word, and the edge represents the dependency relationship between words.
[0283] In this embodiment, the server receives the health survey text data to be screened, such as "I feel a little tired recently, my sleep quality is not very good, I often wake up at night, and my appetite has decreased. I don't have cold symptoms, but I occasionally have headaches. I suspect that it is because of too much work pressure."
[0284] The first semantic feature extraction unit starts working. It first analyzes each word in the text. For the word "I", it recognizes it as a pronoun, and according to the semantic feature extraction algorithm, it may be given a basic semantic vector representation. For example, in the predefined semantic space, "I" may be represented as [v_{I}=[0.1, 0.0, 0.0]] (this is just a simple example, the actual vector dimension and value are obtained according to model training).
[0285] For the keyword "fatigue", it will take into account the semantic importance of this word in the context of health surveys, and generate a more semantically representative vector through the semantic analysis mechanism within the feature extraction unit (which may be based on the statistical information of a large-scale corpus or a pre-trained semantic model). Assume that "fatigue" is represented as [v_{fatigue}=[0.3, 0.2, -0.1]], where each element may represent a different semantic dimension, such as a semantic dimension related to physical state, emotional state, duration, etc.
[0286] In this way, feature extraction is performed on each meaningful word in the text, and finally the vectors corresponding to these words are combined to obtain the keyword node vector of the health survey text data to be screened at the basic semantic understanding depth. For example, this vector may be {[v_{I}], [v_{tired}], [v_{poor sleep quality}], ...}.
[0287] Next, the second semantic feature extraction unit takes the keyword node vector at the basic semantic understanding depth generated by the first semantic feature extraction unit as input. For the vector corresponding to the keyword "fatigue" [v_{fatigue}=[0.3, 0.2, -0.1]], it will further mine the derived semantics.
[0288] Assuming that a potential correlation is found between "fatigue" and "high work pressure" in the basic semantic analysis, the second semantic feature extraction unit will further adjust the semantic vector representation of "fatigue" based on this correlation. For example, it may consider the impact of work pressure on the degree of fatigue and the importance of semantic associations in health surveys, and adjust [v_{fatigue}] to [v_{fatigue}'=[0.35, 0.25, -0.15]].
[0289] Similarly, for other keywords, such as "poor sleep quality", if it is found to have a causal relationship with "fatigue" (mining at the derived semantic level), its vector will be adjusted accordingly. By performing such derived semantic mining and adjustment on each keyword vector in the keyword node vector at the basic semantic understanding depth, the keyword node vector of the health survey text data to be screened at the derived semantic understanding depth is finally obtained. For example, this vector may be {[v_{I}'], [v_{fatigue}'], [v_{poor sleep quality}'], ...}.
[0290] The server combines the keyword node vectors of the health survey text data to be screened at the basic semantic understanding depth and the keyword node vectors at the derived semantic understanding depth to form a semantic encoding vector sequence. This sequence completely contains information from basic semantics to derived semantics, providing rich semantic representation for subsequent processing.
[0291] Next, the server processes the input health survey text data to be screened. First, the text "I feel a little tired recently, my sleep quality is not very good, I often wake up at night, and my appetite has decreased. I don't have cold symptoms, but I occasionally have headaches. I wonder if it's because of too much work pressure." is split into independent words, and words such as "I", "recently", "feeling", "body", "some", and "tired" are obtained.
[0292] Then mark the part of speech for each word, "I" as a pronoun, "recently" as a time adverb, "feeling" as a verb, etc. Then restore each word to its basic form. For example, "feeling" does not need special restoration, and modifiers such as "some" may have specific treatment methods in dependency syntactic analysis after restoring the basic form (such as marking the modification relationship of the adjective "tired").
[0293] Based on the marked parts of speech, the server identifies the core words. In this sentence, "feel" is a key verb that expresses the subject's feelings, so "feel" is identified as the central node of the dependency relationship. This is because the semantics of many other words are dependent on "feeling", such as "tired" is the object of "feeling", and "poor sleep quality" also expresses the subject's perception of it through "feeling".
[0294] According to the preset dependency relationship rules, an initial dependency edge is established for each word. For example, "tired" depends on "feeling", so a dependency edge is established from "tired" to "feeling", indicating that the word "tired" is semantically dependent on the verb "feeling". Similarly, "poor sleep quality" also depends on "feeling", and the corresponding dependency edge is established.
[0295] For each word, the server extracts its feature data as well as its position information and context information in the sentence. Taking "tired" as an example, its part of speech feature is an adjective, its morphological feature may not have special changes (if there are deformations in other texts, such as comparatives, superlatives, etc., the morphological features will be reflected), and its semantic role feature is to describe the body state; its position in the sentence is after "feeling", and the context is to describe the body's feelings.
[0296] For the word "night", the part of speech feature is a noun, the morphological feature has no special circumstances, and the semantic role feature may be indicating time; its position in the sentence is in the description related to "poor sleep quality", and the context is describing the time information related to sleep.
[0297] The server inputs the feature data, location information and context information of each word into the trained logical structure model for inference. The logical structure model is trained based on a large amount of corpus data, and it can learn the dependency pattern between words based on the input information.
[0298] After reasoning, an initial dependency graph is generated. In this graph, nodes represent words and edges represent dependency relationships. Then, a dependency syntax tree is constructed based on this initial dependency graph. For example, "feeling" is the root node, and it is connected to nodes such as "tired" and "poor sleep quality" by edges. These edges represent different dependency relationships, such as subject-predicate relationship, verb-object relationship, etc. Finally, this dependency syntax tree is the dependency syntax analysis result of the health survey text data to be screened.
[0299] Then, after the server obtains the dependency syntax analysis result (dependency syntax tree), it processes it according to the structured mapping table, which stores the mapping relationship between different dependency relations and structure encoding vectors.
[0300] For example, in the dependency syntax tree, the dependency relationship between "feeling" and "tired" is a subject-object relationship. In the structured mapping table, the subject-object relationship may correspond to a structural encoding vector [s_{subject-object}=[0.5, -0.1, 0.3]]. For the dependency relationship between "feeling" and "poor sleep quality", it may correspond to another structural encoding vector [s_{another relationship}=[0.4, 0.2, -0.1]].
[0301] In this way, for each dependency relationship in the dependency syntax tree, the corresponding structural encoding vector is found according to the structured mapping table, and these structural encoding vectors are combined to generate a structural encoding vector sequence of the health survey text data to be screened.
[0302] Figure 2The AI-based health data screening device 100 shown includes: a processor 1001 and a memory 1003. Among them, the processor 1001 and the memory 1003 are connected, such as through a bus 1002. Optionally, the AI-based health data screening device 100 may also include a transceiver 1004, which can be used for data interaction between the server and other servers, such as data sending and / or data receiving. It should be noted that the transceiver 1004 is not limited to one in actual scheduling, and the structure of the AI-based health data screening device 100 does not constitute a limitation on the embodiments of the present application.
[0303] Processor 1001 may be a CPU (Central Processing Unit), a general-purpose processor, a DSP (Digital Signal Processor), an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array) or other programmable logic devices, transistor logic devices, hardware components or any combination thereof. It may implement or execute various exemplary logic blocks, modules and circuits described in conjunction with the disclosure of this application. Processor 1001 may also be a combination that implements computing functions, such as a combination of one or more microprocessors, a combination of a DSP and a microprocessor, etc.
[0304] The bus 1002 may include a path to transmit information between the above components. The bus 1002 may be a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus. The bus 1002 may be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 2 Only one thick line is used in the diagram, but this does not mean that there is only one bus or only one type of bus.
[0305] The memory 1003 may be a ROM (Read Only Memory) or other types of static storage devices that can store static information and instructions, a RAM (Random Access Memory) or other types of dynamic storage devices that can store information and instructions, or an EEPROM (Electrically Erasable Programmable Read Only Memory), a CD-ROM (Compact Disc Read Only Memory) or other optical disk storage, optical disk storage (including compressed optical disk, laser disk, optical disk, digital versatile disk, Blu-ray disk, etc.), magnetic disk storage medium, other magnetic storage devices, or any other medium that can be used to carry or store program code and can be read by a computer, without limitation herein.
[0306] The memory 1003 is used to store program codes for executing the embodiments of the present application, and the execution is controlled by the processor 1001. The processor 1001 is used to execute the program codes stored in the memory 1003 to implement the steps shown in the above method embodiments.
[0307] In addition, an embodiment of the present application also provides a readable storage medium, in which computer executable instructions are pre-installed. When the processor executes the computer executable instructions, the above-mentioned AI-based health data screening method is implemented.
[0308] Similarly, it should be noted that in order to simplify the description disclosed in this application, thereby helping to understand one or more embodiments of the invention, in the foregoing description of the embodiments of the present application, multiple features are sometimes combined into one embodiment, drawings, or descriptions thereof. Similarly, it should be noted that in order to simplify the description disclosed in this application, thereby helping to understand one or more embodiments of the invention, in the foregoing description of the embodiments of the present application, multiple features are sometimes combined into one embodiment, drawings, or descriptions thereof.
Claims
1. A health data screening method based on AI, It is characterized in that The method comprises: Obtaining a semantic coding vector sequence and a structural coding vector sequence of the health survey text data to be screened; the semantic coding vector sequence includes a keyword node vector of the health survey text data to be screened at a basic semantic understanding depth and a keyword node vector of the health survey text data to be screened at a derived semantic understanding depth, and the structural coding vector sequence refers to a sequence of converting the syntactic structure and paragraph structure of the text data into a numerical vector; Generate a text feature vector sequence of the health survey text data to be screened based on the semantic coding vector sequence and the structural coding vector sequence; the text feature vector sequence includes a keyword node vector of the health survey text data to be screened at a basic semantic understanding depth and a reinforced semantic understanding vector of the health survey text data to be screened at X semantic understanding depths, and the reinforced semantic understanding vector of the health survey text data to be screened at any one semantic understanding depth is obtained based on the keyword node vector of the health survey text data to be screened at a derived semantic understanding depth and the structural coding vector sequence; Performing an authenticity evaluation on the text feature vector sequence through an authenticity evaluation model to generate an authenticity evaluation result of the health survey text data to be screened, and screening the health survey text data to be screened according to the authenticity evaluation result; wherein the authenticity evaluation result includes an authenticity evaluation result of each text feedback unit in the health survey text data to be screened; The step of generating a text feature vector sequence of the health survey text data to be screened based on the semantic coding vector sequence and the structural coding vector sequence includes: Strengthening and blending the keyword node vectors of the health survey text data to be screened at the derived semantic understanding depth and the structural coding vector sequence to generate a strengthened structural coding vector sequence; The keyword node vector of the health survey text data to be screened at the derived semantic understanding depth is merged with the enhanced structure coding vector sequence to generate the enhanced semantic understanding vector of the health survey text data to be screened at the derived semantic understanding depth; Generate a text feature vector sequence of the health survey text data to be screened based on the keyword node vector of the health survey text data to be screened at the basic semantic understanding depth and the enhanced semantic understanding vector of the health survey text data to be screened at the derived semantic understanding depth; The step of reinforcing and blending the keyword node vectors of the health survey text data to be screened at the derived semantic understanding depth and the structure coding vector sequence to generate a reinforcing structure coding vector sequence includes: Performing feature combination on the keyword node vector of the health survey text data to be screened at the derived semantic understanding depth and the structural coding vector sequence according to the semantic association focusing strategy to generate a first feature combination result; Performing feature combination on the keyword node vector of the health survey text data to be screened at the derived semantic understanding depth and the structure coding vector sequence according to the logical structure focusing strategy to generate a second feature combination result; Performing feature combination on a combination result of the first feature combination result and the second feature combination result according to a recurrent neural network unit to generate a reinforced structure encoding vector sequence; The step of fusing the keyword node vector of the health survey text data to be screened at the derived semantic understanding depth with the enhanced structure coding vector sequence to generate the enhanced semantic understanding vector of the health survey text data to be screened at the derived semantic understanding depth includes: Recursively operating the enhanced structure coding vector sequence through a target recursive model to generate a recursive result of the enhanced structure coding vector sequence; The recursive result of the enhanced structural encoding vector sequence is accumulated on an element-by-element basis with the keyword node vector of the health survey text data to be screened at the derived semantic understanding depth to generate an enhanced semantic understanding vector of the health survey text data to be screened at the derived semantic understanding depth.
2. The AI-based health data screening method according to claim 1, It is characterized in that The method further comprises: Iteratively extract semantic features of the enhanced semantic understanding vector of the health survey text data to be screened at the derived semantic understanding depth by Y semantic feature extraction units to generate the enhanced semantic understanding vector of the health survey text data to be screened at Y semantic understanding depths, where Y is a positive integer; The method of generating a text feature vector sequence of the health survey text data to be screened based on the keyword node vector of the health survey text data to be screened at the basic semantic understanding depth and the enhanced semantic understanding vector of the health survey text data to be screened at the derived semantic understanding depth includes: The keyword node vectors of the health survey text data to be screened at the basic semantic understanding depth, the enhanced semantic understanding vectors of the health survey text data to be screened at the derived semantic understanding depth, and the enhanced semantic understanding vectors of the health survey text data to be screened at Y semantic understanding depths are loaded into a sequence to generate a text feature vector sequence of the health survey text data to be screened.
3. The AI-based health data screening method according to claim 1, It is characterized in that The method further comprises: Semantically decompose the keyword node vectors of the health survey text data to be screened at the basic semantic understanding depth and the enhanced semantic understanding vectors of the health survey text data to be screened at X semantic understanding depths to generate X+1 core component vector sets and X+1 auxiliary component vector sets; According to the adaptive focusing strategy, the X+1 core component vector sets and the X+1 auxiliary component vector sets are respectively feature enhanced to generate X+1 enhanced core component vector sets and X+1 enhanced auxiliary component vector sets as component vector sets at the multi-level semantic understanding level; Using the component vector set of the multi-level semantic understanding level to enhance the text feature vectors in the text feature vector sequence, to generate an enhanced text feature vector sequence; The process of performing semantic decomposition on the keyword node vectors of the health survey text data to be screened at the basic semantic understanding depth and the enhanced semantic understanding vectors of the health survey text data to be screened at X semantic understanding depths to generate X+1 core component vector sets and X+1 auxiliary component vector sets includes: Recursively operate the keyword node vectors of the health survey text data to be screened at the basic semantic understanding depth through the first context knowledge space and the second context knowledge space respectively to generate the core recursive vector and the auxiliary recursive vector of the health survey text data to be screened at the basic semantic understanding depth; Performing semantic compression processing on the core recursive vector and the auxiliary recursive vector of the health survey text data to be screened at the basic semantic understanding depth, respectively, to generate a first core component vector set and a first auxiliary component vector set; Recursively operate the enhanced semantic understanding vector of the health survey text data to be screened at the a+1th semantic understanding depth according to the first contextual knowledge space and the second contextual knowledge space respectively, to generate a core recursive vector and an auxiliary recursive vector of the health survey text data to be screened at the a+1th semantic understanding depth, where a is a positive integer not greater than X; The core recursive vectors and the auxiliary recursive vectors of the health survey text data to be screened at the a+1th semantic understanding depth are semantically compressed to generate an a+1th core component vector set and an a+1th auxiliary component vector set.
4. The AI-based health data screening method according to claim 3, It is characterized in that The adaptive focusing strategy includes a recursive unit, a semantic association focusing unit and a logical structure focusing unit; the step of performing feature enhancement on the X+1 core component vector sets and the X+1 auxiliary component vector sets according to the adaptive focusing strategy to generate X+1 enhanced core component vector sets and X+1 enhanced auxiliary component vector sets includes: According to the recursive unit, feature combination is performed on the b-th core component vector set and the b-th auxiliary component vector set respectively to generate a feature combination result of the b-th core component vector set and a feature combination result of the b-th auxiliary component vector set, where b is a positive integer not greater than X+1; According to the feature combination result of the b-th core component vector set and the feature combination result of the b-th auxiliary component vector set by the logical structure focusing unit, the logical structure dimension is strengthened to generate the b-th core reference feature and the b-th auxiliary reference feature; According to the semantic association focusing unit, the b-th core reference feature and the b-th auxiliary reference feature are respectively enhanced in terms of semantic association dimensions to generate a b-th enhanced core component vector set and a b-th enhanced auxiliary component vector set.
5. The AI-based health data screening method according to claim 3, characterized in that, the component vector set of the multi-semantic understanding level includes X + 1 enhanced core component vector sets and X + 1 enhanced auxiliary component vector sets; the step of using the component vector set of the multi-semantic understanding level to perform enhancement processing on the text feature vectors in the text feature vector sequence to generate an enhanced text feature vector sequence includes: merging the X + 1 enhanced core component vector sets to generate a core semantic enhancement map, and merging the X + 1 enhanced auxiliary component vector sets to generate an auxiliary semantic enhancement map; performing enhanced fusion on the keyword node vectors of the core semantic enhancement map and the health survey text data to be screened at the basic semantic understanding depth through a recursive model to generate a first enhanced fusion result; performing enhanced fusion on the enhanced semantic understanding vectors of the auxiliary semantic enhancement map and the health survey text data to be screened at the Xth semantic understanding depth according to the recursive model to generate a second enhanced fusion result; wherein, the process of merging the X + 1 enhanced core component vector sets to generate a core semantic enhancement map includes: using a semantic merging network model to merge the cth loaded feature vector data and the X - c + 1th enhanced core component vector set to generate a cth merging result; wherein, c is a positive integer not greater than X; when c = 1, the cth loaded feature vector data is the X + 1th enhanced core component vector set; when 1 < c < X + 1, the cth loaded feature vector data is the (c - 1)th merging result; wherein, the semantic merging network model includes a context-based mapping model; the step of using the semantic merging network model to merge the cth loaded feature vector data and the X - c + 1th enhanced core component vector set to generate a cth merging result includes: calibrating the cth loaded feature vector data with the semantic compression result according to the context-based mapping model to generate a calibration result; the semantic compression result is obtained by performing semantic compression processing on the X - c + 1th enhanced core component vector set; adjusting the position adjustment coefficient and the weight adjustment coefficient according to the calibration result to generate an adjusted position adjustment coefficient and an adjusted weight adjustment coefficient; calculating the dot product of the X - c + 1th enhanced core component vector set and the first recursively extended result to generate a reference semantic vector sequence; the first recursively extended result is obtained by recursively extending the adjusted weight adjustment coefficient; performing element-level accumulation on the reference semantic vector sequence and the second recursively extended result to generate a cth merging result; the second recursively extended result is obtained by recursively extending the adjusted position adjustment coefficient.
6. The AI-based health data screening method according to claim 1, characterized in that, the obtaining of the semantic coding vector sequence and the structure coding vector sequence of the health survey text data to be screened includes: Performing feature extraction on the health survey text data to be screened by a first semantic feature extraction unit to generate a keyword node vector of the health survey text data to be screened at a basic semantic understanding depth; Performing feature extraction on the keyword node vector of the health survey text data to be screened at the basic semantic understanding depth by a second semantic feature extraction unit to generate the keyword node vector of the health survey text data to be screened at the derived semantic understanding depth; The keyword node vector of the health survey text data to be screened at the basic semantic understanding depth and the keyword node vector of the health survey text data to be screened at the derived semantic understanding depth are used as the semantic encoding vector sequence of the health survey text data to be screened; Performing dependency syntactic analysis on the health survey text data to be screened according to the logical structure model to generate dependency syntactic analysis results of the health survey text data to be screened; Performing structured mapping processing on the dependency syntax analysis results of the health survey text data to be screened according to the structured mapping table to generate a structured encoding vector sequence of the health survey text data to be screened; The step of performing dependency syntactic analysis on the health survey text data to be screened according to the logical structure model to generate the dependency syntactic analysis result of the health survey text data to be screened includes: The health survey text data to be screened is divided into independent words, and after marking the part of speech for each word, each word is restored to its basic form; According to the part of speech marked for each word, the core words in the health survey text data to be screened are identified as the central nodes of the dependency relationship; According to a preset dependency relationship rule, an initial dependency edge is established for each word in the health survey text data to be screened, wherein the dependency edge connects the word with the word on which it depends; Extracting feature data of each word and position information and context information of each word in the sentence, wherein the feature data includes part-of-speech feature, morphological feature, and semantic role feature; Inputting the feature data of each word and the position information and context information of each word in the sentence into the trained logical structure model for reasoning to generate an initial dependency graph; According to the initial dependency graph, a dependency syntax tree is constructed as the dependency syntax analysis result, wherein each node in the dependency syntax tree represents a word, and the edges represent the dependency relationship between words.
7. An AI-based health data screening device, It is characterized in that The AI-based health data screening device includes a processor and a machine-readable storage medium, wherein the machine-readable storage medium stores machine-executable instructions, and the machine-executable instructions are loaded and executed by the processor to implement the AI-based health data screening method described in any one of claims 1-6.
