Information processing method based on big data mining and population fertility survey service system

Abstract

The embodiment of the present application provides an information processing method and a population fertility survey service system based on big data mining, which semantically understands the survey text and generates corresponding semantic understanding data, and then encodes the semantic understanding data into a semantic encoding vector sequence, and then determines X survey topic knowledge points, and predicts the user's tendency for each survey topic knowledge point, and generates a user tendency prediction score, so as to gain in-depth insight into the user's attitude and preference for each survey topic. In addition, an attention analysis mechanism is introduced to calculate the attention coefficient corresponding to each survey topic knowledge point, further improving the accuracy and pertinence of the tendency prediction. Finally, combined with the user tendency prediction score and the attention coefficient, the survey topic tendency distribution of the target survey user is accurately determined, which provides data support for research, investigation and analysis in related fields, and greatly improves the efficiency and accuracy of information processing.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an information processing method based on big data mining and a population fertility survey service system. Background Art

[0002] In the era of information explosion, big data has become an indispensable and important resource for all walks of life. Especially in the fields of market research and user behavior analysis, by mining and analyzing the massive survey text data generated by users (such as population fertility intention survey text data, medical treatment survey text data, social security application survey text data, etc.), we can deeply understand the user's views, preferences and needs, and provide strong support for corporate decision-making. However, traditional information processing methods often face many challenges when facing large-scale and diverse text data.

[0003] On the one hand, text data has semantic complexity and diversity, and how to accurately understand and extract key information from it has become a major challenge. Traditional text processing methods often rely on manually set rules or features, which are difficult to adapt to the ever-changing text expressions and contexts, resulting in limited accuracy and efficiency of information extraction.

[0004] On the other hand, even if we can achieve a preliminary understanding of text data, it is still a very challenging task to accurately identify the focus of users among many topics and knowledge points. Traditional analysis methods may not fully consider the inherent correlation of text data and the personalized tendencies of users, resulting in a lack of pertinence and accuracy in the analysis results. Summary of the invention

[0005] 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 information processing method based on big data mining and a population fertility survey service system.

[0006] According to one aspect of the present application, there is provided an information processing method based on big data mining, the method comprising:

[0007] Performing text semantic understanding on the survey text big data of the target survey user to generate text semantic understanding data corresponding to the survey text big data;

[0008] Performing semantic coding representation on the text semantic understanding data to generate a semantic coding vector sequence corresponding to the text semantic understanding data;

[0009] Determine X survey subject knowledge points corresponding to the survey text big data, perform user tendency prediction on the semantic coding vector sequence according to each survey subject knowledge point, and generate a user tendency prediction score corresponding to each survey subject knowledge point; X is a positive integer;

[0010] Performing attention analysis on the text semantic understanding data to generate an attention coefficient corresponding to each survey topic knowledge point;

[0011] The survey topic tendency distribution of the target survey user is determined based on the user tendency prediction score corresponding to each survey topic knowledge point and the attention coefficient corresponding to each survey topic knowledge point.

[0012] According to one aspect of the present application, a population fertility survey service system is provided, which 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 above-mentioned information processing method based on big data mining.

[0013] In the technical solutions provided in some embodiments of the present application, the embodiments of the present application realize efficient and accurate analysis of the target survey user survey text big data. First, by semantically understanding the survey text and generating corresponding semantic understanding data, and then encoding the semantic understanding data into a semantic encoding vector sequence, by determining X survey topic knowledge points, and predicting the user's tendency for each survey topic knowledge point, a user tendency prediction score is generated, so that the user's attitude and preference for each survey topic can be deeply understood. In addition, an attention analysis mechanism is introduced to calculate the attention coefficient corresponding to each survey topic knowledge point, further improving the accuracy and pertinence of the tendency prediction. Finally, combined with the user tendency prediction score and the attention coefficient, the survey topic tendency distribution of the target survey user is accurately determined, which provides data support for research, investigation and analysis in related fields, and greatly improves the efficiency and accuracy of information processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] 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.

[0015] Figure 1 A flowchart of an information processing method based on big data mining provided in an embodiment of the present application;

[0016] Figure 2 This is a schematic block diagram of the architecture of the population fertility survey service system provided in an embodiment of the present application. DETAILED DESCRIPTION

[0017] 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.

[0018] Figure 1 This is a flow chart of an information processing method based on big data mining provided by an embodiment of the present application. The information processing method based on big data mining is introduced in detail below.

[0019] Step S110 , performing text semantic understanding on the survey text big data of the target survey user, and generating text semantic understanding data corresponding to the survey text big data.

[0020] In detail, the target survey user refers to a specific individual or group for which user tendency prediction is required. For example, the target survey user may be a nurse of childbearing age, i.e., a group of nurses at the childbearing age stage. The survey text big data is composed of a large amount of text data from the target survey users, which usually contains information such as user's views, opinions, needs and concerns.

[0021] In this embodiment, taking the childbearing intention and demand survey project of childbearing nurses in a hospital as an example, the server receives a large amount of survey text big data from childbearing nurses, which may include survey texts from multiple survey platforms. These survey texts contain various thoughts, concerns and expectations of nurses on topics related to childbearing in their daily work and life.

[0022] First, we conduct a content logical structure analysis on these survey text big data. For example, a nurse’s survey text is as follows: “As a nurse of childbearing age, I face tremendous pressure at work. The irregular schedule makes it difficult for me to have a stable life rhythm to consider having children. Moreover, there are often emergencies in the hospital that require overtime, which makes me worried about the balance between family and work after giving birth. In addition, from an economic perspective, it is very expensive to raise a child, and our nurses’ salaries are slightly insufficient under the current cost of living. At the same time, I am also worried that having children will affect my career development, and I have seen some nurses around me who have given birth encounter difficulties in job promotion.”

[0023] For such survey text, the first prediction sub-model in the target user tendency prediction model is used for operation. First, the long text is segmented, and words such as "childbearing age", "nurse", "fertility", "work pressure", "scheduling", "family and work balance", "economic cost", and "career development" are segmented. Then, feature extraction is performed on each segmented vocabulary unit to obtain preliminary text features, and these preliminary text features are input into the first prediction sub-model to convert each vocabulary unit into a vector representation in the semantic space.

[0024] The first prediction sub-model is used to conduct a preliminary semantic analysis of the survey text, analyzing the semantic similarity between vocabulary units and the semantic roles of the vocabulary units in the sentence. For example, there is a semantic similarity of causal relationship between "irregularity of shifts" and "work pressure", and "work pressure" is a key influencing factor in the sentence, that is, a semantic role. The preliminary semantic analysis results represent the semantic relationship between vocabulary units in the form of a semantic graph, where "nurse" is the subject, and "work pressure", "economic cost", "career development" and other factors related to the subject that affect fertility intentions. The semantic relationship between these factors is represented by different lines and arrows.

[0025] According to the preliminary semantic analysis results, a candidate set of semantic logic units is constructed. The candidate elements in the candidate set may be grouped vocabulary units, such as "work stress-related vocabulary (work stress, irregular shifts, overtime)"; or they may be sentence combinations, such as "Moreover, there are often emergencies in the hospital that require overtime, which makes me worry about the balance between family and work after giving birth." Each candidate element has corresponding semantic features and potential semantic relationships with other candidate elements.

[0026] Semantic analysis is performed on each candidate element in the candidate set. For candidate elements composed of lexical units, the ambiguity of the lexical unit is analyzed. For example, "pressure" here clearly refers to the pressure brought by work, rather than other pressures. For candidate elements composed of sentences or paragraphs, the influence of the nested structure and rhetoric on the semantics is analyzed. In this example, there are no complex rhetoric, but the causal structure in the sentence clearly expresses the influence of various factors on the desire to have children. For the theme coherence and logical development in the paragraph, the whole paragraph revolves around the factors affecting the desire to have children of nurses of childbearing age, and is logically coherent.

[0027] For the candidate elements in the candidate set, the semantic coherence between the candidate element and other candidate elements is analyzed. For example, there is a direct logical coherence between "work stress related words (work stress, irregular shifts, overtime)" and "family and work balance" because work stress factors will directly affect family and work balance, and then affect fertility intentions. Through analysis, the candidate elements that meet the preset semantic coherence requirements are retained and determined as the semantic logical units corresponding to the survey text big data.

[0028] According to the selected semantic logic units, the logical positions and context information of these semantic logic units are relocated in the survey text big data, and a semantic logic network is constructed with each semantic logic unit as a node. For example, "work pressure" is a node, which is connected to nodes such as "irregular scheduling" and "overtime"; "family and work balance" is another node associated with nodes such as "work pressure".

[0029] In this semantic logic network, for each semantic logic unit, the connection relationship between the semantic logic unit and other semantic logic units is determined. For example, the connection relationship between "work pressure" and "career development" is determined based on the logical relationship in the survey text, because work pressure may hinder the career development of nurses after childbirth. A semantic logic network data containing semantic logic units and the connection relationship between semantic logic units is obtained.

[0030] Perform basic knowledge analysis on each semantic logic unit in the semantic logic network data. For semantic logic units composed of words, such as "salary", analyze the semantic category of the words (belonging to economic words) and the semantic hierarchy of the words (one of the economic factors that affect the willingness to have children). For semantic logic units composed of sentences and paragraphs, analyze their basic semantic propositions. For example, "From an economic perspective, the cost of raising a child is very high, and our nurses' salaries are slightly insufficient under the current cost of living." The semantic proposition of this sentence can be expressed as (nurse salary, child-raising cost, and economic pressure affect the willingness to have children).

[0031] The basic knowledge analysis results are represented in a structured form in the semantic logic network data. The semantic categories of the words are marked with predefined semantic category labels, such as "salary" is marked as "economic factors". The semantic propositions are represented in the form of triples, such as the triple form of the sentence mentioned above.

[0032] Based on the results of basic knowledge analysis, deep knowledge extraction is performed on semantic logic network data. For the connection relationship between semantic logic units, the logical deduction relationship between cause and effect, the strength of causal relationship, and the existence of intermediate factors are analyzed. For example, in the relationship chain of "work pressure->family and work balance->fertility intention", the intensity of the impact of work pressure on family and work balance is analyzed, and whether there are other intermediate factors (such as family support, etc.). Analyze the common characteristics and differences between parallel elements and the role of parallel elements in the global semantics. For example, "insufficient salary" and "hindered career development" are both factors that affect fertility intention. Their common feature is that they are both negative influencing factors. The difference is that one is economic and the other is professional. They both play a role in hindering fertility intention in the global semantics.

[0033] The deep knowledge extraction results and basic knowledge analysis results are integrated to generate semantic logic data corresponding to the semantic logic network data. Finally, the server performs such processing on the survey texts of each nurse in the entire survey text big data to generate text semantic understanding data corresponding to the survey text big data.

[0034] Step S120, semantically encode the text semantic understanding data to generate a semantic encoding vector sequence corresponding to the text semantic understanding data.

[0035] In this embodiment, after obtaining the text semantic understanding data, the text semantic encoder in the first prediction sub-model is used to semantically encode and represent the data. Taking the text semantic understanding data generated by the survey text of the nurse mentioned above as an example, the text semantic encoder will convert various semantic information in the text semantic understanding data into vector form.

[0036] The semantic information representing the semantic logic unit "work pressure" may be represented as a specific vector after encoding, and each element in this vector represents a different dimension of the semantic information of "work pressure", such as the source of pressure, the magnitude of pressure, the degree of pressure's impact on life, etc. Similarly, the semantic information corresponding to the semantic logic units such as "family and work balance", "economic cost", and "career development" will also be encoded into corresponding vectors.

[0037] These vectors are combined in a certain order to form a semantic coding vector sequence. This sequence fully represents the semantic information of the entire survey text big data after text semantic understanding, providing a data basis for subsequent user tendency prediction.

[0038] Step S130, determining X survey subject knowledge points corresponding to the survey text big data, performing user tendency prediction on the semantic coding vector sequence according to each survey subject knowledge point, and generating a user tendency prediction score corresponding to each survey subject knowledge point. X is a positive integer.

[0039] In the scenario of analyzing the reproductive intentions and needs of nurses of childbearing age, the server identified X survey topic knowledge points, such as "time arrangement between work and childbirth", "financial support after childbirth", "balance between career development and childbirth", etc.

[0040] A second prediction sub-model is determined from the target user tendency prediction model, and the second prediction sub-model includes a multi-layer perceptron corresponding to each survey topic knowledge point.

[0041] Taking the survey topic of "work and childbearing schedule" as an example, the corresponding multi-layer perceptron predicts user preferences for the semantic coding vector sequence. The information about work pressure (such as irregular shifts, overtime, etc.), family and work balance, etc. contained in each vector in the semantic coding vector sequence will be processed by this multi-layer perceptron.

[0042] If in the semantic coding vector sequence, the value of the vector element indicating high work pressure and irregular shift scheduling is high, then after calculation by the multi-layer perceptron, it may be concluded that the nurse has a low tendency score for the survey topic knowledge point "work and childbirth time arrangement". Assuming that this score is 30 points (out of 100 points), it means that she is less satisfied with the current work and childbirth time arrangement, and her willingness to have children may be greatly affected.

[0043] For the survey topic knowledge point of "economic support after childbirth", the multi-layer perceptron analyzes the vector elements related to economic costs (such as child-rearing costs, nurses' salaries, etc.) in the semantic coding vector sequence. If the semantic coding vector sequence indicates that nurses' salaries are low and child-rearing costs are high, then it may be concluded that the nurse's tendency prediction score for the survey topic knowledge point of "economic support after childbirth" is 20 points, indicating that she is more concerned about the economic support after childbirth and has a low willingness to have children.

[0044] Similarly, for the survey topic knowledge point of "balance between career development and childbearing", the multi-layer perceptron performs calculations based on the vector information related to career development obstacles in the semantic coding vector sequence. If the vector elements indicating career development obstacles (such as difficulty in job promotion after childbearing) are obvious, the possible tendency prediction score may be 25 points, indicating that she is not satisfied with the current balance between career development and childbearing, and her desire to have children will also be suppressed.

[0045] In this way, the semantic encoding vector sequence is processed according to the multi-layer perceptron corresponding to each survey topic knowledge point to generate a user tendency prediction score corresponding to each survey topic knowledge point.

[0046] Step S140, performing attention analysis on the text semantic understanding data to generate an attention coefficient corresponding to each survey topic knowledge point.

[0047] In this embodiment, the server determines the third prediction sub-model from the target user tendency prediction model. Taking the text semantic understanding data generated by the survey text of the nurse mentioned above as an example, the third prediction sub-model is used to perform attention analysis on it.

[0048] The guidance unit in the third prediction sub-model performs guidance feature analysis on the text semantic understanding data. For the survey topic knowledge point of "Time arrangement of work and childbirth", the guidance unit will focus on the features related to work pressure (such as scheduling, overtime, etc.) and family and work balance in the text semantic understanding data. For example, when analyzing the work scheduling of nurses, the guidance unit will focus on features such as the scheduling cycle and the frequency of overtime. The manifestation of these features in the text semantic understanding data will be extracted as guidance features.

[0049] Then, the attention unit is used to perform attention analysis on the guiding features. If the description of irregular work schedules is very detailed and mentioned many times in the text semantic understanding data, then a higher attention coefficient may be generated for the survey topic knowledge point "time arrangement of work and childbirth", assuming it is 0.8. This means that when analyzing the nurse's fertility intention, the time arrangement of work and childbirth is a very important factor that needs to be focused on.

[0050] For the survey topic knowledge point of "economic support after childbirth", the guidance unit will focus on features related to economic costs (such as the cost of raising children, nurses' salaries, etc.). If in the text semantic understanding data, the nurse elaborated in detail that the cost of raising children is very high and there is a big gap compared with her own salary, then the attention unit may generate an attention coefficient of 0.7 for this survey topic knowledge point, indicating that this is also a relatively important factor.

[0051] For the survey topic knowledge point of "balance between career development and childbearing", the guidance unit will focus on the characteristics related to career development obstacles (such as difficulty in job promotion after childbearing, etc.). If there are many descriptions in the text about seeing nurses around them having difficulty in job promotion after childbearing, the attention unit may generate an attention coefficient of 0.6 for this survey topic knowledge point, indicating that this factor is also important in analyzing childbearing intentions.

[0052] In this way, a corresponding attention coefficient is generated for each survey topic knowledge point.

[0053] Step S150, determining the survey topic tendency distribution of the target survey user according to the user tendency prediction score corresponding to each survey topic knowledge point and the attention coefficient corresponding to each survey topic knowledge point.

[0054] In this embodiment, the server uses the attention coefficient corresponding to each survey topic knowledge point to perform a fusion calculation on the user tendency prediction score corresponding to each survey topic knowledge point.

[0055] Taking the previous example, for the survey theme knowledge point of "work and childbirth time arrangement", the user's tendency prediction score is 30 points and the attention coefficient is 0.8; for the survey theme knowledge point of "financial support after childbirth", the user's tendency prediction score is 20 points and the attention coefficient is 0.7; for the survey theme knowledge point of "balance between career development and childbirth", the user's tendency prediction score is 25 points and the attention coefficient is 0.6.

[0056] First, calculate the fusion user tendency prediction score. For "time arrangement between work and childbirth", the fusion user tendency prediction score = 30×0.8 = 24 points; for "financial support after childbirth", the fusion user tendency prediction score = 20×0.7 = 14 points; for "balance between career development and childbirth", the fusion user tendency prediction score = 25×0.6 = 15 points.

[0057] Based on these fusion user tendency prediction scores, the survey topic tendency distribution of the target survey user (i.e., the nurse of childbearing age) is determined. It can be seen that in the analysis of this nurse's fertility intention and needs, the factor of "time arrangement between work and childbearing" accounts for a relatively large proportion in her consideration (the fusion score is 24 points), and "economic support after childbearing" and "balance between career development and childbearing" also have a certain influence (the fusion scores are 14 points and 15 points respectively). Through such a survey topic tendency distribution, hospital managers or relevant policy makers can better understand the fertility intention and needs of nurses of childbearing age, and thus take corresponding measures, such as adjusting the work scheduling system, providing fertility economic subsidies, and ensuring the career development of nurses after childbearing.

[0058] For the survey text big data of the entire group of nurses of childbearing age, the server processes the survey text of each nurse in the above steps to obtain the survey topic tendency distribution of the entire group, providing a strong basis for a comprehensive understanding of the fertility intentions and needs of nurses of childbearing age.

[0059] Based on the above steps, the embodiment of the present application realizes efficient and accurate analysis of the target survey user survey text big data. First, by semantically understanding the survey text and generating corresponding semantic understanding data, and then encoding the semantic understanding data into a semantic encoding vector sequence, by determining X survey topic knowledge points, and predicting the user's tendency for each survey topic knowledge point, a user tendency prediction score is generated, so that the user's attitude and preference for each survey topic can be deeply understood. In addition, an attention analysis mechanism is introduced to calculate the attention coefficient corresponding to each survey topic knowledge point, further improving the accuracy and pertinence of the tendency prediction. Finally, combined with the user tendency prediction score and the attention coefficient, the survey topic tendency distribution of the target survey user is accurately determined, which provides data support for research, investigation and analysis in related fields, and greatly improves the efficiency and accuracy of information processing.

[0060] In a possible implementation, step S110 includes:

[0061] Step S111 , performing content logical structure analysis on the survey text big data to generate content logical structure data of the survey text big data.

[0062] Step S112: performing logical semantic understanding on the content logical structure data to generate text semantic understanding data of the content logical structure data.

[0063] In a possible implementation, step S111 includes:

[0064] Step S1111, when the survey text big data of the target survey user is obtained, a target user tendency prediction model for predicting the user tendency of the target survey user is obtained. The target user tendency prediction model includes a first prediction sub-model.

[0065] Step S1112, using the first prediction sub-model to determine the semantic logic unit corresponding to the survey text big data, extracting the semantic logic network data of the survey text big data based on the semantic logic unit, and performing knowledge extraction on the semantic logic network data to generate semantic logic data corresponding to the semantic logic network data.

[0066] Step S1113: Use the first prediction sub-model to perform structural processing on the semantic logic data corresponding to the semantic logic network data to generate content logical structure data of the survey text big data.

[0067] In this embodiment, the content logical structure analysis is first performed on the survey text big data. When a large amount of survey text big data of nurses of childbearing age is obtained, the server will obtain a target user tendency prediction model for predicting user tendencies for these nurses, and the model contains a first prediction sub-model. For example, a nurse's survey text mentioned, "As a nurse of childbearing age, my working hours are unstable and I often work shifts. This working status affects my consideration of having children. Financially, the mortgage and living expenses are high, and having children will add more financial pressure. Moreover, in the hospital promotion system, having children may cause me to miss promotion opportunities."

[0068] The server uses the first prediction sub-model to determine the semantic logic unit. It first performs word segmentation on the text, such as "nurses of childbearing age", "unstable working hours", "shift work", "fertility considerations", "economic pressure", "mortgage", "hospital promotion system" and other words are segmented. Then these words are feature extracted and converted into vector representations in the semantic space. Through preliminary semantic analysis, a candidate set of semantic logic units is constructed, which may include candidate elements such as "work-related factors (unstable working hours, shift work)", "economic factors (mortgage, high living expenses, childbirth increases economic pressure)", and "career promotion factors (hospital promotion system, childbirth affects promotion opportunities)".

[0069] The server performs semantic analysis on these candidate elements, analyzing the polysemy of vocabulary units, the impact of sentence structure on semantics, etc. For example, "pressure" here clearly refers to economic and professional pressure. Then the semantic coherence between candidate elements is analyzed. For example, "work-related factors" and "fertility considerations" are logically related because work status affects fertility considerations. Candidate elements that meet the preset semantic coherence requirements are determined as semantic logical units.

[0070] Based on these semantic logic units, the server extracts the semantic logic network data of the survey text big data. Taking "work-related factors" as a node, it is connected to the "fertility considerations" node; the "economic factors" node is associated with the "fertility considerations" node. In this semantic logic network, the connection relationship between each node (semantic logic unit) is determined based on the logical relationship in the text, such as cause and effect.

[0071] The server extracts knowledge from the semantic logic network data. For semantic logic units composed of words, such as "housing loan", analyze its semantic category (belongs to economic vocabulary) and semantic hierarchy (is part of economic pressure); for semantic logic units composed of sentences, analyze its semantic proposition, such as "Working hours are unstable and I often work shifts. This working state affects my consideration of fertility." The semantic proposition can be expressed as (working state, fertility consideration, influence relationship). Then the basic knowledge analysis results are expressed in a structured form, such as "housing loan" is marked as "economic factors", and the semantic proposition is expressed in the form of triples. On this basis, deep knowledge extraction is performed to analyze the logical deduction relationship of the connection relationship between semantic logic units, such as in the "working state->fertility consideration" relationship, analyze the intensity of the influence of working state on fertility consideration, etc. The deep knowledge extraction results and the basic knowledge analysis results are integrated to obtain the semantic logic data corresponding to the semantic logic network data.

[0072] Finally, the server uses the first prediction sub-model to perform structural processing on the semantic logic data corresponding to the semantic logic network data, thereby generating content logic structure data of the survey text big data. This content logic structure data contains information such as the structural relationship and semantic features of various semantic logic units in the survey text, laying the foundation for further generation of text semantic understanding data. Afterwards, the server performs logical semantic understanding on the content logic structure data, thereby generating text semantic understanding data of the content logic structure data. This process will deeply mine the semantic information in the content logic structure data, consider more complex semantic relationships and semantic levels between semantic logic units, and finally obtain complete text semantic understanding data for subsequent analysis and processing of various related analyses and processing of childbearing intentions and needs of nurses of childbearing age.

[0073] In a possible implementation, step S1112 includes:

[0074] Step S1112-1, after performing word segmentation operation on the survey text big data, feature extraction is performed on each word-segmented vocabulary unit to obtain preliminary text features, and the preliminary text features are input into the first prediction sub-model to convert each vocabulary unit into a vector representation in the semantic space.

[0075] Step S1112-2, performing preliminary semantic analysis on the survey text big data through the first prediction sub-model to obtain preliminary semantic analysis results, wherein the preliminary semantic analysis includes analyzing the semantic similarity between vocabulary units and the semantic roles of the unit vocabulary in the sentence. The preliminary semantic analysis results represent the semantic relationship between vocabulary units in the form of a semantic graph, or represent the semantic role of the vocabulary unit in the form of a tag.

[0076] Step S1112-3, constructing a candidate set of semantic logic units based on the preliminary semantic analysis results, the candidate set includes multiple candidate elements, the candidate elements are grouped vocabulary units, sentences or paragraph combinations, each candidate element has corresponding semantic features and potential semantic relationships with other candidate elements.

[0077] Step S1112-4, performing semantic analysis on each candidate element in the candidate set, and updating the semantic information of each candidate element according to the semantic analysis result, wherein, for candidate elements composed of vocabulary units, the ambiguity of the vocabulary units is analyzed; for candidate elements composed of sentences or paragraphs, the influence of the nested structure and rhetoric in the sentence on the semantics, as well as the thematic coherence and logical development in the paragraph are analyzed.

[0078] Step S1112-5, for the candidate elements in the candidate set, analyze the semantic coherence between the candidate element and other candidate elements, and retain the candidate elements that meet the preset semantic coherence requirements to determine them as the semantic logical units corresponding to the survey text big data, specifically including analyzing the expression of the semantic coherence between vocabulary units in sentences and paragraphs, and the expression of the semantic coherence between sentences and paragraphs in the survey text big data.

[0079] Step S1112-6, based on the filtered semantic logic units, relocate the logical position and context information of the semantic logic units in the survey text big data, and construct a semantic logic network with each semantic logic unit as a node.

[0080] Step S1112-7, in the semantic logic network, for each semantic logic unit, determine the connection relationship between the semantic logic unit and other semantic logic units, the connection relationship is determined based on the grammatical structure, semantic reference and logical relationship in the survey text big data, and obtain a semantic logic network data including semantic logic units and the connection relationship between semantic logic units.

[0081] Step S1112-8, performing basic knowledge analysis on each semantic logic unit in the semantic logic network data to obtain basic knowledge analysis results, wherein for the semantic logic unit composed of vocabulary, the semantic categories and semantic hierarchical structures of the vocabulary are analyzed; for the semantic logic unit composed of sentences and paragraphs, its basic semantic propositions are analyzed.

[0082] Step S1112-9: The basic knowledge analysis result is represented in the semantic logic network data in a structured form, wherein the semantic category of the vocabulary is marked with a predefined semantic category label, and the semantic proposition is represented in the form of a triple.

[0083] Step S1112-10, based on the basic knowledge analysis results, deep knowledge extraction is performed on the semantic logic network data to obtain deep knowledge extraction results, wherein the connection relationship between the semantic logic units is analyzed, the logical deduction relationship between the cause and the effect, the strength of the causal relationship and the existence of intermediate factors, as well as the common characteristics and differences between the parallel elements and the role of the parallel elements in the global semantics.

[0084] Step S1112-11, integrating the deep knowledge extraction result and the basic knowledge analysis result to generate semantic logic data corresponding to the semantic logic network data.

[0085] In this embodiment, first, the survey text big data is segmented, for example, for a survey text of a nurse of childbearing age, "I am very busy at work in the hospital and often need to work night shifts, which makes me very entangled when considering the issue of childbearing. From an economic point of view, the cost of raising a child is very high, and my income is limited, which is also a factor that hinders me from having children. In addition, the hospital does not have a clear support policy for the career development of nurses after childbirth, which makes me worry that having children will affect my career advancement." After word segmentation, vocabulary units such as "hospital", "busy work", "night shift", "fertility issues", "economic aspects", "child raising costs", "limited income", "career development", and "career advancement" are obtained. Then, feature extraction is performed on each word-segmented vocabulary unit to obtain preliminary text features. For example, "busy work" may extract features such as working hours and work intensity, and "limited income" may extract features such as income value range and stability of income source. After that, the preliminary text features are input into the first prediction sub-model, and each vocabulary unit is converted into a vector representation in the semantic space, so that each vocabulary unit has a quantitative representation at the semantic level, which is convenient for subsequent analysis.

[0086] The first prediction sub-model is used to conduct a preliminary semantic analysis of the survey text big data, analyzing the semantic similarity between vocabulary units and the semantic roles of the unit vocabulary in the sentence. For example, there is a semantic similarity between "night shift" and "busy at work" because night shift is a factor that leads to busy work; in the sentence "This makes me very entangled when considering the issue of fertility", "this" refers to the previously mentioned "busy at work" and "night shift", and its semantic role is to be the cause of the subsequent "entanglement with fertility issues". When the preliminary semantic analysis results represent the semantic relationship between vocabulary units in the form of a semantic graph, vocabulary units such as "busy at work" and "night shift" may point to "fertility issues" through directed arrows, indicating the influence relationship on fertility issues; or the semantic role of the vocabulary unit is represented in the form of a marker, such as "this" is marked as a demonstrative pronoun referring to the previous content and has a causal role in the sentence.

[0087] According to the preliminary semantic analysis results, a candidate set of semantic logic units is constructed. The candidate elements in the candidate set have various forms, such as grouped vocabulary units, such as "work-related vocabulary (busy work, night shift)", which has the semantic feature of busy work and has a potential semantic relationship with other candidate elements such as "fertility problems", because busy work can affect fertility problems; sentence combinations are also candidate elements, such as "from an economic point of view, the cost of raising a child is very high, and my income is limited, which is also a factor that hinders me from having children", the semantic feature of this sentence combination is to explain the obstacles of economic factors to fertility, and has a potential relationship with other candidate elements related to work or career development, because they are all different aspects that affect the willingness to have children; paragraph combinations can also be candidate elements.

[0088] Perform semantic analysis on each candidate element in the candidate set and update the semantic information. For candidate elements composed of vocabulary units, analyze their ambiguity. For example, the vocabulary unit "busy" in this context clearly refers to being busy with many work tasks and tight schedules, rather than being busy in other senses. For candidate elements composed of sentences or paragraphs, analyze the impact of the nested structure and rhetoric on semantics in the sentence. There are no complex rhetoric in the above sentence combination, but there is a nested structure. For example, the nested structure of "the cost of raising a child is high, and my income is limited" shows the contrast between the two, which affects the understanding of factors that hinder fertility; analyze the theme coherence and logical development in the paragraph. The entire paragraph revolves around different factors that affect the willingness to have children, from work to economy to career development, with logical coherence and clear themes.

[0089] Then, the semantic coherence between candidate elements is analyzed to determine semantic logical units. For the expression of semantic coherence between lexical units in sentences and paragraphs, for example, in the sentence "I am very busy at work in the hospital and often have to work night shifts, which makes me very entangled when considering fertility issues", "busy work" and "night shifts" are semantically coherent within the sentence, and this coherence is combined with the semantic coherence of "fertility issues" in the paragraph, expressing the impact of work factors on fertility issues; semantic coherence between sentences and paragraphs is also very important. For example, different paragraphs explain the impact of work, economy, and career development on fertility respectively. The semantic coherence between these paragraphs together constitutes an expression of the factors affecting overall fertility intention. Candidate elements that meet the preset semantic coherence requirements are determined as semantic logical units corresponding to the survey text big data, such as "work-related factors (busy work, night shifts)", "economic factors (child-rearing costs, limited income)", "career development factors (career promotion, hospital support policies)", etc. are determined as semantic logical units.

[0090] According to the selected semantic logic units, their logical positions and context information are relocated in the survey text big data to construct a semantic logic network. For example, the semantic logic unit "work-related factors" is located near the beginning of the text, and its context is the nurse explaining the impact of her work status on fertility; with each semantic logic unit as a node, a semantic logic network is constructed, and the "work-related factors" node is connected to the "fertility problem" node, and the "economic factors" node is also connected to the "fertility problem" node.

[0091] In the semantic logic network, the connection relationship between each semantic logic unit and other semantic logic units is determined. The connection relationship is determined based on the grammatical structure, semantic reference and logical relationship in the survey text big data. For example, in the sentence "In addition, the hospital does not have a clear support policy for the career development of nurses after childbirth, which makes me worry that childbirth will affect my career advancement", from the grammatical structure, "this" refers to the previous hospital policy, and there is a causal relationship in semantics, so it is determined that there is a causal connection relationship between the semantic logic unit "hospital support policy" and the semantic logic unit "career advancement", and the semantic logic network data containing semantic logic units and the connection relationship between them is obtained.

[0092] Perform basic knowledge analysis on each semantic logic unit in the semantic logic network data. For semantic logic units composed of words, such as "limited income", the semantic category of the words is analyzed as economic words, and the semantic hierarchy is the personal income aspect of the economic factors affecting fertility intention; for semantic logic units composed of sentences and paragraphs, analyze their basic semantic propositions. For example, for the sentence "From an economic point of view, the cost of raising a child is very high, and my income is limited, which is also a factor that hinders me from having children", its semantic proposition can be expressed as (child-raising cost, limited income, hindering fertility).

[0093] The results of the basic knowledge analysis are represented in a structured form in the semantic logic network data. The semantic categories of the words are marked with predefined semantic category labels, and "limited income" is marked as "economic factors"; the semantic propositions are represented in the form of triples, such as the triple form of the above sentence.

[0094] Based on the results of basic knowledge analysis, deep knowledge extraction is performed on semantic logic network data. For the connection relationship between semantic logic units, the logical deduction relationship between cause and effect, the strength of causal relationship, and the existence of intermediate factors are analyzed. For example, in the relationship of "busy work->fertility problems", the rationality of the logical deduction between busy work and fertility problems, the intensity of the impact of busy work on fertility problems, and whether there are intermediate factors such as family support; analyze the common characteristics and differences between parallel elements and their role in the global semantics. For example, "child-raising costs" and "limited income" are parallel economic factors. The common feature is that they are both economic factors that affect fertility. The difference is that one is expenditure and the other is income. In the global semantics, they together constitute the impact of economic factors on fertility intentions.

[0095] Finally, the results of deep knowledge extraction and basic knowledge analysis are integrated to generate semantic logic data corresponding to semantic logic network data. This semantic logic data comprehensively covers semantic information from vocabulary units to semantic logic units, from semantic relations to logical deductions, and provides a rich data basis for further analysis of the fertility intentions and needs of nurses of childbearing age.

[0096] In a possible implementation, the semantic encoding vector sequence is generated by a text semantic encoder in a first prediction sub-model performing semantic encoding representation on the text semantic understanding data. The first prediction sub-model belongs to a target user tendency prediction model for predicting user tendency of the target survey user.

[0097] The method further comprises:

[0098] Step A110, obtaining a first template survey text data sequence. The first template survey text included in the first template survey text data sequence is a template survey text that does not carry training supervision data.

[0099] Step A120, obtaining Y sample learning processes for the first basic prediction sub-model, where Y is a positive integer.

[0100] Step A130, using the first basic prediction sub-model to perform text semantic understanding on the first template survey text, generating a sample semantic coding vector sequence corresponding to each sample learning process, and determining Y model error information corresponding to the Y sample learning processes based on the sample semantic coding vector sequence. One sample learning process corresponds to one model error information.

[0101] Step A140, determining first target model error information based on the Y model error information, training the first basic prediction sub-model based on the first target model error information, and outputting the trained first basic prediction sub-model as the first prediction sub-model.

[0102] In this embodiment, the semantic encoding vector sequence is generated by the text semantic encoder in the first prediction sub-model by semantically encoding the text semantic understanding data. In this scenario, the text semantic understanding data has been obtained by processing the survey text big data of nurses of childbearing age. For example, these data contain semantic information about factors related to fertility intention, such as nurses' work pressure, economic conditions, and career development. The first prediction sub-model is part of the target user tendency prediction model used to predict the user tendency of the target survey user (nurses of childbearing age).

[0103] Next, the server needs to obtain the first template survey text data sequence. These first template survey texts are template survey texts that do not carry training supervision data. For example, these template survey texts may be some general descriptions about the occupation and fertility of women of childbearing age, such as "In a professional environment, women's fertility is often accompanied by risks in career development, and economic factors also have an impact on fertility decisions."

[0104] Then the server obtains Y sample learning processes (Y is a positive integer) for the first basic prediction sub-model. For example, Y may be 3, and these three sample learning processes respectively optimize and train the first basic prediction sub-model from different angles.

[0105] In the first sample learning process, the first basic prediction sub-model is used to perform text semantic understanding on the first template survey text. Just like the template survey text mentioned above, the model will first perform similar processing steps on the actual survey text, such as word segmentation, feature extraction, preliminary semantic analysis, etc., and then generate a sample semantic encoding vector sequence corresponding to each sample learning process. It is assumed that the vector elements in this sample semantic encoding vector sequence represent the quantitative features of different semantic concepts in the text (such as occupational risks, economic impact, etc.). Based on this sample semantic encoding vector sequence, the model error information corresponding to this sample learning process is determined. This process may involve comparing with a predefined standard semantic encoding vector sequence (if it exists) or a reasonable range determined based on some prior knowledge, and calculating the difference between the two as the model error information.

[0106] In the second sample learning process, the first template survey text is also processed. This template survey text may be "In some jobs, women may face job adjustments after giving birth, and changes in economic income are also closely related to childbirth." The first basic prediction sub-model performs text semantic understanding again to generate a new sample semantic encoding vector sequence. This sequence reflects the encoding of the semantic concepts in the text (job adjustments, changes in economic income, etc.). Then, based on this sample semantic encoding vector sequence, the model error information corresponding to this sample learning process is determined, and the determination method is also to compare with the expected or a priori reasonable encoding situation.

[0107] In the third sample learning process, for another template survey text, such as "After giving birth, changes in work intensity and increased family financial burden are common phenomena for working women", the first basic prediction sub-model repeats the text semantic understanding operation to obtain the corresponding sample semantic encoding vector sequence, and determines the model error information corresponding to this sample learning process.

[0108] Afterwards, the first target model error information is determined based on the model error information corresponding to the Y (here 3) sample learning processes. This determination process may be a weighted average of the three model error information or a comprehensive calculation according to a predefined algorithm. Finally, the first basic prediction sub-model is trained based on the first target model error information, and the error is reduced by adjusting the parameters of the model. The trained first basic prediction sub-model is output as the first prediction sub-model, and this first prediction sub-model can more accurately process the survey text of nurses of childbearing age, such as more accurately performing semantic encoding, and provide a more reliable basis for subsequent operations such as user tendency prediction.

[0109] For example, in a possible implementation, the Y sample learning processes include a semantic hidden learning subprocess.

[0110] Step A130 includes:

[0111] The first basic prediction sub-model is used to perform content logical structure analysis on the first template survey text to generate first template content logical structure data of the first template survey text. The first template content logical structure data includes template semantic logic data on multiple logical nodes.

[0112] A first logical node to be semantically hidden is randomly extracted from a plurality of logical nodes corresponding to the first template content logical structure data.

[0113] The template semantic logic data on the first logic node in the first template content logical structure data is semantically hidden according to the semantic hiding information to generate first semantic hiding learning data.

[0114] The first basic prediction sub-model is used to perform logical semantic understanding on the first semantic hidden learning data, to generate first sample text semantic understanding data of the first semantic hidden learning data, to perform semantic encoding representation on the first sample text semantic understanding data, to generate a first sample semantic encoding vector sequence corresponding to the first sample text semantic understanding data. The first sample semantic encoding vector sequence belongs to the sample semantic encoding vector sequence.

[0115] The prediction sub-model corresponding to the semantic hidden learning sub-process is used to perform hidden semantic prediction on the first sample semantic coding vector sequence to generate predicted semantic logic data corresponding to the semantic hidden information in the first semantic hidden learning data.

[0116] According to the template semantic logic data on the first logic node in the first template content logical structure data and the predicted semantic logic data, the model error information corresponding to the semantic hidden learning subprocess is determined.

[0117] In this embodiment, the server obtains the first template survey text, which is related to the fertility intention of nurses of childbearing age, such as "Women of childbearing age working in hospitals face problems such as high work intensity and mismatch between salary and child-rearing costs, which affect their fertility intention. At the same time, the unclear career development prospects also make fertility decisions complicated". The server uses the first basic prediction sub-model to perform content logical structure analysis on the first template survey text. This process is similar to the previous processing of the actual survey text. First, the text is segmented, feature extracted, and then the semantic logic unit is determined, and the semantic logic network is constructed to generate the first template content logical structure data of the first template survey text. This data contains template semantic logic data on multiple logical nodes. For example, the template semantic logic data on the logical node "high work intensity" may contain information such as working hours and work pressure; the template semantic logic data on the logical node "mismatch between salary and child-rearing costs" may contain information such as salary level and composition of child-rearing costs; the template semantic logic data on the logical node "uncertain career development prospects" may contain information such as promotion opportunities and career planning.

[0118] Then, from the multiple logical nodes corresponding to the logical structure data of the first template content, the first logical node to be semantically hidden is randomly extracted. Assume that the logical node "salary does not match child-rearing costs" is randomly selected. Then, semantic hiding is performed on the template semantic logic data on this logical node based on the semantic hiding information. For example, the key information such as the specific value of the salary and the specific composition of the child-rearing costs in this logical node is hidden to generate the first semantic hiding learning data.

[0119] The first basic prediction sub-model is used to perform logical semantic understanding on the first semantic hidden learning data. In this process, the model will try to understand the semantics from the text with hidden partial information and generate the first sample text semantic understanding data. For example, the model may infer the semantic impact of the hidden part based on the information of other related logical nodes (such as high work intensity and unclear career development prospects). Then, the first sample text semantic understanding data is semantically encoded to generate the first sample semantic encoding vector sequence corresponding to the first sample text semantic understanding data. The elements in this vector sequence may represent the quantitative characteristics of each semantic concept in the text after hiding some information.

[0120] The prediction sub-model corresponding to the semantic hidden learning sub-process is used to perform hidden semantic prediction on the first sample semantic coding vector sequence. For example, the prediction sub-model attempts to predict the semantic logic data of the hidden logic node "salary does not match child-rearing costs" based on other relevant semantic features in the vector sequence. Generate predicted semantic logic data corresponding to the semantic hidden information in the first semantic hidden learning data. Then, based on the original template semantic logic data and the predicted semantic logic data on the first logic node in the first template content logical structure data, determine the model error information corresponding to the semantic hidden learning sub-process. This error information may be determined by comparing the differences between the original data and the predicted data, such as a comprehensive evaluation of the gap between the predicted value and the actual value of the salary level, the predicted accuracy of the child-rearing cost structure, etc.

[0121] For example, in a possible implementation, the Y sample learning processes include a semantic change learning subprocess.

[0122] Step A130 includes:

[0123] The first basic prediction sub-model is used to perform content logical structure analysis on the first template survey text to generate first template content logical structure data of the first template survey text. The first template content logical structure data includes template semantic logic data on multiple logical nodes.

[0124] A second logical node to be semantically changed is randomly extracted from a plurality of logical nodes corresponding to the first template content logical structure data.

[0125] The template semantic logic data at the second logic node in the first template content logical structure data is subjected to a semantic change operation according to random semantic information to generate second semantic hidden learning data. The random semantic information is different from the template semantic logic data at the second logic node.

[0126] The first basic prediction sub-model is used to perform logical semantic understanding on the second semantic hidden learning data, to generate second sample text semantic understanding data of the second semantic hidden learning data, to perform semantic encoding representation on the second sample text semantic understanding data, to generate a second sample semantic encoding vector sequence corresponding to the second sample text semantic understanding data. The second sample semantic encoding vector sequence belongs to the sample semantic encoding vector sequence.

[0127] The prediction sub-model corresponding to the semantic change learning sub-process is used to predict the semantic change logic nodes of the second sample semantic coding vector sequence, and the predicted semantic change logic nodes corresponding to the multiple logic nodes are generated.

[0128] According to the predicted semantic change logical nodes corresponding to the second logical node and the multiple logical nodes, model error information corresponding to the semantic change learning sub-process is determined.

[0129] For the same first template survey text "Women of childbearing age working in hospitals face problems such as high work intensity and mismatch between salary and child-rearing costs, which affect their willingness to have children. At the same time, unclear career prospects also complicate fertility decisions", the server uses the first basic prediction sub-model to analyze the content logical structure and generates the first template content logical structure data of the first template survey text, which includes template semantic logical data on each logical node (such as high work intensity, mismatch between salary and child-rearing costs, and unclear career prospects).

[0130] From the multiple logical nodes corresponding to the logical structure data of the first template content, randomly extract the second logical node to be semantically changed. Assume that the logical node "uncertain career development prospects" is selected. Perform semantic change operations on the template semantic logic data on this logical node based on random semantic information. For example, the original unclear career development prospects may refer to situations such as few promotion opportunities and lack of vocational training. Now, based on random semantic information, it is changed to situations such as there are certain plans for career development but difficulties in implementation, and the second semantic hidden learning data is generated.

[0131] The first basic prediction sub-model is used to perform logical semantic understanding on the second semantic hidden learning data to generate the second sample text semantic understanding data. In this process, the model needs to re-understand the text content after the semantic change and consider the semantic relationship of the impact of the changed career development on the overall fertility intention. Then the second sample text semantic understanding data is semantically encoded to generate the second sample semantic encoding vector sequence corresponding to the second sample text semantic understanding data. This vector sequence reflects the quantitative characteristics of the text semantic concept after the semantic change.

[0132] The prediction sub-model corresponding to the semantic change learning sub-process is used to predict the semantic change logic nodes of the second sample semantic coding vector sequence. The prediction sub-model attempts to predict the predicted semantic change logic nodes corresponding to multiple logic nodes under this semantic change situation based on the semantic features in the vector sequence. For example, predict the status of other relevant factors of career development (such as the relationship with salary, the impact on work input, etc.) after the semantic change. Then, based on the second logic node (the original "uncertain career development prospects") and the predicted semantic change logic nodes corresponding to the multiple logic nodes, the model error information corresponding to the semantic change learning sub-process is determined. The determination of this error information may involve comparing the differences between the original logic node and the predicted semantic change logic node in terms of semantic relationship, logical coherence, etc.

[0133] For example, in one possible implementation, the Y sample learning processes include a perturbation semantic learning sub-process.

[0134] Step A130 includes:

[0135] Semantic disturbance information is randomly injected into the semantic logic network data of the first template survey text to generate a disturbance template survey text, and content logical structure analysis is performed on the disturbance template survey text to generate disturbance content logical structure data of the disturbance template survey text.

[0136] A content logical structure analysis is performed on the first template survey text to generate first template content logical structure data of the first template survey text.

[0137] The first basic prediction sub-model is used to perform logical semantic understanding on the disturbance content logical structure data, generate third sample text semantic understanding data of the disturbance content logical structure data, perform semantic coding representation on the third sample text semantic understanding data, and generate a third sample semantic coding vector sequence corresponding to the third sample text semantic understanding data. The third sample semantic coding vector sequence belongs to the sample semantic coding vector sequence.

[0138] The first basic prediction sub-model is used to perform logical semantic understanding on the first template content logical structure data, generate fourth sample text semantic understanding data of the first template content logical structure data, perform semantic coding representation on the fourth sample text semantic understanding data, and generate a fourth sample semantic coding vector sequence corresponding to the fourth sample text semantic understanding data. The fourth sample semantic coding vector sequence belongs to the sample semantic coding vector sequence.

[0139] Model error information corresponding to the perturbation semantic learning subprocess is determined based on the third sample semantic coding vector sequence and the fourth sample semantic coding vector sequence.

[0140] For the first template survey text "Women of childbearing age working in hospitals face problems such as high work intensity and mismatch between salary and child-rearing costs, which affect their willingness to have children. At the same time, uncertain career prospects also make fertility decisions complicated", the server randomly injects semantic perturbation information into its semantic logic network data to generate a perturbation template survey text. For example, injecting perturbation information into the semantic relationship of the semantic logic unit "high work intensity" may change the influence relationship between work intensity and fertility intention to a certain extent (such as a positive correlation that becomes a negative correlation after perturbation). Then, the perturbation template survey text is analyzed for content logic structure to generate perturbation content logic structure data of the perturbation template survey text. At the same time, the original first template survey text is analyzed for content logic structure to generate the first template content logic structure data of the first template survey text.

[0141] The first basic prediction sub-model is used to perform logical semantic understanding on the perturbed content logical structure data, and generate third sample text semantic understanding data of the perturbed content logical structure data. In this process, the model needs to understand the text semantic relationship after the perturbation information is injected. Then the third sample text semantic understanding data is semantically encoded and represented, and a third sample semantic encoding vector sequence corresponding to the third sample text semantic understanding data is generated. Similarly, the first basic prediction sub-model is used to perform logical semantic understanding on the first template content logical structure data, and the fourth sample text semantic understanding data is semantically encoded and represented, and a fourth sample semantic encoding vector sequence corresponding to the fourth sample text semantic understanding data is generated.

[0142] According to the third sample semantic coding vector sequence and the fourth sample semantic coding vector sequence, the model error information corresponding to the perturbation semantic learning subprocess is determined. This error information may be determined by comparing the difference between the two vector sequences, such as the degree of change of the semantic concepts represented by the vector elements before and after the perturbation, such as the difference in the quantitative representation of the relationship between work intensity and fertility intention in the two vector sequences, etc.

[0143] For example, in a possible implementation, the Y sample learning processes include a fuzzy labeling subprocess.

[0144] Step A130 includes:

[0145] A content logical structure analysis is performed on the first template survey text to generate first template content logical structure data of the first template survey text.

[0146] The first basic prediction sub-model is used to perform logical semantic understanding on the first template content logical structure data, generate fourth sample text semantic understanding data of the first template content logical structure data, perform semantic coding representation on the fourth sample text semantic understanding data, and generate a fourth sample semantic coding vector sequence corresponding to the fourth sample text semantic understanding data. The fourth sample semantic coding vector sequence belongs to the sample semantic coding vector sequence.

[0147] According to the prediction sub-model corresponding to the fuzzy labeling sub-process, user tendency prediction is performed on the fourth sample semantic coding vector sequence to generate a sample user tendency prediction score corresponding to the first template survey text.

[0148] A reference prediction model corresponding to the first basic prediction sub-model is obtained, and user tendency prediction is performed on the first template survey text according to the reference prediction model to generate a fuzzy user tendency prediction score corresponding to the first template survey text.

[0149] The model error information corresponding to the fuzzy labeling subprocess is determined according to the sample user propensity prediction score and the fuzzy user propensity prediction score.

[0150] For the first template survey text "Women of childbearing age working in hospitals face problems such as high work intensity and mismatch between salary and child-rearing costs, which affect their willingness to have children. At the same time, uncertain career prospects also make fertility decisions complicated", the server first performs content logical structure analysis on it and generates the first template content logical structure data of the first template survey text. Then, the first basic prediction sub-model is used to perform logical semantic understanding on the first template content logical structure data, generate the fourth sample text semantic understanding data of the first template content logical structure data, perform semantic encoding representation on the fourth sample text semantic understanding data, and generate the fourth sample semantic encoding vector sequence corresponding to the fourth sample text semantic understanding data.

[0151] According to the prediction sub-model corresponding to the fuzzy labeling sub-process, the user tendency is predicted for the fourth sample semantic encoding vector sequence to generate a sample user tendency prediction score corresponding to the first template survey text. For example, this score may reflect the degree of fertility intention of nurses of childbearing age based on the description of the current text as predicted by the model.

[0152] Obtain a reference prediction model corresponding to the first basic prediction sub-model, perform user tendency prediction on the first template survey text according to the reference prediction model, and generate a fuzzy user tendency prediction score corresponding to the first template survey text. This fuzzy user tendency prediction score may be a less accurate but reference-based score based on the reference prediction model. Then, based on the sample user tendency prediction score and the fuzzy user tendency prediction score, determine the model error information corresponding to the fuzzy annotation sub-process. This model error information may be determined by calculating the difference between the two scores or a comparison algorithm to evaluate the difference between the first basic prediction sub-model and the reference model in terms of user tendency prediction.

[0153] In a possible implementation, the semantic coding vector sequence is generated by semantically coding the text semantic understanding data by a text semantic encoder in a first prediction sub-model. The first prediction sub-model belongs to a target user tendency prediction model for predicting user tendency of the target survey user. The target user tendency prediction model also includes a second prediction sub-model connected after the text semantic encoder.

[0154] Step S130 includes:

[0155] Step S131, determining a second prediction sub-model from the target user tendency prediction model. The second prediction sub-model includes a multi-layer perceptron corresponding to each survey topic knowledge point.

[0156] Step S132, using the multi-layer perceptron corresponding to each survey subject knowledge point, predicting the user tendency of the semantic coding vector sequence, and generating the user tendency prediction score corresponding to each survey subject knowledge point. A multi-layer perceptron is used to determine the user tendency prediction score corresponding to a survey subject knowledge point.

[0157] In this embodiment, a second prediction sub-model is determined in the target user tendency prediction model. When analyzing the fertility intention and needs of nurses of childbearing age, this target user tendency prediction model aims to predict the nurses' tendency towards different survey subject knowledge points. The second prediction sub-model contains a multi-layer perceptron corresponding to each survey subject knowledge point. For example, the survey subject knowledge points in this scenario may include "the impact of work scheduling on fertility intention", "the impact of the economic cost of fertility on fertility intention", "the impact of fertility on career development", etc. Each such survey subject knowledge point has a corresponding multi-layer perceptron. These multi-layer perceptrons are specially constructed to process the semantic encoding vector sequences related to the respective survey subject knowledge points.

[0158] Then, the multi-layer perceptron corresponding to each survey topic knowledge point is used to predict the user tendency of the semantic coding vector sequence, and generate the user tendency prediction score corresponding to each survey topic knowledge point. Taking the survey topic knowledge point "the impact of work schedule on fertility intention" as an example, the corresponding multi-layer perceptron receives the semantic coding vector sequence as input. This semantic coding vector sequence is generated by the text semantic encoder in the first prediction sub-model to semantically encode the text semantic understanding data obtained previously. Assume that this semantic coding vector sequence contains semantic information about the working conditions of nurses of childbearing age (such as working hours, overtime frequency, etc.) and their thoughts on fertility. The multi-layer perceptron analyzes the relationship between these semantic information and "the impact of work schedule on fertility intention" through its internal neuron structure and weight calculation. If the elements in the semantic coding vector sequence indicate that the nurses' work schedule is very irregular (such as frequent night shifts, long continuous work, etc.), then after the calculation of the multi-layer perceptron, a lower user tendency prediction score may be generated, such as 30 points (out of 100 points), which means that the nurses have a lower willingness to have children due to work schedule problems. Similarly, for the multi-layer perceptron corresponding to the survey theme knowledge point "The impact of the economic cost of childbirth on the willingness to give birth", when the elements in the semantic encoding vector sequence reflect that the economic cost is very high (such as high childcare costs, low nurses' salaries, etc.), a lower user tendency prediction score, such as 25 points, may be obtained, indicating that nurses are not very willing to give birth due to the economic cost of childbirth. For the multi-layer perceptron corresponding to the survey theme knowledge point "The impact of childbirth on career development", if the information in the semantic encoding vector sequence shows that nurses face obstacles to career advancement after childbirth, etc., a user tendency prediction score of 20 points may be obtained, indicating that the negative impact of childbirth on career development leads to a low willingness of nurses to give birth.

[0159] In a possible implementation, the method further includes:

[0160] Step B110, obtaining a second template survey text data sequence. The second template survey text included in the second template survey text data sequence carries survey subject annotation data. The survey subject annotation data includes X survey subject identifiers corresponding to the X survey subject knowledge points, one survey subject knowledge point corresponds to one survey subject identifier, and the X survey subject identifiers are determined based on the survey subject knowledge points to which the second template survey text belongs.

[0161] Step B120, obtain the first prediction sub-model that has completed training, use the first prediction sub-model to perform text semantic understanding on the second template survey text, generate fifth sample text semantic understanding data of the second template survey text, perform semantic encoding representation on the fifth sample text semantic understanding data, and generate a fifth sample semantic encoding vector sequence corresponding to the fifth sample text semantic understanding data.

[0162] Step B130, obtaining X basic multilayer perceptrons corresponding to the X survey subject knowledge points, performing knowledge point prediction on the fifth sample semantic coding vector sequence according to the X basic multilayer perceptrons, and generating X knowledge point prediction confidences corresponding to the X basic multilayer perceptrons. One survey subject knowledge point corresponds to one basic multilayer perceptron, and one basic multilayer perceptron is used to determine one knowledge point prediction confidence.

[0163] Step B140, determining the second target model error information according to the X knowledge point prediction confidences and the X survey topic identifiers, training the X basic multilayer perceptrons according to the second target model error information, outputting the trained X basic multilayer perceptrons as the X multilayer perceptrons, and determining the second prediction sub-model according to the X multilayer perceptrons.

[0164] In this embodiment, for example, a second template survey text may be "As a nurse of childbearing age, the high cost of raising children makes me afraid to give birth easily, and the hospital does not have a financial subsidy policy for childbearing nurses." The survey topic annotation data of this text contains the survey topic identifier corresponding to the survey topic knowledge point "The impact of the economic cost of childbirth on the willingness to give birth." Another second template survey text may be "I am worried that I will miss the promotion opportunity in the hospital after giving birth, because I have seen that the promotion of colleagues who gave birth before has been affected." Its survey topic annotation data contains the survey topic identifier corresponding to the survey topic knowledge point "The impact of childbirth on career development." The server obtains the first prediction sub-model that has completed training. This first prediction sub-model has undergone the previous training process and can better understand the text semantics of the survey text.

[0165] Next, taking the second template survey text on the economic cost of childbirth mentioned above as an example, the first prediction sub-model will perform operations such as word segmentation, feature extraction, and construction of semantic logic units on the text to generate the fifth sample text semantic understanding data of the second template survey text. This data contains semantic information about the relationship between the economic cost of childbirth (such as childcare expenses, economic subsidy policies, etc.) and the willingness to have children. Then, the semantic encoding representation of the fifth sample text semantic understanding data is performed to generate the fifth sample semantic encoding vector sequence corresponding to the fifth sample text semantic understanding data. The elements in this vector sequence represent the quantitative representation of each semantic concept in the text (such as the level of childcare expenses, the presence or absence of economic subsidy policies, etc.) in the semantic space.

[0166] For the basic multi-layer perceptron corresponding to the survey theme knowledge point "The impact of the economic cost of childbirth on the willingness to have children" mentioned earlier, it is used to predict the knowledge point of the fifth sample semantic coding vector sequence. This basic multi-layer perceptron calculates a prediction result about this survey theme knowledge point based on the semantic information in the vector sequence, and generates the corresponding knowledge point prediction confidence. Assume that for this survey theme knowledge point about the economic cost of childbirth, the basic multi-layer perceptron calculates a lower prediction result of the willingness to have children based on the information such as high childcare costs and no economic subsidy policy in the fifth sample semantic coding vector sequence, and generates a knowledge point prediction confidence of 0.7, indicating the degree of confidence in this prediction result. Similarly, for the basic multi-layer perceptron corresponding to the survey theme knowledge point "The impact of childbirth on career development", based on the information such as the obstacles to promotion of colleagues after childbirth in the fifth sample semantic coding vector sequence, a prediction result of a lower willingness to have children is calculated, and a knowledge point prediction confidence of 0.8 is generated.

[0167] Assume that for the survey topic knowledge point "The impact of the economic cost of childbirth on the willingness to bear children", its survey topic identifier indicates that the text is indeed mainly around this topic, and the knowledge point prediction confidence is 0.7. If there is an ideal confidence (for example, 0.9 based on prior knowledge or other accurate models), then the gap between them can be calculated as part of the error information. Similar calculations are also performed for the survey topic knowledge point "The impact of childbirth on career development". Combining these error information, the error information of the second target model is determined. Then, based on the error information of the second target model, X basic multi-layer perceptrons are trained. By adjusting the weights, neuron connections and other parameters of the basic multi-layer perceptron, the error is reduced. The trained X basic multi-layer perceptrons are output as X multi-layer perceptrons, and the second prediction sub-model is determined based on these multi-layer perceptrons. In this way, the trained second prediction sub-model can more accurately predict the user tendency of different survey topic knowledge points based on the semantic encoding vector sequence, so as to better analyze the willingness and demand of nurses of childbearing age.

[0168] In a possible implementation, the text semantic understanding data is generated by the semantic logic unit in the first prediction sub-model performing logical semantic understanding on the content logical structure data. The first prediction sub-model belongs to a target user tendency prediction model for predicting the user tendency of the target survey user. The target user tendency prediction model also includes a third prediction sub-model connected after the semantic logic unit.

[0169] Step S140 includes:

[0170] Step S141, determining a third prediction sub-model from the target user tendency prediction model.

[0171] Step S142: Use the third prediction sub-model to perform attention analysis on the text semantic understanding data to generate an attention coefficient corresponding to each survey topic knowledge point.

[0172] In this embodiment, in the target user tendency prediction model, the server determines a third prediction sub-model. This third prediction sub-model is connected after the semantic logic unit in the first prediction sub-model, and is used to perform attention analysis on the text semantic understanding data, thereby generating an attention coefficient corresponding to each survey topic knowledge point. Taking the previously processed text related to the fertility intention of nurses of childbearing age as an example, the text semantic understanding data is generated by the semantic logic unit in the first prediction sub-model performing logical semantic understanding on the content logical structure data. These text semantic understanding data contain semantic information related to fertility intention, such as work pressure, economic cost, and career development.

[0173] The guidance unit in the third prediction sub-model first analyzes the guidance features of the text semantic understanding data. For survey subject knowledge points such as "the impact of work schedules on fertility intentions", the guidance unit will focus on the features related to work schedules in the text semantic understanding data. For example, the text mentions "I have to work three night shifts a week, and the frequent alternation of day and night shifts makes me very tired. It is difficult for me to consider having children under this working state." The guidance unit will identify "three night shifts a week" and "frequent alternation of day and night shifts" as guidance features related to work schedules. These guidance features are important information parts directly related to the survey subject knowledge point "the impact of work schedules on fertility intentions".

[0174] Then, the attention unit in the third prediction sub-model performs attention analysis on the guiding features. If the description of work schedules in the text semantic understanding data is very detailed and frequently appears, such as multiple mentions of the impact of night shifts on the body and the disruption of the rhythm of life, then for the survey theme knowledge point "The impact of work schedules on fertility intentions", the attention unit may generate a higher attention coefficient, assuming it is 0.8. This means that when analyzing the nurse's fertility intention, work schedules are a very critical factor and need to be focused on. Similarly, for the survey theme knowledge point "The impact of fertility economic costs on fertility intentions", the guiding unit will look for features related to economic costs in the text semantic understanding data, such as "Raising a child requires at least 3,000 yuan per month, but my salary is only 5,000 yuan, and I have to pay for mortgages and daily expenses" in "Monthly cost of raising a child", "salary level", "mortgage and daily expenses", etc. The attention unit may generate an attention coefficient of 0.7 based on the presentation of these features, indicating that this is also a relatively important factor. For the survey theme knowledge point "The impact of childbearing on career development", the guidance unit will focus on relevant features such as the promotion status of colleagues after childbearing, such as "I see that my colleagues who have given birth are ignored in the promotion assessment, and I am worried that I will face the same situation after giving birth." The attention unit may generate an attention coefficient of 0.6 for this survey theme knowledge point, indicating that this factor is also important in analyzing the willingness to have children.

[0175] In a possible implementation, the target user propensity prediction model also includes a second prediction sub-model for determining a user propensity prediction score corresponding to each survey topic knowledge point.

[0176] The method further comprises:

[0177] Step C110, obtaining a third template survey text data sequence. The third template survey text included in the third template survey text data sequence carries training supervision data.

[0178] Step C120, obtain the first prediction sub-model that has completed training, use the first prediction sub-model to perform text semantic understanding on the third template survey text, generate sixth sample text semantic understanding data of the third template survey text, perform semantic encoding representation on the sixth sample text semantic understanding data, and generate a sixth sample semantic encoding vector sequence corresponding to the sixth sample text semantic understanding data.

[0179] Step C130, obtaining the trained second prediction sub-model, using the X multi-layer perceptrons in the second prediction sub-model to perform user tendency prediction on the sixth sample semantic encoding vector sequence, and generating a sample user tendency prediction score corresponding to each survey theme knowledge point. A multi-layer perceptron is used to determine a sample user tendency prediction score corresponding to a survey theme knowledge point.

[0180] Step C140, using the third basic prediction sub-model to perform attention analysis on the sixth sample text semantic understanding data, and generate a sample attention coefficient corresponding to each survey topic knowledge point.

[0181] Step C150, determining the sample survey topic tendency distribution of the template survey user corresponding to the third template survey text based on the sample user tendency prediction score corresponding to each survey topic knowledge point and the sample attention coefficient corresponding to each survey topic knowledge point.

[0182] Step C160, determine the third target model error information based on the training supervision data and the sample survey topic tendency distribution, train the third basic prediction sub-model based on the third target model error information, output the trained third basic prediction sub-model as the third prediction sub-model, and determine the target user tendency prediction model based on the first prediction sub-model, the second prediction sub-model and the third prediction sub-model.

[0183] Wherein, the third prediction sub-model includes a guidance unit and an attention unit.

[0184] Step S142 includes:

[0185] Step S1421: Utilize the guidance unit to perform guidance feature analysis on the text semantic understanding data to generate guidance features corresponding to the text semantic understanding data.

[0186] Step S1422, using the attention unit to perform attention analysis on the guiding feature to generate an attention coefficient corresponding to each survey topic knowledge point.

[0187] For example, a third template survey text may be "I am a nurse of childbearing age. The work intensity in the hospital is high and I often have to work overtime. This makes me worry about whether I can balance family and work after giving birth. The economic cost of childbirth is very high, and my current salary is difficult to afford. In addition, I have seen that nurses after giving birth are restricted in their career development, which also affects my willingness to have children." The training supervision data in this text may include the predetermined distribution of the nurse's actual fertility intention tendency, such as the work scheduling factor accounting for 0.3 influence weight, the economic cost factor accounting for 0.4, and the career development factor accounting for 0.3.

[0188] Next, the trained first prediction sub-model is obtained, and the text semantics of the third template survey text is understood using the first prediction sub-model to generate the sixth sample text semantic understanding data of the third template survey text. This process is similar to the previous semantic understanding of other texts. By performing operations such as word segmentation, feature extraction, and construction of semantic logic units on the text, the sixth sample text semantic understanding data containing semantic information related to fertility intentions such as work intensity, overtime, economic costs, and career development is obtained. Then, the sixth sample text semantic understanding data is semantically encoded to generate the sixth sample semantic encoding vector sequence corresponding to the sixth sample text semantic understanding data.

[0189] Then, the server obtains the trained second prediction sub-model, uses the X multi-layer perceptrons in the second prediction sub-model to predict the user tendency of the sixth sample semantic coding vector sequence, and generates a sample user tendency prediction score corresponding to each survey theme knowledge point. For the multi-layer perceptron corresponding to the survey theme knowledge point "the impact of work schedule on fertility intention", according to the information such as high work intensity and frequent overtime in the sixth sample semantic coding vector sequence, a lower sample user tendency prediction score may be obtained, assuming 30 points. For the multi-layer perceptron corresponding to the survey theme knowledge point "the impact of fertility economic cost on fertility intention", according to the information of low salary and high fertility economic cost in the vector sequence, a sample user tendency prediction score of 25 points may be obtained. For the multi-layer perceptron corresponding to the survey theme knowledge point "the impact of fertility on career development", according to the information such as limited career development of nurses after childbirth, a sample user tendency prediction score of 20 points may be obtained.

[0190] At the same time, the third basic prediction sub-model is used to perform attention analysis on the semantic understanding data of the sixth sample text to generate the sample attention coefficient corresponding to each survey theme knowledge point. For example, for the survey theme knowledge point "the impact of work schedules on fertility intentions", the sample attention coefficient is 0.7 through analysis; for "the impact of the economic cost of fertility on fertility intentions", it is 0.8; for "the impact of fertility on career development", it is 0.6.

[0191] According to the sample user tendency prediction score and sample attention coefficient corresponding to each survey topic knowledge point, the sample survey topic tendency distribution of the template survey user corresponding to the third template survey text is determined. For "the impact of work schedule on fertility intention", the influence degree after fusion is calculated to be 30×0.7 = 21; for "the impact of economic cost of fertility on fertility intention", it is 25×0.8 = 20; for "the impact of fertility on career development", it is 20×0.6 = 12. Based on these results, the sample survey topic tendency distribution can be determined, for example, the work schedule factor accounts for 21 / (21+20+12)≈0.35, the economic cost factor accounts for 20 / (21+20+12)≈0.33, and the career development factor accounts for 12 / (21+20+12)≈0.20 (here, in order to simplify the calculation, no precise normalization is performed).

[0192] The error information of the third target model is determined based on the training supervision data (such as the actual distribution of fertility intentions mentioned above) and the distribution of sample survey subject tendencies. For example, if the actual work schedule factor influence weight is 0.3, and the calculated sample result is 0.35, then the difference between the two is part of the error information. The error information of the third target model is determined by combining the error information of each survey subject knowledge point.

[0193] According to the error information of the third target model, the third basic prediction sub-model is trained. The error is reduced by adjusting the parameters of the third basic prediction sub-model, such as the weights in the guide unit and the attention unit. The trained third basic prediction sub-model is output as the third prediction sub-model. Finally, the target user tendency prediction model is determined based on the first prediction sub-model, the second prediction sub-model and the third prediction sub-model. This target user tendency prediction model can more accurately process the survey text of nurses of childbearing age, including user tendency prediction and attention analysis, so as to better analyze the fertility intention and needs of nurses of childbearing age.

[0194] For example, in one possible implementation, the X survey subject knowledge points include target survey subject knowledge points.

[0195] The method further comprises:

[0196] Step D110 , obtaining a target user tendency prediction model for predicting the user tendency of a target survey user, and obtaining a basic knowledge imitation model for performing knowledge imitation training from the target user tendency prediction model.

[0197] Step D120, obtaining a fourth template survey text data sequence corresponding to the target survey subject knowledge point. The fourth template survey text data sequence includes a fourth template survey text.

[0198] Step D130, using the target user tendency prediction model to perform text semantic understanding on the fourth template survey text, generate a survey topic tendency distribution corresponding to the fourth template survey text, and output the survey topic tendency distribution corresponding to the fourth template survey text as fuzzy training supervision data for training the basic knowledge imitation model.

[0199] Step D140, using the basic knowledge imitation model to perform text semantic understanding on the fourth template survey text, and generating a reference survey topic tendency distribution corresponding to the fourth template survey text.

[0200] Step D150, training the basic knowledge emulation model according to the reference survey topic tendency distribution and the fuzzy training supervision data, and outputting the trained basic knowledge emulation model as the target knowledge emulation model of the target user tendency prediction model. The target knowledge emulation model is used to predict the user tendency of the survey text big data under the target survey topic knowledge point.

[0201] For example, in a possible implementation, the fourth template survey text includes a theme template survey text generated under the target survey theme knowledge point.

[0202] Step D120 includes:

[0203] Step D121, obtaining the subject template survey text under the target survey subject knowledge point.

[0204] Step D122, determining a candidate template survey text from the template survey text sequence.

[0205] Step D123 , performing semantic relevance analysis on the candidate template survey text and the subject template survey text, and generating a semantic relevance between the subject template survey text and the candidate template survey text.

[0206] Step D124, if the semantic association meets the semantic derivation requirement, the candidate template survey text is output as a derived template survey text corresponding to the target survey subject knowledge point.

[0207] Step D125, outputting the derived template survey text and the subject template survey text as a fourth template survey text corresponding to the target survey subject knowledge point.

[0208] In this embodiment, the server obtains a target user tendency prediction model for predicting the user tendency of the target survey user (nurses of childbearing age), and simultaneously obtains a basic knowledge imitation model for knowledge imitation training from the target user tendency prediction model. Assume that the target survey subject knowledge point is "the impact of post-childbirth career development support on childbearing intention". Then the server obtains the fourth template survey text data sequence corresponding to this target survey subject knowledge point, and the fourth template survey text in this sequence is the text related to the subject knowledge point.

[0209] In detail, first obtain the subject template survey text under the target survey subject knowledge point. For example, the subject template survey text may be "As a nurse of childbearing age, I am very concerned about whether the hospital has a career development support plan for nurses after childbirth, such as special training courses, fair promotion opportunities, etc. Without these supports, my willingness to have children will be very low." Then determine the candidate template survey text from the template survey text sequence. These candidate template survey texts may be some general texts related to the career development of nurses, such as "Working in a hospital, career development is the focus of nurses, and training and promotion opportunities at different stages affect the work enthusiasm of nurses."

[0210] Then, the candidate template survey text and the theme template survey text are subjected to semantic relevance analysis. For the candidate template survey text and theme template survey text mentioned above, the semantic relevance analysis will consider factors such as vocabulary overlap and the relevance of semantic concepts. For example, both mentioned words such as "nurse" and "career development", and both involved the semantic concept of the impact of career development on nurses. If the semantic relevance meets the semantic derivation requirements (this requirement may be based on a predefined threshold or rule, such as the vocabulary overlap reaches a certain proportion, the core part of the semantic concept is the same, etc.), the candidate template survey text is output as a derivative template survey text corresponding to the target survey theme knowledge point. In this example, if the requirements are met, the candidate template survey text mentioned above becomes a derivative template survey text. Then the derivative template survey text and the theme template survey text are output as a fourth template survey text corresponding to the target survey theme knowledge point.

[0211] Next, the target user tendency prediction model is used to perform text semantic understanding on the fourth template survey text. For the information in the subject template survey text, the target user tendency prediction model will perform a series of operations, such as word segmentation, feature extraction, and construction of semantic logic units. Taking the subject template survey text as an example, the model will identify semantic logic units such as "post-childbirth career development support (such as training courses, promotion opportunities)" and "childbearing intention", and analyze the relationship between them. Similar operations are performed on the derivative template survey text, and then the survey topic tendency distribution corresponding to the fourth template survey text is generated by combining this information. For example, based on the information in the subject template survey text, it may be concluded that if there is no career development support, the childbearing intention is very low. In the survey topic tendency distribution, the weight of the factor "the impact of post-childbearing career development support on childbearing intention" may be very high, such as 0.8. This survey topic tendency distribution is output as fuzzy training supervision data for training the basic knowledge imitation model.

[0212] Then, the basic knowledge imitation model is used to perform text semantic understanding on the fourth template survey text. The basic knowledge imitation model also performs operations such as word segmentation and feature extraction on the fourth template survey text, but because its initial structure and parameters may be different from the target user tendency prediction model, the reference survey topic tendency distribution corresponding to the fourth template survey text generated by it may be different from the fuzzy training supervision data. For example, it may be concluded that the weight of the factor "the impact of post-childbirth career development support on fertility intention" is 0.6. Then the basic knowledge imitation model is trained based on the reference survey topic tendency distribution and fuzzy training supervision data. If the weight of "the impact of post-childbirth career development support on fertility intention" in the fuzzy training supervision data is 0.8, while it is 0.6 in the reference survey topic tendency distribution, then the parameters of the basic knowledge imitation model are adjusted according to this difference, such as adjusting the weights related to semantic understanding and tendency calculation. The trained basic knowledge imitation model is output as the target knowledge imitation model of the target user tendency prediction model, and this target knowledge imitation model can be used to make more accurate user tendency predictions for the survey text big data under the target survey topic knowledge point (the impact of post-childbirth career development support on fertility intention). For example, when processing a large number of survey texts from nurses of childbearing age regarding the relationship between post-childbirth career development support and fertility intentions, the target knowledge emulation model can better analyze the nurses' attitude tendencies, thereby providing a basis for hospitals to formulate relevant policies (such as career development support plans).

[0213] This knowledge imitation training is of great significance in the whole analysis system of nurses' fertility intention and needs during childbearing age. By performing the above operations on the fourth template survey text related to the target survey subject knowledge point, the prediction ability of the model can be continuously optimized. The target knowledge imitation model can better adapt to the text analysis needs under the specific survey subject knowledge point. For different survey subject knowledge points, such as "the impact of the economic cost of childbearing on the willingness to bear children" and "the impact of work scheduling on the willingness to bear children", the model can be optimized through a similar process. In this way, the target user tendency prediction model and its related target knowledge imitation model can analyze the willingness and needs of nurses of childbearing age more comprehensively and accurately. In practical applications, this helps hospital managers to deeply understand the ideas of nurses and formulate reasonable human resources policies, fertility support policies, etc., to improve the job satisfaction and quality of life of nurses, and also helps to ensure the stability and development of the hospital's nursing team. For example, if it is found that post-childbirth career development support has a great impact on nurses' willingness to bear children, the hospital can formulate a special training and promotion plan to provide more development opportunities for nurses after childbirth, thereby improving nurses' willingness to bear children and solving potential problems such as nursing staff shortage.

[0214] For example, in a possible implementation, step S150 includes:

[0215] Step S151, using the attention coefficient corresponding to each survey theme knowledge point, the user tendency prediction score corresponding to each survey theme knowledge point is fused and calculated to generate a fused user tendency prediction score corresponding to each survey theme knowledge point.

[0216] Step S152, determining the survey topic tendency distribution of the target survey user according to the fused user tendency prediction score corresponding to each survey topic knowledge point.

[0217] In this embodiment, the previously mentioned survey topic knowledge points related to the fertility intention of nurses of childbearing age are taken as an example, such as "the impact of work schedules on fertility intentions", "the impact of the economic cost of fertility on fertility intentions", "the impact of fertility on career development", etc.

[0218] For the survey topic knowledge point "The impact of work schedule on fertility intention", it is assumed that the corresponding user tendency prediction score is 30 points (out of 100 points), which means that based on the previous analysis, the nurse's willingness to have children is at a low level of inclination due to the status of work schedule. For example, the nurse's work schedule may be multiple night shifts per week, and the day and night shifts alternate frequently. This scheduling method causes her physical fatigue and disrupts her rhythm of life, which affects her willingness to have children, so a score of 30 points is obtained in the user tendency prediction model. The corresponding attention coefficient is 0.8, which means that among the many factors in analyzing the nurse's willingness to have children, the factor of work schedule is relatively important and needs to be given a higher degree of attention.

[0219] The server uses this attention coefficient to perform a fusion calculation on the user's tendency prediction score, that is, to calculate the fusion user tendency prediction score. The specific calculation is: 30×0.8 = 24 points. This 24 points is the fusion user tendency prediction score corresponding to the survey topic knowledge point "The impact of work schedules on fertility intentions". This score comprehensively considers the user's tendency towards this factor and the importance of this factor in the overall analysis.

[0220] For the survey topic knowledge point "The impact of the economic cost of childbirth on the willingness to have children", assume that its user tendency prediction score is 25 points. This may be because the salary level of nurses is limited, and the economic cost of raising a child is very high, including the child's living expenses, education expenses, etc., which makes her have greater financial pressure, thus affecting her willingness to have children. Its corresponding attention coefficient is 0.7, indicating that the economic cost of childbirth is also a relatively important factor in analyzing the willingness to have children. Through fusion calculation, 25×0.7 = 17.5 points, which is the fusion user tendency prediction score corresponding to the survey topic knowledge point.

[0221] For the survey topic knowledge point "The impact of childbirth on career development", assume that the user tendency prediction score is 20 points. This is because in the hospital environment, nurses who have given birth may be restricted in career advancement and career training opportunities, which leads to nurses worrying that childbirth will have a negative impact on their career development, so their willingness to give birth is low. The corresponding attention coefficient is 0.6, and the fusion calculation results in 20×0.6 = 12 points, which is the fusion user tendency prediction score corresponding to the survey topic knowledge point.

[0222] Based on the fused user tendency prediction scores corresponding to each survey topic knowledge point, the survey topic tendency distribution of the target survey user (the childbearing-age nurse) is determined. First, the sum of the fused user tendency prediction scores of all survey topic knowledge points is calculated, that is, 24 + 17.5 + 12 = 53.5 points (this is just an example calculation, and there may be other survey topic knowledge points in actual situations).

[0223] Then, the proportion of each survey topic knowledge point in the total is calculated to determine the survey topic tendency distribution. For "the impact of work schedules on fertility intentions", its proportion in the survey topic tendency distribution is 24 / 53.5≈0.45 (two decimal places are retained here); for "the impact of the economic cost of fertility on fertility intentions", the proportion is 17.5 / 53.5≈0.33; for "the impact of fertility on career development", the proportion is 12 / 53.5≈0.22.

[0224] In this way, the survey topic tendency distribution of the target survey users (nurses of childbearing age) can be determined. This survey topic tendency distribution can clearly reflect the relative role of different survey topic knowledge points in the nurse's fertility intention. Hospital managers or relevant researchers can formulate targeted measures or conduct further research based on this survey topic tendency distribution. For example, if the proportion of "the impact of work scheduling on fertility intention" is high, the hospital can consider adjusting the nurse scheduling system to improve the nurse's fertility intention; if the proportion of "the impact of the economic cost of childbearing on fertility intention" is high, the hospital can consider providing some fertility subsidies or welfare measures to alleviate the economic pressure of nurses; if the proportion of "the impact of childbearing on career development" is high, the hospital can formulate some policies to ensure the career development of nurses after childbearing, such as providing special vocational training opportunities or fair promotion mechanisms, etc., which will help to deeply understand the fertility intention and needs of nurses of childbearing age, so as to adopt effective response strategies to ensure the stability and development of the nursing team, and also help to improve the quality of life and job satisfaction of nurses.

[0225] Figure 2 The population fertility survey service system 100 shown includes: a processor 1001 and a memory 1003. The processor 1001 and the memory 1003 are connected, such as through a bus 1002. Optionally, the population fertility survey service system 100 may also include a transceiver 1004, which may be used for data interaction between the server and other servers, such as data transmission and / or data reception. It should be noted that in actual scheduling, the transceiver 1004 is not limited to one, and the structure of the population fertility survey service system 100 does not constitute a limitation on the embodiments of the present application.

[0226] 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.

[0227] 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 the convenience of diagram coding features, Figure 2 The feature is encoded using only one thick line, but it does not encode the feature of only one bus or one type of bus.

[0228] 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.

[0229] 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.

[0230] 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 information processing method based on big data mining is implemented.

[0231] 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. An information processing method based on big data mining, characterized in that: The method comprises: Performing text semantic understanding on the survey text big data of the target survey user to generate text semantic understanding data corresponding to the survey text big data; Performing semantic coding representation on the text semantic understanding data to generate a semantic coding vector sequence corresponding to the text semantic understanding data; Determine X survey subject knowledge points corresponding to the survey text big data, perform user tendency prediction on the semantic coding vector sequence according to each survey subject knowledge point, and generate a user tendency prediction score corresponding to each survey subject knowledge point; X is a positive integer; Performing attention analysis on the text semantic understanding data to generate an attention coefficient corresponding to each survey topic knowledge point; Determine the survey topic tendency distribution of the target survey user according to the user tendency prediction score corresponding to each survey topic knowledge point and the attention coefficient corresponding to each survey topic knowledge point; The performing text semantic understanding on the survey text big data of the target survey user to generate text semantic understanding data corresponding to the survey text big data includes: Performing content logical structure analysis on the survey text big data to generate content logical structure data of the survey text big data; Performing logical semantic understanding on the content logical structure data to generate text semantic understanding data of the content logical structure data; The step of performing content logical structure analysis on the survey text big data to generate content logical structure data of the survey text big data includes: When the survey text big data of the target survey user is obtained, a target user tendency prediction model for predicting the user tendency of the target survey user is obtained; the target user tendency prediction model includes a first prediction sub-model; Determine the semantic logic unit corresponding to the survey text big data by using the first prediction sub-model, extract the semantic logic network data of the survey text big data according to the semantic logic unit, perform knowledge extraction on the semantic logic network data, and generate semantic logic data corresponding to the semantic logic network data; The first prediction sub-model is used to perform structural processing on the semantic logic data corresponding to the semantic logic network data to generate content logic structure data of the survey text big data.

2. The information processing method based on big data mining according to claim 1 is characterized in that: The method of using the first prediction sub-model to determine the semantic logic unit corresponding to the survey text big data, extracting the semantic logic network data of the survey text big data according to the semantic logic unit, and performing knowledge extraction on the semantic logic network data to generate semantic logic data corresponding to the semantic logic network data includes: After performing a word segmentation operation on the survey text big data, performing feature extraction on each word segmented vocabulary unit to obtain preliminary text features, and inputting the preliminary text features into the first prediction sub-model to convert each vocabulary unit into a vector representation in a semantic space; Performing a preliminary semantic analysis on the survey text big data by using the first prediction sub-model to obtain a preliminary semantic analysis result, wherein the preliminary semantic analysis includes analyzing the semantic similarity between vocabulary units and the semantic roles of the unit vocabulary in the sentence, and the preliminary semantic analysis result represents the semantic relationship between the vocabulary units in the form of a semantic graph, or represents the semantic role of the vocabulary unit in the form of a tag; According to the preliminary semantic analysis results, a candidate set of semantic logic units is constructed, wherein the candidate set includes a plurality of candidate elements, wherein the candidate elements are grouped vocabulary units, sentences or paragraph combinations, and each candidate element has corresponding semantic features and potential semantic relationships with other candidate elements; Performing semantic analysis on each candidate element in the candidate set, and updating semantic information of each candidate element according to the semantic analysis result, wherein, for candidate elements composed of vocabulary units, analyzing the polysemy of the vocabulary units, and for candidate elements composed of sentences or paragraphs, analyzing the influence of nested structures and rhetoric devices in sentences on semantics, as well as thematic coherence and logical development in paragraphs; For a candidate element in the candidate set, analyzing the semantic coherence between the candidate element and other candidate elements, and retaining the candidate elements that meet the preset semantic coherence requirements to determine them as semantic logical units corresponding to the survey text big data, specifically including analyzing the expression of semantic coherence between vocabulary units in sentences and paragraphs, and the expression of semantic coherence between sentences and paragraphs in the survey text big data; According to the screened semantic logic units, relocate the logical position and context information of the semantic logic units in the survey text big data, and construct a semantic logic network with each semantic logic unit as a node; In the semantic logic network, for each semantic logic unit, the connection relationship between the semantic logic unit and other semantic logic units is determined, and the connection relationship is determined based on the grammatical structure, semantic reference and logical relationship in the survey text big data, so as to obtain a semantic logic network data including the semantic logic units and the connection relationship between the semantic logic units; Performing basic knowledge analysis on each semantic logic unit in the semantic logic network data to obtain basic knowledge analysis results, wherein for the semantic logic unit composed of words, analyzing the semantic categories and semantic hierarchical structures of the words; for the semantic logic unit composed of sentences and paragraphs, analyzing the basic semantic propositions; The basic knowledge analysis result is represented in the semantic logic network data in a structured form, wherein the semantic category of the vocabulary is marked with a predefined semantic category label; and the semantic proposition is represented in the form of a triple; Based on the basic knowledge analysis results, deep knowledge extraction is performed on the semantic logic network data to obtain deep knowledge extraction results, wherein the connection relationship between semantic logic units is analyzed, the logical deduction relationship between cause and effect, the strength of the causal relationship and the existence of intermediate factors, and the common characteristics and differences between parallel elements and the role of parallel elements in global semantics are analyzed; The deep knowledge extraction result and the basic knowledge analysis result are integrated to generate semantic logic data corresponding to the semantic logic network data.

3. The information processing method based on big data mining according to claim 1 is characterized in that: The semantic coding vector sequence is generated by semantic coding representation of the text semantic understanding data by a text semantic encoder in a first prediction sub-model; the first prediction sub-model belongs to a target user tendency prediction model for predicting user tendency of the target survey user; The method further comprises: Acquire a first template survey text data sequence; the first template survey text included in the first template survey text data sequence is a template survey text that does not carry training supervision data; Obtain Y sample learning processes for the first basic prediction sub-model; Y is a positive integer; Using the first basic prediction sub-model to perform text semantic understanding on the first template survey text, generating a sample semantic coding vector sequence corresponding to each sample learning process, and determining Y model error information corresponding to the Y sample learning processes according to the sample semantic coding vector sequence; one sample learning process corresponds to one model error information; Determine first target model error information based on the Y model error information, train the first basic prediction sub-model based on the first target model error information, and output the trained first basic prediction sub-model as the first prediction sub-model.

4. The information processing method based on big data mining according to claim 1 is characterized in that: The semantic coding vector sequence is generated by semantic coding representation of the text semantic understanding data by the text semantic encoder in the first prediction sub-model; the first prediction sub-model belongs to a target user tendency prediction model for predicting the user tendency of the target survey user; the target user tendency prediction model also includes a second prediction sub-model connected to the text semantic encoder; The step of performing user tendency prediction on the semantic coding vector sequence according to each survey subject knowledge point to generate a user tendency prediction score corresponding to each survey subject knowledge point includes: Determining a second prediction sub-model from the target user tendency prediction model; the second prediction sub-model includes a multi-layer perceptron corresponding to each survey topic knowledge point; The multi-layer perceptron corresponding to each survey topic knowledge point is used to predict the user tendency of the semantic coding vector sequence, and generate the user tendency prediction score corresponding to each survey topic knowledge point; a multi-layer perceptron is used to determine the user tendency prediction score corresponding to a survey topic knowledge point.

5. The information processing method based on big data mining according to claim 4 is characterized in that: The method further comprises: Acquire a second template survey text data sequence; the second template survey text included in the second template survey text data sequence carries survey subject annotation data; the survey subject annotation data includes X survey subject identifiers corresponding to the X survey subject knowledge points, one survey subject knowledge point corresponds to one survey subject identifier, and the X survey subject identifiers are determined according to the survey subject knowledge points to which the second template survey text belongs; Acquire the trained first prediction sub-model, use the first prediction sub-model to perform text semantic understanding on the second template survey text, generate fifth sample text semantic understanding data of the second template survey text, perform semantic encoding representation on the fifth sample text semantic understanding data, and generate a fifth sample semantic encoding vector sequence corresponding to the fifth sample text semantic understanding data; Obtaining X basic multi-layer perceptrons corresponding to the X survey subject knowledge points, performing knowledge point prediction on the fifth sample semantic encoding vector sequence according to the X basic multi-layer perceptrons, and generating X knowledge point prediction confidences corresponding to the X basic multi-layer perceptrons; one survey subject knowledge point corresponds to one basic multi-layer perceptron, and one basic multi-layer perceptron is used to determine one knowledge point prediction confidence; Determine the second target model error information according to the X knowledge point prediction confidences and the X survey topic identifiers, train the X basic multilayer perceptrons according to the second target model error information, output the trained X basic multilayer perceptrons as X multilayer perceptrons, and determine the second prediction sub-model according to the X multilayer perceptrons.

6. The information processing method based on big data mining according to claim 1 is characterized in that: The text semantic understanding data is generated by the semantic logic unit in the first prediction sub-model performing logical semantic understanding on the content logical structure data; the first prediction sub-model belongs to a target user tendency prediction model for predicting the user tendency of the target survey user; the target user tendency prediction model also includes a third prediction sub-model connected to the semantic logic unit; The performing attention analysis on the text semantic understanding data to generate the attention coefficient corresponding to each survey topic knowledge point includes: Determining a third prediction sub-model from the target user tendency prediction model; The third prediction sub-model is used to perform attention analysis on the text semantic understanding data to generate an attention coefficient corresponding to each survey topic knowledge point.

7. The information processing method based on big data mining according to claim 6 is characterized in that: The target user tendency prediction model also includes a second prediction sub-model for determining the user tendency prediction score corresponding to each survey topic knowledge point; The method further comprises: Acquire a third template survey text data sequence; the third template survey text included in the third template survey text data sequence carries training supervision data; Acquire a first prediction sub-model that has been trained, use the first prediction sub-model to perform text semantic understanding on the third template survey text, generate sixth sample text semantic understanding data of the third template survey text, perform semantic encoding representation on the sixth sample text semantic understanding data, and generate a sixth sample semantic encoding vector sequence corresponding to the sixth sample text semantic understanding data; Acquire the trained second prediction sub-model, use X multi-layer perceptrons in the second prediction sub-model to perform user tendency prediction on the sixth sample semantic encoding vector sequence, and generate a sample user tendency prediction score corresponding to each survey subject knowledge point; one multi-layer perceptron is used to determine the sample user tendency prediction score corresponding to one survey subject knowledge point; Using the third basic prediction sub-model to perform attention analysis on the sixth sample text semantic understanding data, and generate a sample attention coefficient corresponding to each survey topic knowledge point; Determine the sample survey topic tendency distribution of the template survey user corresponding to the third template survey text according to the sample user tendency prediction score corresponding to each survey topic knowledge point and the sample attention coefficient corresponding to each survey topic knowledge point; Determine the third target model error information according to the training supervision data and the sample survey topic tendency distribution, train the third basic prediction sub-model according to the third target model error information, output the trained third basic prediction sub-model as a third prediction sub-model, and determine the target user tendency prediction model according to the first prediction sub-model, the second prediction sub-model and the third prediction sub-model; Wherein, the third prediction sub-model includes a guidance unit and an attention unit; The using the third prediction sub-model to perform attention analysis on the text semantic understanding data to generate an attention coefficient corresponding to each survey topic knowledge point includes: Using the guidance unit to perform guidance feature analysis on the text semantic understanding data to generate guidance features corresponding to the text semantic understanding data; Using the attention unit to perform attention analysis on the guidance feature to generate an attention coefficient corresponding to each survey topic knowledge point; The tropism prediction score determines the survey topic tendency distribution of the target survey user.

8. A population fertility survey service system, characterized in that: The population fertility survey service system 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 information processing method based on big data mining as described in any one of claims 1-7.

Images