Public health survey text analysis method and system based on natural language processing

Abstract

The present application relates to the field of text processing, and specifically provides a method and system for analyzing public health research texts based on natural language processing. The method includes: shielding style text paragraphs in public health research texts to obtain style shielded texts; extracting core information from research subject content in public health research texts to obtain research core information paragraphs; style text paragraphs include research content style text paragraphs in research subject content; replacing style text paragraphs in public health research texts with placeholders to obtain paragraph placeholder sequences; and reconstructing style texts on style shielded texts based on target style texts, research core information paragraphs, and paragraph placeholder sequences to obtain style-integrated research texts that incorporate the text style contained in the target style text. The present application can improve the accuracy of text style replacement.

Description

Technical Field

[0001] The present application relates to the field of text processing technology, and in particular to a public health research text analysis method and system based on natural language processing. Background Art

[0002] In the field of public health, research work generates a large amount of text data, which contains rich information, such as disease prevalence, effectiveness of prevention and control measures, allocation of public health resources, etc. Effective analysis of these public health research texts is an important basis for public health management decision-making, scientific research and policy formulation.

[0003] Natural language processing technology has made significant progress in the field of text analysis, including text classification, information extraction, sentiment analysis and other technologies are widely used. In the analysis of public health research texts, existing technologies still have some limitations. When processing public health research texts, an important requirement is to adjust the text style. Different audiences or application scenarios may have different styles of text expression. For example, in scientific research reports, the text style is rigorous and formal, while in health education propaganda for the general public, the style may be easy to understand and vivid.

[0004] Current natural language processing technology has shortcomings in text style conversion. Existing style conversion methods often lack an accurate understanding of the content of texts in specific fields (such as public health), and it is difficult to ensure the integrity and accuracy of the content while converting the style. For example, simply replacing vocabulary or sentence patterns can easily lead to semantic distortion or loss of important information. In addition, in public health research texts, the research topic content is the core part, and existing style conversion technologies often do not fully consider how to ensure that the research topic content is not changed or missing during the style conversion process. For example, when converting a more colloquial public health research text into a formal scientific research report style, existing technologies may ignore key research data or conclusions and other core information because of excessive focus on style conversion. Moreover, there is a lack of effective distinction mechanism for the processing of style text paragraphs and non-style text paragraphs in public health research texts. When performing style conversion, existing technologies find it difficult to accurately convert style text paragraphs without affecting non-style text paragraphs, resulting in problems in the coherence of style and content of the converted text. Summary of the invention

[0005] In view of this, the present application provides a method and system for analyzing public health research text based on natural language processing. The technical solution of the present application is implemented as follows:

[0006] In a first aspect, the present application provides a public health survey text analysis method based on natural language processing, the method comprising: shielding style text paragraphs in a public health survey text to obtain style shielded text; extracting core information from a survey subject content in the public health survey text to obtain a survey core information paragraph; the style text paragraphs include survey content style text paragraphs in the survey subject content; replacing style text paragraphs in the public health survey text with placeholders to obtain a paragraph placeholder sequence; and reconstructing style text on the style shielded text according to a target style text, the survey core information paragraph and the paragraph placeholder sequence to obtain a style-integrated survey text that incorporates the text style contained in the target style text.

[0007] In one embodiment, the style-shielded text is reconstructed according to the target style text, the survey core information paragraph and the paragraph placeholder sequence to obtain the style-integrated survey text that incorporates the text style contained in the target style text, including: embedding and mapping the target style text, the style-shielded text and the survey core information paragraph respectively to obtain a first text implicit representation of the target style text, a second text implicit representation of the style-shielded text and a third text implicit representation of the survey core information paragraph; based on the first text implicit representation, the second text implicit representation, the third text implicit representation and the paragraph placeholder sequence, feature generation is performed step by step on a random placeholder sequence to obtain a style-integrated text implicit representation; and restoration mapping is performed on the style-integrated text implicit representation to obtain a style-integrated survey text that incorporates the text style contained in the target style text.

[0008] In one embodiment, based on the first text implicit representation, the second text implicit representation, the third text implicit representation and the paragraph placeholder sequence, feature generation is gradually performed on a random placeholder sequence to obtain a style-integrated text implicit representation, including: implicit representation encoding is performed on the first text implicit representation based on a first machine learning model to obtain an implicit representation encoding result; using the second text implicit representation, the third text implicit representation and the paragraph placeholder sequence as feature generation constraints, based on a second machine learning model, feature generation is gradually performed on the random placeholder sequence according to the implicit representation encoding result to obtain the style-integrated text implicit representation.

[0009] In one embodiment, the first machine learning model includes a t-layer first weighted focusing network, and the second machine learning model includes a t-layer second weighted focusing network; the implicit representation encoding result includes the first focused implicit representation of each first weighted focusing network input, wherein t≥2; the second text implicit representation, the third text implicit representation and the paragraph placeholder sequence are used as feature generation constraints, and based on the second machine learning model, the random placeholder sequence is gradually generated with features according to the implicit representation encoding result to obtain the style-integrated text implicit representation, including: using the second text implicit representation, the third text implicit representation and the paragraph placeholder sequence as feature generation constraints, based on the second A machine learning model, which performs x feature generation on the random placeholder sequence in sequence according to the implicit representation encoding result, and integrates the output text implicit representation obtained by the x-th feature generation into the text implicit representation as the style; x≥2; for the s-th feature generation, it includes the following operation process: obtaining the input text implicit representation of the s-th feature generation, wherein the input text implicit representation of the first feature generation is obtained by integrating the second text implicit representation, the third text implicit representation, the paragraph placeholder sequence and the random text implicit representation corresponding to the random placeholder sequence; if s is a positive integer greater than 1 and not greater than x, the input text implicit representation of the s-th feature generation is the output text implicit representation of the s-1-th feature generation. Formula representation; in the r-th layer second weight focusing network, weight focusing is performed according to the first focused implicit representation input by the r-th layer first weight focusing network and the second focused implicit representation generated by the r-th layer second weight focusing network relative to the s-th feature, and the integrated text implicit representation output by the r-th layer second weight focusing network relative to the s-th feature generation is obtained; wherein r is a positive integer not greater than t; the second focused implicit representation generated by the first layer second weight focusing network relative to the s-th feature is determined based on the input text implicit representation generated by the s-th feature; feature generation is performed on the integrated text implicit representation output by the r-th layer second weight focusing network to obtain the r-th temporary feature generation implicit representation relative to the s-th feature generation; if r<t, based on the r-th temporary feature generation implicit representation relative to the s-th feature generation, determine the second focused implicit representation of the r+1-th layer second weight focusing network relative to the s-th feature generation, jump to the r-th layer second weight focusing network, perform weight focusing based on the first focused implicit representation input by the r-th layer first weight focusing network and the second focused implicit representation of the r-th layer second weight focusing network relative to the s-th feature generation, and iteratively execute the step of obtaining the integrated text implicit representation output by the r-th layer second weight focusing network relative to the s-th feature generation; if r=t, use the r-th temporary feature generation implicit representation relative to the s-th feature generation as the output text implicit representation of the s-th feature generation.

[0010] In one embodiment, the first focused implicit representation includes a first reference array and a first content array; the second focused implicit representation includes a second reference array, a second content array and a target search array; in the r-th layer second weight focusing network, weight focusing is performed based on the first focused implicit representation input by the r-th layer first weight focusing network and the second focused implicit representation generated by the r-th layer second weight focusing network relative to the s-th feature, to obtain the integrated text implicit representation output by the r-th layer second weight focusing network relative to the s-th feature generation, including: completing the following process in the r-th layer second weight focusing network: integrating the first reference array and the second reference array generated by the r-th layer second weight focusing network relative to the s-th feature to obtain the The target reference array generated by the r-th layer second weight focusing network relative to the s-th feature is obtained; the first content array and the second content array generated by the r-th layer second weight focusing network relative to the s-th feature are integrated to obtain the target content array generated by the r-th layer second weight focusing network relative to the s-th feature; according to the target search array generated by the r-th layer second weight focusing network relative to the s-th feature and the target reference array, the focusing weight of the r-th layer second weight focusing network relative to the s-th feature is determined; according to the target content array corresponding to the focusing weight generated by the r-th layer second weight focusing network relative to the s-th feature, the integrated text implicit representation output by the r-th layer second weight focusing network relative to the s-th feature is obtained.

[0011] In one embodiment, the method further includes: obtaining a training style text, a training style shielding text, a training paragraph placeholder sequence, and a training survey core information paragraph; the training style text is determined based on a training style text paragraph in a first training text, the training style shielding text is obtained by shielding a training style text paragraph in the first training text, the training paragraph placeholder sequence is a placeholder sequence of the training style text paragraph, and the training survey core information paragraph is obtained by extracting core information from a training survey subject content in the first training text; the training style text paragraph includes a survey content style text paragraph in the training survey subject content; embedding and mapping the training style text, the training style shielding text, and the training survey core information paragraph are performed separately to obtain a first training text implicit representation of the training style text, the The second training text implicit representation of the training style-shielded text and the third training text implicit representation of the training survey core information paragraph are trained; based on the first machine learning model, the first training text implicit representation is implicitly encoded to obtain the training implicit representation encoding result; with the second training text implicit representation, the third training text implicit representation and the training paragraph placeholder sequence as feature generation constraints, based on the second machine learning model, the training random placeholder sequence is gradually generated according to the training implicit representation encoding result to obtain the training style integrated text implicit representation; the training style integrated text implicit representation is restored and mapped to obtain the training style integrated survey text; according to the error between the training style integrated survey text and the first training text, the first machine learning model and the second machine learning model are updated until the first convergence requirement is met.

[0012] In one embodiment, the method of obtaining training style text, training style shielding text, training paragraph placeholder sequence and training survey core information paragraph includes: performing style text paragraph recognition on the first training text to determine the training style text paragraph in the first training text; performing placeholder replacement on the training style text paragraph to obtain the training paragraph placeholder sequence; performing a shielding operation on the training style text paragraph in the first training text based on the training paragraph placeholder sequence to obtain the training style shielding text; processing the first training text based on the training paragraph placeholder sequence to obtain the training style text; performing core information extraction on the training survey subject content in the first training text to obtain the training survey core information paragraph.

[0013] In one embodiment, the method of performing style text paragraph recognition on the first training text and determining the training style text paragraphs in the first training text includes: performing local context implicit representation encoding on the first training text based on the first sub-model to obtain a training context text implicit representation; performing semantic implicit representation encoding on the first training text based on the second sub-model to obtain a training semantic text implicit representation; performing a pooling operation on the training context text implicit representation to obtain a pooled context text implicit representation; performing feature enhancement on the training semantic text implicit representation to obtain an enhanced semantic text implicit representation; merging the training semantic text implicit representation with the pooled context text implicit representation to obtain a first training merged text implicit representation; merging the training context text implicit representation with the enhanced semantic text implicit representation to obtain a second training merged text implicit representation; integrating the first training merged text implicit representation and the second training merged text implicit representation to obtain a training text implicit representation; and determining the training style text paragraphs in the first training text based on the training text implicit representation.

[0014] In a second aspect, the present application provides a computer system, comprising a memory and a processor, wherein the memory stores a computer program executable on the processor, and the processor implements the steps in the above-described method when executing the program.

[0015] In a third aspect, the present application provides a computer-readable storage medium having a computer program stored thereon, which implements the steps in the above method when executed by a processor.

[0016] The beneficial effects of the present application are as follows: the present application provides a public health survey text analysis method and system based on natural language processing, the present application reconstructs the style text of the style-shielded text based on the target style text, the survey core information paragraphs and the paragraph placeholder sequence, and obtains the style-integrated survey text that integrates the text style contained in the target style text, and based on the present application, the text style in the public health survey text can be changed, that is, the text style in the public health survey text is changed to the text style in the target style text. Because the language style of the target style text is more prominent (for example, the language is more rigorous and formal), it can accurately describe the characteristics of the text style, so that the style-integrated survey text generated based on the target style text can reflect the style of the target text, and the effect of style integration is more accurate; in addition, the core information paragraph of the survey can accurately represent the survey topic content in the style-shielded text, and the paragraph placeholder sequence can accurately express the position of the style text paragraph in the style-shielded text. In this way, based on the target style text, the core information paragraph of the survey and the paragraph placeholder sequence, the style-shielded text is reconstructed. Not only can the text style of the style-integrated survey text be consistent with the text style in the target style text based on the position to be reconstructed represented by the paragraph placeholder sequence and the text style of the target style text, but the non-style text paragraph in the style-integrated survey text is consistent with the non-style text paragraph in the style-shielded text. In addition, the core information paragraph of the survey is used as a restriction to make the survey topic content in the generated style-integrated survey text consistent with the main content of the core information paragraph of the survey, to prevent the loss or change of the topic information after changing the text style, and to ensure the fidelity of the information. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The drawings herein are incorporated into the specification and constitute a part of the specification. These drawings illustrate embodiments consistent with the present application and are used together with the specification to illustrate the technical solution of the present application.

[0018] Figure 1 A schematic diagram of the implementation flow of a public health research text analysis method based on natural language processing provided in an embodiment of the present application.

[0019] Figure 2 A hardware entity diagram of a computer system provided in an embodiment of the present application. DETAILED DESCRIPTION

[0020] The embodiment of the present application provides a method for analyzing public health research text based on natural language processing, which can be executed by a processor of a computer system. The computer system can refer to a device with data processing capabilities such as a server, a laptop, a tablet computer, a desktop computer, a mobile device, etc.

[0021] Figure 1A schematic diagram of the implementation process of a public health research text analysis method based on natural language processing provided in an embodiment of the present application is as follows: Figure 1 As shown, the method includes:

[0022] Step S100: shielding the style text paragraphs in the public health research text to obtain style shielded text.

[0023] In the embodiment of the present application, the public health survey text is the text content of the survey results or records on public health-related topics, such as a survey on the prevalence of infectious diseases in a certain area, which includes descriptions of disease transmission routes, number of infections, prevention and control measures, etc. These descriptions together constitute the public health survey text.

[0024] Stylistic text paragraphs refer to the parts of public health research texts that have a specific style. This style can be reflected in the way of language expression, word usage habits, etc. For example, in some research texts, there may be some paragraphs that are highly descriptive and have subjective emotional colors, such as "The spread of this infectious disease is extremely fast, which is simply terrifying." Here, "simply terrifying" reflects a style with subjective emotions; or in some texts, there are more colloquial expression styles, such as "When we went to investigate, we found that those residents did not pay much attention to hygiene." The expression "When we went to investigate, we found that those residents" is more colloquial. These paragraphs can be regarded as stylistic text paragraphs.

[0025] Rules can be set to identify style text paragraphs. For example, the rules are set to identify paragraphs containing specific emotional words (such as "panic", "surprise" and other emotional words) or specific colloquial words (such as "ah", "what") as style text paragraphs. Once these paragraphs are identified, a shielding operation is performed. The shielding operation can be performed by replacing the identified style text paragraphs with specific tags or directly deleting them. Suppose we replace the style text paragraph with the tag "[STYLE_BLOCK]", then when the computer system identifies the style text paragraph "The spread of this infectious disease is extremely fast, which is simply terrifying", it will replace it with "[STYLE_BLOCK]". After such an operation, all parts of the original public health survey text that are identified as style text paragraphs are replaced or deleted, thereby obtaining style shielded text.

[0026] Another implementation method is based on machine learning. The computer system can use the annotated public health survey text data (in which the style text paragraphs are annotated) to train a classification model, such as a naive Bayes classifier or a support vector machine. During the training process, the features of the text (such as vocabulary features, grammatical structure features, etc.) are used as input, and whether it is a style text paragraph is used as output for training. After the model is trained, for new public health survey texts, the model can predict which paragraphs are style text paragraphs, and then perform a masking operation. For example, the text is converted into a vector representation. Assuming that the bag-of-words model is used, for a sentence "The spread of this infectious disease is extremely fast, which is simply terrifying", it is represented as a vector, and each dimension in the vector corresponds to a word. If the word exists, the value of the corresponding dimension is 1, and if it does not exist, it is 0. Then this vector is input into the trained support vector machine model. If the model determines that the paragraph where the sentence is located is a style text paragraph, the masking operation is performed in the manner mentioned above, and finally the style masked text is obtained.

[0027] In actual public health survey texts, there may be a mixture of multiple styles, and the computer system accurately identifies and blocks these style text paragraphs. For example, a survey text on the use of public health facilities may contain both colloquial style paragraphs (such as a description of residents' intuitive feelings about the use of public health facilities) and style paragraphs with certain emotional tendencies (such as expressions of dissatisfaction with the maintenance of public health facilities). The computer system comprehensively and accurately identifies and blocks these paragraphs through the above implementation method to ensure that the obtained style-shielded text can provide accurate basic data for subsequent analysis steps. In this process, the accuracy and efficiency of the computer system are crucial to the entire public health survey text analysis method, because if the identification and blocking of style text paragraphs are inaccurate, it will affect the extraction of core survey information and reconstruction of style text in subsequent steps, and thus affect the results of the entire public health survey text analysis.

[0028] In addition, during the shielding operation, the computer system also takes into account the coherence and integrity of the text. Although the style text paragraphs are shielded, the basic structure and logical relationship of the entire public health research text cannot be destroyed. For example, in a longer description of the development process of a public health event, if there are style text paragraphs interspersed in it, the computer system ensures that the remaining text can still clearly express the context of the development of the event when shielding these paragraphs. This requires a more sophisticated design of the shielding operation. For example, when using the replacement with a mark method, the setting of the mark should not affect the subsequent text processing operations as much as possible, and it should be easy to distinguish between the marked part and the normal text part when performing text analysis.

[0029] Step S200: extract core information from the research subject content in the public health research text to obtain the research core information paragraph; the style text paragraph includes the research content style text paragraph in the research subject content.

[0030] In the embodiment of the present application, within the scope of public health research, the research subject content refers to the description of relevant information around the core issues of public health. For example, in a research text on the influenza epidemic in a certain area, the research subject content may cover the incidence of influenza (such as the number of infected people in different age groups, the time distribution of the incidence rate, etc.), transmission factors (such as the impact of crowd gathering places on transmission, seasonal factors, etc.), prevention and control measures (such as vaccination status, health prevention and control measures in public places, etc.) and pressure on the local medical and health system. These core information related to the influenza epidemic together constitute the research subject content.

[0031] Core information extraction is a key operation performed by the computer system on the above research topic content. The computer system must accurately identify and extract the key information part that best represents the research topic content from a large amount of text data, that is, the core information paragraph of the research. This process uses a variety of implementation methods.

[0032] One feasible implementation method is rule-based information extraction. The computer system identifies and extracts core information according to pre-set rules. For example, for the description of the incidence of diseases in public health survey texts, the rules are set to identify sentences or paragraphs containing specific words (such as "onset", "infection", "number of cases", etc.) and their related numerical information as possible core information parts. Take "In this influenza epidemic, the number of infected people aged 18-30 reached 500, and the incidence rate showed a significant upward trend from November to December" as an example. Since it contains key words such as "infected people" and "incidence rate" and corresponding numerical information, the computer system identifies it as a possible core information part according to the rules. However, relying solely on simple vocabulary matching rules may have limitations because text expressions may be diverse. For example, although the wording of the expression "the total number of infected people is 500, mainly concentrated in the young group, and the incidence rate began to rise at the turn of autumn and winter" is different, it also expresses the core information related to the incidence. In order to cover this situation more comprehensively, the rules are further refined, such as adding synonym expansion rules to treat "total" and "reach" as synonymous expressions.

[0033] Another implementation method is an information extraction method based on machine learning. The computer system can use a large amount of existing publicly available health research text data that has been annotated for model training. For example, sequence labeling models in supervised learning (such as Hidden Markov Models or Conditional Random Field Models) can be adopted. During the training process, each word in the text is labeled as whether it belongs to the core information part (for example, labeled as "core information" or "non-core information"). Taking a simple sentence "The influenza vaccination rate is 30% locally and has played a certain role in controlling the epidemic" as an example, for each word in this sentence, it is labeled according to whether it is related to the core information (here, the relationship between the vaccination rate and epidemic control). When training the model, the text is converted into feature vectors. The features can include lexical features (such as word stems, part-of-speech, etc.), context features (such as the part-of-speech and vocabulary of the several words before and after), etc. Assuming that the lexical features are represented by one-hot encoding, for the word "influenza vaccine" in the sentence, if there are 1000 words in the vocabulary, the one-hot encoding vector corresponding to the word "influenza vaccine" may be a 1000-dimensional vector, where only the dimension corresponding to the word "influenza vaccine" is 1, and the rest are 0. The context features can be the part-of-speech of the two words before and after. For example, the word before "influenza vaccine" is "in", and its part-of-speech is a preposition, and the word after is "vaccination rate", and its part-of-speech is a noun. These part-of-speech information are also encoded as part of the feature vector. Taking these feature vectors as the input and the labeling results as the output, after training the model, the computer system can extract the core information from new public health research texts.

[0034] In public health research texts, the research topic content may be scattered in different parts of the text and may have multiple expressions. For example, in a research text on the community public health status, the core information regarding environmental hygiene may mention the frequency of community garbage cleaning at the beginning, the implementation of garbage classification in the middle, and the cleaning results of the sanitation dead corners in public areas at the end. The computer system comprehensively uses the above implementation methods to accurately extract these scattered and differently expressed core information related to environmental hygiene and form a complete core information paragraph regarding environmental hygiene in the research.

[0035] In addition, when extracting core information, the computer system also considers the completeness and accuracy of the information. Completeness means that important core information should not be missed as much as possible. For example, in a survey on infectious disease prevention and control measures, it is not possible to only extract information related to vaccination while ignoring information on isolation policies, health publicity and education, etc. Accuracy requires that the extracted information is indeed the core content closely related to the survey topic, avoiding the mistaken extraction of irrelevant or misleading information. For example, in a survey text on the allocation of public health resources, some descriptions of local cultural activities may have a certain indirect connection with public health, but they should not be mistakenly extracted as core information.

[0036] Step S300: Replace the style text paragraphs in the public health research text with placeholders to obtain a paragraph placeholder sequence.

[0037] In the embodiment of the present application, the public health survey text is a text record formed after conducting a survey on public health-related topics, covering various aspects such as disease epidemic situation, prevention and control measures, and allocation of public health resources. There are style text paragraphs in these texts. For example, when describing an epidemic of an infectious disease, there may be a paragraph like this: "This epidemic is so fierce that it caught people off guard." This statement has a strong subjective emotional color and belongs to a style text paragraph.

[0038] Style text paragraphs have unique language features. These features may include specific vocabulary, grammatical structure or semantic expression. For example, the use of some emotional words (such as "coming with all force" and "caught off guard"), or some more colloquial and literary expressions. The computer system identifies the elements in these style text paragraphs in order to perform placeholder replacement.

[0039] Placeholder replacement means that the computer system replaces the content in the style text paragraph with a specific placeholder. A placeholder is a marker symbol used to indicate the location of the original style text paragraph. For example, a simple symbol such as "[PLACEHOLDER]" can be used as a placeholder. For the style text paragraph about the epidemic mentioned above, "This epidemic came with great force and caught us off guard." The computer system will replace it with "[PLACEHOLDER]".

[0040] When performing placeholder replacement, the computer system can adopt a rule-based implementation method. For example, a set of rules are predefined to determine which paragraphs containing words or phrases are replaced with placeholders. The rule can be set as follows: if the number of sentiment words contained in a paragraph exceeds a certain threshold (for example, the number of sentiment words exceeds 2 for every 10 words), then the paragraph is marked as a style text paragraph and placeholder replacement is performed. For each word in a paragraph, a sentiment word dictionary can be established, and the number of sentiment words can be counted by querying this dictionary. Assume that the sentiment word dictionary is , the sequence of words in the paragraph is , the sentiment vocabulary counter is c=0. For each word ,if , then c=c+1. (that is, when the number of sentiment words exceeds 2 in every 10 words), the paragraph is replaced with a placeholder.

[0041] In addition to rule-based methods, machine learning-based methods can also be used. Computer systems can be trained using a large amount of annotated public health survey text data. Text paragraphs are represented as feature vectors. For example, a bag-of-words model can be used to convert the words in a paragraph into vector form. Assuming the vocabulary size is V, for the words in a paragraph , if its index in the vocabulary is j, then the value of the jth position in the bag-of-words model vector is the number of times the word appears in the paragraph. At the same time, other features can be added, such as part-of-speech information, paragraph length, etc. These feature vectors are used as input, and whether it is a style text paragraph (marked as 1 or 0) is used as output for training. The trained model can predict whether a new paragraph is a style text paragraph, and then replace the style text paragraph with a placeholder. The paragraph placeholder sequence refers to the sequence formed after all style text paragraphs are replaced with placeholders in the entire public health survey text. For example, a public health survey text originally contains multiple paragraphs, including three style text paragraphs. After the placeholder replacement, the entire text will have placeholders at these three positions, forming a new text sequence containing placeholders. This sequence is the paragraph placeholder sequence. This sequence records the position information of the style text paragraph in the original text, which is of great significance for subsequent style text reconstruction and other operations.

[0042] In the actual public health survey text analysis, it is very important to accurately identify and replace style text paragraphs with placeholders. Because different public health survey texts may have different styles and structures, style text paragraphs can be expressed in various forms. For example, in a survey text on satisfaction with the use of public health facilities, there may be a style text paragraph like this: "When we walked into those public toilets, the unpleasant smell was simply unbearable, it was really bad." This statement expresses the views on public health facilities by describing personal sensory experience, and has obvious style characteristics. The computer system accurately identifies such paragraphs and replaces them with placeholders to form a correct paragraph placeholder sequence. In addition, the placeholder replacement process considers the coherence and logical relationship of the text. Although the style text paragraph is replaced with a placeholder, the overall coherence of the text cannot be destroyed. For example, in a text that explains the development process of a public health event, if the style text paragraph is unreasonably replaced with a placeholder, it may cause a logical break in the text. When the computer system replaces the placeholder, it ensures that the logical structure of the text is still clearly discernible, so that subsequent operations such as accurate style text reconstruction based on the paragraph placeholder sequence can be performed.

[0043] Step S400: reconstructing the style text of the style shielding text according to the target style text, the research core information paragraph and the paragraph placeholder sequence, and obtaining the style-integrated research text incorporating the text style included in the target style text.

[0044] In the embodiment of the present application, the target style text is a text with specific style characteristics, and this style is predetermined. For example, in the field of public health research, the target style text may be a text with a rigorous, scientific, and objective style. For example, "In a survey of the public health status of a certain area, it was found that the incidence of infectious diseases in the area showed seasonal fluctuations, which was closely related to the local population mobility pattern, environmental sanitation conditions, and the level of health awareness of residents. Through analysis of the data, the incidence of infectious diseases in winter is significantly higher than that in summer. This phenomenon may be attributed to factors such as increased indoor activities and poor ventilation conditions in winter." This text explains the survey results in an objective and rigorous manner, without too much subjective emotions or colloquial expressions, and can be used as a target style text.

[0045] The style shielding text is the text obtained by shielding the style text paragraph in the public health survey text in step S100. It removes the style part of the original survey text and retains the core survey content-related part, but the text lacks specific style features at this time. For example, an original public health survey text may contain a sentence like this: "Oh, the sanitation condition of this place is really bad, and there is garbage everywhere. It's uncomfortable to look at it." After the style shielding operation, it may become "The sanitation condition of this place is poor, and there is garbage accumulation." The core information paragraph of the survey is the core information part extracted from the survey subject content in the public health survey text in step S200. It contains the most important and essential information content in the survey. For example, in a survey on community public health service satisfaction, the core information paragraph of the survey may contain key information such as the equipment of community health service facilities, the professional level of service personnel, and residents' satisfaction evaluation of services, such as "The community health service center is equipped with basic medical testing equipment, and medical staff have corresponding professional qualifications, but residents are less satisfied with the service response speed."

[0046] The paragraph placeholder sequence is a sequence obtained by replacing the style text paragraph with a placeholder in step S300, which marks the position information of the original style text paragraph in the text. For example, in a public health survey text after placeholder replacement, there may be a placeholder such as "[PLACEHOLDER]", and multiple placeholders constitute a paragraph placeholder sequence in the order in the original text.

[0047] The process of style text reconstruction by the computer system uses a series of implementation methods. One possible implementation method is based on the statistical language model. The computer system can use a large-scale corpus to learn the language patterns of texts of different styles. For example, for the target style text, it can count the frequency of occurrence and collocation relationship of different words and phrases in this style. Taking the target style as a rigorous scientific expression as an example, some professional terms (such as "morbidity", "epidemiology", "risk factors", etc.) appear relatively frequently in this style, and there is a specific collocation relationship between them (such as "disease-morbidity-influencing factors", etc.). The computer system constructs a probability-based language model by analyzing these statistical features of the target style text.

[0048] For style-shielded text, the computer system regards it as the basic framework of the style text to be generated. According to the positions indicated by the paragraph placeholder sequence, the computer system gradually inserts text content that conforms to the target style at these positions. The inserted content is generated based on the language model of the target style. For example, if the placeholder is in the part describing the influencing factors of the disease, the computer system generates sentences that conform to the target style, such as "Studies have found that the main risk factors for the disease include bad living habits, genetic susceptibility, and exposure to environmental pollutants, etc." to replace the placeholder based on the vocabulary and collocation relationships of the description of the influencing factors of the disease in the target style language model. At the same time, in order to ensure that the research topic content integrated into the generated style is consistent with the main content of the research core information paragraph, the computer system uses the research core information paragraph as a constraint. For example, if the research core information paragraph indicates that the incidence of a certain disease is high in a specific population, in the process of style text reconstruction, the computer system must ensure that this core information is accurately reflected in the generated text, and this information cannot be lost or distorted due to changes in style.

[0049] Another implementation method is based on neural network. The computer system can build a neural network model, such as a generative adversarial network (GAN) or a variational autoencoder (VAE).

[0050] For the generative adversarial network, it consists of two parts: the generator and the discriminator. The task of the generator is to generate style-integrated survey text based on the target style text, the core survey information paragraph, and the paragraph placeholder sequence. The task of the discriminator is to determine whether the generated text conforms to the target style. During the training process, the generator continuously tries to generate text, and the discriminator evaluates the generated text. The two compete with each other, and finally enable the generator to generate style-integrated survey text that conforms to the target style. For example, the generator receives style-shielded text, target style text, core survey information paragraph, and paragraph placeholder sequence as input. It first converts these inputs into vector representations suitable for neural network processing. Assume that the style-shielded text is represented as a vector S b , the target style text is represented as vector T s , the core information paragraph of the survey is represented by vector C i , the paragraph placeholder sequence is represented as a vector P s The generator contains multiple hidden layers, which gradually generate a text represented by a vector with the same length as the style masked text through a series of nonlinear transformations (such as using activation functions such as ReLU or Tanh). This generated text vector is then converted into text in natural language form through an output layer.

[0051] The discriminator takes the generated text and the real target style text as input and converts them into vector representation. Then, a discriminant score is calculated through a series of neural network layers to indicate whether the input text is the target style text. For example, if the discriminant score is close to 1, it means that the input text is likely to be the target style text; if it is close to 0, it means that it is not the target style text.

[0052] In the variational autoencoder, the computer system encodes the style mask text, target style text, research core information paragraphs, and paragraph placeholder sequences into a vector representation in a latent space. The decoder then decodes this latent vector into the style-integrated research text. During the encoding and decoding process, the model learns how to fuse the various input information and transform it into text that conforms to the target style and contains the research core information.

[0053] In actual public health research text analysis, for example, when reconstructing the style text of a research text on emergency response to public health emergencies, if the target style is concise, clear and official. Based on the style-shielded text, the computer system generates the corresponding text content according to the requirements of the target style according to the position marked by the paragraph placeholder sequence. For example, in the part describing emergency response measures, a concise and formal sentence is generated to replace the placeholder, such as "The emergency management department quickly launched the emergency plan, deployed medical resources, and implemented personnel isolation and testing measures." At the same time, it ensures that the core information about emergency response measures in the entire text is consistent with the content in the core information paragraph of the research.

[0054] As an implementation mode, step S400, based on the target style text, the research core information paragraph and the paragraph placeholder sequence, reconstructs the style shielding text to obtain the style integrated research text integrated with the text style included in the target style text, including:

[0055] Step S410: embedding and mapping the target style text, the style shielding text, and the research core information paragraph respectively to obtain a first text implicit representation of the target style text, a second text implicit representation of the style shielding text, and a third text implicit representation of the research core information paragraph;

[0056] Step S420: based on the first text implicit representation, the second text implicit representation, the third text implicit representation and the paragraph placeholder sequence, gradually generate features for the random placeholder sequence to obtain a style-integrated text implicit representation;

[0057] Step S430: restore and map the implicit representation of the style-integrated text to obtain a style-integrated survey text that incorporates the text style contained in the target style text.

[0058] In step S410, the computer system performs embedding mapping on the target style text, the style shielding text and the core information paragraph of the survey, respectively, so as to obtain a first text implicit representation of the target style text, a second text implicit representation of the style shielding text and a third text implicit representation of the core information paragraph of the survey. Embedding mapping is an implementation method of converting text into a low-dimensional vector representation, which can capture the semantic information in the text. When performing embedding mapping, a pre-trained word vector model, such as Word2Vec or GloVe, can be used. These models are trained based on a large-scale corpus and can map each word in the text to a vector space of a fixed dimension.

[0059] Taking the target style text as an example, suppose the target style text is "In public health research, rigorous scientific methods should be followed, data should be accurately recorded, and results should be objectively analyzed." The computer system first performs word segmentation on this text to obtain words such as "in", "public health", "research", "should", "follow", "rigorous", "scientific methods", "accurately", "record", "data", "objectively", "analyze", and "results" (for the sake of convenience, the conventional word segmentation method is used here, which may be more complicated in practice). Then, using the pre-trained Word2Vec model, each word will be mapped to a vector. For example, the word "rigorous" may be mapped to a 100-dimensional vector. (This is just the form of an example vector). By combining the word vectors in the entire target style text, we can get a vector representation of the target style text. This vector representation is the first text implicit representation. It contains the semantic information of the target style text to a certain extent. For example, the vector representation of words such as "rigorous", "precise", and "objective" will reflect the rigorous and objective style characteristics of the target style text. For style-shielded text and core research information paragraphs, the computer system also uses a similar method to embed and map. For example, the style-shielded text "There are problems with the public health situation in the area, and sanitary facilities are insufficient." After word segmentation and embedding mapping, a vector representation will also be obtained, namely the second text implicit representation. This vector representation reflects the semantic information about public health and sanitary facilities in the style-shielded text. The core research information paragraphs such as "The incidence of infectious diseases in the area is high, and the main influencing factors are poor environmental sanitation and frequent population mobility." will also obtain the third text implicit representation through embedding mapping. This representation contains the semantic representation of key information such as the incidence of infectious diseases and their influencing factors.

[0060] In step S420, the computer system gradually generates features for the random placeholder sequence based on the first text implicit representation, the second text implicit representation, the third text implicit representation and the paragraph placeholder sequence, thereby obtaining the style-integrated text implicit representation. The random placeholder sequence is a placeholder representation temporarily generated by the computer system in order to construct the style-integrated text. Its function is similar to a template framework when constructing a new text. The computer system generates features for it based on various text implicit representations and paragraph placeholder sequences obtained previously.

[0061] Feature generation here refers to the process by which a computer system generates new semantic features based on existing information. One possible implementation is based on the recurrent neural network (RNN) or its variants in the neural network, such as the long short-term memory network (LSTM) or the gated recurrent unit (GRU). These network structures can process sequence data and are very suitable for processing sequential information such as text.

[0062] Assume that a simple public health survey text is used as an example. The target style text emphasizes scientificity and rigor. The style shielding text describes the basic public health status (such as the number of health facilities, health status of personnel, etc.). The core information paragraph of the survey contains key public health data (such as disease incidence, etc.). The paragraph placeholder sequence marks the original position of the style text paragraph. The computer system generates features for the random placeholder sequence based on this information.

[0063] For example, the sequence of random placeholders might be p=[p1, p2,…, p n ] (where p i Represents a single placeholder), the computer system uses RNN to gradually generate new feature vectors based on the rigorous characteristics of the target style text in the first text implicit representation (such as the vector representation relationship of certain professional vocabulary), the current situation description semantic information of the style masking text in the second text implicit representation, and the key data semantic information in the third text implicit representation. In each step of the generation process, RNN will generate new feature vectors based on the current input (such as the placeholder p i The next output is calculated by combining the intermediate results generated in the previous step and the hidden state (which contains the semantic information of the previous step). This output is the new features generated step by step, and finally these features are combined to get the implicit representation of style into text.

[0064] In step S430 , the computer system performs a restoration mapping on the implicit representation of the style-integrated text, thereby obtaining a style-integrated survey text that incorporates the text style included in the target style text.

[0065] Restoration mapping is an operation corresponding to the previous embedding mapping. The computer system converts the style-integrated text implicit representation (low-dimensional vector representation) obtained in step S420 back to the natural language text form. This process may involve some implementation methods, such as an inverse mapping algorithm or a decoder-based neural network structure.

[0066] Assume that the style obtained in step S420 is implicitly represented as a vector v si If a decoder-based neural network structure is used (such as the decoder part in a sequence-to-sequence model), this decoder will take the style-infused text implicit representation vector as input and gradually generate a word sequence.

[0067] For example, the decoder might operate on a vocabulary that contains all the words that might appear in a public health survey text. The decoder computes the vector v si Calculate the probability of each word appearing at the current position, and then select the word with the highest probability as the output. For example, when generating a description of public health measures, if the target style is rigorous, the decoder calculates the probability of words such as "implementation" and "promotion" at the current position based on the style-integrated text implicit representation vector, selects the word "implementation" as the first output word, and then continues to calculate the probability of the next word based on the new input (including the generated words and the original vector representation) and selects the output, thus gradually generating the entire style-integrated survey text, such as "implement strict public health prevention and control measures, including strengthening the construction of health facilities and personnel health monitoring."

[0068] As an implementation mode, step S420, based on the first text implicit representation, the second text implicit representation, the third text implicit representation and the paragraph placeholder sequence, gradually generates features for the random placeholder sequence to obtain the style-integrated text implicit representation, including:

[0069] Step S421: performing implicit representation encoding on the first text implicit representation based on the first machine learning model to obtain an implicit representation encoding result;

[0070] Step S422: Using the second text implicit representation, the third text implicit representation and the paragraph placeholder sequence as feature generation constraints, based on the second machine learning model, gradually generate features for the random placeholder sequence according to the implicit representation encoding result to obtain the style-integrated text implicit representation.

[0071] In step S421, the computer system performs implicit representation encoding on the first text implicit representation based on the first machine learning model to obtain an implicit representation encoding result.

[0072] The first machine learning model is a model that is designed and trained to perform specific encoding operations on the implicit representation of text. This model can be a deep neural network model, such as a multi-layer perceptron (MLP) or a variant of a convolutional neural network (CNN). These model structures have different characteristics and advantages and are suitable for processing different types of text feature representations.

[0073] Implicit representation encoding is the process of further transforming and extracting features from the implicit representation of the first text (the vector representation obtained by embedding the target style text). Assume that the implicit representation of the first text is V target , which is a vector of dimension d, namely V target =[v1, v2, … , v d ].

[0074] Let’s take a multi-layer perceptron (MLP) as an example to explain this process. MLP consists of multiple neurons arranged in layers, including input layer, hidden layer, and output layer. For implicit representation encoding, the number of neurons in the input layer is equal to the dimension d of the first text implicit representation, that is, each neuron receives V target An element in is used as input.

[0075] Assume that the MLP has one hidden layer with h neurons. The calculation method of each hidden layer neuron is as follows: ;in is the weighted sum of the inputs to the jth hidden layer neuron, is the weight connecting the i-th neuron in the input layer and the j-th neuron in the hidden layer, is the bias of the jth hidden layer neuron. Then, through the activation function f (for example, the commonly used ReLU function f(x)=max(0,x)) Perform activation operation to obtain the output of hidden layer neurons .

[0076] After the hidden layer is calculated, the output layer will further calculate the implicit representation encoding result based on the output of the hidden layer. This result is a new feature representation after encoding the implicit representation of the target style text, which may contain deeper semantic information or style features of the target style text. For example, if the style features of the target style text are scattered in the original implicit representation, after implicit representation encoding, these features may be concentrated or presented in a way that is more conducive to subsequent operations (such as feature generation).

[0077] In step S422, the computer system uses the second text implicit representation, the third text implicit representation and the paragraph placeholder sequence as feature generation constraints, and based on the second machine learning model, gradually generates features for the random placeholder sequence according to the implicit representation encoding results, thereby obtaining an implicit representation of the text with style integrated into it.

[0078] The second machine learning model is also a model specially designed for feature generation tasks, which can be built based on the principles of generative adversarial networks (GANs) or variational autoencoders (VAEs). The random placeholder sequence is a placeholder representation initially set in the process of style text reconstruction in order to gradually build a style-integrated text. Each of its elements represents a position to be filled, which will eventually be filled with content that conforms to the target style and contains the core information of the survey. The second text implicit representation (obtained from the style masked text) and the third text implicit representation (obtained from the core information paragraph of the survey) and the paragraph placeholder sequence are used as constraints for feature generation, which means that in the process of feature generation, the computer system considers these factors to ensure that the generated style-integrated text implicit representation is reasonable and meets the requirements. Take the variational autoencoder (VAE) as an example to explain this process. VAE consists of two parts: an encoder and a decoder. In this scenario, the encoder part will integrate the constraints such as the second text implicit representation, the third text implicit representation, and the paragraph placeholder sequence with the implicit representation encoding result (from step S421) as input.

[0079] Assume that the implicit representation encoding result is C, and the second text is implicitly represented as , the third text is implicitly represented as , the paragraph placeholder sequence is P. The encoder combines this information in some way (such as simple concatenation or weighted summation) to obtain a comprehensive input vector , where g represents the combination operation function. Then, the encoder converts the input vector I into a latent space representation Z, which has certain distribution characteristics (such as Gaussian distribution). In the latent space, the computer system can perform sampling operations based on this distribution characteristic to obtain different latent vectors z. The decoder takes these sampled latent vectors z as input and gradually generates implicit representations of style integrated into the text. Each step of the generation process of the decoder is similar to the operation in the neural network mentioned above. It will calculate the eigenvalues ​​at each position based on the latent vector z and the parameters inside the model (such as the weight matrix, etc.). These eigenvalues ​​are gradually combined to form an implicit representation of style integrated into the text.

[0080] For example, assuming that there are n placeholders in a random placeholder sequence, the decoder will gradually generate corresponding feature vectors for these n placeholder positions when generating the implicit representation of style integrated into text. For the first placeholder position, the decoder calculates a feature vector F1 based on the latent vector z and internal parameters. This feature vector contains semantic features related to the target style and core information of the survey. Then, based on the generated F1 and the latent vector z (which may be dynamically adjusted according to the generation process), continue to generate a feature vector F2 for the next placeholder position, and so on, until feature vectors are generated for all n placeholder positions. Finally, these feature vectors are combined to obtain the implicit representation of style integrated into text.

[0081] This feature generation process based on the second machine learning model, by combining multiple constraints with the implicit representation encoding results, can effectively utilize the information of the target style text, style-shielded text, survey core information paragraphs, and paragraph placeholder sequences to generate an implicit representation of the style-integrated text that meets the requirements, laying the foundation for the subsequent restoration mapping to obtain the style-integrated survey text.

[0082] As an implementation method, the first machine learning model includes a t-layer first weighted focusing network, and the second machine learning model includes a t-layer second weighted focusing network; the implicit representation encoding result includes the first focused implicit representation of each first weighted focusing network input, where t≥2. Based on this, step S422, using the second text implicit representation, the third text implicit representation and the paragraph placeholder sequence as feature generation constraints, based on the second machine learning model, gradually generates features for the random placeholder sequence according to the implicit representation encoding result, and obtains the style-integrated text implicit representation, including:

[0083] Step S4221: taking the second text implicit representation, the third text implicit representation and the paragraph placeholder sequence as feature generation constraints, based on the second machine learning model, performing x times of feature generation on the random placeholder sequence in turn according to the implicit representation encoding result, and integrating the output text implicit representation obtained by the xth feature generation into the text implicit representation as the style; x≥2;

[0084] For the sth feature generation, the following operations are included:

[0085] Step S42211: Obtain the implicit representation of the input text generated by the s-th feature, wherein the implicit representation of the input text generated by the first feature is obtained by integrating the implicit representation of the second text, the implicit representation of the third text, the paragraph placeholder sequence, and the implicit representation of the random text corresponding to the random placeholder sequence; if s is a positive integer greater than 1 and not greater than x, the implicit representation of the input text generated by the s-th feature is the implicit representation of the output text generated by the s-1-th feature;

[0086] Step S42212: In the r-th layer second weighted focusing network, weight focusing is performed based on the first focused implicit representation input by the r-th layer first weighted focusing network and the second focused implicit representation generated by the r-th layer second weighted focusing network relative to the s-th feature, to obtain the integrated text implicit representation generated by the r-th layer second weighted focusing network relative to the s-th feature output; wherein r is a positive integer not greater than t; the second focused implicit representation generated by the first layer second weighted focusing network relative to the s-th feature is determined based on the input text implicit representation generated by the s-th feature;

[0087] Step S42213: performing feature generation on the integrated text implicit representation output by the second weighted focusing network of the rth layer, and obtaining the rth temporary feature generation implicit representation relative to the sth feature generation;

[0088] Step S42214: if r<t, generate an implicit representation based on the r-th temporary feature generated relative to the s-th feature, determine the second focused implicit representation generated relative to the s-th feature by the second weighted focusing network of the r+1-th layer, jump to the second weighted focusing network of the r-th layer, perform weight focusing based on the first focused implicit representation input by the r-th layer first weighted focusing network and the second focused implicit representation generated relative to the s-th feature by the r-th layer second weighted focusing network, and iteratively execute the step of obtaining the integrated text implicit representation generated relative to the s-th feature by the second weighted focusing network of the r-th layer;

[0089] Step S42215: If r=t, the implicit representation of the r-th temporary feature generation relative to the s-th feature generation is used as the implicit representation of the output text of the s-th feature generation.

[0090] In step S4221, the computer system uses the second text implicit representation, the third text implicit representation, and the paragraph placeholder sequence as feature generation constraints, and based on the second machine learning model, performs x feature generation on the random placeholder sequence in turn according to the implicit representation encoding result, and uses the output text implicit representation obtained by the xth feature generation as the style to be integrated into the text implicit representation. Here, x is a pre-set number of times, indicating that in order to obtain a satisfactory style to be integrated into the text implicit representation, the computer system performs multiple feature generation operations.

[0091] Taking the analysis of infectious disease prevention and control in public health research text as an example, suppose our goal is to convert a research text with a more casual style into a text with a rigorous and scientific style. In this process, the implicit representation of the second text (from the style-shielded text, which contains basic public health information after the style is removed, such as the basic description of prevention and control measures, etc.), the implicit representation of the third text (from the core information paragraph of the research, such as the incidence data of infectious diseases, transmission routes and other key information) and the paragraph placeholder sequence (marking the position of the original style text) together constitute the constraints.

[0092] The second machine learning model plays a core role here. This model can be a neural network model based on a sequence-to-sequence architecture, which is good at processing input sequences and generating corresponding output sequences, and is very suitable for generating new text implicit representations based on existing text representations and constraints.

[0093] A random placeholder sequence is a set of temporarily constructed placeholders that represent the locations where the final style is integrated into the filled content of the text. For example, a random placeholder sequence can be represented as p=[p1, p2, … , p n ], where p i Represents the i-th placeholder. The computer system gradually generates features for this random placeholder sequence based on the implicit representation encoding result (from the processing of the target style text in step S421). Each feature generation is based on the current input (including the random placeholder sequence, the implicit representation encoding result, and the constraints), and generates an intermediate output text implicit representation according to the internal mechanism of the second machine learning model. After x such iterative operations, the output text implicit representation obtained for the xth time is considered to be the result that best meets the requirements of style integration, that is, the style is integrated into the text implicit representation.

[0094] In step S42211, the computer system obtains the implicit representation of the input text generated by the sth feature. Here, s indicates the number of feature generation operations currently being performed. .

[0095] For the first feature generation (s=1), the implicit representation of the input text is obtained by integrating the implicit representation of the second text, the implicit representation of the third text, the paragraph placeholder sequence, and the implicit representation of the random text corresponding to the random placeholder sequence.

[0096] Assume that the second text is implicitly represented as , which is a vector with a dimension of d_1, representing the semantic information of the style masked text after embedding mapping; the third text is implicitly represented as , the dimension is , containing the semantic features of the core information paragraph of the survey; the paragraph placeholder sequence is , where p_i is a placeholder; the random text corresponding to the random placeholder sequence is implicitly represented as , the dimension is .

[0097] Implicit representation of input text for first feature generation , the computer system can use a simple vector concatenation operation to integrate this information, namely , where different vectors are concatenated together in order to form a new input vector. This input vector contains all the information related to style integration and is the basis for the first feature generation.

[0098] If s is a positive integer greater than 1 and not greater than x, the implicit representation of the input text of the sth feature generation is the implicit representation of the output text of the s-1th feature generation. This means that in each subsequent feature generation process, the output result of the previous time will be used as the input of the next time, thereby realizing an iterative feature generation process and gradually optimizing the generated implicit representation of the text.

[0099] For example, during the second feature generation (s=2), the computer system directly uses the output text implicit representation obtained from the first feature generation as the input text implicit representation I2. This iterative mechanism enables each feature generation to be improved based on the previous results, gradually moving closer to the ideal style that meets the target style and contains the core information of the survey and integrates the text implicit representation.

[0100] In step S42212, in the r-th layer second weighted focusing network, the computer system performs weighted focusing (i.e., performs attention calculation) based on the first focused implicit representation input by the r-th layer first weighted focusing network and the second focused implicit representation generated by the r-th layer second weighted focusing network relative to the s-th feature, and obtains the integrated text implicit representation output by the r-th layer second weighted focusing network relative to the s-th feature generation. Here, r is a positive integer not greater than t, indicating the number of layers in the second weighted focusing network, and t is the total number of layers of the second weighted focusing network set in advance, t≥2.

[0101] Weight focusing (attention calculation) is a mechanism used to make the model pay more attention to certain parts of the input information in the process of generating implicit representations of text. The first focused implicit representation and the second focused implicit representation are the key representations used to calculate the attention weights in the weight focusing network.

[0102] Assume that the first focus is implicitly expressed as ,in and are sub-vectors with different semantic information; the second focus implicitly expressed as ,in It also contains different semantic information.

[0103] In the second weighted focus network of the rth layer, the computer system first calculates the attention weight. A feasible formula for calculating the attention weight can be: ; where · represents the vector dot product operation, and d is the dimension of the vector (to prevent the value from being too large, it is divided by , the softmax function converts the calculation result into a weight vector in the form of probability distribution .

[0104] Then, the corresponding semantic information is weighted and summed according to this attention weight to obtain the implicit representation of the integrated text of the second weighted focusing network of the rth layer relative to the sth feature generation output. For example, if the implicit representation of the integrated text is I, then ,in etc. are the semantic information vectors corresponding to the attention weights. This weighted focusing mechanism enables the computer system to dynamically adjust the importance of different semantic information when generating implicit text representations, thereby better integrating various input information and generating more satisfactory implicit text representations.

[0105] For the first-layer second weighted focusing network, the second focused implicit representation generated by the s-th feature is determined based on the input text implicit representation generated by the s-th feature. This means that in the first-layer weighted focusing network, the second focused implicit representation is directly extracted or converted from the input text implicit representation, providing the initial semantic information basis for the subsequent weighted focusing calculation.

[0106] In step S42213, the computer system performs feature generation on the integrated text implicit representation output by the r-th layer second weight focusing network to obtain the r-th temporary feature generation implicit representation relative to the s-th feature generation.

[0107] Feature generation is a process of further creating new semantic features based on the integration of implicit text representations. Computer systems can use some operations in neural networks to implement this process. For example, a linear transformation layer (which can be represented as a weight matrix W and a bias vector b) is used to operate on the integrated text implicit representation I.

[0108] Assume that the dimension of the integrated text implicit representation I is , the dimension of the weight matrix W of the linear transformation layer is (in is the dimension of the implicit representation generated by the desired temporary feature), the dimension of the bias vector b is Then, the rth temporary feature generates an implicit representation It can be calculated by the following formula: This temporary feature generates an implicit representation that includes the new semantic features generated for the s-th feature generation under the second weight focusing network at the r-th layer. It is obtained through operations such as linear transformation on the basis of integrating the text implicit representation, providing further semantic information for subsequent iterative operations or the generation of the final style-integrated text implicit representation.

[0109] In step S42214, if r < t, the computer system generates an implicit representation based on the r-th temporary feature with respect to the s-th feature generation, determines the second focused implicit representation of the second weight focusing network at the (r + 1)-th layer with respect to the s-th feature generation, and then jumps to the step of performing weight focusing in the second weight focusing network at the r-th layer based on the first focused implicit representation input by the first weight focusing network at the r-th layer and the second focused implicit representation of the second weight focusing network at the r-th layer with respect to the s-th feature generation, to obtain the integrated text implicit representation output by the second weight focusing network at the r-th layer with respect to the s-th feature generation, and perform iterative execution.

[0110] This step reflects the iterative process in the multi-layer weight focusing network. For example, in a model with a three-layer (t = 3) second weight focusing network, when r = 1, the computer system first determines the second focused implicit representation of the second weight focusing network at the second layer with respect to the s-th feature generation according to the implicit representation generated from the first temporary feature. This determination process can be an operation based on function mapping. For example, the implicit representation generated from the temporary feature is transformed through a non-linear function f (such as f(x) = tanh(x)) to obtain the second focused implicit representation.

[0111] Then, the computer system performs the weight focusing operation again (as described in step S42212) to obtain the integrated text implicit representation output by the second weight focusing network at the second layer with respect to the s-th feature generation. Then, feature generation is performed based on this integrated text implicit representation (as described in step S42213) to obtain the implicit representation generated from the second temporary feature. If r = 2 (still less than t = 3), the computer system will continue to determine the second focused implicit representation of the second weight focusing network at the third layer with respect to the s-th feature generation according to the implicit representation generated from the second temporary feature, and then repeat the above weight focusing and feature generation operations, and so on, to implement the iterative feature generation process in the multi-layer weight focusing network. This iterative process helps to gradually and deeply explore the semantic relationships in the input information, so as to generate a more accurate style-integrated text implicit representation that better conforms to the target style and the core information of the research.

[0112] In step S42215, if r = t, the computer system generates an implicit representation for the r-th temporary feature generated with respect to the s-th feature generation as the output text implicit representation of the s-th feature generation. For example, when the second weight focusing network has a total of 3 layers (t = 3), after the previous steps (determining the second focusing implicit representation, weight focusing, and feature generation to obtain the implicit representation of the 3rd temporary feature generation) in the 3rd layer (r = 3), this implicit representation of the 3rd temporary feature generation is determined as the output text implicit representation of the s-th feature generation. This output text implicit representation contains the semantic information processed by the multi-layer weight focusing network during the s-th feature generation process. It will be used as the input text implicit representation for the next feature generation (if s < x) as described in step S42211, or when s = x, as the final style integrated into the text implicit representation, thus completing the entire feature generation process and providing input for the subsequent reduction mapping operation to finally obtain the style integrated research text containing the text style of the target style integrated text.

[0113] As an implementation, the first focusing implicit representation includes a first reference array and a first content array; the second focusing implicit representation includes a second reference array, a second content array, and a target search array. Based on this, in step S42214, in the r-th layer of the second weight focusing network, weight focusing is performed based on the first focusing implicit representation input by the r-th layer of the first weight focusing network and the second focusing implicit representation of the r-th layer of the second weight focusing network with respect to the s-th feature generation, to obtain the integrated text implicit representation output by the r-th layer of the second weight focusing network with respect to the s-th feature generation, including:

[0114] The following process is completed in the r-th layer of the second weight focusing network:

[0115] Step S422141: Integrate the first reference array and the second reference array of the r-th layer of the second weight focusing network with respect to the s-th feature generation to obtain the target reference array of the r-th layer of the second weight focusing network with respect to the s-th feature generation;

[0116] Step S422142: Integrate the first content array and the second content array of the r-th layer of the second weight focusing network with respect to the s-th feature generation to obtain the target content array of the r-th layer of the second weight focusing network with respect to the s-th feature generation;

[0117] Step S422143: Determine the focusing weight of the r-th layer of the second weight focusing network with respect to the s-th feature generation according to the target search array and the target reference array of the r-th layer of the second weight focusing network with respect to the s-th feature generation;

[0118] Step S422144: Weight the target content array corresponding to the focusing weight of the r-th layer second weight focusing network relative to the s-th feature generation to obtain the integrated text implicit representation of the r-th layer second weight focusing network relative to the s-th feature generation output.

[0119] In step S422141, the computer system integrates the first reference array and the second reference array of the r-th layer second weight focusing network relative to the s-th feature generation to obtain the target reference array of the r-th layer second weight focusing network relative to the s-th feature generation.

[0120] First, the concepts of the first reference array and the second reference array are clarified. The reference array is a data structure that stores specific semantic information or feature representation, and it plays an important role in the calculation process of the weighted focusing network.

[0121] Assume that the first reference array is are elements with specific semantic meanings. For example, in the analysis of public health research texts, if the target style is scientific rigor, May represent semantic information related to scientific terms and quantitative data.

[0122] The second reference array of the second weight focusing network in the rth layer relative to the sth feature generation is assumed to be , this array contains specific semantic information at the current network layer (layer r) and feature generation times (times s). For example, it may be related to the semantic features of the core information paragraph and style masked text processed by the current layer.

[0123] When a computer system performs an integration operation, multiple implementation methods can be used. One feasible method is vector concatenation. For example, the first reference array and the second reference array are concatenated together in sequence to obtain the target reference array R target = t = This concatenation operation enables the semantic information in the two reference arrays to be merged together, providing a more comprehensive information basis for subsequent calculations.

[0124] Another possible way is weighted summation. The computer system assigns corresponding weights w1 and w2 to each element in the first reference array and the second reference array (these weights can be preset or obtained through model training), and then uses the formula R target [i]= (i represents the element index in the array) Each element in the target reference array is calculated to obtain the integrated target reference array.

[0125] In step S422142, the computer system integrates the first content array and the second content array generated by the r-th layer second weight focusing network relative to the s-th feature to obtain a target content array generated by the r-th layer second weight focusing network relative to the s-th feature.

[0126] The content array is also a data structure that stores information related to the semantic content of the text. The first content array is assumed to be ,in Contains semantic content elements corresponding to the target style text or other related information. For example, it may contain semantic information of some typical expression patterns or common words in the target style text.

[0127] The second content array of the second weighted focusing network of the rth layer relative to the sth feature generation is assumed to be , this array contains the specific content-related semantic information at the current network layer and feature generation number. For example, it may be related to the semantic features of the current layer after processing a random placeholder sequence.

[0128] The way the computer system performs the integration operation is similar to step S422141. For the vector concatenation method, the two content arrays are concatenated together in sequence to obtain the target content array C target = .

[0129] If the weighted summation method is adopted, a weight is assigned to each element in the first content array and the second content array. , through formula C target [j]= (j represents the element index in the array) Calculate each element in the target content array.

[0130] In step S422143, the computer system determines the focusing weight of the r-th layer second weight focusing network relative to the s-th feature generation based on the target search array and the target reference array of the r-th layer second weight focusing network relative to the s-th feature generation.

[0131] The target search array is a special array structure that plays a key role in determining the focus weight. Assume that the target search array is S=[s1, s2,…, s k ], which contains information used to search and match related elements in the target reference array.

[0132] One way for the computer system to determine the focus weight is by calculating the similarity between the target search array and the target reference array. For example, the cosine similarity formula can be used:

[0133] ;

[0134] in is the target reference array R target This similarity value can be used directly as the focus weight, or it can be further transformed (such as through a nonlinear function f, for example Get the final focus weight .

[0135] In step S422144, the computer system weights the target content array corresponding to the focus weight of the second weighted focus network of the rth layer relative to the sth feature generation, and obtains the integrated text implicit representation of the output of the second weighted focus network of the rth layer relative to the sth feature generation. Assume that the focus weight obtained in step S422143 is (Here it is assumed that the target content array C target The target content array is C target = .

[0136] The computer system obtains the integrated text implicit representation I through weighted operations, and the calculation formula is: This integrated text implicit representation is the result of weighting the target content array with focus weights at the current network layer (layer r) and feature generation times (s). It combines the content array information integrated in the previous step and the focus weight information determined based on the target search array and target reference array, providing a comprehensive semantic representation for subsequent operations (such as further feature generation or next layer processing in a multi-layer network).

[0137] In one implementation, the method further includes a model training process, which specifically includes the following steps:

[0138] Step S10: obtaining a training style text, a training style shielding text, a training paragraph placeholder sequence, and a training survey core information paragraph; the training style text is determined based on the training style text paragraph in the first training text, the training style shielding text is obtained by shielding the training style text paragraph in the first training text, the training paragraph placeholder sequence is a placeholder sequence of the training style text paragraph, and the training survey core information paragraph is obtained by extracting core information from the training survey subject content in the first training text; the training style text paragraph includes the survey content style text paragraph in the training survey subject content;

[0139] Step S20: embedding and mapping the training style text, the training style shielding text, and the training survey core information paragraph respectively to obtain a first training text implicit representation of the training style text, a second training text implicit representation of the training style shielding text, and a third training text implicit representation of the training survey core information paragraph;

[0140] Step S30: performing implicit representation encoding on the first training text implicit representation based on the first machine learning model to obtain a training implicit representation encoding result;

[0141] Step S40: using the second training text implicit representation, the third training text implicit representation and the training paragraph placeholder sequence as feature generation constraints, based on the second machine learning model, and according to the encoding result of the training implicit representation, gradually generate features for the training random placeholder sequence to obtain the implicit representation of the training style integrated into the text;

[0142] Step S50: restore and map the implicit representation of the training style integrated into the text to obtain the research text integrated with the training style;

[0143] Step S60: updating the first machine learning model and the second machine learning model according to the error between the survey text and the first training text integrated into the training style until the first convergence requirement is met.

[0144] In step S10, the computer system obtains training style text, training style shielding text, training paragraph placeholder sequence and training survey core information paragraph. These different text elements play an important role in model training, and each contains different information related to the public health survey text. The training style text is determined based on the training style text paragraph in the first training text. The first training text is the original public health survey text material used to train the model. The training style text paragraph is a part of the content with a specific style in the first training text. For example, in a first training text about the epidemic situation of infectious diseases in a certain area, there may be some paragraphs that are highly descriptive and subjective, such as "The spread of this infectious disease is beyond imagination, and people are in great panic." Such paragraphs may be determined as training style text paragraphs, and then constitute training style text.

[0145] The training style masked text is obtained by masking the training style text paragraphs in the first training text. The masking operation is intended to remove the style part in the first training text so that the subsequent model can focus on learning the core information and style reconstruction. For example, a computer system may obtain the training style masked text by replacing the identified training style text paragraphs with specific tags or directly deleting them. If there is a description of the public health incident in the original first training text as "This incident happened extremely suddenly and caught people off guard" (assuming that this is a training style text paragraph), after the masking operation, this part may be replaced with "[STYLE_BLOCK]", thereby obtaining the training style masked text.

[0146] The training paragraph placeholder sequence is a placeholder sequence of the training style text paragraph. The placeholder sequence marks the position information of the training style text paragraph in the original first training text. For example, if there are three training style text paragraphs in the first training text, located in the 2nd, 5th and 8th paragraphs respectively, the computer system can replace these paragraphs with specific placeholders (such as "[PLACEHOLDER_1]", "[PLACEHOLDER_2]", "[PLACEHOLDER_3]"), and these placeholders constitute the training paragraph placeholder sequence in sequence.

[0147] The core information paragraph of the training survey is obtained by extracting the core information of the training survey theme content in the first training text. The training survey theme content revolves around the core issues of public health research. For example, in the first training text on the allocation of public health resources, the training survey theme content may involve the types of resources (such as medical equipment, medical staff, etc.), regional differences in allocation, and access to different groups of people. The computer system extracts the most critical information part through a specific implementation method to form the core information paragraph of the training survey. For example, in a first training text on community public health services, the core information may be "the number of medical staff in the community health service center is X, mainly providing basic medical services and preventive health care services, and the services cover Y people in the community." This is the extracted core information paragraph of the training survey.

[0148] The process of obtaining these elements by a computer system involves a variety of implementation methods. For identifying training style text paragraphs, rule-based and machine learning-based methods can be used. The rule-based method may be to set some predefined rules, such as paragraphs containing specific emotional words (such as "panic", "suddenly", etc.) or specific expression structures (such as colloquial expressions) as training style text paragraphs. The machine learning-based method can be to use a large amount of annotated public health survey text data for model training, such as using a naive Bayes classifier or a support vector machine, etc., using the characteristics of the text (such as vocabulary features, grammatical structure features, etc.) as input, and whether it is a training style text paragraph as output for training.

[0149] For core information extraction, rule-based information extraction and machine learning-based information extraction methods can be used. For example, rule-based information extraction sets rules to identify sentences or paragraphs containing specific words (such as "medical staff", "medical equipment", "number of people served", etc.) and their related numerical information as possible core information parts. Machine learning-based methods use sequence labeling models (such as hidden Markov models or conditional random field models) to convert text into feature vectors (such as bag-of-words models to represent lexical features, adding contextual features such as part of speech), and use whether it belongs to the core information part (marked as "core information" or "non-core information") as output for training.

[0150] In step S20, the computer system embeds and maps the training style text, the training style shielding text, and the training survey core information paragraph respectively, so as to obtain a first training text implicit representation of the training style text, a second training text implicit representation of the training style shielding text, and a third training text implicit representation of the training survey core information paragraph.

[0151] Embedding mapping is a way to convert text into a low-dimensional vector representation in order to capture the semantic information in the text. The computer system can use pre-trained word vector models such as Word2Vec or GloVe.

[0152] Taking the training style text as an example, suppose the training style text is "The outbreak of this infectious disease has had a huge impact on the lives of local residents, and everyone is very worried." The computer system first performs word segmentation on this text to obtain words such as "this," "infectious disease," "outbreak," "on," "local," "residents," "life," "produced," "huge," "impact," "everyone," "very," and "worried" (here is a simple example). Then, using the pre-trained Word2Vec model, each word is mapped to a vector. By combining the word vectors in the entire training style text, a vector representation of the training style text can be obtained, namely, the implicit representation of the first training text. This vector representation contains the semantic information of the training style text to a certain extent. For example, the vector representation of words such as "huge" and "worry" will reflect the style characteristics of the training style text, such as the subjective color. For the training style shielding text and the core information paragraphs of the training survey, the computer system also uses a similar method to embed and map. For example, the training style shielded text "Local areas have taken some prevention and control measures, such as isolating patients, etc.", after word segmentation and embedding mapping, obtains the implicit representation of the second training text. This vector representation reflects the semantic information about prevention and control measures in the training style shielded text. The core information paragraph of the training survey, such as "The incidence of infectious diseases in the area is X%, and the main influencing factors are population density and environmental sanitation," will also obtain the implicit representation of the third training text through embedding mapping. This representation contains the semantic representation of key information such as the incidence of infectious diseases and their influencing factors.

[0153] In step S30, the computer system performs implicit representation encoding on the first training text implicit representation based on the first machine learning model, thereby obtaining a training implicit representation encoding result.

[0154] The first machine learning model can be a deep neural network model, such as a multi-layer perceptron (MLP) or a variant of a convolutional neural network (CNN). These model structures have different characteristics and advantages and are suitable for processing different types of text feature representations.

[0155] Assume that the first training text is implicitly represented as V_{train-target}, which is a vector of dimension d, that is, . Let’s take Multi-Layer Perceptron (MLP) as an example to explain this process.

[0156] MLP consists of multiple neurons arranged in layers, including input layer, hidden layer and output layer. For implicit representation encoding, the number of neurons in the input layer is equal to the dimension d of the implicit representation of the first training text, that is, each neuron receives An element in is used as input.

[0157] Assume that the MLP has one hidden layer with h neurons. The calculation method of each hidden layer neuron is as follows: ;in is the weighted sum of the inputs to the jth hidden layer neuron, is the weight connecting the i-th neuron in the input layer and the j-th neuron in the hidden layer, is the bias of the jth hidden layer neuron. Then, z_j is activated by the activation function f (for example, the commonly used ReLU function f(x)=max(0,x)) to obtain the output a of the hidden layer neuron j =f(z j ).

[0158] After the hidden layer is calculated, the output layer will further calculate the training implicit representation encoding result based on the output of the hidden layer. This result is a new feature representation after encoding the implicit representation of the training style text, which may contain deeper semantic information or style features of the training style text. For example, if the style features of the training style text are scattered in the original implicit representation, after the implicit representation encoding, these features may be concentrated or presented in a way that is more conducive to subsequent operations (such as feature generation).

[0159] In step S40, the computer system uses the second training text implicit representation, the third training text implicit representation and the training paragraph placeholder sequence as feature generation constraints, and based on the second machine learning model, gradually generates features for the training random placeholder sequence according to the encoding results of the training implicit representation, so as to obtain the implicit representation of the training style integrated into the text.

[0160] The second machine learning model can be built based on the principles of generative adversarial networks (GANs) or variational autoencoders (VAEs).

[0161] The training random placeholder sequence is a placeholder representation temporarily generated by the computer system in order to construct a text with training style. It acts like a template framework when constructing new text.

[0162] The second training text implicit representation (obtained from the training style masked text), the third training text implicit representation (obtained from the training survey core information paragraphs) and the training paragraph placeholder sequence are used as constraints for feature generation, which means that during the feature generation process, the computer system considers these factors to ensure that the generated training style is integrated into the text implicit representation in a reasonable and satisfactory manner.

[0163] This process is explained by taking a variational autoencoder (VAE) as an example. VAE consists of two parts: an encoder and a decoder. In this scenario, the encoder part integrates the constraints such as the second training text implicit representation, the third training text implicit representation, and the training paragraph placeholder sequence with the training implicit representation encoding result (from step S30) as input.

[0164] Then, the encoder converts the input vector I into a latent space representation Z, which has certain distribution characteristics (such as Gaussian distribution). In the latent space, the computer system can perform sampling operations based on this distribution characteristic to obtain different latent vectors z. The decoder uses these sampled latent vectors z as input to gradually generate implicit representations of the training style integrated into the text. Each step of the decoder's generation process is similar to the operation in the neural network mentioned above. It will calculate the eigenvalues ​​at each position based on the latent vector z and the parameters inside the model (such as the weight matrix, etc.). These eigenvalues ​​are gradually combined to form the implicit representation of the training style integrated into the text.

[0165] For example, assuming that there are n placeholders in the training random placeholder sequence, the decoder will gradually generate corresponding feature vectors for these n placeholder positions when generating the implicit representation of the training style integrated into the text. For the first placeholder position, the decoder calculates a feature vector F1 based on the latent vector z and internal parameters. This feature vector contains semantic features related to the training style and the core information of the training survey. Then, based on the generated F1 and the latent vector z (which may be dynamically adjusted according to the generation process), continue to generate a feature vector F2 for the next placeholder position, and so on, until feature vectors are generated for all n placeholder positions. Finally, these feature vectors are combined to obtain the implicit representation of the training style integrated into the text.

[0166] In step S50, the computer system performs restoration mapping on the implicit representation of the training style integrated into the text, thereby obtaining the survey text integrated into the training style.

[0167] The restoration mapping is an operation corresponding to the previous embedding mapping. The computer system converts the training style obtained in step S40 into the text implicit representation (low-dimensional vector representation) back to the natural language text form. This process may involve some implementation methods, such as an inverse mapping algorithm or a decoder-based neural network structure.

[0168] Assuming that the training style obtained in step S40 is implicitly represented as a vector, if a decoder-based neural network structure is adopted (such as the decoder part in the sequence-to-sequence model), this decoder will use the training style-integrated text implicit representation vector as input and gradually generate a word sequence.

[0169] For example, the decoder may operate based on a vocabulary that contains all the words that may appear in the public health survey text. The decoder calculates the probability of each word appearing at the current position based on the vector, and then selects the word with the highest probability as the output. For example, when generating a description of public health measures, if the training style is rigorous, the decoder calculates the probability of words such as "implementation" and "promotion" at the current position based on the implicit representation vector of the text integrated into the training style, selects the word "implementation" as the first output word, and then continues to calculate the probability of the next word based on the new input (including the generated words and the original vector representation) and selects the output, thus gradually generating the entire training style integrated into the survey text, such as "implement strict public health prevention and control measures, including strengthening the construction of health facilities and personnel health monitoring."

[0170] In step S60, the computer system updates the first machine learning model and the second machine learning model according to the error between the survey text and the first training text in the training style until the first convergence requirement is met.

[0171] The error here refers to the degree of difference between the semantics, style, content, etc. between the training style integrated into the survey text and the first training text. Computer systems can measure this error in many ways. For example, the mean square error (MSE) formula can be used: ; where N is the number of samples, is the element in the first training text (such as the vector representation of words, sentences, etc.), It is the corresponding element of training style integrated into the research text.

[0172] Based on this error, the computer system uses the back propagation algorithm to update the parameters in the first machine learning model and the second machine learning model. The back propagation algorithm is based on the chain rule, calculates the gradient of the error relative to the model parameters, and then updates the parameters according to the gradient descent algorithm. For example, for the weight parameter w in the model, the update formula is: ;in is the learning rate, which controls the step size of parameter updates.

[0173] The first convergence requirement is a pre-set standard used to determine whether the model has been trained to a good enough degree. For example, it can be set that when the mean square error is less than a certain threshold (such as 0.01), the model is considered to have converged. During the training process, the computer system continuously repeats steps S10-S50, calculates the error and updates the model until the first convergence requirement is met. In this way, after multiple iterations of training, the first machine learning model and the second machine learning model can gradually learn how to accurately reconstruct the style text, so that the generated style integration survey text is as close as possible to the original first training text, while achieving style conversion and retention of core information.

[0174] As an implementation mode, in step S10, obtaining the training style text, the training style shielding text, the training paragraph placeholder sequence and the training survey core information paragraph includes:

[0175] Step S11: performing style text paragraph recognition on the first training text to determine the training style text paragraphs in the first training text;

[0176] Step S12: performing placeholder replacement on the training style text paragraph to obtain a training paragraph placeholder sequence;

[0177] Step S13: performing a shielding operation on the training style text paragraph in the first training text according to the training paragraph placeholder sequence to obtain the training style shielded text;

[0178] Step S14: Processing the first training text according to the training paragraph placeholder sequence to obtain a training style text;

[0179] Step S15: extract core information from the training and research subject content in the first training text to obtain training and research core information paragraphs.

[0180] In step S11 , the computer system performs style text paragraph recognition on the first training text, thereby determining the training style text paragraphs in the first training text.

[0181] One feasible implementation method for identifying style text paragraphs is a rule-based identification method. The computer system pre-sets a series of rules to determine which paragraphs belong to style text paragraphs. For example, for public health research texts, the rules can be set to determine that if a paragraph contains more emotional words or more colloquial and casual expressions, it will be determined as a style text paragraph. Another implementation method is a recognition method based on statistical features. The computer system can perform statistical analysis on the first training text, such as counting the distribution of parts of speech and the frequency of words in each paragraph. Certain parts of speech (such as a specific combination of interjections and adjectives) or abnormalities in word frequency may suggest that the paragraph has a specific style. For example, if the frequency of adjectives in a paragraph is too high, and these adjectives are mostly words that express subjective feelings (such as "terrible" and "amazing", etc.), the computer system can judge that the paragraph is a training style text paragraph. A machine learning-based method can also be used to identify style text paragraphs. The computer system uses a large amount of annotated public health research text data (these data have marked which paragraphs are style text paragraphs) to train a classification model, such as a decision tree model or a neural network model. Taking the decision tree model as an example, the computer system converts the paragraphs of the first training text into feature vectors. These features may include vocabulary features (such as whether a specific word exists), grammatical features (such as the complexity of sentence structure), etc. These feature vectors are then used as input, and whether it is a style text paragraph (marked as 1 or 0) is used as output for training. The trained decision tree model can classify the paragraphs in the new first training text and determine which paragraphs are training style text paragraphs.

[0182] In step S12, the computer system replaces the placeholders of the training style text paragraphs, thereby obtaining a training paragraph placeholder sequence.

[0183] The placeholder replacement is to mark the position of the training style text paragraph in the original first training text and to facilitate subsequent operations. The computer system replaces the content in the training style text paragraph with a specific placeholder.

[0184] For example, suppose the training style text paragraph is "This public health incident is really too serious and makes people worried." The computer system can use a simple marker "[PLACEHOLDER]" to replace this paragraph, and the result is that "[PLACEHOLDER]" appears at the position of the paragraph in the original first training text. If there are multiple training style text paragraphs in the first training text, after such replacement, these placeholders will form a training paragraph placeholder sequence in the order in the original text.

[0185] When performing placeholder replacement, the computer system accurately identifies the boundary of the training style text paragraph to ensure the accuracy of the replacement. This can be achieved by the start and end position information of the style text paragraph determined in step S11. At the same time, in order to avoid confusion with other content in the original text, the placeholder selection should be unique and can be easily identified and processed in the subsequent processing.

[0186] In step S13, the computer system performs a masking operation on the training style text paragraphs in the first training text according to the training paragraph placeholder sequence, thereby obtaining training style masked text.

[0187] The purpose of the masking operation is to remove the style part of the first training text so that the subsequent learning can focus more on the core information and style reconstruction. The computer system processes the corresponding training style text paragraph according to the position marked by the training paragraph placeholder sequence.

[0188] One processing method is to directly delete the training style text paragraph. For example, if the training paragraph placeholder sequence indicates that a paragraph is a training style text paragraph, the computer system directly deletes the paragraph from the first training text. Another method is to replace the training style text paragraph with a specific tag, which can be the same as or different from the placeholder in step S12, for example, also using "[STYLE_BLOCK]" for replacement.

[0189] Assume that the first training text is "This public health event has caused a great panic in the community. Everyone is trying to find ways to deal with it, and the community has organized some prevention and control measures." Among them, "This public health event has caused a great panic in the community." is a training style text paragraph. According to the training paragraph placeholder sequence, the computer system blocks it. If the method of replacing it with a marker is adopted, the obtained training style block text is "[STYLE_BLOCK]Everyone is trying to find ways to deal with it, and the community has organized some prevention and control measures."

[0190] In step S14, the computer system processes the first training text according to the training paragraph placeholder sequence to obtain the training style text.

[0191] The computer system extracts the contents of the paragraphs from the first training text according to the position information of the training style text paragraphs marked by the training paragraph placeholder sequence, thereby forming the training style text.

[0192] For example, suppose the first training text is "The public health situation in the community has not been very good. This problem has existed for a long time, and everyone is very worried. The community has begun to take some measures to improve it." Among them, "This problem has existed for a long time, and everyone is very worried." is a training style text paragraph, and the training paragraph placeholder sequence marks the position of this paragraph. The computer system extracts this paragraph from the first training text based on this placeholder sequence and obtains the training style text "This problem has existed for a long time, and everyone is very worried."

[0193] In this process, the computer system accurately locates the training style text paragraphs based on the placeholder sequence to ensure the accuracy of the acquired training style text. This is crucial for subsequent model training, because the training style text will serve as an important basis for the model to learn style features.

[0194] In step S15, the computer system extracts core information from the training survey topic content in the first training text, thereby obtaining a training survey core information paragraph.

[0195] Core information extraction is to screen out the most critical information that best represents the core of the research from the training research topic content.

[0196] The computer system can use a rule-based core information extraction method. For example, for public health survey texts, the rules are set to identify sentences or paragraphs containing specific public health-related vocabulary (such as "disease incidence", "prevention and control measures", "health resources", etc.) and their related digital information or key descriptions as possible core information parts. Suppose in the first training text about an infectious disease survey in a certain area, there is such a description: "The recent incidence of infectious diseases in the area is 10%. The main prevention and control measures include isolating patients and strengthening environmental sanitation and disinfection. However, health resources are relatively tight and the number of medical staff is insufficient." According to the rules, the computer system will identify this description containing the incidence of infectious diseases, prevention and control measures, and health resources as the core information paragraph of the training survey.

[0197] Machine learning-based methods can also be used to extract core information. The computer system uses a large amount of annotated public health survey text data (these data have already marked the core information part) to train models, such as support vector machine models or neural network models. The paragraphs in the first training text are converted into feature vectors. These features may include lexical features (such as stems, parts of speech, etc.), contextual features (such as the parts of speech and vocabulary of the previous and next words, etc.). These feature vectors are used as input, and whether it is the core information part (marked as 1 or 0) is used as output for training. The trained model can extract the core information of the training survey topic content in the new first training text to obtain the core information paragraphs of the training survey.

[0198] As an implementation manner, step S11, performing style text paragraph recognition on the first training text to determine the training style text paragraph in the first training text, includes:

[0199] Step S111: encoding the local context implicit representation of the first training text based on the first sub-model to obtain the implicit representation of the training context text;

[0200] Step S112: performing semantic implicit representation encoding on the first training text based on the second sub-model to obtain implicit representation of the training semantic text;

[0201] Step S113: performing a pooling operation on the implicit representation of the training context text to obtain a pooled implicit representation of the context text;

[0202] Step S114: performing feature enhancement on the implicit representation of the training semantic text to obtain an enhanced implicit representation of the semantic text;

[0203] Step S115: merging the training semantic text implicit representation with the pooled context text implicit representation to obtain a first training merged text implicit representation;

[0204] Step S116: merging the training context text implicit representation and the enhanced semantic text implicit representation to obtain a second training merged text implicit representation;

[0205] Step S117: Integrate the first training merged text implicit representation and the second training merged text implicit representation to obtain the training text implicit representation;

[0206] Step S118: determining a training style text paragraph in the first training text according to the implicit representation of the training text.

[0207] In step S111, the computer system encodes the local context implicit representation of the first training text based on the first sub-model, thereby obtaining an implicit representation of the training context text.

[0208] The first sub-model is a model structure specifically used to encode the local context implicit representation of text. It can be a neural network-based model, such as a variant of convolutional neural network (CNN). CNN has the characteristics of local connection and weight sharing, which can effectively capture the local features in the text.

[0209] For the first training text, the computer system regards it as a sequence of words. Assume that the first training text is "In the public health survey, we found that there are some problems with the community's sanitation facilities, such as untimely garbage removal, which may affect the health of residents." The computer system first preprocesses the text, such as word segmentation, to obtain words such as "in" "public health" "survey" "we" "found" "community" "sanitation facilities" "exist" "some" "problems" "for example" "garbage removal" "untimely" "this" "may" "will" "affect" "residents" "health" and so on.

[0210] Then, these words are converted into vector representations, and a pre-trained word vector model (such as Word2Vec or GloVe) can be used.

[0211] Based on the first sub-model (assuming it is a simple one-dimensional convolutional neural network), its convolution kernel slides on this word vector sequence to perform a convolution operation. The size and step size of the convolution kernel are pre-set parameters. For example, the convolution kernel size is 3 (indicating that it processes 3 consecutive word vectors at a time) and the step size is 1. For each word vector within the convolution window, a weighted sum operation is performed, and a bias term is added to obtain a new feature value. This process can be expressed by the formula: ;in is the jth convolution result (i.e., the new eigenvalue), f is the activation function (such as the ReLU function f(x)=max(0,x)), w i is the weight in the convolution kernel, x i+j is the input word vector, k is the convolution kernel size (here k=3), m is the starting index, and b is the bias term.

[0212] After the convolution operation, the computer system obtains a series of feature values, which constitute the implicit representation of the training context text. This representation contains the local context information in the first training text, such as the semantic relationship and combination features between adjacent words. For example, in the above text about public health research, the convolution operation may capture the local context relationship between "sanitation facilities" and "existing problems" and encode it into the implicit representation of the training context text.

[0213] In step S112, the computer system performs semantic implicit representation encoding on the first training text based on the second sub-model, thereby obtaining an implicit representation of the training semantic text.

[0214] The second sub-model is also a model designed for semantic implicit representation encoding, which may be a recurrent neural network (RNN) or its variants, such as long short-term memory network (LSTM) or gated recurrent unit (GRU). These model structures can handle long-distance dependencies in sequence data and are suitable for encoding semantic information of text.

[0215] Take the first training text "In the public health survey, it was found that there were some problems with the community's sanitation facilities, such as untimely garbage removal, which may affect the health of residents." as an example. The computer system inputs the word vector sequence of this text into the second sub-model (assuming it is LSTM). The basic unit of LSTM contains input gate, forget gate and output gate, which control the flow and update of information. At each time step (corresponding to each word vector), the input gate determines which new information can enter the unit, the forget gate determines which old information should be forgotten, and the output gate determines how much state information of the current unit should be output.

[0216] For example, for the word "community", the LSTM unit will update the state of the unit based on the state information of the previous word (such as "we found") and the vector representation of the current word "community" through the calculation of the input gate, forget gate, and output gate. After the entire LSTM network processes the entire first training text, the hidden state of the last time step or the entire hidden state sequence can be used as the implicit representation of the training semantic text. This representation contains the semantic information of the first training text, such as the semantic associations between the words in the text and the overall semantic meaning. For example, it can capture the semantic relationship between "there are problems with sanitation facilities" and "affecting the health of residents" and encode it into the implicit representation of the training semantic text.

[0217] In step S113, the computer system performs a pooling operation on the training context text implicit representation to obtain a pooled context text implicit representation.

[0218] Pooling is a method for reducing data dimensions and extracting key features. When pooling the implicit representation of training context text, the computer system can use different types of pooling methods, such as maximum pooling or average pooling.

[0219] Assume that the implicit representation of the training context text is a two-dimensional matrix (for example, the feature map obtained after the convolution operation) with a dimension of m×n, where m represents the number of rows of the feature map and n represents the number of columns.

[0220] For the maximum pooling operation, the computer system will select the maximum value in a preset pooling window (for example, a size of 2×2) as the result of pooling. Specifically, the pooling window is slid on the feature map, and the maximum value in the window is taken each time. For example, in the 2×2 window in the upper left corner of a 4×4 training context text implicit representation matrix, if the elements are [0.1, 0.2; 0.3, 0.4], then the result after maximum pooling is 0.4. For the average pooling operation, the computer system will calculate the average value in the pooling window as the result of pooling. For example, for the above 2×2 window, the result after average pooling is (0.1+0.2+0.3+0.4) / 4=0.25.

[0221] Through the pooling operation, the computer system obtains the implicit representation of the pooled context text. This representation retains the main features in the implicit representation of the training context text while reducing the dimension of the data, making subsequent calculations more efficient. For example, if the original implicit representation of the training context text contains too much local detail information, the pooling operation can extract the most significant features, such as the main semantic tendency in the local context relationship in the text.

[0222] In step S114, the computer system performs feature enhancement on the training semantic text implicit representation, thereby obtaining an enhanced semantic text implicit representation.

[0223] Feature enhancement is to further highlight the key semantic features in the implicit representation of the training semantic text, so as to better identify the style text paragraph. The computer system can use a variety of methods to perform feature enhancement. One method is to perform a nonlinear transformation on the vector elements in the implicit representation of the training semantic text. For example, a power transformation is used to transform each element v in the implicit representation of the training semantic text. i , transform it into v i =v i k (where k is a constant greater than 1, such as k=2). Suppose an element in the implicit representation of the training semantic text is 0.5, and after the power transformation, it becomes 0.5 2 = 0.25. This transformation can amplify the differences between elements and thus highlight the differences between different semantic features.

[0224] Another approach is to introduce additional semantic information for weighted combination. For example, the computer system can use an external semantic knowledge base (such as a terminology dictionary in the field of public health). If a word has important semantic associations in the semantic knowledge base, its corresponding vector element in the implicit representation of the training semantic text is weighted. Assume that the word "health facilities" has relevant semantic associations in the semantic knowledge base, and its vector element in the implicit representation of the training semantic text is v 卫生设施, the computer system can assign a weight w (such as w=1.5) to the semantic association according to its importance, and then update the vector element to v 卫生设施 '=w×v 卫生设施 Through these feature enhancement operations, the computer system obtains an enhanced semantic text implicit representation, which further strengthens the key semantic features in the training semantic text implicit representation and helps to more accurately identify the style text paragraphs in the subsequent steps.

[0225] In step S115, the computer system merges the training semantic text implicit representation with the pooled context text implicit representation, thereby obtaining a first training merged text implicit representation.

[0226] The merging operation is to integrate the information in two different text implicit representations to obtain a more comprehensive text feature representation. The computer system can use a simple vector concatenation method to merge. This merging method makes the first training merged text implicit representation contain both the main local context features in the pooled context text implicit representation and the semantic information in the training semantic text implicit representation, providing a richer information basis for subsequent operations and helping to more accurately determine the style text paragraphs in the first training text.

[0227] In step S116, the implicit representation of the training context text is merged with the implicit representation of the enhanced semantic text to obtain a second implicit representation of the training merged text. Similarly, the computer system uses vector splicing to merge. This merging operation combines the local context information in the implicit representation of the training context text with the enhanced semantic features in the implicit representation of the enhanced semantic text, providing another comprehensive text feature representation for subsequent style text paragraph recognition, which complements the first implicit representation of the training merged text and reflects the features of the first training text from different angles.

[0228] In step S117, the computer system integrates the first training merged text implicit representation and the second training merged text implicit representation to obtain the training text implicit representation.

[0229] The integration operation can be a weighted summation of the implicit representation of the first training merged text and the implicit representation of the second training merged text, or other more complex combination operations. Through this integration operation, the implicit representation of the training text integrates the information of the implicit representations of the two merged texts, forming a more comprehensive representation that better reflects the characteristics of the first training text, which provides an important basis for finally determining the style text paragraph in the first training text.

[0230] In step S118 , the computer system determines the training style text paragraph in the first training text according to the implicit representation of the training text.

[0231] The computer system can determine the style text paragraph according to the implicit representation of the training text by setting a threshold or using a classifier. One method is to set a threshold. It is assumed that some feature values ​​in the implicit representation of the training text have a certain correlation with the style text paragraph. For example, if the feature value of a certain dimension is greater than a preset threshold T, the corresponding paragraph is considered to be a style text paragraph. For example, in the implicit representation of the training text, if a certain dimension is related to emotional expression, when the value of this dimension is greater than 0.8 (assuming T=0.8), the computer system determines that the paragraph containing this feature is a training style text paragraph. Another method is to use a classifier. The computer system can use a large amount of annotated public health survey text data (these data have marked which paragraphs are style text paragraphs) to train a classifier, such as a support vector machine (SVM) or a neural network classifier. The implicit representation of the training text is used as the input feature vector, and whether it is a style text paragraph (marked as 1 or 0) is used as the output for training. The trained classifier can classify the new first training text (the implicit representation of the training text is obtained through the previous step) and determine the training style text paragraph therein.

[0232] Figure 2 A hardware entity diagram of a computer system provided in an embodiment of the present application is shown in FIG. Figure 2 As shown, the hardware entity of the computer system 1000 includes: a processor 1001 and a memory 1002, wherein the memory 1002 stores a computer program that can be run on the processor 1001, and the processor 1001 implements the steps in the method of any of the above embodiments when executing the program.

[0233] An embodiment of the present application also provides a computer storage medium, which stores one or more programs. The one or more programs can be executed by one or more processors to implement the steps of the method in any of the above embodiments.

Claims

1. A public health research text analysis method based on natural language processing, characterized in that: The method comprises: Shielding the style text paragraphs in the public health research text to obtain the style shielded text; Extracting core information from the research subject content in the public health research text to obtain a research core information paragraph; the style text paragraph includes the research content style text paragraph in the research subject content; Performing placeholder replacement on the style text paragraphs in the public health research text to obtain a paragraph placeholder sequence; Embedding and mapping the target style text, the style shielding text, and the research core information paragraph respectively to obtain a first text implicit representation of the target style text, a second text implicit representation of the style shielding text, and a third text implicit representation of the research core information paragraph; According to the first text implicit representation, the second text implicit representation, the third text implicit representation and the paragraph placeholder sequence, gradually generate features for the random placeholder sequence to obtain a style-integrated text implicit representation; The implicit representation of the style-integrated text is restored and mapped to obtain a style-integrated survey text that is integrated with the text style contained in the target style text.

2. The method according to claim 1, characterized in that The step of gradually generating features for a random placeholder sequence based on the first text implicit representation, the second text implicit representation, the third text implicit representation and the paragraph placeholder sequence to obtain a style-integrated text implicit representation includes: Performing implicit representation encoding on the first text implicit representation based on a first machine learning model to obtain an implicit representation encoding result; The second text implicit representation, the third text implicit representation and the paragraph placeholder sequence are used as feature generation constraints, based on a second machine learning model, and according to the implicit representation encoding result, feature generation is gradually performed on the random placeholder sequence to obtain the style-integrated text implicit representation.

3. The method according to claim 2, characterized in that The first machine learning model includes a t-layer first weighted focusing network, and the second machine learning model includes a t-layer second weighted focusing network; the implicit representation encoding result includes the first focused implicit representation of each first weighted focusing network input, wherein t≥2; The method of using the second text implicit representation, the third text implicit representation and the paragraph placeholder sequence as feature generation constraints, based on a second machine learning model, and gradually generating features for the random placeholder sequence according to the implicit representation encoding result to obtain the style-integrated text implicit representation includes: The second text implicit representation, the third text implicit representation and the paragraph placeholder sequence are used as feature generation constraints, and based on the second machine learning model, x features are generated for the random placeholder sequence in sequence according to the implicit representation encoding result, and the output text implicit representation obtained by the xth feature generation is used as the style to be integrated into the text implicit representation; x≥2; For the sth feature generation, the following operations are included: Obtaining an implicit representation of the input text generated by the s-th feature, wherein the implicit representation of the input text generated by the first feature is obtained by integrating the second implicit representation of the text, the third implicit representation of the text, the paragraph placeholder sequence, and the implicit representation of the random text corresponding to the random placeholder sequence; if s is a positive integer greater than 1 and not greater than x, the implicit representation of the input text generated by the s-th feature is the implicit representation of the output text generated by the s-1-th feature; In the r-th layer second weight focusing network, weight focusing is performed according to the first focused implicit representation input by the r-th layer first weight focusing network and the second focused implicit representation generated by the r-th layer second weight focusing network relative to the s-th feature, so as to obtain the integrated text implicit representation generated by the r-th layer second weight focusing network relative to the s-th feature output; wherein r is a positive integer not greater than t; the second focused implicit representation generated by the first layer second weight focusing network relative to the s-th feature is determined based on the input text implicit representation generated by the s-th feature; Performing feature generation on the integrated text implicit representation output by the second weighted focusing network of the r-th layer to obtain an r-th temporary feature generation implicit representation relative to the s-th feature generation; If r<t, according to the r-th temporary feature generation relative to the s-th feature generation, an implicit representation is generated, a second focused implicit representation of the second weighted focusing network of the r+1th layer relative to the s-th feature generation is determined, and the second weighted focusing network of the r-th layer is jumped to, and weight focusing is performed according to the first focused implicit representation input by the r-th layer first weighted focusing network and the second focused implicit representation generated by the r-th layer second weighted focusing network relative to the s-th feature generation, and the step of obtaining the integrated text implicit representation output by the r-th layer second weighted focusing network relative to the s-th feature generation is iteratively executed; If r=t, the implicit representation generated by the r-th temporary feature relative to the s-th feature generation is used as the implicit representation of the output text of the s-th feature generation.

4. The method according to claim 3, characterized in that The first focus implicit representation includes a first reference array and a first content array; the second focus implicit representation includes a second reference array, a second content array and a target search array; In the r-th layer second weight focusing network, weight focusing is performed according to the first focused implicit representation input by the r-th layer first weight focusing network and the second focused implicit representation generated by the r-th layer second weight focusing network relative to the s-th feature, to obtain the integrated text implicit representation output by the r-th layer second weight focusing network relative to the s-th feature generation, including: The following process is completed in the second weight focusing network of the rth layer: Integrate the first reference array and the second reference array of the r-th layer second weight focusing network relative to the s-th feature generation to obtain a target reference array of the r-th layer second weight focusing network relative to the s-th feature generation; Integrate the first content array and the second content array generated by the r-th layer second weighted focusing network relative to the s-th feature to obtain a target content array generated by the r-th layer second weighted focusing network relative to the s-th feature; Determine the focusing weight of the second weighted focusing network of the rth layer relative to the sth feature generation according to the target search array of the second weighted focusing network of the rth layer relative to the sth feature generation and the target reference array; The target content array corresponding to the focusing weight of the r-th layer second weight focusing network relative to the s-th feature generation is weighted to obtain an integrated text implicit representation of the r-th layer second weight focusing network relative to the s-th feature generation output.

5. The method according to any one of claims 2 to 4, characterized in that: The method further comprises: Acquire a training style text, a training style shielding text, a training paragraph placeholder sequence, and a training survey core information paragraph; the training style text is determined based on a training style text paragraph in a first training text, the training style shielding text is obtained by shielding a training style text paragraph in the first training text, the training paragraph placeholder sequence is a placeholder sequence of the training style text paragraph, and the training survey core information paragraph is obtained by extracting core information from the training survey subject content in the first training text; the training style text paragraph includes a survey content style text paragraph in the training survey subject content; Embedding and mapping the training style text, the training style shielding text, and the training survey core information paragraph respectively to obtain a first training text implicit representation of the training style text, a second training text implicit representation of the training style shielding text, and a third training text implicit representation of the training survey core information paragraph; Performing implicit representation encoding on the first training text implicit representation based on the first machine learning model to obtain a training implicit representation encoding result; Taking the second training text implicit representation, the third training text implicit representation and the training paragraph placeholder sequence as feature generation constraints, based on the second machine learning model, and according to the encoding result of the training implicit representation, gradually generate features for the training random placeholder sequence to obtain the training style integrated into the text implicit representation; Performing restoration mapping on the implicit representation of the training style integrated into the text to obtain the research text integrated with the training style; According to the error between the survey text integrated with the training style and the first training text, the first machine learning model and the second machine learning model are updated until a first convergence requirement is met.

6. The method according to claim 5, characterized in that The obtaining of the training style text, the training style shielding text, the training paragraph placeholder sequence and the training survey core information paragraph includes: Performing style text paragraph recognition on the first training text to determine training style text paragraphs in the first training text; Performing placeholder replacement on the training style text paragraph to obtain the training paragraph placeholder sequence; According to the training paragraph placeholder sequence, a shielding operation is performed on the training style text paragraph in the first training text to obtain a training style shielded text; Processing the first training text according to the training paragraph placeholder sequence to obtain the training style text; Core information is extracted from the training and research subject content in the first training text to obtain the training and research core information paragraph.

7. The method according to claim 6, characterized in that The step of performing style text paragraph recognition on the first training text to determine the training style text paragraphs in the first training text includes: Performing local context implicit representation encoding on the first training text based on the first sub-model to obtain implicit representation of the training context text; Performing semantic implicit representation encoding on the first training text based on the second sub-model to obtain implicit representation of the training semantic text; Performing a pooling operation on the training context text implicit representation to obtain a pooled context text implicit representation; Performing feature enhancement on the implicit representation of the training semantic text to obtain an enhanced implicit representation of the semantic text; Merging the training semantic text implicit representation with the pooled context text implicit representation to obtain a first training merged text implicit representation; Merging the training context text implicit representation with the enhanced semantic text implicit representation to obtain a second training merged text implicit representation; Integrating the first training merged text implicit representation and the second training merged text implicit representation to obtain a training text implicit representation; According to the implicit representation of the training text, a training style text paragraph in the first training text is determined.

8. A computer system comprising a memory and a processor, wherein the memory stores a computer program executable on the processor, wherein: When the processor executes the program, the steps in the method according to any one of claims 1 to 7 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps in the method according to any one of claims 1 to 7 are implemented.

Images