Online health impact assessment method, device, computer equipment and storage medium

Abstract

The embodiment of the present invention discloses an online health impact evaluation method, device, computer equipment and storage medium. The method includes: obtaining a public policy document to be evaluated issued by a government worker; identifying health impact keywords in the public policy document; sending the health impact keywords to a second terminal so that the second terminal evaluates the public policy document; obtaining evaluation content from the second terminal; obtaining matching related documents from a document abstract library based on the evaluation content to obtain an evaluation report. By implementing the method of the embodiment of the present invention, the difficulty of manuscript review can be reduced, and it is easy to identify potential health-impacting content in the document, with less manpower workload and a short evaluation cycle, thereby ensuring the timeliness of the evaluation.

Description

Technical Field

[0001] The present invention relates to computers, and more specifically to an online health impact assessment method, device, computer equipment and storage medium. Background Art

[0002] There are two main forms of health impact assessment in China: one is to refer to the existing technical process abroad, which requires on-site investigation of relevant stakeholders for evaluation; the other is health impact assessment by soliciting opinions; the first is rigorous but inefficient and cannot meet the pace of government documents. The second is efficient but lacks rigor.

[0003] At present, there is no systematic health impact assessment system in China, so the domestic health impact assessment mainly refers to foreign literature and relevant cases. Due to the limited number of relevant foreign literature and cases, the current health impact assessment often encounters the problem of insufficient available data. The health impact assessment team may make a decision in the evaluation process to devote part of the time to collecting new data, which makes it difficult to form large-scale relevant data, and also explains why health impact assessment is a time-consuming and labor-intensive task. At present, the health impact assessment process for domestic professionals and managers needs to go through four processes: screening, scope definition, evaluation, and reporting. The current expert manual health impact assessment process has the following problems: there is a lack of online health impact assessment system, and the evaluation rules are unclear, which increases the difficulty of expert review; the policy documents cover a wide range, and due to the professional limitations of experts in different fields, it is difficult to organize experts in different fields to identify potential health-related content in the documents; the process of consulting relevant literature as a reference when conducting a health impact assessment consumes a lot of manpower and workload, resulting in a longer evaluation cycle and difficulty in ensuring timeliness.

[0004] Therefore, it is necessary to design a new method to reduce the difficulty of review and facilitate the identification of potential health-impacting content in documents, with less manpower workload and shorter evaluation cycle to ensure the timeliness of the evaluation. Summary of the invention

[0005] The purpose of the present invention is to overcome the defects of the prior art and provide an online health impact assessment method, device, computer equipment and storage medium.

[0006] To achieve the above object, the present invention adopts the following technical solution: an online health impact assessment method, comprising:

[0007] Access public policy documents issued by government staff for review;

[0008] Identify health impact keywords within said public policy documents;

[0009] Sending the health impact keyword to a second terminal so that the second terminal can evaluate the public policy document;

[0010] Acquiring evaluation content from the second terminal;

[0011] According to the evaluation content, matching relevant documents are obtained from the document abstract library to obtain an evaluation report.

[0012] A further technical solution is: the identifying of health impact keywords in the public policy document includes:

[0013] Inputting the public policy document into the self-attention mechanism model for identification to obtain health impact keywords;

[0014] Among them, the self-attention mechanism model is obtained by training the BI-GRU-CRF model by using relevant policy documents with keyword labels as the sample set.

[0015] The further technical solution is: the self-attention mechanism model is obtained by training the BI-GRU-CRF model using relevant policy documents with keyword tags as sample sets, including:

[0016] Obtain relevant policy texts, and preprocess the policy texts to obtain policy texts;

[0017] Using a word segmentation tool to segment the policy text to obtain a word segmentation result;

[0018] Labeling the word segmentation results to obtain labeling results;

[0019] Obtaining a corresponding word vector for the annotation result;

[0020] Build a BI-GRU-CRF model;

[0021] The word vector is input into the BI-GRU-CRF model for training, and the trained BI-GRU-CRF model is evaluated. The BI-GRU-CRF model is improved accordingly according to the evaluation results to obtain a self-attention mechanism model.

[0022] Its further technical solution is: the BI-GRU-CRF model includes a bidirectional GRU neural network model, a self-attention mechanism layer, a CRF layer and an output layer.

[0023] A further technical solution is: obtaining the corresponding word vector for the annotation result includes:

[0024] Get the word embedding vector dictionary;

[0025] The annotation result is matched with the word embedding vector dictionary to obtain the corresponding word vector.

[0026] A further technical solution is: obtaining matching relevant documents from a document abstract library according to the evaluation content to obtain an evaluation report, including:

[0027] Build a health impact dataset and literature summary library through crawler technology and semantic similarity deduplication technology;

[0028] According to the evaluation content, the matching relevant documents in the document abstract library are screened out to obtain an evaluation report.

[0029] A further technical solution is: the step of screening out matching relevant documents in the document abstract library according to the evaluation content to obtain an evaluation report includes:

[0030] Performing word segmentation processing on the evaluation content and the text in the literature abstract library respectively, and removing stop words to obtain two texts after word segmentation;

[0031] Match the two texts after word segmentation and find the same words as keywords;

[0032] Merge the words on the left and right of the keyword into the corresponding text to obtain input content;

[0033] Input the input content into the KBS-GRU model to perform text similarity analysis to obtain an analysis result;

[0034] An evaluation report is determined according to the analysis results.

[0035] The present invention also provides an online health impact assessment device, comprising:

[0036] A document acquisition unit, used to acquire public policy documents to be evaluated issued by government staff;

[0037] an identification unit, for identifying health impact keywords in the public policy document;

[0038] A sending unit, configured to send the health impact keyword to a second terminal, so that the second terminal can evaluate the public policy document;

[0039] A content acquisition unit, used to acquire evaluation content from the second terminal;

[0040] The matching unit is used to obtain matching related documents from the document abstract library according to the evaluation content to obtain an evaluation report.

[0041] The present invention further provides a computer device, which includes a memory and a processor. The memory stores a computer program, and the processor implements the above method when executing the computer program.

[0042] The present invention also provides a storage medium, wherein the storage medium stores a computer program, and the computer program can implement the above method when executed by a processor.

[0043] Compared with the prior art, the beneficial effects of the present invention are as follows: the present invention uses natural language processing technology in artificial intelligence to construct a self-attention mechanism model to extract health impact keywords from public policy documents, and after obtaining the evaluation content of the second terminal, uses the KBS-GRU model in the document abstract library to perform text similarity analysis on the document abstract and the evaluation content to determine the evaluation report, thereby reducing the difficulty of manuscript review and facilitating the identification of potential health-impacting content in the document, with less manpower workload and a short evaluation cycle, thereby ensuring the timeliness of the evaluation.

[0044] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying any creative work.

[0046] Figure 1 A schematic diagram of an application scenario of the online health impact assessment method provided by an embodiment of the present invention;

[0047] Figure 2 A schematic diagram of a flow chart of an online health impact assessment method provided by an embodiment of the present invention;

[0048] Figure 3 A schematic diagram of sub-processes of an online health impact assessment method provided by an embodiment of the present invention;

[0049] Figure 4 A schematic diagram of sub-processes of an online health impact assessment method provided by an embodiment of the present invention;

[0050] Figure 5 A schematic diagram of sub-processes of an online health impact assessment method provided by an embodiment of the present invention;

[0051] Figure 6 A schematic diagram of sub-processes of an online health impact assessment method provided by an embodiment of the present invention;

[0052] Figure 7A schematic diagram of the structure of the Bi-GRU-CRF model provided in an embodiment of the present invention;

[0053] Figure 8 A schematic diagram of the structure of a multi-head attention mechanism provided by an embodiment of the present invention;

[0054] Fig. 9 A schematic diagram of the structure of a hybrid model of Jaro-Winkler and LSA provided in an embodiment of the present invention;

[0055] Fig.10 A schematic diagram of the structure of the KBS-GRU model provided in an embodiment of the present invention;

[0056] Fig.11 A schematic diagram of the architecture of a health impact assessment auxiliary system provided by an embodiment of the present invention;

[0057] Fig.12 A schematic block diagram of an online health impact assessment device provided by an embodiment of the present invention;

[0058] Fig.13 A schematic block diagram of a matching unit of an online health impact assessment device provided by an embodiment of the present invention;

[0059] Fig.14 A schematic block diagram of a file screening subunit of an online health impact assessment device provided by an embodiment of the present invention;

[0060] Fig.15 A schematic block diagram of a computer device provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0061] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0062] It should be understood that when used in this specification and the appended claims, the terms "include" and "comprises" indicate the presence of described features, integers, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or combinations thereof.

[0063] It should also be understood that the terms used in this specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. As used in the specification of the present invention and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include plural forms.

[0064] It should be further understood that the term "and / or" used in the present description and the appended claims refers to any and all possible combinations of one or more of the associated listed items, and includes these combinations.

[0065] See also Figure 1 and Figure 2 , Figure 1 A schematic diagram of an application scenario of the online health impact assessment method provided in an embodiment of the present invention. Figure 2 A schematic flow chart of an online health impact assessment method provided by an embodiment of the present invention. The online health impact assessment method is applied in a server. The server interacts with a first terminal and a second terminal for data exchange, wherein the first terminal is held by a government worker and is used to publish a public policy document to be evaluated, and the second terminal is held by an expert and is used to evaluate the health impact keywords extracted from the public policy document and upload the evaluation content to the server, which performs text similarity calculation on the evaluation content and determines matching documents to form an evaluation report.

[0066] Figure 2 Schematic diagram of the process of the online health impact assessment method provided by the embodiment of the present invention. Figure 2 As shown, the method includes the following steps S110 to S150.

[0067] S110. Obtain public policy documents for evaluation issued by government staff.

[0068] In this embodiment, the public policy document refers to a policy document issued by a government official.

[0069] After the government staff uploads the original file of the public policy document through the first terminal, the server will create a new evaluation task and select an expert group to distribute the evaluation task.

[0070] S120. Identify health impact keywords in the public policy document.

[0071] In this embodiment, the health impact keywords refer to the core words that have an impact on health in public policy documents.

[0072] Specifically, the public policy document is input into the self-attention mechanism model for identification to obtain health impact keywords; wherein the self-attention mechanism model is obtained by training the BI-GRU-CRF model using relevant policy documents with keyword labels as a sample set.

[0073] In one embodiment, see Figure 3 The above-mentioned self-attention mechanism model is obtained by training the BI-GRU-CRF model by using relevant policy documents with keyword tags as sample sets, and may include steps S121 to S126.

[0074] S121. Obtain relevant policy texts, and preprocess the policy texts to obtain policy texts.

[0075] In this embodiment, the policy text refers to the policy content including only the text part.

[0076] Specifically, organize the relevant policy texts, remove the tables, graphics and other data in the policy texts, and only keep the text.

[0077] S122. Use a word segmentation tool to segment the policy text to obtain a word segmentation result.

[0078] In this embodiment, the word segmentation result refers to each word obtained after the policy text is segmented using a word segmentation tool.

[0079] S123: label the word segmentation result to obtain a labeling result.

[0080] In this embodiment, the labeling result refers to labeling the words belonging to the health impact keywords.

[0081] Specifically, keywords related to health impacts were sorted out from relevant policy documents based on expert opinions. There are 56 categories of health impact factors and the corresponding types under each category. Each type has several keywords and descriptions of potential health impacts. Named entity recognition technology is used to identify and expand keywords in policy document texts, and the resulting health impact keywords are shown in Table 1.

[0082] Table 1. Health impact keywords

[0083]

[0084]

[0085] S124. Obtain corresponding word vectors for the annotation results.

[0086] In this embodiment, the word vector refers to a word vector with a health impact keyword label.

[0087] In one embodiment, see Figure 4 , the above-mentioned step S124 may include steps S1241 to S1242.

[0088] S1241, obtaining a word embedding vector dictionary;

[0089] S1242. Match the annotation result with the word embedding vector dictionary to obtain the corresponding word vector.

[0090] Specifically, the word embedding vector dictionary obtained by the Chinese word embedding hybrid model is matched with the segmented text to obtain the corresponding word embedding vector, i.e., the word vector.

[0091] S125. Build a BI-GRU-CRF model.

[0092] Specifically, Figure 7 As shown, the BI-GRU-CRF model includes a bidirectional GRU neural network model, a self-attention mechanism layer, a CRF layer and an output layer.

[0093] The BI-GRU-CRF model is based on GRU (Bidirectional Gated Recurrent Unit), and its general formula is as follows: t =σ(W z [h t-1 ,x t ]); r t =σ(W r [h t-1 ,x t ]); Among them, z t represents the update gate of GRU, r t Represents the reset gate of GRU. The value of the former represents the degree to which the information at time t-1 is brought into the current information; the latter can control the candidate set of the current state that can be written into the previous state The amount of information.

[0094] Although the GRU unit is capable of fast data encoding, it has the disadvantage of insufficient ability to learn dependencies on long texts. Therefore, this embodiment adds a multi-head attention mechanism to the neural network structure. The multi-head attention mechanism can not only solve the problem of the model's insufficient ability to learn long texts, but also make full use of contextual information to focus on selecting partial input information and fast labeling, making the entire model easier to implement and understand. The multi-head attention mechanism can capture multi-part and multi-level feature information of the input, and pay attention to the contextual information of different parts of the input respectively, which allows more information to be paid attention to in the final encoding. The multi-head attention mechanism is as follows Figure 8The principle of the multi-head attention mechanism is to use the scaled dot product attention mechanism for different parts of the input, and then concatenate the multiple output results to finally form an output matrix. Its mathematical expression is as follows: multi-head(Q,K,V)=concat(head1,…,head h );in, concat means concatenating multiple output results.

[0095] The last layer of the BI-GRU-CRF model is the CRF layer, which uses the adjacent position sequence of the target sequence for conditional probability modeling, making the serialization annotation more accurate. Define z = [z1, z2, z3, ..., z t ] represents an input sequence, where z i Represents the feature vector of the i-th word in the sequence. y=[y1,y2,y3,…,y t ] is the label sequence of sequence z, y i is the label of word i. Φ(z) represents the set of possible label sequences for sequence z. CRF defines a conditional probability model p(y|z; W,b) under all possible label sequences, as shown below: in, represents the potential energy function, where and b y′,y represents the parameter matrix and bias term.

[0096] This embodiment uses the conditional maximum likelihood estimation method to train the conditional random field CRF. The following formula gives the training set {(z i ,y i )}, the larger the value of L(W,b), the better the training effect. L(W,b)=∑ i logp(y|z; W, b); The following formula represents the sequence y* with the highest conditional probability, y*=argmax y∈Φ(z) p(y|z;W,b).

[0097] S126. Input the word vector into the BI-GRU-CRF model for training, evaluate the trained BI-GRU-CRF model, and improve the BI-GRU-CRF model accordingly according to the evaluation results to obtain a self-attention mechanism model.

[0098] The word vector is input into the bidirectional GRU neural network model for training to obtain vocabulary feature information; the attention mechanism is added to the hidden state layer of the non-key information word part of the network output, and the weight is assigned. Finally, it passes through the CRF layer, calculates the joint probability and outputs the result; the BI-GRU-CRF model is evaluated and improved accordingly based on the evaluation results.

[0099] After obtaining the word vectors with richer semantic information output by the Chinese word embedding hybrid model, the present invention proposes a Bi-GRU-CRF keyword recognition model with a self-attention mechanism, which utilizes the GRU unit with a simpler structure and smaller computational complexity, and combines the self-attention mechanism to extract deeper features of the contextual text from multiple levels and angles to achieve effective recognition of small-scale training data sets.

[0100] S130, sending the health impact keyword to the second terminal, so that the second terminal evaluates the public policy document;

[0101] Specifically, the health impact keywords are provided for expert reference, and the experts evaluate the corresponding public policy texts based on the health impact keywords, thereby forming the evaluation content.

[0102] S140, obtaining evaluation content from the second terminal;

[0103] S150. Acquire matching related documents from a document abstract library according to the evaluation content to obtain an evaluation report.

[0104] In this embodiment, the evaluation report refers to the document that matches the evaluation content. The health impact data set and document abstract library are constructed by crawler technology and semantic similarity deduplication technology, and the text to be analyzed, i.e., the evaluation content, is matched to the most relevant N documents from the document abstract library through text similarity analysis.

[0105] In one embodiment, see Figure 5 , the above-mentioned step S150 may include steps S151 to S152.

[0106] S151. Construct a health impact dataset and literature abstract library through crawler technology and semantic similarity deduplication technology.

[0107] In this embodiment, the health impact data set refers to a data set consisting of health impact keywords; the literature abstract library refers to a database consisting of a large number of health impact-related literature abstracts crawled based on health impact keywords.

[0108] Specifically, we use crawler technology to crawl a large number of literature abstracts related to health impacts based on the health impact keywords, that is, the content of Table 1, and build a literature abstract library. In the process of screening literature abstracts, we found that there were many literature abstracts with repeated meanings or even the same content. This shows that there are many abstract texts with similar semantics in the literature abstract library. Therefore, we use the hybrid model of Jaro-Winkler and LSA for literature matching. When the semantic information and character information of two documents in the literature abstract library are similar, we perform deduplication. The framework of the hybrid model of Jaro-Winkler and LSA is as follows: Fig. 9 As shown. The Jaro-Winkler similarity formula is as follows: sim w =sim j +L×P(1-sim j ); LSA (latent semantic analysis) model is decomposed using SVD (singular value decomposition), and its formula is: D = UΣV T ; Among them, U is the matrix composed of eigenvectors, Σ is a diagonal matrix, and V is the matrix composed of eigenvectors.

[0109] S152. Filter out matching relevant documents in the document abstract library according to the evaluation content to obtain an evaluation report.

[0110] For the evaluation content and the document abstracts in the document abstract library, the evaluation content is usually a short text with less than 50 characters, while the document abstracts in the document abstract library are longer texts with 200-500 characters. For the similarity matching task between short text and long text, this embodiment proposes a KBS-GRU model, the model structure is as follows Fig.10 shown.

[0111] In one embodiment, see Figure 6 , the above-mentioned step S152 may include steps S1521 to S1525.

[0112] S1521. Perform word segmentation on the evaluation content and the text in the document abstract library respectively, and remove stop words to obtain two texts after word segmentation.

[0113] In this embodiment, the two texts after word segmentation refer to words formed by removing stop words after word segmentation processing of the evaluation content and the document abstracts in the document abstract library.

[0114] Specifically, Jieba is used to segment the evaluation content and the text in the document abstract library and to remove stop words.

[0115] S1522. Match the two texts after word segmentation and search for the same words as keywords.

[0116] S1523: Merge several words on the left and right of the keyword into the corresponding text to obtain input content.

[0117] In this embodiment, the input content refers to the content formed by merging a number of words on the left and right of the keyword into the corresponding text.

[0118] S1524: input the input content into the KBS-GRU model to perform text similarity analysis to obtain analysis results.

[0119] The analysis result refers to the textual similarity between the review content and different document elements in the literature abstract library.

[0120] In this embodiment, the Word2Vec model is used for word embedding representation, and the two embedded vector representations are used to learn the text representation information using the BiGRU-Attention model. Finally, the output vectors of the two sub-networks of the model are merged, and a fully connected neural network is used to learn deeper context information. Among them, BiGRU-Attention is the key part of this model. The GRU model is the same as the formula in the previous chapter, while the Attention model represents the importance of a word by adding weight to each word. Its formula is: i =tanh(W h h i +b h );

[0121] To solve the similarity matching problem between short text and long text, this paper proposes a KBS-GRU text similarity matching model, which uses multiple modules such as keyword matching technology, self-attention mechanism and fully connected neural network to extract deep semantic information in the text, and proves the reliability of the model on the processed WebMedQA dataset.

[0122] S1525. Determine an evaluation report based on the analysis results.

[0123] The system constructed according to the method of this embodiment is as follows Fig.11As shown in the figure, in this system, the whole process of health impact assessment can be represented as government staff uploading the original policy document and creating a new task with the document as the core, and then selecting appropriate experts in related fields in the system to distribute the task; expert users will receive task information after logging in. After receiving the task, the expert can open the original policy text online on the system and use the automatic identification of potential impact keywords to automatically display potential health impact keywords and their categories. On this basis, the transfer can be manually annotated, and the expert can select the text that he thinks may affect health and write the corresponding health impact. The system summarizes the health impact evaluation of the expert and generates a report. At the same time, this evaluation is stored in the database as a case. After the final report is generated, the government staff gets the evaluation report to complete the entire online health impact assessment process. Due to the recent introduction of the concept of microservices, considering that microservices can reduce the coupling between functional modules, improve the response speed of server-side request processing, and more importantly, facilitate the writing of code in subsequent software upgrades and new functional modules, this system adopts a microservice architecture for design and development. The health impact assessment auxiliary system based on microservice architecture is generally divided into three layers, namely the interaction layer, service layer and data layer. The interaction layer mainly uses Vue.js technology to obtain the JSON data transmitted by the back-end page service through the request sent by the front-end web page, including the new distribution tasks of government staff users. The distribution task operation includes processes such as uploading files, file descriptions, specifying experts, and checking the progress of ongoing tasks; the expert role mainly includes functions such as automatic keyword recognition, manual annotations, and automatic generation of evaluation reports. The service layer is mainly built through SpringCloud for unified management and deployment. The main services are evaluation summary services, user services, file services, data transmission services, gateway services, time services and other main functional services. They are classified by client and server, and their specific functions are as follows:

[0124] Eureka Client:

[0125] User service: provides member data interface for page service, including user addition, modification and login; File service: provides original policy document upload, download and online opening interface for page service; Data transmission service: provides interface for page service to convert original document into txt file format and input it into keyword recognition model and literature retrieval model for recognition, and returns the recognized specific text and category data; Page service: provides data interface for front-end page; Evaluation summary service: provides health impact evaluation summary interface in tabular form for page service; Time service: provides page service with unified time interface.

[0126] Eureka Server:

[0127] Registration center: All services are registered in the registration center, and their interface calls are centrally coordinated by the registration center; Configuration center: Provides the configuration required by all services in the system; Control center: Controls services and manages the start and stop operations of services.

[0128] The data layer uses Docker containers to encapsulate data storage tools including MySQL, Redis, RabbitMQ message queues, etc., which facilitates port management and calls. Finally, the project integrates the Git data warehouse, and their functions and effects are as follows:

[0129] MySQL database: stores information including user information, keyword library, health impact assessment case library, file information library, etc.; Redis database: stores session information to ensure session consistency between services and solve the problem of inconsistent session reading after cross-domain between the front-end and back-end; RabbitMQ message queue: ensures the unification of file upload and download threads under high traffic; Git data warehouse: project version control and code update to ensure the real-time and consistency of code development.

[0130] The system provides a platform for government workers and experts to participate in health impact assessment online. By using computer software related technologies and the artificial intelligence-assisted evaluation results designed by the present invention, the system realizes functions such as automatic identification of health impact keywords, keyword library expansion, file upload and download, manual annotation, and generation of health impact assessment reports, which is of practical significance.

[0131] In order to prove the reliability of the Bi-GRU-CRF model based on the self-attention mechanism in the artificial intelligence-assisted keyword extraction model, the policy documents of health impact assessment were tested. Since the current health impact assessment is carried out manually by relevant experts, there is no public data set. The data sets used in the present invention are all obtained by manual annotation by team members based on the standard health impact keywords given by experts. The files used are 22 documents involving environmental factors in policy documents issued by a government from 2017 to 2020. Since the length of each policy document varies, in order to facilitate model input, the present invention uniformly intercepts all files with sub-text segments of about 500 words, totaling 329 txt files. The present invention compares the currently popular model with the model of the present invention, and the F1 scores (%) of each model on the manually annotated health impact assessment keyword recognition task data set of the present invention are shown in Table 2.

[0132] Table 2. Comparison of manual annotation of health impact assessment keyword recognition tasks

[0133]

[0134] In order to prove the reliability of the text similarity analysis model KBS-GRU in the artificial intelligence-assisted literature retrieval model, this embodiment uses the processed Chinese medical quality assurance data WebMedQA dataset for testing. Since the model proposed in the present invention mainly focuses on the similarity between short texts and longer texts, this embodiment extracts question-answer pairs with question texts less than 50 characters from the WebMedQA dataset for testing. The present invention compares the currently popular models with the model of the present invention, and the F1 scores (%) and MAP (%) of each model on the processed Chinese medical quality assurance data WebMedQA dataset are shown in Table 3.

[0135] Table 3. Experimental comparison of text similarity analysis task on WebMedQA dataset

[0136] Model F1 score (%) MAP (%) BM25 58.1 49.2 Siamese-LSTM 70.3 67.7 Multi-CNN 72.4 68.8 Model of the present invention 74.2 70.3

[0137] The above-mentioned online health impact assessment method uses natural language processing technology in artificial intelligence to build a self-attention mechanism model to extract health impact keywords from public policy documents, and after obtaining the evaluation content of the second terminal, uses the KBS-GRU model in the document abstract library to perform text similarity analysis on the document abstract and the evaluation content to determine the evaluation report, thereby reducing the difficulty of review and facilitating the identification of potential health-impacting content in the document. It requires less manpower and has a short evaluation cycle, thereby ensuring the timeliness of the evaluation.

[0138] Fig.12 is a schematic block diagram of an online health impact assessment device 300 provided by an embodiment of the present invention. Fig.12 As shown, corresponding to the above online health impact assessment method, the present invention also provides an online health impact assessment device 300. The online health impact assessment device 300 includes a unit for executing the above online health impact assessment method, and the device can be configured in a server. Fig.12 The online health impact assessment device 300 includes a file acquisition unit 301, an identification unit 302, a sending unit 303, a content acquisition unit 304 and a matching unit 305.

[0139] The file acquisition unit 301 is used to acquire a public policy document to be evaluated issued by a government worker; the identification unit 302 is used to identify health impact keywords in the public policy document; the sending unit 303 is used to send the health impact keywords to the second terminal so that the second terminal can evaluate the public policy document; the content acquisition unit 304 is used to acquire evaluation content from the second terminal; the matching unit 305 is used to acquire matching related documents from the document abstract library according to the evaluation content to obtain an evaluation report.

[0140] In one embodiment, the recognition unit 302 is used to input the public policy document into a self-attention mechanism model for recognition to obtain health impact keywords.

[0141] In one embodiment, the online health impact assessment device 300 also includes a first model acquisition unit, which is used to train a BI-GRU-CRF model using relevant policy documents with keyword tags as a sample set to obtain a self-attention mechanism model.

[0142] In one embodiment, the first model acquisition unit includes a text acquisition subunit, a first word segmentation subunit, a tagging subunit, a vector acquisition subunit, a model construction subunit, and a training subunit.

[0143] The text acquisition subunit is used to acquire relevant policy texts and preprocess the policy texts to obtain policy texts; the first word segmentation subunit is used to segment the policy texts using a word segmentation tool to obtain word segmentation results; the labeling subunit is used to label the word segmentation results to obtain labeling results; the vector acquisition subunit is used to acquire corresponding word vectors for the labeling results; the model construction subunit is used to construct a BI-GRU-CRF model; the training subunit is used to input the word vectors into the BI-GRU-CRF model for training, and evaluate the trained BI-GRU-CRF model, and make corresponding improvements to the BI-GRU-CRF model according to the evaluation results to obtain a self-attention mechanism model.

[0144] In one embodiment, the vector acquisition subunit includes a dictionary acquisition module and a dictionary matching module.

[0145] The dictionary acquisition module is used to acquire the word embedding vector dictionary; the dictionary matching module is used to match the annotation result with the word embedding vector dictionary to obtain the corresponding word vector.

[0146] In one embodiment, if Fig.13 As shown, the matching unit 305 includes a database building subunit 3051 and a screening subunit 3052 .

[0147] The database construction subunit 3051 is used to construct a health impact data set and a literature abstract library through crawler technology and semantic similarity deduplication technology; the screening subunit 3052 is used to screen out matching relevant documents in the literature abstract library according to the evaluation content to obtain an evaluation report.

[0148] In one embodiment, if Fig.14As shown, the screening subunit 3052 includes a second word segmentation module 30521 , a word segmentation matching module 30522 , a merging module 30523 , an analysis module 30524 and a report determination module 30525 .

[0149] The second word segmentation module 30521 is used to perform word segmentation on the evaluation content and the text in the document abstract library respectively, and remove stop words to obtain two texts after word segmentation; the word segmentation matching module 30522 is used to match the two texts after word segmentation and find the same words as keywords; the merging module 30523 is used to merge several words on the left and right of the keyword into the corresponding text to obtain input content; the analysis module 30524 is used to input the input content into the KBS-GRU model for text similarity analysis to obtain analysis results; the report determination module 30525 is used to determine the evaluation report according to the analysis results.

[0150] It should be noted that those skilled in the art can clearly understand that the specific implementation process of the above-mentioned online health impact assessment device 300 and each unit can refer to the corresponding description in the aforementioned method embodiment, and for the convenience and brevity of description, it will not be repeated here.

[0151] The above-mentioned online health impact assessment device 300 can be implemented in the form of a computer program. The computer program can be used in Fig.15 Runs on the computer device shown.

[0152] See also Fig.15 , Fig.15 5 is a schematic block diagram of a computer device provided in an embodiment of the present application. The computer device 500 may be a server, wherein the server may be an independent server or a server cluster composed of multiple servers.

[0153] See also Fig.15 The computer device 500 includes a processor 502 , a memory and a network interface 505 connected via a system bus 501 , wherein the memory may include a non-volatile storage medium 503 and an internal memory 504 .

[0154] The non-volatile storage medium 503 can store an operating system 5031 and a computer program 5032. The computer program 5032 includes program instructions, and when the program instructions are executed, the processor 502 can execute an online health impact assessment method.

[0155] The processor 502 is used to provide computing and control capabilities to support the operation of the entire computer device 500 .

[0156] The internal memory 504 provides an environment for the operation of the computer program 5032 in the non-volatile storage medium 503. When the computer program 5032 is executed by the processor 502, the processor 502 can execute an online health impact assessment method.

[0157] The network interface 505 is used to communicate with other devices over the network. Fig.15 The structure shown in the figure is merely a block diagram of a portion of the structure related to the solution of the present application, and does not constitute a limitation on the computer device 500 to which the solution of the present application is applied. The specific computer device 500 may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0158] The processor 502 is used to run the computer program 5032 stored in the memory to implement the following steps:

[0159] Obtain a public policy document to be evaluated issued by a government worker; identify health impact keywords in the public policy document; send the health impact keywords to a second terminal so that the second terminal evaluates the public policy document; obtain evaluation content from the second terminal; obtain matching related documents from a document abstract library based on the evaluation content to obtain an evaluation report.

[0160] In one embodiment, when implementing the step of identifying the health impact keywords in the public policy document, the processor 502 specifically implements the following steps:

[0161] The public policy document is input into the self-attention mechanism model for identification to obtain health impact keywords; wherein the self-attention mechanism model is obtained by training the BI-GRU-CRF model using relevant policy documents with keyword labels as a sample set.

[0162] In one embodiment, when the processor 502 implements the step of obtaining the self-attention mechanism model by training the BI-GRU-CRF model using the relevant policy documents with keyword tags as the sample set, the processor 502 specifically implements the following steps:

[0163] Obtain relevant policy texts and preprocess the policy texts to obtain policy texts; use a word segmentation tool to segment the policy texts to obtain word segmentation results; label the word segmentation results to obtain labeling results; obtain corresponding word vectors for the labeling results; build a BI-GRU-CRF model; input the word vectors into the BI-GRU-CRF model for training, and evaluate the trained BI-GRU-CRF model, and make corresponding improvements to the BI-GRU-CRF model according to the evaluation results to obtain a self-attention mechanism model.

[0164] The BI-GRU-CRF model includes a bidirectional GRU neural network model, a self-attention mechanism layer, a CRF layer and an output layer.

[0165] In one embodiment, when the processor 502 implements the step of obtaining the corresponding word vector for the annotation result, the processor 502 specifically implements the following steps:

[0166] Obtain a word embedding vector dictionary; match the annotation result with the word embedding vector dictionary to obtain a corresponding word vector.

[0167] In one embodiment, when the processor 502 implements the step of acquiring matching related documents from the document abstract library according to the evaluation content to obtain the evaluation report, the processor 502 specifically implements the following steps:

[0168] A health impact data set and a literature abstract library are constructed by using crawler technology and semantic similarity deduplication technology; and matching relevant literature in the literature abstract library is screened out according to the evaluation content to obtain an evaluation report.

[0169] In one embodiment, when the processor 502 implements the step of screening out matching related documents in the document abstract library according to the evaluation content to obtain an evaluation report, the processor 502 specifically implements the following steps:

[0170] The evaluation content and the text in the literature abstract library are segmented respectively, and stop words are removed to obtain two texts after segmentation; the two texts after segmentation are matched to find the same words as keywords; several words on the left and right of the keywords are merged into the corresponding text to obtain input content; the input content is input into the KBS-GRU model for text similarity analysis to obtain analysis results; and the evaluation report is determined according to the analysis results.

[0171] It should be understood that in the embodiment of the present application, the processor 502 may be a central processing unit (CPU), and the processor 502 may also be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among them, the general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0172] It can be understood by those skilled in the art that all or part of the processes in the method for implementing the above embodiment can be completed by instructing the relevant hardware through a computer program. The computer program includes program instructions, and the computer program can be stored in a storage medium, which is a computer-readable storage medium. The program instructions are executed by at least one processor in the computer system to implement the process steps of the embodiment of the above method.

[0173] Therefore, the present invention also provides a storage medium. The storage medium may be a computer-readable storage medium. The storage medium stores a computer program, wherein when the computer program is executed by a processor, the processor executes the following steps:

[0174] Obtain a public policy document to be evaluated issued by a government worker; identify health impact keywords in the public policy document; send the health impact keywords to a second terminal so that the second terminal evaluates the public policy document; obtain evaluation content from the second terminal; obtain matching related documents from a document abstract library based on the evaluation content to obtain an evaluation report.

[0175] In one embodiment, when the processor executes the computer program to implement the step of identifying the health impact keywords in the public policy document, the processor specifically implements the following steps:

[0176] Inputting the public policy document into the self-attention mechanism model for identification to obtain health impact keywords;

[0177] Among them, the self-attention mechanism model is obtained by training the BI-GRU-CRF model by using relevant policy documents with keyword labels as the sample set.

[0178] In one embodiment, when the processor executes the computer program to implement the step that the self-attention mechanism model is obtained by training the BI-GRU-CRF model with relevant policy documents with keyword tags as a sample set, the processor specifically implements the following steps:

[0179] Obtain relevant policy texts and preprocess the policy texts to obtain policy texts; use a word segmentation tool to segment the policy texts to obtain word segmentation results; label the word segmentation results to obtain labeling results; obtain corresponding word vectors for the labeling results; build a BI-GRU-CRF model; input the word vectors into the BI-GRU-CRF model for training, and evaluate the trained BI-GRU-CRF model, and make corresponding improvements to the BI-GRU-CRF model according to the evaluation results to obtain a self-attention mechanism model.

[0180] The BI-GRU-CRF model includes a bidirectional GRU neural network model, a self-attention mechanism layer, a CRF layer and an output layer.

[0181] In one embodiment, when the processor executes the computer program to implement the step of obtaining the corresponding word vector for the annotation result, the processor specifically implements the following steps:

[0182] Obtain a word embedding vector dictionary; match the annotation result with the word embedding vector dictionary to obtain a corresponding word vector.

[0183] In one embodiment, when the processor executes the computer program to implement the step of obtaining matching related documents from the document abstract library according to the evaluation content to obtain an evaluation report, the processor specifically implements the following steps:

[0184] A health impact data set and a literature abstract library are constructed by using crawler technology and semantic similarity deduplication technology; and matching relevant literature in the literature abstract library is screened out according to the evaluation content to obtain an evaluation report.

[0185] In one embodiment, when the processor executes the computer program to implement the step of screening out matching relevant documents in the document abstract library according to the evaluation content to obtain an evaluation report, the processor specifically implements the following steps:

[0186] The evaluation content and the text in the literature abstract library are segmented respectively, and stop words are removed to obtain two texts after segmentation; the two texts after segmentation are matched to find the same words as keywords; several words on the left and right of the keywords are merged into the corresponding text to obtain input content; the input content is input into the KBS-GRU model for text similarity analysis to obtain analysis results; and the evaluation report is determined according to the analysis results.

[0187] The storage medium may be a USB flash drive, a mobile hard disk, a read-only memory (ROM), a magnetic disk, or an optical disk, etc., which are computer-readable storage media that can store program codes.

[0188] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in terms of function in the above description. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0189] In the several embodiments provided by the present invention, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of each unit is only a logical function division, and there may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed.

[0190] The steps in the method of the embodiment of the present invention can be adjusted in order, combined and deleted according to actual needs. The units in the device of the embodiment of the present invention can be combined, divided and deleted according to actual needs. In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0191] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, terminal, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention.

[0192] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of various equivalent modifications or replacements within the technical scope disclosed by the present invention, and these modifications or replacements should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. An online health impact assessment method, characterized in that: include: Access public policy documents issued by government staff for review; Identify health impact keywords within said public policy documents; Sending the health impact keyword to a second terminal so that the second terminal can evaluate the public policy document; Acquiring evaluation content from the second terminal; Acquire matching relevant documents from the literature abstract library according to the evaluation content to obtain an evaluation report; The identification of health impact keywords in the public policy document includes: Inputting the public policy document into the self-attention mechanism model for identification to obtain health impact keywords; The self-attention mechanism model is obtained by training the BI-GRU-CRF model using relevant policy documents with keyword tags as sample sets; The self-attention mechanism model is obtained by training the BI-GRU-CRF model using relevant policy documents with keyword labels as sample sets, including: Obtain relevant policy texts, and preprocess the policy texts to obtain policy texts; Using a word segmentation tool to segment the policy text to obtain a word segmentation result; Labeling the word segmentation results to obtain labeling results; Obtaining a corresponding word vector for the annotation result; Build a BI-GRU-CRF model; The word vector is input into the BI-GRU-CRF model for training, and the trained BI-GRU-CRF model is evaluated, and the BI-GRU-CRF model is improved accordingly according to the evaluation results to obtain a self-attention mechanism model; The BI-GRU-CRF model includes a bidirectional GRU neural network model, a self-attention mechanism layer, a CRF layer and an output layer; The step of obtaining matching relevant documents from a document abstract library according to the evaluation content to obtain an evaluation report includes: Build a health impact dataset and literature summary library through crawler technology and semantic similarity deduplication technology; Screening out matching relevant documents in the literature abstract library according to the evaluation content to obtain an evaluation report; The step of screening out matching relevant documents in the document abstract library according to the evaluation content to obtain an evaluation report includes: Performing word segmentation processing on the evaluation content and the text in the literature abstract library respectively, and removing stop words to obtain two texts after word segmentation; Match the two texts after word segmentation and find the same words as keywords; Merge the words on the left and right of the keyword into the corresponding text to obtain input content; Input the input content into the KBS-GRU model to perform text similarity analysis to obtain an analysis result; An evaluation report is determined according to the analysis results.

2. The online health impact assessment method according to claim 1, characterized in that: The obtaining a corresponding word vector for the annotation result includes: Get the word embedding vector dictionary; The annotation result is matched with the word embedding vector dictionary to obtain the corresponding word vector.

3. An online health impact assessment device, characterized in that: include: A document acquisition unit, used to acquire public policy documents to be evaluated issued by government staff; an identification unit, for identifying health impact keywords in the public policy document; A sending unit, configured to send the health impact keyword to a second terminal, so that the second terminal can evaluate the public policy document; A content acquisition unit, used to acquire evaluation content from the second terminal; A matching unit, used to obtain matching related documents from a document abstract library according to the evaluation content to obtain an evaluation report; The identification of health impact keywords in the public policy document includes: Inputting the public policy document into the self-attention mechanism model for identification to obtain health impact keywords; The self-attention mechanism model is obtained by training the BI-GRU-CRF model using relevant policy documents with keyword tags as sample sets; The self-attention mechanism model is obtained by training the BI-GRU-CRF model using relevant policy documents with keyword labels as sample sets, including: Obtain relevant policy texts, and preprocess the policy texts to obtain policy texts; Using a word segmentation tool to segment the policy text to obtain a word segmentation result; Labeling the word segmentation results to obtain labeling results; Obtaining a corresponding word vector for the annotation result; Build a BI-GRU-CRF model; The word vector is input into the BI-GRU-CRF model for training, and the trained BI-GRU-CRF model is evaluated, and the BI-GRU-CRF model is improved accordingly according to the evaluation results to obtain a self-attention mechanism model; The BI-GRU-CRF model includes a bidirectional GRU neural network model, a self-attention mechanism layer, a CRF layer and an output layer; The step of obtaining matching relevant documents from a document abstract library according to the evaluation content to obtain an evaluation report includes: Build a health impact dataset and literature summary library through crawler technology and semantic similarity deduplication technology; Screening out matching relevant documents in the document abstract library according to the evaluation content to obtain an evaluation report; The step of screening out matching relevant documents in the document abstract library according to the evaluation content to obtain an evaluation report includes: Performing word segmentation processing on the evaluation content and the text in the literature abstract library respectively, and removing stop words to obtain two texts after word segmentation; Match the two texts after word segmentation and find the same words as keywords; Merge the words on the left and right of the keyword into the corresponding text to obtain input content; Input the input content into the KBS-GRU model to perform text similarity analysis to obtain an analysis result; An evaluation report is determined according to the analysis results.

4. A computer device, characterized in that: The computer device comprises a memory and a processor, the memory stores a computer program, and the processor implements the method according to any one of claims 1 to 2 when executing the computer program.

5. A storage medium, characterized in that: The storage medium stores a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 2 can be implemented.

Images