Enterprise brand reputation intelligent diagnosis management system, electronic equipment and medium

Abstract

The invention provides an enterprise brand reputation intelligent diagnosis management system, electronic equipment and a medium, and the system comprises a data collection module which is configured to collect basic semantic data which comprises a plurality of first text vocabularies; the data processing module is configured to receive the basic semantic data and label the first text vocabulary based on a preset labeling strategy; the data processing module is further configured to extract a first entity of the basic semantic data; the analysis module is configured to receive the labels, obtain semantics of the first text vocabularies corresponding to the labels on the basis of a semantic model, determine whether the semantics are positive or negative, and analyze the labels according to the number of the first text vocabularies corresponding to the negative and on the basis of a preset model; obtaining the increment of the quantity of the basic semantic data corresponding to the negative values in a future preset time period; and the diagnosis decision module is configured to output a decision strategy based on a preset case library according to the increment and the first entity.

Description

Technical Field

[0001] This invention relates to the field of corporate brand reputation management technology, and more specifically, to a corporate brand reputation intelligent diagnostic management system, electronic device, and medium. Background Technology

[0002] The core pain point in corporate reputation management lies in the inability to accurately pinpoint the root cause of a crisis and make predictive decisions, specifically manifested in:

[0003] ① Information noise drowns out key issues. Traditional public opinion systems lack the ability to label dimensions, mixing up product quality complaints (such as "abnormal brake noise"), service complaints (such as "customer service shirking responsibility"), and negative brand information (such as "false advertising"), making it difficult for companies to identify the core drivers of reputation decline.

[0004] ② Lack of entity association. General sentiment analysis only outputs "positive / negative" statistics, which cannot link negative public opinion with specific entities, resulting in resource misallocation.

[0005] ③ Passive response is delayed. Relying on manual statistics of changes in negative sentiment makes it impossible to quantify and predict future growth trends (such as predicting that the topic of "brake failure" will spread by 200% within 48 hours), thus missing the golden window for intervention. Summary of the Invention

[0006] In view of this, one of the objectives of this application is to provide an intelligent diagnostic management system for corporate brand reputation, which can improve the problem of being unable to accurately locate the root cause of a crisis and make predictive decisions.

[0007] To achieve the above technical objectives, the technical solution adopted in this application is as follows:

[0008] In a first aspect, embodiments of this application provide an intelligent diagnostic management system for enterprise brand reputation, including:

[0009] The data acquisition module is configured to collect basic semantic data, which includes several first text words;

[0010] The data processing module is configured to receive the basic semantic data and annotate the first text words based on a preset annotation strategy. The annotations include product quality, after-sales service, and / or brand image. The data processing module is also configured to extract the first entity from the basic semantic data. The first entity includes the brand name and / or product name.

[0011] The analysis module is configured to receive the annotation, obtain the semantics of the first text word corresponding to the annotation based on the semantic model, determine whether the semantics is positive or negative, and obtain the growth rate of the amount of basic semantic data corresponding to the negative within a future preset time period based on the number of first text words corresponding to the negative.

[0012] The diagnostic decision module is configured to output a decision strategy based on the growth amount and the first entity, using a preset case library.

[0013] Furthermore, the data processing module is also configured as follows:

[0014] Using the first text vocabulary, a second text vocabulary with the same semantics as the first text vocabulary is determined in a preset vocabulary library. The annotation is determined based on the second text vocabulary. The preset vocabulary library includes several first mappings, and the first mappings represent the correspondence between the second text vocabulary and the annotation.

[0015] Furthermore, the data processing module is also configured to input the basic semantic data into the trained BERT-Base Chinese pre-trained model so that the trained BERT-Base Chinese pre-trained model can extract the first entity from the basic semantic data.

[0016] Furthermore, the analysis module is also configured as follows:

[0017] The basic semantic data with the annotations is input into the semantic model so that the semantic model determines the semantics of the first text word based on the type of the annotations according to its built-in semantic database. The semantic database includes several second mappings, which are used to represent the correspondence between the type of annotation, the first text word and the semantics.

[0018] Furthermore, the data acquisition module is also configured to acquire time-series data corresponding to the basic semantic data;

[0019] The analysis module is also configured to:

[0020] Receive the timing data;

[0021] The time-series data and the number of corresponding first text words with negative semantics are input into the trained LSTM model so that the trained LSTM model outputs the growth amount.

[0022] Furthermore, the diagnostic decision module is also configured as follows:

[0023] Based on the growth rate, the severity level of the accident is determined according to preset rules;

[0024] Based on the accident severity level and the first entity, a decision strategy is obtained by querying the preset case library. The preset case library includes several third mappings, which are used to represent the correspondence between the accident severity level, the second entity, the label corresponding to the growth amount, and the decision strategy. The similarity between the second entity and the first entity is greater than the similarity threshold.

[0025] Furthermore, the analysis module is also configured as follows:

[0026] Receive the first entity, output the semantic vector of the first entity based on the semantic model, and input the semantic vector of the first entity into the diagnostic decision module;

[0027] The diagnostic decision module is further configured to: calculate the similarity between the semantic vector of the first entity and the semantic vectors of all the second entities, and obtain the semantic vector of the second entity whose similarity is greater than the similarity threshold.

[0028] An electronic device includes a processor and a memory coupled to each other, wherein the memory stores the system described above.

[0029] A computer-readable storage medium storing the above-described system.

[0030] The invention employing the above technical solution has the following advantages:

[0031] In the technical solution provided in this application, a data processing module is set up after the data acquisition module. This module annotates the first text words to separate mixed information. Simultaneously, the data processing module extracts the first entity to pinpoint the specific target to be managed. The analysis module determines the semantics of the annotated first text words, thereby determining the number of negative first text words and predicting their growth rate. Finally, a diagnostic decision-making module outputs a decision-making strategy. Through this system, the first text words generating negative emotions are categorized into specific product quality, after-sales service, or brand image, accurately locating the root cause of the crisis. Furthermore, after extracting the first entity and predicting its growth rate, the system outputs a decision-making strategy based on the first entity, its growth rate, and the annotations, improving the accuracy of the output strategy and overcoming the problem of not being able to accurately locate the root cause of the crisis and make predictive decisions. Attached Figure Description

[0032] This application can be further illustrated by the non-limiting embodiments given in the accompanying drawings. It should be understood that the following drawings only illustrate certain embodiments of this application and should not be considered as limiting the scope. For those skilled in the art, other related drawings can be obtained from these drawings without any inventive effort.

[0033] Figure 1A block diagram of the system provided in the embodiments of this application.

[0034] Figure 2 The system architecture and data flow diagram provided for the embodiments of this application.

[0035] Figure 3 A flowchart of a system-based method provided in an embodiment of this application.

[0036] Icons: 100 - Data Acquisition Module; 200 - Data Processing Module; 300 - Analysis Module; 400 - Diagnostic Decision Module. Detailed Implementation

[0037] The present application will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that similar or identical parts are referred to by the same reference numerals in the drawings or description. Implementations not shown or described in the drawings are forms known to those skilled in the art. In the description of this application, terms such as "first" and "second" are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0038] This application provides an electronic device that may include a processing module and a storage module. The storage module stores the following enterprise brand reputation intelligent diagnostic management system.

[0039] Please refer to Figure 1 and Figure 2 This application also provides an intelligent diagnostic management system for enterprise brand reputation. The intelligent diagnostic management system for enterprise brand reputation includes a data acquisition module 100, a data processing module 200, an analysis module 300, and a diagnostic decision module 400. The acquisition module 100, data processing module 200, analysis module 300, and diagnostic decision module 400 refer to software functional modules that can be stored in the memory of electronic devices / servers or embedded in the operating system (OS) in the form of software or firmware.

[0040] The data acquisition module 100 is configured to collect basic semantic data, which includes several first text words.

[0041] The basic semantic data can come from internal data, such as the company's official website, product reviews, customer service feedback, sales data, and official social media accounts.

[0042] External data, such as public opinion on platforms like Weibo, WeChat, Douyin, and Xiaohongshu.

[0043] News media, such as mainstream news websites and industry media reports.

[0044] User reviews on e-commerce platforms such as Taobao, JD.com, and Pinduoduo.

[0045] Forums and Q&A platforms, such as Zhihu, Douban groups, and industry forums.

[0046] Regulatory and third-party sources, such as government announcements, industry reports, and consumer association complaint data.

[0047] Competitive data, such as competitor reputation dynamics and user comparison reviews.

[0048] Basic semantic data can be understood as user comments on news media and e-commerce platforms. For example, a comment on an e-commerce platform might say, "The clothes from brand XX are really nice!", or a comment on a forum might say, "The chassis of brand XX is unstable" or "Brand XX's after-sales service is shirking responsibility."

[0049] The first text term refers to the text terms that constitute the basic semantic data. These terms can be Chinese, English, or other languages. Taking Chinese as an example, in this embodiment, the first text term can be an adjective or adverb describing a certain entity in the basic semantic data. For instance, in the basic semantic data “XX brand clothes are really nice!”, “XX brand clothes” is the entity, and “really nice” is the first text term, used to describe how the style of XX brand clothes satisfies the user. Similarly, in “XX brand’s after-sales service exhibits a tendency to shirk responsibility”, “XX brand’s after-sales service” is the entity, and “troublemaking” is the first text term, indicating that the after-sales service of XX brand is unsatisfactory to the user.

[0050] In this embodiment, basic semantic data can be collected by using web crawler (distributed crawler framework, such as Scrapy+Redis) API interface (social media open platform, e-commerce data interface) and enterprise internal system data interface (CRM, ERP, OA system), in real time (public opinion hotspots) and at regular intervals (regular data).

[0051] The data processing module 200 is configured to receive basic semantic data and, based on a preset annotation strategy, annotate the first text words, including product quality, after-sales service, and / or brand image.

[0052] The preset annotation strategy can be as follows: using the first text vocabulary, determine the second text vocabulary with the same semantics as the first text vocabulary in the preset vocabulary library, and determine the annotation based on the second text vocabulary. The preset vocabulary library includes several first mappings, and the first mappings represent the correspondence between the second text vocabulary and the annotations.

[0053] Based on the above, the first text vocabulary in this embodiment is an adjective or adverb used to describe entities. Therefore, according to a preset vocabulary library, the first text vocabulary is labeled and categorized as product quality, after-sales service, or brand image. For example, the phrase "XX brand's after-sales service has a tendency to shirk responsibility" essentially indicates that XX brand's after-sales service has problems, so "shirk responsibility" is labeled as a tag to characterize after-sales service. Similarly, the phrase "XX car's chassis is unstable" essentially indicates that XX car has product quality problems, so "chassis instability" is labeled as a tag to characterize product quality.

[0054] In this embodiment, the applicant can set up a preset vocabulary library according to the type of the enterprise. By querying existing publicly available online datasets, the applicant can obtain adjectives that are generally used to describe the products produced by the enterprise. For example, if the enterprise's main products are mobile phones and digital products, then positive adjectives used to describe mobile phones and digital products are generally "smooth", "long standby time" and "professional after-sales service", while negative adjectives are generally "lag". Therefore, the applicant can integrate the labels according to the product type of the enterprise to obtain the preset vocabulary library.

[0055] In this embodiment, since netizens' comments online are related to their personal expression, for example, to describe a poor-quality phone, they generally use words like "laggy," while some comments prefer to say the opposite, such as "This phone isn't laggy at all, it just takes two hours to boot up," which actually means the phone is laggy. Based on this, a pre-trained large language model can be set up to select second text words in a preset vocabulary library that are semantically similar to the first text words. For example, the pre-trained large language model traverses all second text words until it selects a second text word that has the same semantics as the first text word. Here, semantic similarity can be understood as the similarity between the semantic vectors of the first text word and the semantic vectors of the second text word reaching a threshold. The threshold can be set manually according to the actual situation.

[0056] The data processing module 200 is also configured to extract a first entity of basic semantic data, the first entity including the brand name and / or product name.

[0057] The large language model for semantic recognition mentioned in this embodiment can be trained using existing methods, such as:

[0058] To train this semantic mapping large language model, a domain-adaptive pre-training and irony-enhanced fine-tuning strategy is adopted: First, the model is initialized based on general Chinese corpus, and then vertical domain text (including ironic sentences and internet slang) is injected. Semantic understanding is optimized through masking training and contrastive learning. Then, a manually constructed irony-true intent pair is used for fine-tuning, and an irony detection head and semantic decoupling loss are incorporated to enable the model to separate literal meaning from true emotion (e.g., mapping "booting up for 2 hours" to "lag"). Finally, through dynamic lexicon alignment technology, non-standard expression vectors are forced to approximate the standard words in the preset lexicon (e.g., mapping "fire dragon" to "overheating"), and an online learning mechanism is deployed for continuous optimization to ensure that the system accurately parses complex expressions such as saying the opposite of what is meant.

[0059] In this embodiment, the data processing module 200 is further configured to extract a first entity of basic semantic data, the first entity including brand name and / or product name.

[0060] In this embodiment, the extraction of the first entity is achieved in the following manner:

[0061] The basic semantic data is input into the trained BERT-Base Chinese pre-trained model so that the trained BERT-Base Chinese pre-trained model can extract the first entity from the basic semantic data.

[0062] The BERT-Base Chinese pre-trained model is based on a 12-layer stacked architecture with a Transformer encoder core structure. Each layer includes a self-attention mechanism and a feedforward neural network module, and performs deep semantic representation of each input character through a 768-dimensional hidden layer. The model is also configured with 12 parallel attention heads to build multi-subspace semantic understanding capabilities and supports sequence processing lengths of up to 512 characters. For the characteristics of Chinese tasks, its input layer uses Chinese characters as the basic processing unit, inserts [CLS] markers at the beginning of the sequence to obtain global semantics, sets [SEP] separators at the end of sentences or between sentences, uses [MASK] markers during training to achieve character-level prediction, and employs learnable position vectors to capture the strict sequential relationships of Chinese character sequences. To enhance entity recognition capabilities, a sequence labeling output head is added to the top of the basic encoder, using the BIOES labeling system (e.g., "B-Brand" for the beginning of the brand name, "I-Product" for the middle section of the product name) to achieve character-by-character entity label prediction of the input sequence. For example, when you input "[CLS]Star Chaser Lumina Brake Noise [SEP]", the output will be the tag sequence "B-Brand I-Brand I-Product OO".

[0063] The model is optimized through a multi-stage training strategy: In the pre-training stage, a masked language modeling objective is used to randomly mask 15% of Chinese characters (e.g., "Xingchi [MASK]umina → Lumina"), and the whole-word masking technique is enabled in the Chinese scenario to ensure entity integrity (masking both "Xingchi" characters simultaneously), and character-radical joint encoding is combined to enhance the glyph parsing ability; In the domain adaptation stage, vertical corpora (product reviews / social media / customer service conversations) are injected, the masking probability of brand entities is increased to 25% and an entity continuity loss constraint is added (forcing the continuous probability of B-Brand and I-Brand > 90%); In the fine-tuning stage, a composite loss function is designed based on 500,000 labeled texts, and the CRF is used to constrain the logical transition of labels (blocking illegal jumps from O to I-Brand) and the boundary loss penalizes entity splitting errors (e.g., ensuring "Lumina Pro" as a single entity); Finally, adversarial training is performed in batches of 256 on 8×A100 hardware (injecting misspelled words such as "Tesla"), completing one million steps of pre-training and three rounds of fine-tuning, achieving a double breakthrough in entity recognition robustness and boundary accuracy.

[0064] In this embodiment, the first entity in the basic semantic data is extracted through the trained BERT-Base Chinese pre-trained model.

[0065] In this embodiment, the analysis module 300 is configured to receive annotations, obtain the semantics of the first text vocabulary corresponding to the annotations based on the semantic model, determine whether the semantics are positive or negative, and based on the number of the first text vocabulary corresponding to the negative, obtain the growth amount of the number of the basic semantic data corresponding to the negative in a future preset time period based on a preset model.

[0066] In this embodiment, for the first text vocabulary, it is necessary to distinguish whether it is positive or negative. This embodiment especially targets the negative first text vocabulary. When there are too many negative evaluations of the products output by an enterprise, strategies are made according to the types targeted by the negative evaluations.

[0067] According to the above, if the first text vocabulary is an adjective or an adverb, first it is necessary to determine whether the emotion expressed by the first text vocabulary is positive or negative, and then classify the expressions of the negative first text vocabulary into product quality, after-sales service or brand image, determine the number of negative evaluations, and predict the growth amount of the negative evaluations after a preset time period.

[0068] In this embodiment, the semantic database can be pre-set in the semantic model. The semantic database can pre-set multiple words, each word providing its semantic meaning under its corresponding label. For example, in the labeling of product quality, "excellent" might mean the product quality is good, a positive evaluation. However, in the labeling of after-sales service, "excellent" might mean the after-sales staff's attitude is too rigid, a negative evaluation. Therefore, the emotional expression type needs to be determined in conjunction with the label. In this embodiment, the semantic database can be an Excel spreadsheet storing multiple second mappings. The semantic model can serve as a software tool including a retrieval function, retrieving the emotion corresponding to the label and the first text word—whether positive or negative—based on the obtained first text word.

[0069] In this embodiment, the data acquisition module 200 is also configured to acquire time-series data corresponding to the basic semantic data. The time-series data represents the time when the basic semantic data is generated. For example, when acquiring basic semantic data on a forum, the time when the basic semantic data is published will be acquired simultaneously. Since the data acquisition module 200 acquires data at a fixed frequency, the acquisition time and the publication time are not significantly different.

[0070] Analysis module 300 is also configured as follows:

[0071] Receive the timing data;

[0072] The time-series data and the number of corresponding first text words with negative semantics are input into the trained LSTM model so that the trained LSTM model outputs the growth amount.

[0073] This LSTM model employs a two-layer hidden layer architecture with 128 neurons per layer. The input layer receives 168 hours of continuous time-series data (hourly negative word count vectors), which are processed cyclically through a memory unit consisting of a forget gate, an input gate, and an output gate to capture the periodic fluctuations in public opinion. The training process uses a sliding window to generate sequence samples, employs mean squared error as the loss function in conjunction with the Adam optimizer, avoids overfitting through temporal cross-validation, and injects adversarial perturbations (such as abnormal traffic surges during holidays) into real social media datasets to enhance the model's robustness to sudden events. Ultimately, it achieves high-precision prediction of the growth of negative words in the next 3 hours.

[0074] The preset model in this embodiment can be an LSTM model.

[0075] The diagnostic decision module 400 is configured to derive a decision strategy based on the growth rate and the first entity, using a pre-defined case library. Specifically, it determines the severity level of the accident based on the growth rate and pre-defined rules.

[0076] Based on the severity level of the accident and the first entity, a decision strategy is obtained by querying a preset case library. The preset case library includes several third mappings, which are used to represent the correspondence between the severity level of the accident, the second entity, and the decision strategy. The similarity between the second entity and the first entity is greater than the similarity threshold.

[0077] In this embodiment, the preset rules can be in tabular form, recording the correspondence between the growth rate and the severity level of the incident. When the number of negative first text words predicted by the LSTM model is large, the severity level of the incident is high, and a more timely strategy should be adopted; when the number of negative first text words predicted by the LSTM model is small, the severity level of the incident is low, and a more moderate strategy should be adopted. Therefore, this embodiment does not limit the number of intervals for the growth rate or the number of incident severity levels.

[0078] In this embodiment, the preset case library contains multiple cases, each case is labeled with the appropriate accident severity level, label, and corresponding second entity.

[0079] The second entity can be the object corresponding to the main body of the case. For example, if Company A proposes a global recall plan for brake pedal failure accidents, the strategy is "global recall plan", the second entity is "Company A", and the overall case represents the type of "product quality".

[0080] Therefore, in this embodiment, there are different case descriptions in the preset case library for different accident levels, different second entities, and different labels. By obtaining information that the first entity and the second entity should belong to the same type of enterprise, such as both being automobile companies or both being cosmetic manufacturing companies, after determining the second entity with a similarity greater than the similarity threshold with the first entity, as well as the accident severity level and label, a strategy is output according to the preset case library.

[0081] The analysis module 300 is also configured as follows:

[0082] Receive the first entity, output the semantic vector of the first entity based on the semantic model, and input the semantic vector of the first entity into the diagnostic decision module 400;

[0083] The diagnostic decision module 400 is also configured to: calculate the similarity between the semantic vector of the first entity and the semantic vectors of all second entities, and obtain the semantic vector of the second entity whose similarity is greater than the similarity threshold.

[0084] The above method can be used to filter out second entities that are similar to the first entity.

[0085] The first entity vector and the second entity vector can both be calculated using the BERT model; the specific calculation process will not be explained here.

[0086] Based on the above description of each layer of the system, such as Figure 3 As shown, the specific implementation method based on this system can be as follows:

[0087] S1: The data acquisition module collects basic semantic data;

[0088] S2: The data processing module receives basic semantic data and annotates the first text words based on a preset annotation strategy;

[0089] S3: The data processing module extracts the first entity from the basic semantic data and transmits the first entity to the diagnostic decision module;

[0090] S4: The analysis module receives the annotations, obtains the semantics of the first text words corresponding to the annotations based on the semantic model, and obtains the growth rate of the basic semantic data corresponding to the negative in the future preset time period based on the number of the first text words corresponding to the negative, and transmits the growth rate to the diagnostic decision module.

[0091] S5: The diagnostic decision module outputs a decision strategy based on the growth rate and the first entity, using a pre-defined case library.

[0092] In this embodiment, the "intelligence" of the system is reflected in the fact that based on the data acquisition module 100, data processing module 200, analysis module 300 and diagnostic decision module 400, the above-mentioned S1-S5 can be automatically operated, that is, the decision strategy can be automatically output, and at the same time, based on the basic semantic data, suitable decision strategies can be recommended, which has the characteristics of intelligence.

[0093] In this embodiment, the processing module can be an integrated circuit chip with signal processing capabilities. The processing module can be a general-purpose processor. For example, the processor can be a Central Processing Unit (CPU), a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components, capable of implementing or executing the methods, steps, and logic block diagrams disclosed in the embodiments of this application.

[0094] The storage module can be, but is not limited to, random access memory, read-only memory, programmable read-only memory, erasable programmable read-only memory, electrically erasable programmable read-only memory, etc.

[0095] Based on the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by hardware or by using software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solution of this application can be embodied in the form of a software product. This software product can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, mobile hard drive, etc.) and includes several instructions to cause a computer device (such as a personal computer, electronic device, or network device, etc.) to execute the methods described in the various implementation scenarios of this application.

[0096] In the embodiments provided in this application, it should be understood that the disclosed system can also be implemented in other ways. The system embodiments described above are merely illustrative. For example, the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of the system and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, program segment, or part of code, which includes one or more executable instructions for implementing a specified logical function. It should also be noted that each block in a block diagram and / or flowchart, and combinations of blocks in block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions. Furthermore, the functional modules in the various embodiments of this application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0097] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. An enterprise brand reputation intelligent diagnosis management system, characterized in that, include: The data acquisition module is configured to collect basic semantic data, which includes several first text words; The data processing module is configured to receive the basic semantic data and annotate the first text words based on a preset annotation strategy. The annotations include product quality, after-sales service, and / or brand image. The data processing module is also configured to extract the first entity from the basic semantic data. The first entity includes the brand name and / or product name. The analysis module is configured to receive the annotation, obtain the semantics of the first text word corresponding to the annotation based on the semantic model, determine whether the semantics is positive or negative, and obtain the growth rate of the amount of basic semantic data corresponding to the negative within a future preset time period based on the number of first text words corresponding to the negative. The diagnostic decision module is configured to output a decision strategy based on the growth amount and the first entity, using a preset case library.

2. The system according to claim 1, characterized in that, The data processing module is also configured to: Using the first text vocabulary, a second text vocabulary with the same semantics as the first text vocabulary is determined in a preset vocabulary library. The annotation is determined based on the second text vocabulary. The preset vocabulary library includes several first mappings, and the first mappings represent the correspondence between the second text vocabulary and the annotation.

3. The system according to claim 1, characterized in that, The data processing module is further configured to input the basic semantic data into the trained BERT-Base Chinese pre-trained model so that the trained BERT-Base Chinese pre-trained model can extract the first entity from the basic semantic data.

4. The system according to claim 1, characterized in that, The analysis module is also configured to: The basic semantic data with the annotations is input into the semantic model so that the semantic model determines the semantics of the first text words based on the type of the annotations according to its built-in semantic database. The semantic database includes several second mappings, which are used to represent the correspondence between the type of annotations, the first text words and the semantics.

5. The system according to claim 1, characterized in that, The data acquisition module is also configured to acquire time-series data corresponding to the basic semantic data; The analysis module is also configured to: Receive the timing data; The time-series data and the number of corresponding first text words with negative semantics are input into the trained LSTM model so that the trained LSTM model outputs the growth amount.

6. The system according to claim 1, characterized in that, The diagnostic decision module is also configured to: Based on the growth rate, the severity level of the accident is determined according to preset rules; Based on the accident severity level and the first entity, a decision strategy is obtained by querying the preset case library. The preset case library includes several third mappings, which are used to represent the correspondence between the accident severity level, the second entity, the label corresponding to the growth amount, and the decision strategy. The similarity between the second entity and the first entity is greater than the similarity threshold.

7. The system according to claim 6, characterized in that, The analysis module is also configured to: Receive the first entity, output the semantic vector of the first entity based on the semantic model, and input the semantic vector of the first entity into the diagnostic decision module; The diagnostic decision module is further configured to: calculate the similarity between the semantic vector of the first entity and the semantic vectors of all the second entities, and obtain the semantic vector of the second entity whose similarity is greater than the similarity threshold.

8. An electronic device, characterized in that, The electronic device includes a processor and a memory coupled together, wherein the memory stores the system according to any one of claims 1-7.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores the system according to any one of claims 1-7.

Images