A community demand intelligent matching service system based on natural language processing

Abstract

The application relates to the technical field of artificial intelligence and community demand matching, and discloses a community demand intelligent matching service system based on natural language processing, which comprises a text vector generation module, a multi-source vector fusion module, a knowledge graph analysis module, a candidate service optimization module and a service matching and pushing module. Compared with the community service recommendation mode in the prior art which depends on keyword or fixed template matching, especially under the condition that the expression of residents is unclear or the context changes frequently, the technical problem that service intention accurate identification and intelligent recommendation are difficult to realize is solved. Since the natural language pre-training model is introduced, the community service knowledge graph is constructed, and the multi-layer semantic review sorting mechanism is constructed, semantic understanding of the service demand of residents and personalized service pushing are realized, so that the problems of large recommendation deviation, low service coverage and poor response efficiency are avoided, and the intelligentization and response accuracy of the community service system are improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of artificial intelligence and community demand matching, and particularly relates to a community demand intelligent matching service system based on natural language processing. BACKGROUND

[0002] At present, community governance and service response systems mostly rely on manual screening, fixed template questionnaires, keyword matching and the like to identify resident demands and recommend services, but these methods have obvious deficiencies in the face of non-structured and fuzzy expressions of residents. For example, in the process of community App, service hotline or visit, residents usually express their appeals through voice, pictures or short messages, and these expressions often have semantic ambiguity, expression jumps or incoherent logic. The traditional service dispatch mechanism based on keyword rules or process node triggering cannot accurately identify the true service intention, resulting in inaccurate service recommendation results, delayed response or even mismatched services, which seriously affects the intelligent level of community services and the satisfaction of residents. The existing technology cannot fully meet the intelligent identification and accurate recommendation needs in community service applications such as "low expression ability resident groups", "complex semantic expression scenarios" and "multi-channel heterogeneous data fusion", especially in the case of strong emotional expression, ambiguous context and service level dependence, the service intention is difficult to be accurately restored, and the adaptation and coverage of the system recommendation result are low. Therefore, there is an urgent need for a community service intelligent matching method that can still realize service intention semantic modeling, service node structure matching and service priority ordering optimization in the case of complex resident expression and dense service structure association, so as to improve the understanding ability, response accuracy and multi-service coordination ability of the community service system to the actual needs of residents. SUMMARY

[0003] In view of the above technical deficiencies, the purpose of the present application is to provide a community demand intelligent matching service system and method based on natural language processing, which aims to solve the technical problems that the existing community service recommendation methods mostly rely on keyword or fixed template matching, especially in the case of unclear expression or frequent context change of residents, it is difficult to realize accurate identification and intelligent recommendation of service intention.

[0004] To solve the above technical problems, the technical scheme adopted by the present application is as follows: the present application provides a community demand intelligent matching service system based on natural language processing,

[0005] The community demand intelligent matching service system based on natural language processing comprises:

[0006] The text vector generation module is configured to obtain an original expression corpus of a current service window from a community server, perform OCR recognition and standardization processing on the original expression corpus, and generate a standard expression sentence; a natural language pre-training model is preset and trained, the standard expression sentence is input into the natural language pre-training model, and a text expression vector is output;

[0007] The multi-source vector fusion module is configured to obtain a current resident portrait feature vector and a current service environment feature vector from the community server, and perform weighted fusion processing on the text expression vector to obtain a fused demand expression vector;

[0008] The knowledge graph analysis module is configured to preset a community service knowledge graph, perform semantic matching on the fused demand expression vector through the community service knowledge graph, and generate a preliminary candidate service set;

[0009] The candidate service optimization module is configured to set a three-layer natural language review model, perform sorting processing on the preliminary candidate service set based on the three-layer natural language review model and a multi-objective sorting optimization method, and generate an optimized candidate service set;

[0010] The service matching and pushing module is configured to generate a service recommendation card according to the optimized candidate service set, and push the service recommendation card to a resident end.

[0011] Preferably, in the text vector generation module, the original expression corpus specifically includes hotline voice transcription text, community App message, WeChat text picture combination data, interview form and visit record, applet search bar input, AI outbound call system recognized open-ended answer, handwritten demand single, medical record and paper application form.

[0012] Preferably, in the text vector generation module, the structure of the natural language pre-training model specifically includes:

[0013] The input encoding layer is configured to encode the standard expression sentence into a vector input that can be processed by the natural language pre-training model;

[0014] The multi-layer Transformer context modeling layer includes a multi-head self-attention mechanism network, a feedforward neural network, and a residual connection and layer normalization network;

[0015] The fine-tuning mechanism processing layer is configured to obtain community domain annotated corpus, and support using the community domain annotated corpus to perform light fine-tuning on the semantic vector output by the multi-layer Transformer context modeling layer;

[0016] The output layer is configured to extract an overall sentence semantic vector from the semantic vector output by the fine-tuning mechanism processing layer light fine-tuning, and output the text expression vector.

[0017] Preferably, in the multi-source vector fusion module, the current resident portrait feature vector and the current service environment feature vector are obtained from the community service side, combined with the text expression vector for weighted fusion processing to obtain the fusion demand expression vector, and the step specifically comprises:

[0018] The current resident portrait feature vector is used to represent the current resident portrait, and the current service environment feature vector is used to represent the environmental characteristics of the current service, including time, channel, weather and emergency state.

[0019] First, the three-vector weighting method is used for preliminary weighted fusion of the text expression vector, the current resident portrait feature vector and the current service environment feature vector to obtain a preliminary demand expression vector.

[0020] Then, the deep neural network Transformer is used for further nonlinear expression expansion processing of the preliminary demand expression vector to obtain the fusion demand expression vector.

[0021] Preferably, in the knowledge graph analysis module, the step of generating a preliminary candidate service set by performing semantic matching on the fusion demand expression vector through the community service knowledge graph comprises:

[0022] The community service knowledge graph G is preset, G = (V, E); wherein V is a service node set, used to represent the fusion demand expression vector, including a "card loss expression vector", an "accompanying medical treatment expression vector" and a "power repair expression vector"; E is an edge between services, used to represent the service relationship, including "preposition dependence", "similar service" and "parallel recommendation";

[0023] The service node similarity score of the semantic vector of the fusion demand expression vector obtained in the multi-source vector fusion module and the service node V in the service knowledge graph G is calculated by the cosine similarity method; the service node propagation influence degree factor is introduced, and the preliminary node matching score is calculated according to the service node propagation influence degree factor and the service node similarity score;

[0024] The propagation relationship between the service nodes of the service knowledge graph G is learned using the graph convolution network GCN to obtain the attention weight between the service nodes of the service knowledge graph G The final node matching score is calculated by weighting the attention weight between the service nodes and the preliminary node matching score;

[0025] The service nodes are sorted in descending order according to the final node matching score, and a preliminary candidate service set is generated based on the top K service nodes.

[0026] Preferably, in the candidate service optimization module, the three-layer natural language review model specifically comprises:

[0027] a context intention reconstruction layer, configured to process the preliminary candidate service set using a multi-task BERT analysis method combined with embedded context prompt labels, output semantic association scores of service nodes, reorder the preliminary candidate service set according to the semantic association scores, and output a one-layer optimized candidate service set;

[0028] a semantic redundancy disambiguation layer, configured to perform semantic aggregation and exclusion on the one-layer optimized candidate service set, and output a two-layer optimized candidate service set;

[0029] a user emotion subjective preference adjustment layer, configured to obtain historical satisfaction corpus, analyze an intention confidence according to the historical satisfaction corpus and a current resident portrait feature vector, reorder the two-layer optimized candidate service set according to the intention confidence, and output a three-layer optimized candidate service set.

[0030] Preferably, in the candidate service optimization module, the three-layer natural language review model is combined with a multi-objective sorting optimization method to sort and process the preliminary candidate service set, and generate the optimized candidate service set, which specifically comprises: obtaining a priority recommendation score by multi-objective weighted calculation based on the one-layer optimized candidate service set, the two-layer optimized candidate service set and the three-layer optimized candidate service set, sorting and processing the preliminary candidate service set according to the priority recommendation score, and generating the optimized candidate service set.

[0031] The formula for calculating the priority recommendation score is:

[0032] ;

[0033] wherein, is the priority recommendation score of the i th service node ; is the priority score of the service node in the one-layer optimized candidate service set, determined according to the position of the service node in the one-layer optimized candidate service set; is the priority score of the service node in the two-layer optimized candidate service set, determined according to the position of the service node in the two-layer optimized candidate service set; is the priority score of the service node in the three-layer optimized candidate service set, determined according to the position of the service node in the three-layer optimized candidate service set; , and are target weighting coefficients, satisfying .

[0034] The application also provides a community demand intelligent matching service method based on natural language processing, comprising:

[0035] Step S10: obtaining an original expression corpus set of a current service window from a community service end, performing OCR identification and standardization processing on the original expression corpus set to generate a standard expression sentence, and pre-setting and training a natural language pre-training model, inputting the standard expression sentence into the natural language pre-training model, and outputting a text expression vector;

[0036] Step S20: obtaining a current resident portrait feature vector and a current service environment feature vector from the community service end, and performing weighted fusion processing on the text expression vector to obtain a fusion demand expression vector;

[0037] Step S30: pre-setting a community service knowledge graph, and performing semantic matching on the fusion demand expression vector through the community service knowledge graph to generate a preliminary candidate service set;

[0038] Step S40: setting a three-layer natural language review model, and performing sorting processing on the preliminary candidate service set based on the three-layer natural language review model and a multi-target sorting optimization method to generate an optimized candidate service set;

[0039] Step S50: generating a service recommendation card according to the optimized candidate service set, and pushing the service recommendation card to a resident end.

[0040] The application also provides a computer program product comprising a community demand intelligent matching service program based on natural language processing, wherein the community demand intelligent matching service program based on natural language processing realizes the community demand intelligent matching service method based on natural language processing when executed by a processor.

[0041] The application has the advantages that, compared with the community service recommendation mode in the prior art which depends on keywords or fixed templates for matching, especially under the condition that the resident expression is unclear or the context changes frequently, the technical problem of accurately identifying service intention and intelligently recommending services is difficult to solve, since the natural language pre-training model is introduced, the community service knowledge graph is constructed, and the multi-layer semantic review sorting mechanism is constructed, the semantic understanding of the resident service demand and the personalized service pushing are realized, thereby avoiding the problems of large recommendation deviation, low service coverage, and poor response efficiency, and improving the intelligence and response accuracy of the community service system. BRIEF DESCRIPTION OF DRAWINGS

[0042] In order to make the technical solutions in the embodiments of the present application or the prior art clearer, the accompanying drawings needed in the embodiments or prior art description will be briefly introduced. Obviously, the accompanying drawings in the following description only represent some embodiments of the present application, and other drawings can be obtained by those of ordinary skill in the art without any creative effort based on these drawings.

[0043] Figure 1 The system schematic diagram of the first embodiment of the community demand intelligent matching service system based on natural language processing.

[0044] Figure 2 The device schematic diagram of the community demand intelligent matching service system based on natural language processing. DETAILED DESCRIPTION

[0045] The technical solutions in the embodiments of the present application will be described clearly and completely with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments only represent some embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without any creative effort belong to the protection scope of the present application.

[0046] Embodiment one: as shown in the flowchart of the first embodiment of the community demand intelligent matching service system based on natural language processing, the first embodiment of the community demand intelligent matching service system based on natural language processing is proposed. Figure 1

[0047] In the first embodiment, the community demand intelligent matching service system based on natural language processing comprises:

[0048] The text vector generation module is used to obtain the original expression corpus set of the current service window from the community service end, perform OCR recognition and standardization processing on the original expression corpus set, and generate a standard expression sentence. A natural language pre-training model is pre-set and trained, the standard expression sentence is input into the natural language pre-training model, and a text expression vector is output.

[0049] ​It should be noted that in the text vector generation module, the original expression corpus specifically includes hotline voice transcription text, community App message, WeChat text picture combination data, interview form and visit record, applet search bar input, AI outbound system recognized open-ended answers, handwritten demand single, medical record and paper application form. In the text vector generation module, the structure of the natural language pre-training model specifically includes: an input encoding layer for encoding standard expression sentences into vector inputs that can be processed by the natural language pre-training model; a multi-layer Transformer context modeling layer including a multi-head self-attention mechanism network, a feedforward neural network, and a residual connection and layer normalization network; a fine-tuning mechanism processing layer for obtaining community domain annotated corpus and supporting the use of community domain annotated corpus to perform light fine-tuning on the semantic vectors output by the multi-layer Transformer context modeling layer; and an output layer for extracting overall sentence semantic vectors from the semantic vectors output by the light fine-tuning mechanism processing layer, and outputting the text expression vectors.

[0050] It can be understood that the design of the text vector generation module breaks through the limitation of traditional keyword extraction or TF-IDF method which can only process "standard language" and "explicit keywords", and can automatically capture implicit emotional expression, role relationship and service scene context in sentences, and is suitable for "incomplete expression", "sentence jump", "ambiguous words" and even "emotional color" resident service request text.

[0051] It should be understood that in order to adapt to the community multi-expression scene, the natural language pre-training model enhances the semantic perception ability of the model to localized service expressions such as service terms through the domain corpus fine-tuning mechanism, and has stronger service-oriented semantic recognition accuracy and multi-round question and answer expression fusion ability than general BERT. Compared with the traditional static bag-of-words model or single-layer RNN model, the Transformer structure of the present scheme supports long-span dependency analysis.

[0052] For example, taking 1000 pieces of original service expression data of residents collected by a city community service platform as an example, the traditional TF-IDF method, the LSTM + attention mechanism model, and the natural language pre-training model described in the application are used for semantic matching test of service items. Test found that, in the resident text processed by the TF-IDF method, only about 524 texts are matched correctly in the Top-1 recommendation, with an accuracy of 52.4%; after using the deep model of LSTM + attention mechanism, the Top-1 accuracy is improved to 63.2%, but there is still a certain recognition deviation when facing colloquial and jump expressions. And by using the natural language pre-training model in the application, through context modeling, domain fine-tuning and multi-layer semantic integration mechanism, the service matching capability can be significantly improved. Among the same 1000 samples, 847 texts are accurately matched in the Top-1 recommendation, with an accuracy of 84.7%, which reflects good semantic generalization ability and actual landing effect.

[0053] The multi-source vector fusion module is configured to obtain a current resident portrait feature vector and a current service environment feature vector from a community service end, combine a text expression vector for weighted fusion processing, and obtain a fused demand expression vector.

[0054] It should be noted that the step of obtaining a current resident portrait feature vector and a current service environment feature vector from a community service end, combining a text expression vector for weighted fusion processing, and obtaining a fused demand expression vector in the multi-source vector fusion module specifically includes: obtaining a current resident portrait feature vector and a current service environment feature vector from a community service end; wherein the current resident portrait feature vector is used to represent the current resident portrait; the current service environment feature vector is used to represent the environmental features of the current service, including time, channel, weather and emergency state; the text expression vector, the current resident portrait feature vector and the current service environment feature vector are preliminarily weighted and fused by a three-vector weighting method to obtain a preliminary demand expression vector; and the preliminary demand expression vector is further processed by a deep neural network Transformer for non-linear expression expansion to obtain the fused demand expression vector.

[0055] It can be understood that the multi-source vector fusion module in the application combines the text expression vector, the resident portrait feature vector and the service environment feature vector by weighted combination to form a more comprehensive demand expression representation. Compared with the traditional service matching method which only relies on text semantics, the demand expression representation more accurately reflects the service demand of residents under specific population characteristics and environmental context. Especially when the resident expression is unclear or ambiguous, the individual portrait information and environmental context can be used as an important basis for semantic completion to accurately depict the real intention, effectively improving the accuracy and robustness of subsequent knowledge graph matching and re-sorting.

[0056] It should be understood that the three-vector weighting only provides a preliminary fusion expression capability, in order to further enhance the effect of nonlinear modeling of demand expression in a high-dimensional semantic space, the application introduces a deep Transformer structure on the basis of weighted fusion to perform nonlinear expression expansion on the preliminary vector. The multi-head attention mechanism in the Transformer structure can automatically capture the cross-dimensional dependency relationship in the fusion vector, enabling the model to identify real service intentions based on the combination of "resident category + environment type + language mode". For example, for the same sentence "card cannot be processed", the system will tilt to "on-site assistance with loss" based on the context of "elderly single + voice channel + night expression", rather than recommending the default "card loss" service item. This mechanism enhances the generalization ability and service matching sensitivity of the model under diverse input contexts.

[0057] For example, during the pilot deployment in a certain street, the system collected 500 resident service requests containing ambiguous expressions, and compared and tested the performance of three models in terms of service matching accuracy. The Top-1 recommendation accuracy of the traditional BERT model (based only on text semantics) for these requests was 68.2%; the model that spliced the resident portrait information as additional labels improved the accuracy to 75.4%; and after using the multi-source vector fusion module described in the application, the "weighted fusion + Transformer nonlinear expansion", the Top-1 recommendation accuracy was improved to 88.6%. Among them, in the sub-sample of "weakly expressed elderly population", the accuracy improved most significantly, indicating that the present scheme has better service intention recognition capability for complex contexts and special groups.

[0058] The knowledge graph analysis module is configured to preset a community service knowledge graph, perform semantic matching through the community service knowledge graph according to the fused demand expression vector, and generate a preliminary candidate service set;

[0059] It should be noted that in the knowledge graph analysis module, the preset community service knowledge graph is matched according to the fusion demand expression vector through the community service knowledge graph to generate a preliminary candidate service set, and the step specifically comprises: presetting a community service knowledge graph G, G = (V, E); wherein V is a service node set, used to represent the fusion demand expression vector, including a "card loss expression vector", an "accompanying medical treatment expression vector" and a "power repair expression vector"; E is a service edge, used to represent a service relationship, including "preceding dependence", "similar service" and "parallel recommendation"; the service node similarity score of the fusion demand expression vector semantic vector obtained in the multi-source vector fusion module and the service node V in the service knowledge graph G is calculated by the cosine similarity method; the service node propagation influence degree factor is introduced, and the preliminary node matching score is calculated according to the service node propagation influence degree factor and the service node similarity score; the propagation relationship between the service nodes of the service knowledge graph G is learned by using the graph convolution network GCN, and the attention weight between the service nodes of the service knowledge graph G is obtained , the final node matching score is calculated according to the attention weight between the service nodes and the preliminary node matching score; the service nodes are sorted in descending order according to the final node matching score, and a preliminary candidate service set is generated based on the first K service nodes.

[0060] It can be understood that in the present application, the knowledge graph analysis module realizes accurate semantic matching between the fusion demand expression vector and the service node by constructing a community service knowledge graph with structural relationship semantics. The service nodes in the knowledge graph not only carry semantic representation information, but also express the logical dependence, parallel co-occurrence and similar substitution relationship between services through the edge structure, so that the structural context information can be effectively utilized in the matching process, the semantic island problem is avoided, and the service recall ability under fuzzy expression conditions is improved. By introducing the service node propagation influence degree and modeling the relationship weight between the service nodes by the graph convolution network, the model not only focuses on the semantic distance of the current node itself, but also considers the surrounding semantic related services based on the graph structure expansion, and realizes the comprehensive recommendation strategy of "semantic similarity + graph structure linkage".

[0061] It should be understood that most of the traditional service matching methods adopt a static matching mechanism based on keyword matching or sentence vector similarity, ignoring the upstream and downstream dependence and structural cooperation between services. Especially in the community governance scene with fuzzy expression or complex service structure, there are often problems such as one-sided matching, misjudgment or recommendation omission. The present application constructs a service knowledge graph, explicitly models the structural knowledge between service nodes, further introduces a graph propagation mechanism after preliminary similarity calculation, learns the structural influence between service nodes by a graph convolution network, and dynamically adjusts the matching score of the node.

[0062] For example, in a simulation experiment, the system selected 300 natural language expressions of service requests of residents, about 28% of which did not directly contain standard service item names, such as "My mother is hospitalized and cannot see a doctor because the card has not been paid". The Top-1 service recommendation accuracy rate under the traditional BERT semantic matching mechanism is less than 55%. When the knowledge graph analysis module in the application is introduced, the system first selects the preliminary candidate nodes according to the semantic similarity between the fused expression vector and the service node, and then propagates the importance through the graph convolution network, and finally successfully matches the Top-1 recommended item. The overall results show that in the service expression containing structural dependence, the Top-1 recommendation accuracy of the application is improved by 29.3 percentage points, and the Top-3 recall rate reaches more than 92%, significantly improving the coverage and depth of service recommendation and intent understanding.

[0063] The candidate service optimization module is configured to set a three-layer natural language review model, and sort the preliminary candidate service set based on the three-layer natural language review model and a multi-objective sorting optimization method to generate an optimized candidate service set.

[0064] It should be noted that in the candidate service optimization module, the step of sorting the preliminary candidate service set based on the three-layer natural language review model and the multi-objective sorting optimization method to generate the optimized candidate service set specifically includes: calculating a priority recommendation score based on the first-layer optimized candidate service set, the second-layer optimized candidate service set and the third-layer optimized candidate service set through multi-objective weighting, sorting the preliminary candidate service set according to the priority recommendation score to generate the optimized candidate service set; wherein the formula used for calculating the priority recommendation score is: wherein, is the priority recommendation score of the i-th service node . is the priority score of the service node in the first-layer optimized candidate service set, which is determined according to the position of the service node in the first-layer optimized candidate service set; is the priority score of the service node in the second-layer optimized candidate service set, which is determined according to the position of the service node in the second-layer optimized candidate service set; is the priority score of the service node in the third-layer optimized candidate service set, which is determined according to the position of the service node in the third-layer optimized candidate service set; , and are target weighting coefficients, satisfying .

[0065] It can be understood that the candidate service optimization module in the application realizes the fusion evaluation mechanism of semantic matching, structure redundancy elimination and personalized preference by introducing a three-layer optimization perspective. The three-layer optimization process respectively focuses on the accuracy of expression semantics, the logical rationality between candidate services and the subjective perception and historical behavior of residents, so that the recommendation not only has language understanding ability, but also reflects service adaptability and perception satisfaction. Compared with the traditional single sorting method based on only semantic or rule scoring, the module has stronger intention review ability and recommendation sorting explainability.

[0066] It should be understood that the traditional recommendation sorting method is often distorted in the processing of "incomplete expression" or "resident preference polarization" scene, and the single intention score may put the "expressions that seem to match but services are not feasible" options in the first place, resulting in inaccurate recommendations. The application introduces intention matching (first layer), structure rationality screening (second layer) and emotion preference adjustment (third layer) into the unified sorting model, constructs a fusion sorting score function, effectively makes up for the problem of local optimal score of each layer but poor overall recommendation, and improves the global optimality and adaptability of service sorting.

[0067] For example, in a group of samples containing expressions such as "card lost, hospital not to see, my mother cried", the traditional model relying only on intention matching is easy to take "card loss" as the Top-1 recommended item, ignoring "reimbursement assistance" or "social worker accompanying medical treatment" and other service items that better meet the current user's real perception needs. When using the candidate service optimization module of the application: the first layer score gives the highest score to "card loss"; the second layer score is reduced due to incomplete dependency chain; the third layer score increases the preference weight of "accompanying medical treatment", "reimbursement assistance" and other service items due to the strong emotion of the user's expression; in the final sorting result, "card loss" ranks second, and "accompanying medical treatment" jumps to the first, the system recommendation is closer to the current perception state of the user, and the Top-1 recommendation accuracy is improved by more than 20%.

[0068] The service matching and pushing module is used for generating a service recommendation card according to the optimized candidate service set, and pushing the service recommendation card to the resident end.

[0069] It should be noted that the application realizes the conversion of the intelligent sorting result in the system to a user understandable expression by constructing a structured service recommendation card, and improves the explainability of the recommendation result and the user trust. Combined with the service matching reason and the handling guide, the recommendation card not only tells the user "what is recommended", but also explains "why it is recommended", effectively reducing the cognitive barriers of the elderly group to the system automatic recommendation. At the same time, combined with the pushing scheduling mechanism, the closed-loop service experience of "person to send", "voice to answer" and "card to handle" can be realized, and the execution efficiency and user reachability of the service recommendation system are improved.

[0070] It should be understood that the recommended information in the traditional community service system is often presented in the form of a list, lacks semantic explanation and service guidance, and is easy to cause the user to "not understand, not dare to use, and not know the next step", especially in the elderly group or weak expression group. The application encapsulates the recommended service content by a structured recommendation card, and adapts to different touch channels as needed, taking into account the visibility, operability and urgency recognition of the recommendation, and overcoming the technical defects of the traditional recommendation system push mode such as "single form, no personalized push, and interactive break".

[0071] Embodiment two: In addition, the application provides a community demand intelligent matching service method based on natural language processing, which adopts the community demand intelligent matching service system based on natural language processing in the above embodiment, and can solve the technical problem of the community demand intelligent matching service based on natural language processing. Compared with the prior art, the beneficial effects of the community demand intelligent matching service method based on natural language processing provided by the application are the same as those of the community demand intelligent matching service system based on natural language processing provided by the above embodiment, and other technical features in the community demand intelligent matching service method based on natural language processing are the same as those disclosed in the above embodiment method, and will not be repeated here.

[0072] Embodiment three: The application provides a community demand intelligent matching service device based on natural language processing, please refer to Figure 2A community demand intelligent matching service device based on natural language processing includes at least one processor, and a memory connected to the at least one processor in communication. The memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method of community demand intelligent matching service based on natural language processing in the above embodiment one. The community demand intelligent matching service device based on natural language processing in the embodiment of the present application can include, but is not limited to, mobile terminals such as mobile phones, notebook computers, digital broadcast receivers, PDAs (Personal Digital Assistant), PADs (Portable Application Description), PMPs (Portable Media Player), vehicle terminals (such as vehicle navigation terminals), and the like, and fixed terminals such as digital TVs, desktop computers, and the like. The community demand intelligent matching service device based on natural language processing is only an example, and should not bring any limitation to the function and use range of the embodiment of the present application. The community demand intelligent matching service device based on natural language processing can include a processing device 1001 (such as a central processor, a graphics processor, etc.), which can perform various appropriate actions and processes according to programs stored in a read-only memory (ROM) 1002 or loaded from a storage device 1003 to a random access memory (RAM) 1004. In the RAM 1004, various programs and data required for the operation of the community demand intelligent matching service device based on natural language processing are also stored. The processing device 1001, the ROM 1002, and the RAM 1004 are connected to each other through a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Generally, the following systems can be connected to the I / O interface 1006: an input device 1007 including, for example, a touch screen, a touch pad, a keyboard, a mouse, an image sensor, a microphone, an accelerometer, a gyroscope, and the like; an output device 1008 including, for example, a liquid crystal display (LCD), a speaker, a vibrator, and the like; a storage device 1003 including, for example, a magnetic tape, a hard disk, and the like; and a communication device 1009. The communication device 1009 can allow the community demand intelligent matching service device based on natural language processing to communicate with other devices wirelessly or by wire to exchange data. Although the community demand intelligent matching service device based on natural language processing with various systems is shown in the figure, it should be understood that all the shown systems are not required to be implemented or possessed. More or less systems can be alternatively implemented or possessed.

[0073] Embodiment Four: The present application also provides a computer program product comprising a computer program which, when executed by a processor, implements the steps of a natural language processing based community demand intelligent matching service method as described above. The computer program product provided by the present application can solve the technical problem of a natural language processing based community demand intelligent matching service. Compared with the prior art, the beneficial effects of the computer program product provided by the present application are the same as those of the natural language processing based community demand intelligent matching service method provided by the above-described embodiments, and are not described here in detail.

[0074] In particular, according to the embodiments disclosed by the present application, the processes described above with reference to the flowcharts can be implemented as a computer software program. For example, the embodiments disclosed by the present application include a computer program product comprising a computer program carried on a computer readable medium, the computer program comprising program code for executing the method shown in the flowchart. In such embodiments, the computer program can be downloaded and installed from a network through a communication device, or installed from a storage device 1003, or installed from a ROM 1002. When the computer program is executed by a processing device 1001, the above-mentioned functions defined in the method of the embodiments disclosed by the present application are executed.

[0075] It should be understood that various parts of the present application can be realized by hardware, software, firmware, or a combination thereof. In the description of the above-described embodiments, specific features, structures, materials or characteristics can be combined in any appropriate manner in any one or more embodiments or examples.

[0076] Obviously, those skilled in the art can make various modifications and variations to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalent technologies, the present application is also intended to include these modifications and variations.

Claims

1. A community demand intelligent matching service system based on natural language processing, characterized in that, The system comprises: a text vector generation module configured to obtain an original expression corpus of a current service window from a community service server, perform OCR recognition and standardization processing on the original expression corpus, and generate a standard expression sentence; a natural language pre-training model is preset and trained, the standard expression sentence is input into the natural language pre-training model, and a text expression vector is output; a multi-source vector fusion module configured to obtain a current resident portrait feature vector and a current service environment feature vector from the community service server, and perform weighted fusion processing on the text expression vector to obtain a fused demand expression vector; a knowledge graph analysis module configured to preset a community service knowledge graph, perform semantic matching on the fused demand expression vector through the community service knowledge graph, and generate a preliminary candidate service set; a candidate service optimization module configured to set a three-layer natural language review model, perform sorting processing on the preliminary candidate service set based on the three-layer natural language review model and a multi-objective sorting optimization method, and generate an optimized candidate service set; in the candidate service optimization module, the three-layer natural language review model specifically comprises: a context intention reconstruction layer configured to use a multi-task BERT analysis method combined with embedded context prompt labels to process the preliminary candidate service set, output a semantic correlation score of a service node, reorder the preliminary candidate service set according to the semantic correlation score, and output a one-layer optimized candidate service set; a semantic redundancy disambiguation layer configured to perform semantic aggregation and differentiation on the one-layer optimized candidate service set, and output a two-layer optimized candidate service set; a user emotion subjective preference adjustment layer configured to obtain a historical satisfaction corpus, analyze an intention confidence according to the historical satisfaction corpus and the current resident portrait feature vector, reorder the two-layer optimized candidate service set according to the intention confidence, and output a three-layer optimized candidate service set; wherein the step of performing sorting processing on the preliminary candidate service set based on the three-layer natural language review model and the multi-objective sorting optimization method to generate the optimized candidate service set specifically comprises: calculating a priority recommendation score based on the one-layer optimized candidate service set, the two-layer optimized candidate service set, and the three-layer optimized candidate service set through multi-objective weighted calculation, performing sorting processing on the preliminary candidate service set according to the priority recommendation score, and generating the optimized candidate service set; wherein the formula used for calculating the priority recommendation score is: ； wherein, is the priority recommendation score of the i-th service node ; is the priority score of the service node in the one-layer optimization candidate service set, determined according to the position of the service node in the one-layer optimization candidate service set; is the priority score of the service node in the two-layer optimization candidate service set, determined according to the position of the service node in the two-layer optimization candidate service set; is the priority score of the service node in the three-layer optimization candidate service set, determined according to the position of the service node in the three-layer optimization candidate service set; , and are target weighting coefficients, satisfying ; a service matching and pushing module configured to generate a service recommendation card according to the optimized candidate service set, and push the service recommendation card to a resident end.

2. The community demand intelligent matching service system based on natural language processing of claim 1, wherein, In the text vector generation module, the original expression corpus specifically comprises hotline voice transcription text, community App messages, WeChat text picture combination data, interview forms and visit records, applet search bar input, AI outbound call system recognized open-ended answers, handwritten demand forms, medical records, and paper application forms.

3. The community demand intelligent matching service system based on natural language processing of claim 1, wherein, In the text vector generation module, the structure of the natural language pre-training model specifically comprises: an input encoding layer configured to encode the standard expression sentence into a vector input processable by the natural language pre-training model; The multi-layer Transformer context modeling layer includes a multi-head self-attention mechanism network, a feedforward neural network, and a residual connection and layer normalization network. The fine-tuning mechanism processing layer is used to obtain community domain labeled corpus and support lightweight fine-tuning of the semantic vector output by the multi-layer Transformer context modeling layer. The output layer is used to extract the overall sentence semantic vector from the lightweight fine-tuning output of the fine-tuning mechanism processing layer, and output the text expression vector.

4. The community demand intelligent matching service system based on natural language processing of claim 1, wherein, In the multi-source vector fusion module, the current resident portrait feature vector and the current service environment feature vector are obtained from the community service side, and the text expression vector is combined for weighted fusion processing to obtain a fused demand expression vector, which specifically includes: The current resident portrait feature vector is used to represent the current resident portrait, and the current service environment feature vector is used to represent the environmental characteristics of the current service, including time, channel, weather, and emergency state. The three-vector weighting method is used to preliminarily weight and fuse the text expression vector, the current resident portrait feature vector, and the current service environment feature vector to obtain a preliminary demand expression vector. The deep neural network Transformer is used to further perform nonlinear expression expansion processing on the preliminary demand expression vector to obtain a fused demand expression vector.

5. The natural language processing based community demand intelligent matching service system of claim 1, wherein, In the knowledge graph analysis module, a preset community service knowledge graph is used to perform semantic matching on the fused demand expression vector to generate a preliminary candidate service set, which specifically includes: The preset community service knowledge graph G is G = (V, E), where V is a service node set, used to represent the fused demand expression vector, including "card loss expression vector", "accompanying medical treatment expression vector", and "power repair expression vector"; E is the edge between services, used to represent the service relationship, including "preposition dependence", "similar service", and "parallel recommendation"; The service node similarity score of the semantic vector of the fused demand expression vector obtained in the multi-source vector fusion module and the service node V in the service knowledge graph G is calculated by the cosine similarity method; the service node propagation influence degree factor is introduced, and the preliminary node matching score is calculated according to the service node propagation influence degree factor and the service node similarity score; The propagation relationship between the service knowledge graph G service nodes is learned by using a graph convolution network GCN, and attention weights between the service knowledge graph G service nodes are obtained The final node matching score is calculated according to the attention weights between the service nodes and the preliminary node matching score The service nodes are sorted in descending order according to the final node matching score, and the preliminary candidate service set is generated based on the top K service nodes. 6.A community demand intelligent matching service method based on natural language processing, applied to the community demand intelligent matching service system based on natural language processing in any one of claims 1-5, characterized in that, The method includes: Step S10: Obtain the original expression corpus set of the current service window from the community service side, perform OCR recognition and standardization processing on the original expression corpus set to generate a standard expression sentence; a natural language pre-training model is pre-set and trained, the standard expression sentence is input into the natural language pre-training model, and a text expression vector is output; Step S20: Obtain the current resident portrait feature vector and the current service environment feature vector from the community service side, and combine the text expression vector for weighted fusion processing to obtain a fused demand expression vector; Step S30: preset a community service knowledge graph, perform semantic matching on the community service knowledge graph according to the fusion demand expression vector, and generate a preliminary candidate service set; Step S40: set a three-layer natural language review model, sort the preliminary candidate service set based on the three-layer natural language review model and a multi-objective sorting optimization method, and generate an optimized candidate service set; wherein the three-layer natural language review model specifically includes: a context intention reconstruction layer, configured to process the preliminary candidate service set by using a multi-task BERT analysis method combined with embedded context prompt labels, output a semantic correlation score of a service node, re-sort the preliminary candidate service set according to the semantic correlation score, and output a one-layer optimized candidate service set; a semantic redundancy resolution layer, configured to perform semantic aggregation and differentiation on the one-layer optimized candidate service set, and output a two-layer optimized candidate service set; a user emotion subjective preference adjustment layer, configured to obtain historical satisfaction corpus, analyze an intention confidence according to the historical satisfaction corpus and a current resident portrait feature vector, re-sort the two-layer optimized candidate service set according to the intention confidence, and output a three-layer optimized candidate service set; wherein the step of sorting the preliminary candidate service set based on the three-layer natural language review model and the multi-objective sorting optimization method to generate the optimized candidate service set specifically includes: obtaining a priority recommendation score by multi-objective weighting calculation based on the one-layer optimized candidate service set, the two-layer optimized candidate service set and the three-layer optimized candidate service set, sorting the preliminary candidate service set according to the priority recommendation score, and generating the optimized candidate service set; wherein the formula used for calculating the priority recommendation score is: ； wherein, is a priority recommendation score of the i-th service node ; is a priority score of the service node in the one-level optimization candidate service set, determined according to a position of the service node in the one-level optimization candidate service set; is a priority score of the service node in the two-level optimization candidate service set, determined according to a position of the service node in the two-level optimization candidate service set; is a priority score of the service node in the three-level optimization candidate service set, determined according to a position of the service node in the three-level optimization candidate service set; , and are target weighting coefficients, satisfying ;​ Step S50: generate a service recommendation card according to the optimized candidate service set, and push the service recommendation card to a resident end.

7. A community demand intelligent matching service device based on natural language processing, characterized in that, The community demand intelligent matching service device based on natural language processing includes a memory, a processor, and a community demand intelligent matching service program based on natural language processing stored on the memory and executable on the processor. When the community demand intelligent matching service program based on natural language processing is executed by the processor, the community demand intelligent matching service system based on natural language processing of any one of claims 1-5 is implemented.

8. A computer program product, characterised in that, The computer program product includes a community demand intelligent matching service program based on natural language processing, and the community demand intelligent matching service program based on natural language processing is executed by the processor to implement the community demand intelligent matching service system based on natural language processing of any one of claims 1-5.

Images