A smart community multi-modal event intelligent processing method based on a large language model

Abstract

This invention relates to the field of multimodal data processing technology, specifically to a smart community multimodal event intelligent processing method based on a large language model, comprising the following steps: S1: Acquire multi-source event information (text, voice, and images) generated within the smart community, and perform unified formatting processing to construct a multimodal event dataset; construct a community-specific knowledge base and combine it with large model retrieval enhancement, constraining model output with facts to resolve decision illusions and biases, achieving evidence-based and highly reliable judgment; simultaneously, based on multimodal event semantic understanding, extract key features such as event type, location, level, scope of impact, and responsible entity, replacing manual sorting and improving information collection efficiency and accuracy; on this basis, form a fully automated and standardized closed loop of judgment, allocation, and dispatch; the large model, combined with the knowledge base, automatically outputs classification, grading, responsible entity, and disposal suggestions, generating structured work orders and accurately dispatching orders.

Description

Technical Field

[0001] This invention relates to the field of multimodal data processing technology, specifically to a smart community multimodal event intelligent processing method based on a large language model. Background Technology

[0002] Multimodal events in smart communities refer to community events reported or collected through various information carriers such as text, voice, images, and video in the context of smart community management. They represent a type of event that integrates multi-source information in smart community governance.

[0003] Existing intelligent community incident handling technologies generally employ general large models or pure algorithms for incident assessment and decision-making. Due to the lack of constraints from community-specific factual knowledge, these technologies are prone to generating decision-making illusions and subjective biases that are inconsistent with management norms and emergency plans. This results in insufficient credibility, interpretability, and traceability of the assessment conclusions, failing to meet the requirements of compliance, authority, and standardized management in community incident handling, and making it difficult to support a standardized and unified closed-loop handling process. Summary of the Invention

[0004] This invention addresses the technical problems existing in the prior art by providing a smart community multimodal event intelligent processing method based on a large language model.

[0005] The technical solution of this invention to solve the above-mentioned technical problems is as follows: A smart community multimodal event intelligent processing method based on a large language model, comprising the following steps: S1: Acquire multi-source event information, including text, voice, and images generated within the smart community, and perform unified formatting to construct a multimodal event dataset; S2: Construct a structured knowledge base containing community management norms, emergency plans, lists of powers and responsibilities, and historical handling cases; establish an index; perform semantic matching between multimodal event information and the knowledge base; and use a retrieval mechanism to impose factual constraints on the reasoning process of the large language model. S3: Based on a large language model, perform in-depth analysis of multimodal event information to extract key feature information such as event type, location, emergency level, scope of impact, and responsible party; S4: Combining the knowledge base and key features, the large language model automatically generates judgment conclusions, completes event classification, emergency level determination and responsible party matching, and outputs standardized handling suggestions; S5: Based on the analysis results, automatically generate structured work orders, match the corresponding handling departments according to the responsible entities to complete accurate work order assignment, and simultaneously output standardized handling scripts and process guidelines.

[0006] In a preferred embodiment, in step S1, raw event data, including text information, voice information, and image information, is obtained from various information channels of the smart community. The different types of data collected are uniformly formatted, including text data processing, voice data processing, and image data processing. The formatted text, voice-to-text, image description text, and associated file index are integrated into a unified multimodal event dataset, specifically including: common fields, content fields, associated fields, and status fields. Each record in the dataset corresponds to a complete event within the community and contains both structured text information and an index of the original multimodal file.

[0007] In a preferred embodiment, during step S2, when constructing the community knowledge base, core knowledge of management regulations, emergency plans, lists of powers and responsibilities, and historical cases are comprehensively collected from community management-related documents, systems, and historical records. The collected knowledge is then decomposed, labeled, and structured. Each knowledge entry is configured with a unified field containing knowledge ID, type tag, applicable scenario, priority, associated clauses, effective time, and version number. Long documents are split into semantically independent knowledge blocks according to the theme. Based on the natural structure of document chapters, titles, and clauses, semantically independent thematic units in the document are identified. The start and end positions of each knowledge block are divided according to the principle of fully expressing the content of a single theme. Without destroying semantic integrity, each knowledge block is controlled within a uniform length range. The association identifiers between the split knowledge blocks and the original document and adjacent knowledge blocks are retained. At the same time, metadata such as keywords, event type, scope of application, responsible entity, and urgency level are labeled.

[0008] In a preferred embodiment, the structured knowledge blocks are processed to generate a searchable vector library, and vector indexes, keyword indexes and full-text indexes are established simultaneously. Vector Index Building: Based on the feature data of each knowledge block in the vector library, construct the association mapping relationship between vectors, group all knowledge block vectors in the vector library according to preset rules, so that semantically similar vectors are grouped into the same set, forming a preliminary vector aggregation structure, and establish an association record between each vector and its neighboring vectors. On the basis of the vector set, construct a multi-level association mapping structure, with the upper level being the abstract identifier of the vector set and the lower level being the specific vectors within the set, forming a coarse-to-fine retrieval path. Bind and store the constructed vector association mapping relationship with the unique identifier and metadata of the corresponding knowledge block to form an index structure that supports similarity matching; Keyword Index Building: Based on the keywords and metadata tags labeled in the knowledge blocks, an inverted mapping relationship between keywords and corresponding knowledge blocks is established. Keywords and metadata tags in the knowledge blocks are extracted, and keywords are uniformly standardized, including synonym merging and format normalization. An index item is created for each unique keyword as the core entry point of the inverted index. At the same time, the attribute information of the keyword is recorded. For each index item, a mapping list of all knowledge block identifiers corresponding to it is established, and the unique identifiers and related metadata of all knowledge blocks containing the keyword are recorded, forming an index structure that can be filtered by keywords. Full-text indexing: Based on the complete text content of knowledge blocks, a mapping between text content and knowledge blocks is constructed. The complete text content of knowledge blocks is uniformly cleaned and standardized to remove redundant symbols, stop words, and invalid information, and the expression format is standardized. Corresponding index entries are generated for the preprocessed text of each knowledge block. The text content is bound to the unique identifier of the knowledge block to form a basic mapping record. Based on the text content of all knowledge blocks, a retrieval structure that supports fuzzy matching and partial matching is established.

[0009] In a preferred embodiment, a hybrid retrieval rule of priority vector semantic matching supplemented by keyword filtering is preset, the retrieval execution order is set, the vector semantic matching process is started, the similarity between event vectors and knowledge base vectors is used as the core basis to recall a set of semantically related knowledge items, and on the basis of vector matching results, event tags and keyword indexes are called to perform secondary filtering on the result set, retaining items that match tags of event type, urgency level and responsible subject, and comprehensively sorting the filtered results according to vector similarity, tag matching degree and knowledge priority, placing relevant and applicable items at the top, and when the vector matching results are insufficient, the full-text index is automatically called to perform fallback retrieval, and a multi-dimensional index and retrieval engine are constructed; This process transforms multimodal event information into event vectors and extracts structured tags for event type, location, urgency level, and responsible party. Multimodal event information from text, speech-to-text, and image descriptions is integrated into a standardized event information set according to a unified format. The integrated event text is then cleaned and standardized, removing invalid characters and redundant information, and unifying the expression format. Based on the preprocessed text content, key information regarding event type, location, urgency level, and responsible party is extracted according to preset rules and converted into standardized structured tags. Finally, the preprocessed event text content and structured tags are converted into feature data in a unified format according to preset rules, generating event... Event vectors are used to perform similarity searches between event vectors and a knowledge vector database. Combined with event tags, multi-dimensional semantic matching is performed based on event type, urgency level, location, and scope of impact. Search results are sorted by similarity, relevance, and priority to filter relevant knowledge entries. The initial set of knowledge entries returned by the similarity search is then validated through multi-dimensional matching based on event tags: the event type, urgency level, location, and scope of impact tags are compared one by one with the corresponding tags in the knowledge database entries. The structured tags of the events are mapped to the tags of the knowledge database entries according to preset fields, establishing a field mapping relationship for tag comparison. For each field's tag value, a comparison is performed according to preset matching rules. Event Type: Determines whether the event type label matches the event type label of the knowledge base entry; Urgency Level: Determines whether the urgency level label of the event matches the applicable urgency level of the knowledge base entry; Location information: Determines whether the event location tag is within the applicable location range of the knowledge base entry; Scope of impact: Determine whether the scope of impact label of the event matches the applicable scope of impact of the knowledge base entry; The system analyzes the matching results of all fields and calculates the overall tag matching degree according to preset rules. If the matching degree reaches the preset threshold, the knowledge base entry is determined to match the event tag and is retained; otherwise, it is excluded.

[0010] In a preferred embodiment, items whose tag matching degree meets the preset requirements are further filtered out. For the items that pass the verification, they are comprehensively sorted according to three dimensions: vector phase velocity, tag matching degree, and knowledge priority, with preset weights, to form an item priority list. The sorted list is then filtered according to the preset similarity threshold and the number of returned items. When extracting key information about event type, location, urgency level, and responsible party according to preset rules, extraction rules are pre-configured for each key information item. These rules include: keyword set for event type, format and expression rules for location information, grading rules for urgency level, and rules for matching the rights and responsibilities of the responsible party. The pre-processed event text is then matched and identified according to these preset rules. Event Type: Matches the text with descriptions corresponding to preset event type keywords to identify the category to which the event belongs; Location information: Extract location information such as building number and area name from the text according to preset format and keywords; Emergency Level: Based on the risk description and urgency statement in the text, and matching the preset classification rules; Responsible Entity: By combining event type and location information, and matching pre-defined rules for corresponding rights and responsibilities, the corresponding responsible department is identified; The initially identified information is verified to check its completeness and rationality. For information with non-standard expressions, the format is normalized according to preset rules to generate unified and standardized key information results. Input the retrieved knowledge entries, event information, and task requirements for analysis, classification, grading, and task assignment into the constraint template to impose factual constraints on the reasoning process of the large model.

[0011] In a preferred embodiment, in step S3, based on the constructed multimodal event dataset as the input source, the system accesses and loads multi-source event information, including text descriptions submitted by residents through App, telephone, and on-site reporting channels, text content converted by speech recognition, and image description text generated by image understanding. During access, the system verifies the integrity of the data and filters out invalid and duplicate information. By using a large language model to perform integrated deep analysis of multi-source event information, basic semantic understanding of individual information, automatic identification of core semantic subjects, key behaviors and contextual background in event descriptions, and pre-configuration of identification rules: core semantic subject identification rules, key behavior identification rules and contextual background identification rules; Core semantic subject identification rules: including keywords and type definitions of objects, roles, and entities involved in the event; Key behavior recognition rules: include keywords and action types that describe actions, behaviors, and state changes; Context recognition rules: Keywords and format rules that include scene information such as time, location, environment, and conditions; The event description text is segmented into sentences and paragraphs, breaking down long texts into independent semantic units. Each semantic unit is then matched according to preset rules. Match core semantic keywords and mark the core objects in the event description; Match key behavioral keywords and tag actions within the event; Match contextual keywords to mark the time, location, and environment in which the event occurred; The marked subjects, behaviors, and background information are linked and integrated to form a complete semantic structure of the event; It performs cross-modal contextual analysis on multi-source information of the same event, integrating complementary information from text, audio, and image descriptions to form a complete understanding of the event. Simultaneously, it automatically disambiguates ambiguous or vague content, identifying ambiguous content in the text according to preset rules, including unreferential pronouns, polysemous words, and incomplete phrases. It marks the location and type of potential ambiguity, retrieves the sentence, paragraph, and other multi-source information related to the same event containing the ambiguous content, constructs a contextual reference set for ambiguity dissolution, obtains the complete semantic background of the event, and applies preset processing rules based on the type of ambiguity: for unreferential content, it restores the referent based on the entity, location, and event description in the context; for polysemous words, it matches a reasonable interpretation based on the event scenario and contextual semantics; for incomplete content, it supplements key information based on the event background and logical relationships; and it performs consistency verification on the processed content. After verification, it replaces the original ambiguous description with the ambiguity-dissolved text, forming complete and clear event semantic content, and supplements key information based on contextual logic.

[0012] In a preferred embodiment, based on the complete event semantics formed after parsing, key feature information is extracted. The key feature information includes: identifying the event type from the event description, locating the specific location of the event from the text description, determining the urgency level by combining the degree of impact and descriptive tendency in the event description, identifying the spatial range of the event's impact, and matching the corresponding responsible parties. Independent extraction rules are pre-configured for each type of key feature, including: keywords and classification rules for the event type, format and recognition rules for location information, grading and determination rules for the urgency level, rules for defining the scope of impact, and rules for matching the rights and responsibilities of the responsible parties. The system performs keyword matching on the event description text, identifies the event category based on a pre-defined event type classification system, and generates standardized event type tags. It also extracts location information such as building number, unit number, floor, and area name from the text according to pre-defined location identification rules, converting it into standardized location tags. Furthermore, it matches the event description with pre-defined emergency level determination rules based on the risk level, scope of impact, and descriptive tendencies, automatically classifying the event into emergency, general, and low levels to generate emergency level tags. Finally, it extracts the spatial scope of the event's impact and the size of the affected population from the text, generates impact scope description information based on pre-defined scope definition rules, and automatically identifies the responsible department based on the event type and location information, generating a responsible entity matching result. The extracted key feature information is organized and standardized to generate structured data that includes event type labels, location labels, emergency level, description of the scope of impact, and matching results of responsible parties.

[0013] In a preferred embodiment, in step S4, relevant entries retrieved from the community knowledge base and key event feature information are accessed to form a complete input context. The large language model combines management regulations, emergency plans, lists of powers and responsibilities, and historical handling cases in the knowledge base to perform rule matching and compliance verification on key event features: matching corresponding handling regulations according to event type, matching graded response requirements according to urgency level, and matching corresponding responsible parties according to location and scope of impact. Based on the rule matching results, the core judgments are automatically completed: event classification, emergency level determination, and matching of responsible parties; After completing the core judgment, the big language model combines historical handling cases and management standards to automatically generate standardized handling suggestions that can be directly executed, including handling procedures, response time limits, precautions and resource allocation suggestions.

[0014] In a preferred embodiment, in step S5, the structured analysis results output, including information on event classification, urgency level, responsible party, and standardized handling suggestions, serve as the core input basis for work order generation and dispatch. According to the preset work order template, the key information in the analysis conclusion is automatically filled into the corresponding fields of the work order to generate a standardized work order containing event number, event type, location, urgency level, scope of impact, responsible party, handling suggestions, and response time limit. Based on the responsible party information in the work order, the system automatically matches the corresponding handling department in the community management system, automatically sets the response priority based on the urgency level, and retrieves the corresponding pre-set handling scripts and standard procedures from the knowledge base according to the event type, urgency level, and handling suggestions in the work order. This generates standardized guidelines that can be directly used for on-site handling, including communication points, operating steps, precautions, and abnormal handling methods. After the work order is dispatched, the system automatically updates the work order status to "dispatched" and simultaneously records the dispatched event, receiving department, and response time limit.

[0015] The beneficial effects of this invention are as follows: It constructs a community-specific knowledge base and combines it with large-scale model retrieval enhancement, constrains the model output with facts, solves decision-making illusions and biases, and achieves judgments that are based on evidence and highly credible. At the same time, based on multimodal event semantic understanding, it extracts key features such as event type, location, level, scope of impact, and responsible parties, replacing manual sorting and improving the efficiency and accuracy of information collection. On this basis, it forms a closed loop of automation and standardization for the entire process of judgment, allocation, and dispatch. The large-scale model, combined with the knowledge base, automatically outputs classification, grading, responsible parties, and handling suggestions, generates structured orders, and accurately dispatches orders, while simultaneously outputting standard handling scripts and processes, achieving unmanned intervention, unified standards, and zero-error dispatch. Attached Figure Description

[0016] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0018] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0019] In the description of this application, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0020] like Figure 1 This embodiment provides: a smart community multimodal event intelligent processing method based on a large language model, including the following steps: S1: Acquire multi-source event information, including text, voice, and images generated within the smart community, and perform unified formatting to construct a multimodal event dataset; S2: Construct a structured knowledge base containing community management norms, emergency plans, lists of powers and responsibilities, and historical handling cases; establish an index; perform semantic matching between multimodal event information and the knowledge base; and use a retrieval mechanism to impose factual constraints on the reasoning process of the large language model. S3: Based on a large language model, perform in-depth analysis of multimodal event information to extract key feature information such as event type, location, emergency level, scope of impact, and responsible party; S4: Combining the knowledge base and key features, the large language model automatically generates judgment conclusions, completes event classification, emergency level determination and responsible party matching, and outputs standardized handling suggestions; S5: Based on the analysis results, automatically generate structured work orders, match the corresponding handling departments according to the responsible entities to complete accurate work order assignment, and simultaneously output standardized handling scripts and process guidelines.

[0021] In step S1, raw event data, including text, voice, and image information, is obtained from various information channels within the smart community. The collected data of different types undergoes unified formatting processing, including text data processing, voice data processing, and image data processing. During text data processing, invalid characters, garbled text, and emoticons are removed, and the data is structured and annotated according to preset fields. It is then converted to a unified encoding format and stored in the text field. During voice data processing, audio noise reduction and volume normalization are performed. The voice content is converted to text using a speech-to-text model, while retaining the index address of the original audio file. The converted text is then structured and annotated in the same way as the original text data. During image data processing, image compression and resolution adjustments are performed. Normalization processing is performed to unify the image to a size that the model can handle. Basic information about the image is extracted: shooting time, device number, and location coordinates. Through the image understanding model, image content description text is automatically generated and added to the time description field. A unique ID is generated for each image, and a mapping is established between image files and structured text information. Text information includes: text reports submitted by residents through the App, event records from the property management work order system, text information from the community bulletin board, and customer service dialogue records. Voice information includes: recordings of resident hotline calls, audio of on-site voice reports, voice clips collected by community monitoring equipment, and walkie-talkie call recordings. Image information includes: on-site photos taken by residents, images captured by surveillance cameras, on-site record images taken by patrol personnel, and snapshots of equipment malfunctions. The formatted text, speech-to-text, image description text, and associated file indexes are integrated into a unified multimodal event dataset, which includes: common fields, content fields, associated fields, and status fields; common fields: event ID, reporting time, reporting channel, location information, device ID, and reporter ID; content fields: original text content, speech-to-text, and image description text; associated fields: speech file storage path, image file storage path, and image feature vector; status fields: data cleaning status, structured processing status, and quality verification status; each record in the dataset corresponds to a complete event within the community and simultaneously contains structured text information and an index of the original multimodal file.

[0022] In S2, when constructing the community knowledge base, the core knowledge of management norms, emergency plans, lists of powers and responsibilities, and historical cases are comprehensively collected from community management-related documents, systems, and historical records. The collected knowledge is decomposed, labeled, and structured. Each knowledge entry is configured with a unified field including knowledge ID, type label, applicable scenario, priority, related clauses, effective time, and version number. Long documents are split into semantically independent knowledge blocks according to the theme. According to the natural structure of document chapters, titles, and clauses, semantically independent theme units in the document are identified. Based on the principle of fully expressing the content of a single theme, the start and end positions of each knowledge block are divided. Without destroying the semantic integrity, each knowledge block is controlled within a uniform length range. The association identifier between the split knowledge blocks and the original document and adjacent knowledge blocks is retained. At the same time, metadata such as keywords, event type, scope of application, responsible subject, and urgency level are labeled. It should be noted that the following steps are taken: Regulations and systems related to daily community management, security maintenance, and facility operation and maintenance; emergency plans for fires and earthquakes; lists of departmental responsibilities; and historical case studies are collected. These are then broken down, labeled, and structured. Each knowledge entry is configured with a unified field containing a knowledge ID, type tag, applicable scenario, priority, related clauses, effective date, and version number. Long documents are broken down into independent knowledge blocks of 200-500 words each, retaining related identifiers, and annotating keywords, event type, scope of application, responsible party, and urgency level metadata. Finally, all knowledge block text, related identifiers, and metadata are stored uniformly to form a standardized and searchable collection of community management knowledge.

[0023] The structured knowledge blocks are processed to generate a searchable vector library, and vector indexes, keyword indexes and full-text indexes are established. Vector Index Building: Based on the feature data of each knowledge block in the vector library, construct the association mapping relationship between vectors, group all knowledge block vectors in the vector library according to preset rules, so that semantically similar vectors are grouped into the same set, forming a preliminary vector aggregation structure, and establish an association record between each vector and its neighboring vectors. On the basis of the vector set, construct a multi-level association mapping structure, with the upper level being the abstract identifier of the vector set and the lower level being the specific vectors within the set, forming a coarse-to-fine retrieval path. Bind and store the constructed vector association mapping relationship with the unique identifier and metadata of the corresponding knowledge block to form an index structure that supports similarity matching; Keyword Index Building: Based on the keywords and metadata tags labeled in the knowledge blocks, an inverted mapping relationship between keywords and corresponding knowledge blocks is established. Keywords and metadata tags in the knowledge blocks are extracted, and keywords are uniformly standardized, including synonym merging and format normalization. An index item is created for each unique keyword as the core entry point of the inverted index. At the same time, the attribute information of the keyword is recorded. For each index item, a mapping list of all knowledge block identifiers corresponding to it is established, and the unique identifiers and related metadata of all knowledge blocks containing the keyword are recorded, forming an index structure that can be filtered by keywords. Full-text indexing: Based on the complete text content of knowledge blocks, a mapping between text content and knowledge blocks is constructed. The complete text content of knowledge blocks is uniformly cleaned and standardized to remove redundant symbols, stop words and invalid information, and the expression format is unified. Corresponding index entries are generated for the preprocessed text of each knowledge block. The text content is bound to the unique identifier of the knowledge block to form a basic mapping record. Based on the text content of all knowledge blocks, a retrieval structure that supports fuzzy matching and partial matching is established. The system employs a hybrid retrieval rule that prioritizes vector semantic matching and is supplemented by keyword filtering. It sets the retrieval execution order and initiates the vector semantic matching process. Based on the similarity between event vectors and knowledge base vectors, it recalls a set of semantically related knowledge entries. On the basis of vector matching results, it calls event tags and keyword indexes to perform a second filtering of the result set, retaining entries that match tags related to event type, urgency level, and responsible entity. The filtered results are then comprehensively sorted according to vector similarity, tag matching degree, and knowledge priority, placing relevant and applicable entries at the top. When vector matching results are insufficient, it automatically supplements the search by calling the full-text index for fallback retrieval and constructs a multi-dimensional index and retrieval engine. It should be noted that the system uses the similarity between event vectors and knowledge base vectors as the core criterion to quickly retrieve semantically relevant knowledge entries through vector indexing. This process does not rely on literal keyword matching but rather compares the overall semantics of the event and the knowledge block. This effectively captures the implicit intent and scenario relevance in event descriptions. Even if the wording used by the event reporter is not standard, colloquial, or ambiguous, the system can still find the semantically most appropriate course of action. After obtaining the initial result set through vector matching, the system calls upon event tags and keyword indexes to perform a secondary filtering. Only entries that completely match or reach a preset matching degree with tags such as event type, urgency level, and responsible party are retained. Knowledge that is semantically similar but does not fit the applicable scenario is removed. For example, for elevator malfunction events, the system will filter out maintenance standards applicable to indoor public areas based on the event location tag, rather than handling clauses applicable to residential units, ensuring the scenario adaptability of the search results. For filtered entries, the system performs a comprehensive ranking based on three dimensions: vector similarity, tag matching degree, and knowledge priority. Vector similarity determines the degree of fit between knowledge content and event description, tag matching degree reflects the adaptability of rules and event attributes, and knowledge priority is pre-set according to the version validity and scope of application of community management regulations. For example, the latest emergency plan has a higher priority than the old version. Ultimately, entries with high relevance and strong applicability are ranked first. When the vector matching results are insufficient or the overall similarity is lower than the preset threshold, the system will automatically supplement and call the full-text index for fallback retrieval. By performing fuzzy matching and partial matching on the text content of knowledge blocks, the system can uncover relevant handling clauses that may have been missed by vector semantics, avoiding the loss of key information due to semantic misunderstanding. This forms a multi-dimensional retrieval engine consisting of accurate vector retrieval, keyword filtering and scenario adaptation, priority ranking result optimization, and full-text index fallback guarantee. This process transforms multimodal event information into event vectors and extracts structured tags for event type, location, urgency level, and responsible party. Multimodal event information from text, speech-to-text, and image descriptions is integrated into a standardized event information set according to a unified format. The integrated event text is cleaned and standardized, removing invalid characters and redundant information, and unifying the expression format. Based on the preprocessed text content, key information such as event type, location, urgency level, and responsible party is extracted according to preset rules and converted into standardized structured tags. The preprocessed event text content and structured tags are then converted into feature data in a unified format according to preset rules to generate event vectors. Similarity retrieval is performed between the event vectors and a knowledge vector base. Multi-dimensional semantic matching is then performed based on event tags according to event type, urgency level, location, and scope of impact. The search results are sorted according to similarity, relevance, and priority, and relevant knowledge entries are selected. The returned preliminary set of knowledge entries undergoes multi-dimensional matching and verification based on event tags: The event's type, urgency level, location, and scope of impact tags are compared one by one with the corresponding tags in the knowledge base entries. The structured tags of the events are mapped to the tags of the knowledge base entries according to preset fields, establishing a field mapping relationship for tag comparison. For each field's tag value, a comparison is performed according to preset matching rules: Event Type: Determines if the event type tag matches the event type tag of the knowledge base entry; Urgency Level: Determines if the event's urgency level tag matches the applicable urgency level of the knowledge base entry; Location Information: Determines if the event's location tag is within the applicable location range of the knowledge base entry; Scope of Impact: Determines if the event's scope of impact tag matches the applicable scope of impact of the knowledge base entry. The matching results for all fields are statistically analyzed, and the overall tag matching degree is calculated according to preset rules. If the matching degree reaches a preset threshold, the knowledge base entry is determined to match the event tag and is retained; otherwise, it is excluded. It should be noted that the process involves converting multimodal event information into event vectors and extracting structured tags for event type, location, urgency level, and responsible party. Multimodal event information from text, speech-to-text, and image descriptions is then integrated into a standardized event information set using a unified format. This process automatically aligns timestamps from different sources with the event subject, ensuring that information from resident app messages, phone calls, and on-site images corresponds to the same event object, avoiding confusion or mismatches between multiple sources. Next, the integrated event text undergoes cleaning and standardization, removing invalid characters and redundant information, and unifying the expression format. For example, colloquial descriptions like "the elevator isn't moving" or "stuck in the middle of the floor" are uniformly converted to standard expressions like "elevator out of service" and "elevator entrapment," conforming to system processing specifications. Meaningless interjections, repetitive expressions, and symbolic noise are filtered out, providing a clean and consistent text foundation for subsequent information extraction. Based on the preprocessed text... This content extracts key information about the event type, location, urgency level, and responsible party according to preset rules, and converts it into standardized structured tags. For example, for the event description "People trapped in the elevator of Unit 2, Building 3, and someone has called the police," the system will extract the tag information as "elevator entrapment malfunction," "elevator of Unit 2, Building 3," "high urgency level," and "elevator maintenance unit and property customer service" as the responsible parties. The values ​​of these tags are all predefined in the community event classification system and the list of rights and responsibilities to ensure uniform format and no ambiguity. Subsequently, the preprocessed event text content and structured tags are converted into feature data in a unified format according to preset rules to generate an event vector. This vector integrates the overall semantics of the event and key tag information, including both the semantic features of the text description and the encoding of structured information such as type and location, which can comprehensively represent the core attributes of the event so as to conduct efficient similarity comparison with the knowledge base vector in the future. The event vector is compared with the knowledge vector database using similarity retrieval. Multi-dimensional semantic matching is then performed using event tags, categorized by event type, urgency level, location, and impact scope. The search results are sorted by similarity, relevance, and priority to filter relevant knowledge entries. The initial set of knowledge entries returned by the similarity retrieval is then validated using multi-dimensional matching based on event tags. The structured tags of the events are mapped to the tags of the knowledge base entries according to preset fields, establishing a field mapping relationship for tag comparison. For each field's tag value, a comparison is performed according to preset matching rules: When comparing the event type field, it is determined whether the event type tag is consistent with the event type tag of the knowledge base entry, or belongs to a subclass under the same preset category. For example, "elevator entrapment" and "elevator malfunction" belong to the same classification system and are considered a match. When comparing the urgency level field, it is determined whether the event's urgency level tag matches the applicable urgency level of the knowledge base entry. For example, a high-urgency "elevator entrapment" event will only match events applicable to high-urgency levels. Prioritized emergency response plans, rather than routine maintenance procedures, are compared using the location information field. The comparison determines whether the event location tag falls within the applicable location range of the knowledge base entry. For example, if the event location is "Building 3, Unit 2," it will match management regulations applicable to that building and its associated public areas, not regulations for other buildings. Similarly, when comparing the impact scope field, it determines whether the event's impact scope tag matches the applicable impact scope of the knowledge base entry. For example, for an elevator malfunction affecting "multiple people in the unit," it will match handling procedures involving unit-level emergency notifications, not maintenance regulations only for a single device. The matching results of all fields are statistically analyzed, and the overall tag matching degree is calculated according to preset rules. If the matching degree reaches a preset threshold, the knowledge base entry is considered to match the event tag and is retained; otherwise, it is excluded. This multi-dimensional tag verification effectively filters out semantically similar but mismatched knowledge entries in terms of scenario, level, or responsibility, ensuring that the final retrieved knowledge entries not only fit the event semantics but also comply with the specific rules and requirements of community management.

[0024] Then, select entries whose tag matching degree meets the preset requirements. For the entries that pass the verification, sort them comprehensively according to the three dimensions of vector phase velocity, tag matching degree, and knowledge priority, and form an entry priority list. Then, filter the sorted list according to the preset similarity threshold and the number of returned entries. When extracting key information about event type, location, urgency level, and responsible party according to preset rules, extraction rules are pre-configured for each key information item. These rules include: a keyword set for the event type, format and expression rules for location information, grading rules for urgency level, and rules for matching the rights and responsibilities of the responsible party. The pre-processed event text is then matched and identified according to these preset rules: Event type: Matching expressions in the text that correspond to preset event type keywords to identify the event category; Location information: Extracting location information such as building number and area name from the text according to preset format and keywords; Urgency level: Matching preset grading rules based on risk descriptions and urgency expressions in the text; Responsible party: Combining event type and location information with preset rules for matching rights and responsibilities to identify the corresponding responsible department. The initially identified information is verified to check its completeness and rationality. For information with non-standard expressions, the format is normalized according to preset rules to generate unified and standardized key information results. Input the retrieved knowledge entries, event information, and task requirements for analysis, classification, grading, and task assignment into the constraint template to impose factual constraints on the reasoning process of the large model; It should be noted that when extracting key information about event type, location, urgency level, and responsible party according to preset rules, extraction rules are pre-configured for each key information item. These rules include: a keyword set for the event type, format and expression rules for location information, rules for determining the urgency level, and rules for matching the rights and responsibilities of the responsible party. These rules are pre-configured based on existing community management regulations, emergency plans, and lists of rights and responsibilities. For example, the event type keyword set includes standard expressions for common community events such as "elevator malfunction," "water pipe leak," and "noise pollution." The location information rules define "building number, unit number, etc." The unified format of "element number, floor, area" is used. Emergency level rules categorize events into high, medium, and low based on their risk level. Responsibility rules clearly define the departments responsible for handling different event types; for example, elevator malfunctions correspond to maintenance companies and property management customer service, while water pipe leaks correspond to engineering repair teams. Pre-processed event text is matched and identified according to preset rules: regarding event type, the text is matched with descriptions corresponding to preset event type keywords to identify the event category. For example, from the description "elevator stuck in the middle of the floor, someone trapped," the keyword "elevator entrapment" is matched, identifying the event type as "elevator entrapment." Regarding the "trapped person malfunction," in terms of location information, the system extracts location information such as building number and area name from the text according to preset formats and keywords. For example, from the description "in the elevator of Unit 2, Building 3," the location tag "elevator of Unit 2, Building 3" is extracted. Regarding the urgency level, preset classification rules are matched based on the risk descriptions and urgency expressions in the text. For example, expressions such as "someone is trapped" and "alarm has been called" will trigger a high-urgency-level judgment rule, marking the event as high priority. Regarding the responsible party, preset rules for matching authority and responsibility are matched based on the event type and location information to identify the corresponding responsible department. For example, "elevator entrapment malfunction" corresponds to "elevator maintenance unit" and "property emergency customer service". The initially identified information is verified to check its completeness and rationality. Information with non-standard expressions is normalized according to preset rules to generate unified and standardized key information results. For example, the colloquial "the elevator is not moving" is uniformly corrected to the standard expression "elevator out of service malfunction", and the vague "over there in Building 3" is standardized to "public area of ​​Building 3". At the same time, it is checked whether key information is missing. If the location information is missing, it is marked as to be supplemented to ensure that the final generated structured labels are complete, standardized and unambiguous.

[0025] In S3, based on the completed multimodal event dataset as the input source, the system accesses and loads multi-source event information, including text descriptions submitted by residents through App, telephone, and on-site reporting channels, text content converted by speech recognition, and image description text generated by image understanding. During access, the system verifies the integrity of the data and filters out invalid and duplicate information. This system employs a large language model to perform integrated deep analysis of multi-source event information. It performs basic semantic understanding on individual pieces of information, automatically identifying the core semantic subject, key behaviors, and contextual background in event descriptions. Pre-configured identification rules include: core semantic subject identification rules, key behavior identification rules, and contextual background identification rules. The core semantic subject identification rules include keywords and type definitions of objects, roles, and entities involved in the event. The key behavior identification rules include keywords and action types describing actions, behaviors, and state changes. The contextual background identification rules include keywords and format rules for scene information such as time, location, environment, and conditions. The system segments and divides the event description text into sentences and paragraphs, breaking down long texts into independent semantic units. For each semantic unit, it matches according to preset rules: matching core semantic subject keywords to mark the core objects in the event description; matching key behavior keywords to mark actions in the event; and matching contextual background keywords to mark the time, location, and environment of the event. The marked subjects, behaviors, and background information are then integrated to form a complete event semantic structure. It should be noted that the pre-configured recognition rules include: core semantic subject recognition rules, key behavior recognition rules, and contextual background recognition rules. Among them, the core semantic subject recognition rules include keywords and type definitions of objects, roles, and entities involved in the event. For example, in a community scenario, entity keywords such as "elevator," "water supply pipe," "resident," and "public lighting" and their corresponding types are preset. The key behavior recognition rules include keywords and action types describing actions, behaviors, and state changes, such as "out of service," "leaking," "damage," and "complaint," which are used to identify the core actions or states of the event. The contextual background recognition rules include keywords and format rules for scene information such as time, location, environment, and conditions, such as the location format of "building number, unit number, floor" and the time expression rules of "weekdays, holidays, and nighttime." The system segments and divides the event description text into sentences and paragraphs, breaking down long texts into independent semantic units. Taking the reported information "The elevator on the 15th floor of Unit 2, Building 3 is stuck in the middle of the floor, and someone is trapped. The property management has been called" as an example, the system will break it down into multiple semantic units such as "elevator on the 15th floor of Unit 2, Building 3", "stuck in the middle of the floor", "someone is trapped", and "the property management has been called". This avoids recognition omissions caused by the complexity of long text structures. For each semantic unit, it matches according to preset rules: matching core semantic subject keywords to mark the core objects in the event description, such as "elevator" and "residents" as core subjects in the above information; matching key behavioral keywords to mark actions in the event, such as "stuck in the middle of the floor" and "trapped"; and matching contextual background keywords to mark the scene information of the time, place, and environment of the event, such as the location information of "15th floor of Unit 2, Building 3" and the implicit time information such as "night / weekday". It performs cross-modal contextual analysis on multi-source information of the same event, integrating complementary information from text, audio, and image descriptions to form a complete understanding of the event. Simultaneously, it automatically disambiguates ambiguous or vague content, identifying ambiguous content in the text according to preset rules, including unreferential pronouns, polysemous words, and incomplete phrases. It marks the location and type of potential ambiguity, retrieves the sentence, paragraph, and other multi-source information related to the same event containing the ambiguous content, constructs a contextual reference set for ambiguity dissolution, obtains the complete semantic background of the event, and applies preset processing rules based on the type of ambiguity: for unreferential content, it restores the referent based on the entity, location, and event description in the context; for polysemous words, it matches a reasonable interpretation based on the event scenario and contextual semantics; for incomplete content, it supplements key information based on the event background and logical relationships; and it performs consistency verification on the processed content. After verification, it replaces the original ambiguous description with the ambiguity-dissolved text, forming complete and clear event semantic content, and supplements key information based on contextual logic.

[0026] Based on the complete event semantics formed after parsing, key feature information is extracted. Key feature information includes: identifying the event type from the event description, locating the specific location of the event from the text description, determining the urgency level by combining the degree of impact and descriptive tendency in the event description, identifying the spatial scope of the event's impact, and matching the corresponding responsible parties. Independent extraction rules are pre-configured for each type of key feature, including: keywords and classification rules for the event type, format and recognition rules for location information, grading and determination rules for the urgency level, rules for defining the scope of impact, and rules for matching the rights and responsibilities of the responsible parties. The system performs keyword matching on the event description text, identifies the event category based on a pre-defined event type classification system, and generates standardized event type tags. It also extracts location information such as building number, unit number, floor, and area name from the text according to pre-defined location identification rules, converting it into standardized location tags. Furthermore, it matches the event description with pre-defined emergency level determination rules based on the risk level, scope of impact, and descriptive tendencies, automatically classifying the event into emergency, general, and low levels to generate emergency level tags. Finally, it extracts the spatial scope of the event's impact and the size of the affected population from the text, generates impact scope description information based on pre-defined scope definition rules, and automatically identifies the responsible department based on the event type and location information, generating a responsible entity matching result. The extracted key feature information is organized and standardized to generate structured data containing event type labels, location labels, emergency level, description of the scope of impact, and matching results of responsible parties; It should be noted that: independent extraction rules are pre-configured for each type of key feature, including: keyword and classification rules for event types, format and recognition rules for location information, grading rules for emergency levels, rules for defining the scope of impact, and rules for matching the responsibilities of the responsible parties; keyword matching is performed on the event description text, and combined with the pre-set event type classification system, the category to which the event belongs is identified. This classification system divides community events into primary categories such as public security events, fire events, facility malfunctions, environmental sanitation, and public services, with multiple secondary subcategories under each primary category. Standardized event type labels are formed by matching keyword combinations and semantic patterns in the event description; according to the pre-set location recognition rules, the location information of building number, unit number, floor, and area name is extracted from the text. This rule supports various common expression formats such as "XX Community, Building X, Unit X, Floor X" and can automatically complete omitted administrative division information, converting it into standardized location labels containing latitude and longitude coordinates; based on the risk level, scope of impact, and descriptive tendency in the event description, the pre-set emergency level is determined. The system automatically categorizes events into emergency, general, and low levels. For example, events containing keywords like "fire" or "injury" and involving personnel safety are automatically classified as emergency, while events involving only localized facility malfunctions are classified as general or low, thus creating an emergency level label. It extracts the spatial scope and affected population from the text and, combined with pre-defined scope definition rules, generates a description of the impact area, such as "the entire Building 3 of XX Community" or "community public areas." Combining event type and location information, it matches pre-defined responsibility rules, establishing a mapping relationship between event type, location, and responsible department. For example, fire incidents are automatically matched with the community fire liaison point and the street emergency office, while facility malfunctions are matched with the property maintenance department, automatically identifying the responsible department and generating a responsible entity matching result. The extracted key feature information is then organized and standardized to generate structured data containing event type labels, location labels, emergency level, impact area description, and responsible entity matching results. This data is stored in JSON format and can be directly used for subsequent event classification and process distribution.

[0027] In S4, relevant entries from the retrieved community knowledge base and key event feature information are accessed to form a complete input context. The large language model combines management regulations, emergency plans, lists of powers and responsibilities, and historical handling cases in the knowledge base to perform rule matching and compliance verification on key event features: matching corresponding handling regulations according to event type, matching graded response requirements according to urgency level, and matching corresponding responsible parties according to location and scope of impact. Based on the rule matching results, the system automatically completes the core judgments: event classification, urgency level determination, and responsible entity matching. Event classification: Based on the event type definitions and handling standards in the knowledge base, the specific category of the event is determined. Urgency level determination: Combining the impact of the event with the grading standards in the contingency plan, the priority of the event response is clarified. Responsible entity matching: Based on the list of responsibilities and location, and event type information, the system automatically determines the department responsible for handling the matter. After completing the core judgments, the big data model, combined with historical handling cases and management specifications, automatically generates standardized handling suggestions that can be directly executed, including handling procedures, response time limits, precautions, and resource allocation suggestions.

[0028] In S5, the structured analysis results output, including event classification, urgency level, responsible party, and standardized handling suggestions, serve as the core input for work order generation and dispatch. According to the preset work order template, the key information in the analysis conclusion is automatically filled into the corresponding fields of the work order to generate a standardized work order containing event number, event type, location, urgency level, scope of impact, responsible party, handling suggestions, and response time limit. Based on the responsible party information in the work order, the system automatically matches the corresponding handling department in the community management system, automatically sets the response priority based on the urgency level, and completes the targeted dispatch of the work order. At the same time, the dispatch results are verified to ensure that the work orders are not mis-dispatched, missed, or duplicated. Based on the event type, urgency level, and handling suggestions in the work order, the system retrieves the corresponding pre-set handling scripts and standard procedures from the knowledge base to generate standardized guidelines that can be directly used for on-site handling. These guidelines include key communication points, operating steps, precautions, and abnormal handling methods. After the work order is dispatched, the system automatically updates the work order status to "dispatched" and simultaneously records the dispatched event, receiving department, and response time limit.

[0029] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0030] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0031] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0032] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0033] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0034] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0035] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A smart community multimodal event intelligent processing method based on a large language model, characterized in that, Includes the following steps: S1: Acquire multi-source event information, including text, voice, and images generated within the smart community, and perform unified formatting to construct a multimodal event dataset; S2: Construct a structured knowledge base containing community management norms, emergency plans, lists of powers and responsibilities, and historical handling cases; establish an index; perform semantic matching between multimodal event information and the knowledge base; and use a retrieval mechanism to impose factual constraints on the reasoning process of the large language model. S3: Based on a large language model, perform in-depth analysis of multimodal event information to extract key feature information such as event type, location, emergency level, scope of impact, and responsible party; S4: Combining the knowledge base and key features, the large language model automatically generates judgment conclusions, completes event classification, emergency level determination and responsible party matching, and outputs standardized handling suggestions; S5: Based on the analysis results, automatically generate structured work orders, match the corresponding handling departments according to the responsible entities to complete accurate work order assignment, and simultaneously output standardized handling scripts and process guidelines.

2. The intelligent processing method for multimodal events in smart communities based on a large language model according to claim 1, characterized in that, In step S1, raw event data, including text, voice, and image information, is obtained from various information channels of the smart community. The different types of data are uniformly formatted, including text data processing, voice data processing, and image data processing. The formatted text, speech-to-text, image description text, and associated file index are integrated into a unified multimodal event dataset, which specifically includes: common fields, content fields, associated fields, and status fields. Each record in the dataset corresponds to a complete event within the community and contains both structured text information and an index of the original multimodal file.

3. The intelligent processing method for multimodal events in smart communities based on a large language model according to claim 1, characterized in that, In S2, when constructing the community knowledge base, core knowledge of management norms, emergency plans, lists of powers and responsibilities, and historical cases are comprehensively collected from community management-related documents, systems, and historical records. The collected knowledge is decomposed, labeled, and structured. Each knowledge entry is configured with a unified field including knowledge ID, type tag, applicable scenario, priority, related clauses, effective time, and version number. Long documents are split into semantically independent knowledge blocks according to the theme. Based on the natural structure of document chapters, titles, and clauses, semantically independent theme units in the document are identified. The start and end positions of each knowledge block are divided according to the principle of fully expressing the content of a single theme. Without destroying semantic integrity, each knowledge block is controlled within a uniform length range. The association identifiers between the split knowledge blocks and the original document and adjacent knowledge blocks are retained. At the same time, metadata such as keywords, event type, scope of application, responsible entity, and urgency level are labeled.

4. The intelligent processing method for multimodal events in smart communities based on a large language model according to claim 3, characterized in that, The structured knowledge blocks are processed to generate a searchable vector library, and vector indexes, keyword indexes and full-text indexes are established. Vector Index Building: Based on the feature data of each knowledge block in the vector library, construct the association mapping relationship between vectors, group all knowledge block vectors in the vector library according to preset rules, so that semantically similar vectors are grouped into the same set, forming a preliminary vector aggregation structure, and establish an association record between each vector and its neighboring vectors. On the basis of the vector set, construct a multi-level association mapping structure, with the upper level being the abstract identifier of the vector set and the lower level being the specific vectors within the set, forming a coarse-to-fine retrieval path. Bind and store the constructed vector association mapping relationship with the unique identifier and metadata of the corresponding knowledge block to form an index structure that supports similarity matching; Keyword Index Building: Based on the keywords and metadata tags labeled in the knowledge blocks, an inverted mapping relationship between keywords and corresponding knowledge blocks is established. Keywords and metadata tags in the knowledge blocks are extracted, and keywords are uniformly standardized, including synonym merging and format normalization. An index item is created for each unique keyword as the core entry point of the inverted index. At the same time, the attribute information of the keyword is recorded. For each index item, a mapping list of all knowledge block identifiers corresponding to it is established, and the unique identifiers and related metadata of all knowledge blocks containing the keyword are recorded, forming an index structure that can be filtered by keywords. Full-text indexing: Based on the complete text content of knowledge blocks, a mapping between text content and knowledge blocks is constructed. The complete text content of knowledge blocks is uniformly cleaned and standardized to remove redundant symbols, stop words, and invalid information, and the expression format is standardized. Corresponding index entries are generated for the preprocessed text of each knowledge block. The text content is bound to the unique identifier of the knowledge block to form a basic mapping record. Based on the text content of all knowledge blocks, a retrieval structure that supports fuzzy matching and partial matching is established.

5. The intelligent processing method for multimodal events in smart communities based on a large language model according to claim 4, characterized in that, The system employs a hybrid retrieval rule that prioritizes vector semantic matching and is supplemented by keyword filtering. It sets the retrieval execution order and initiates the vector semantic matching process. Based on the similarity between event vectors and knowledge base vectors, it recalls a set of semantically related knowledge entries. On the basis of vector matching results, it calls event tags and keyword indexes to perform a second filtering of the result set, retaining entries that match tags related to event type, urgency level, and responsible entity. The filtered results are then comprehensively sorted according to vector similarity, tag matching degree, and knowledge priority, placing relevant and applicable entries at the top. When vector matching results are insufficient, it automatically supplements the search by calling the full-text index for fallback retrieval and constructs a multi-dimensional index and retrieval engine. This process transforms multimodal event information into event vectors and extracts structured tags for event type, location, urgency level, and responsible party. Multimodal event information from text, speech-to-text, and image descriptions is integrated into a standardized event information set according to a unified format. The integrated event text is then cleaned and standardized, removing invalid characters and redundant information, and unifying the expression format. Based on the preprocessed text content, key information regarding event type, location, urgency level, and responsible party is extracted according to preset rules and converted into standardized structured tags. Finally, the preprocessed event text content and structured tags are converted into feature data in a unified format according to preset rules, generating event... Event vectors are used to perform similarity searches between event vectors and a knowledge vector database. Combined with event tags, multi-dimensional semantic matching is performed based on event type, urgency level, location, and scope of impact. Search results are sorted by similarity, relevance, and priority to filter relevant knowledge entries. The initial set of knowledge entries returned by the similarity search is then validated through multi-dimensional matching based on event tags: the event type, urgency level, location, and scope of impact tags are compared one by one with the corresponding tags in the knowledge database entries. The structured tags of the events are mapped to the tags of the knowledge database entries according to preset fields, establishing a field mapping relationship for tag comparison. For each field's tag value, a comparison is performed according to preset matching rules. Event Type: Determines whether the event type label matches the event type label of the knowledge base entry; Urgency Level: Determines whether the urgency level label of the event matches the applicable urgency level of the knowledge base entry; Location information: Determines whether the event location tag is within the applicable location range of the knowledge base entry; Scope of impact: Determine whether the scope of impact label of the event matches the applicable scope of impact of the knowledge base entry; The system analyzes the matching results of all fields and calculates the overall tag matching degree according to preset rules. If the matching degree reaches the preset threshold, the knowledge base entry is determined to match the event tag and is retained; otherwise, it is excluded.

6. The method for intelligent processing of multimodal events in smart communities based on a large language model according to claim 5, characterized in that, Then, select entries whose tag matching degree meets the preset requirements. For the entries that pass the verification, sort them comprehensively according to the three dimensions of vector phase velocity, tag matching degree, and knowledge priority, and form an entry priority list. Then, filter the sorted list according to the preset similarity threshold and the number of returned entries. When extracting key information about event type, location, urgency level, and responsible party according to preset rules, extraction rules are pre-configured for each key information item. These rules include: keyword set for event type, format and expression rules for location information, grading rules for urgency level, and rules for matching the rights and responsibilities of the responsible party. The pre-processed event text is then matched and identified according to these preset rules. Event Type: Matches the text with descriptions corresponding to preset event type keywords to identify the category to which the event belongs; Location information: Extract location information such as building number and area name from the text according to preset format and keywords; Emergency Level: Based on the risk description and urgency statement in the text, and matching the preset classification rules; Responsible Entity: By combining event type and location information, and matching pre-defined rules for corresponding rights and responsibilities, the corresponding responsible department is identified; The initially identified information is verified to check its completeness and rationality. For information with non-standard expressions, the format is normalized according to preset rules to generate unified and standardized key information results. Input the retrieved knowledge entries, event information, and task requirements for analysis, classification, grading, and task assignment into the constraint template to impose factual constraints on the reasoning process of the large model.

7. The intelligent processing method for multimodal events in smart communities based on a large language model according to claim 1, characterized in that, In S3, based on the completed multimodal event dataset as the input source, the system accesses and loads multi-source event information, including text descriptions submitted by residents through App, telephone, and on-site reporting channels, text content converted by speech recognition, and image description text generated by image understanding. During access, the system verifies the integrity of the data and filters out invalid and duplicate information. By using a large language model to perform integrated deep analysis of multi-source event information, basic semantic understanding of individual information, automatic identification of core semantic subjects, key behaviors and contextual background in event descriptions, and pre-configuration of identification rules: core semantic subject identification rules, key behavior identification rules and contextual background identification rules; Core semantic subject identification rules: including keywords and type definitions of objects, roles, and entities involved in the event; Key behavior recognition rules: include keywords and action types that describe actions, behaviors, and state changes; Context recognition rules: Keywords and format rules that include scene information such as time, location, environment, and conditions; The event description text is segmented into sentences and paragraphs, breaking down long texts into independent semantic units. Each semantic unit is then matched according to preset rules. Match core semantic keywords and mark the core objects in the event description; Match key behavioral keywords and tag actions within the event; Match contextual keywords to mark the time, location, and environment in which the event occurred; The marked subjects, behaviors, and background information are linked and integrated to form a complete semantic structure of the event; It performs cross-modal contextual analysis on multi-source information of the same event, integrating complementary information from text, audio, and image descriptions to form a complete understanding of the event. Simultaneously, it automatically disambiguates ambiguous or vague content, identifying ambiguous content in the text according to preset rules, including unreferential pronouns, polysemous words, and incomplete phrases. It marks the location and type of potential ambiguity, retrieves the sentence, paragraph, and other multi-source information related to the same event containing the ambiguous content, constructs a contextual reference set for ambiguity dissolution, obtains the complete semantic background of the event, and applies preset processing rules based on the type of ambiguity: for unreferential content, it restores the referent based on the entity, location, and event description in the context; for polysemous words, it matches a reasonable interpretation based on the event scenario and contextual semantics; for incomplete content, it supplements key information based on the event background and logical relationships; and it performs consistency verification on the processed content. After verification, it replaces the original ambiguous description with the ambiguity-dissolved text, forming complete and clear event semantic content, and supplements key information based on contextual logic.

8. The intelligent processing method for multimodal events in smart communities based on a large language model according to claim 7, characterized in that, Based on the complete event semantics formed after parsing, key feature information is extracted. Key feature information includes: identifying the event type from the event description, locating the specific location of the event from the text description, determining the urgency level by combining the degree of impact and descriptive tendency in the event description, identifying the spatial scope of the event's impact, and matching the corresponding responsible parties. Independent extraction rules are pre-configured for each type of key feature, including: keywords and classification rules for the event type, format and recognition rules for location information, grading and determination rules for the urgency level, rules for defining the scope of impact, and rules for matching the rights and responsibilities of the responsible parties. The system performs keyword matching on the event description text, identifies the event category based on a pre-defined event type classification system, and generates standardized event type tags. It also extracts location information such as building number, unit number, floor, and area name from the text according to pre-defined location identification rules, converting it into standardized location tags. Furthermore, it matches the event description with pre-defined emergency level determination rules based on the risk level, scope of impact, and descriptive tendencies, automatically classifying the event into emergency, general, and low levels to generate emergency level tags. Finally, it extracts the spatial scope of the event's impact and the size of the affected population from the text, generates impact scope description information based on pre-defined scope definition rules, and automatically identifies the responsible department based on the event type and location information, generating a responsible entity matching result. The extracted key feature information is organized and standardized to generate structured data that includes event type labels, location labels, emergency level, description of the scope of impact, and matching results of responsible parties.

9. The intelligent processing method for multimodal events in smart communities based on a large language model according to claim 1, characterized in that, In S4, relevant entries from the retrieved community knowledge base and key event feature information are accessed to form a complete input context. The large language model combines management regulations, emergency plans, lists of powers and responsibilities, and historical handling cases in the knowledge base to perform rule matching and compliance verification on key event features: matching corresponding handling regulations according to event type, matching graded response requirements according to urgency level, and matching corresponding responsible parties according to location and scope of impact. Based on the rule matching results, the core judgments are automatically completed: event classification, emergency level determination, and matching of responsible parties; After completing the core judgment, the big language model combines historical handling cases and management standards to automatically generate standardized handling suggestions that can be directly executed, including handling procedures, response time limits, precautions and resource allocation suggestions.

10. The intelligent processing method for multimodal events in smart communities based on a large language model according to claim 1, characterized in that, In S5, the structured analysis results output, including event classification, urgency level, responsible party, and standardized handling suggestions, serve as the core input for work order generation and dispatch. According to the preset work order template, the key information in the analysis conclusion is automatically filled into the corresponding fields of the work order to generate a standardized work order containing event number, event type, location, urgency level, scope of impact, responsible party, handling suggestions, and response time limit. Based on the responsible party information in the work order, the system automatically matches the corresponding handling department in the community management system, automatically sets the response priority based on the urgency level, and retrieves the corresponding pre-set handling scripts and standard procedures from the knowledge base according to the event type, urgency level, and handling suggestions in the work order. This generates standardized guidelines that can be directly used for on-site handling, including communication points, operating steps, precautions, and abnormal handling methods. After the work order is dispatched, the system automatically updates the work order status to "dispatched" and simultaneously records the dispatched event, receiving department, and response time limit.

Images