Active work order decomposition method and system, electronic equipment and medium

Abstract

The invention discloses an active work order decomposition method and system, electronic equipment and a medium. The method comprises the steps of obtaining a dialogue text input by a user; subject classification and keyword measurement are conducted on the dialogue text input by the user, a plurality of task units are obtained, and each task unit is associated with at least one preset subject knowledge base; each task unit matches the task unit with knowledge in a preset theme knowledge base associated with the task unit to obtain first knowledge; generating first structured information based on the task unit and the first knowledge corresponding to the task unit; and generating a work order preview corresponding to each task unit according to the first structured information of each task unit. According to the method, the multi-theme text is introduced to be pre-decomposed and recognized, so that the problem that the accuracy is greatly reduced when a single model processes a scene complex text is effectively solved, and the accuracy of work order generation is improved.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence, natural language processing, and large language model application technology, and more specifically, to a proactive work order decomposition method, system, electronic device, and medium. Background Technology

[0002] In customer service and internal management, it is often necessary to generate work orders based on user conversations in instant messaging tools (such as WeChat Work).

[0003] Currently, relevant technical solutions can be divided into two types: passive and active. Passive solutions rely on human assistance (such as a "butler"). The butler needs to first understand the user's multiple dialogues, consciously filter out the key content that needs to be generated into a work order, and then manually copy and paste the text for subsequent model processing. Although this method reduces the requirements for model capabilities, it is not fully automated and requires a lot of manpower for preliminary screening and basic classification.

[0004] Traditional proactive processing solutions typically use a single model to directly process the raw dialogue from the user. When the user's dialogue content is complex, such as containing multiple independent requests within a single sentence, the model's recognition and processing accuracy drops significantly, potentially falling below 80% in small-batch tests. This results in a high error rate in ticket generation, failing to meet the needs of practical applications.

[0005] LLM performs well across various domains, but its high inference costs make it unprofitable in practical applications.

[0006] Therefore, how to effectively process complex dialogue texts containing multiple topics and accurately decompose them into structured work orders under the premise of full automation is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0007] This invention provides a proactive work order decomposition method, system, electronic device, and medium, which solves the technical problem of low accuracy in the processing of complex texts with multiple topics in existing automated work order systems.

[0008] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: The first embodiment of the present invention provides a proactive work order decomposition method, comprising the following steps: Get the dialogue text entered by the user; The user-input dialogue text is categorized by topic and evaluated by keywords to obtain multiple task units; For each task unit, the knowledge in the task unit is matched with the knowledge in multiple preset topic knowledge bases to obtain the first knowledge; Based on the task unit and the first knowledge corresponding to that task unit, generate the first structured information; Based on the first structured information of each task unit, generate a work order preview corresponding to each task unit.

[0009] Furthermore, the user-input dialogue text is categorized by topic and evaluated for keywords, including: The user-input dialogue text is split into multiple sentence segments and word segments; By evaluating the degree of correlation between the sentence segments and different preset topic knowledge bases, several candidate sentences and candidate preset topic knowledge bases corresponding to the candidate sentences are obtained based on the degree of correlation. Candidate keywords are selected based on the importance of keywords in the word segments; Based on the candidate statements, their topics, and candidate keywords, multiple task units are obtained, wherein each task unit is associated with at least one preset topic knowledge base.

[0010] Furthermore, the matching of the task unit with the knowledge in the preset topic knowledge base associated with the task unit to obtain the first knowledge includes: Calculate the similarity between the task unit and the knowledge in the associated preset topic knowledge base, and obtain the knowledge in the preset topic knowledge base whose similarity is greater than the similarity threshold as the first knowledge.

[0011] Furthermore, based on the importance of keywords in the aforementioned word segments, candidate keywords are selected, including: Based on each keyword in the word segmentation and the preset word importance table, keywords with an importance greater than the preset importance threshold are selected as candidate keywords. The preset word importance table includes words and their importance. The importance is calculated based on the number of times the word appears in a large amount of corpus. The more times it appears, the lower the importance of the word. The fewer times it appears, the higher the importance of the word.

[0012] Furthermore, based on the task unit and the first knowledge corresponding to that task unit, first structured information is generated, including: The task unit and the corresponding first knowledge are input into the large language model, and the large language model outputs the first structured information.

[0013] Furthermore, the large language model is obtained by fine-tuning or targeted training based on the business scenario.

[0014] A second embodiment of the present invention provides an active work order decomposition system, comprising: The acquisition module acquires the dialogue text input by the user; The first processing module performs topic classification and keyword evaluation on the dialogue text input by the user to obtain multiple task units; The first matching module matches each task unit with knowledge in multiple preset topic knowledge bases to obtain the first knowledge. The second processing module generates first structured information based on the task unit and the first knowledge corresponding to the task unit. The third processing module generates a work order preview for each task unit based on the first structured information of each task unit.

[0015] Furthermore, the first processing module performs topic classification and keyword evaluation on the user-input dialogue text, including: The user-input dialogue text is split into multiple sentence segments and word segments; By evaluating the degree of correlation between the sentence segments and different preset topic knowledge bases, several candidate sentences and candidate preset topic knowledge bases corresponding to the candidate sentences are obtained based on the degree of correlation. Candidate keywords are selected based on the importance of keywords in the word segments; Based on the candidate statements, their topics, and candidate keywords, multiple task units are obtained, wherein each task unit is associated with at least one preset topic knowledge base.

[0016] A second embodiment of the present invention provides an electronic device, including a processor and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, it implements the steps of the proactive work order decomposition method provided in the first embodiment of the present invention.

[0017] A second embodiment of the present invention provides a computer storage medium storing a computer program. When the computer program is executed by a processor, it implements the steps of the proactive work order decomposition method provided in the first embodiment of the present invention.

[0018] Compared with the prior art, the beneficial effects of the technical solution of the present invention are: 1. Significantly improves accuracy: This invention effectively solves the problem of a significant decrease in accuracy when a single model processes complex text by introducing pre-decomposition and recognition of multi-topic text, thereby improving the accuracy of work order generation; 2. Reduced manual workload: This invention transforms the passive process that previously required housekeepers to consciously select, copy, and paste into a process that the system actively understands and automatically generates, greatly reducing the manual workload and time costs for housekeepers; 3. Improve processing efficiency: The entire work order generation process is automated and streamlined, from users directly submitting natural language requests to generating executable work order information, the response speed is faster, improving the overall efficiency of customer service or internal management; 4. Significantly reduce the cost of using LLM in practical fields. Through architectural design and the design of the front-end topic model, improve the accuracy of small LLM, reduce the number of tokens it needs to process, and accurately provide "problems" and "knowledge", making it possible to apply small LLM in professional fields. Attached Figure Description

[0019] Figure 1 A flowchart illustrating a proactive work order decomposition method provided in an embodiment of the present invention; Figure 2 A schematic diagram illustrating the framework of a proactive work order decomposition method provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of a proactive work order decomposition system provided in an embodiment of the present invention. Detailed Implementation

[0020] The accompanying drawings are for illustrative purposes only and should not be construed as limiting the scope of this patent. To better illustrate this embodiment, some parts in the accompanying drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions; It will be understood by those skilled in the art that certain well-known structures and their descriptions may be omitted in the accompanying drawings.

[0021] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0022] The first embodiment of the present invention provides a proactive work order decomposition method, such as... Figure 1 As shown, it includes the following steps: Get the dialogue text entered by the user; The user-input dialogue text is categorized by topic and evaluated by keywords to obtain multiple task units; For each task unit, the knowledge in the task unit is matched with the knowledge in multiple preset topic knowledge bases to obtain the first knowledge; Based on the task unit and the first knowledge corresponding to that task unit, generate the first structured information; Based on the first structured information of each task unit, generate a work order preview corresponding to each task unit.

[0023] In this embodiment of the invention, the complete technical solution of the proactive work order decomposition method is combined with the appendix. Figure 2 Explanation: Data input and synchronization: Users input dialogue text containing one or more tasks or unrelated to tasks through a client (such as WeChat Work). Here, the user's input is synchronized to the database and processing model in real time. 2. Multi-Function Agent Preprocessing: Unlike directly feeding text into a single model, this embodiment of the invention first uses a "topic classification module & keyword measurement module" to perform preliminary processing on the input text. The core responsibility of this module is to act as the first layer, identifying and distinguishing complex text with multiple topics. It can understand the user's macro-intention, or associate the user's dialogue text with possible topics, and decompose the originally mixed text into different, independent task units (which may include multiple tasks or no tasks).

[0024] 3. Enhanced Understanding Through Knowledge Base: The decomposed task units are sent to the next processing level. The system has built-in multiple knowledge bases (knowledge books) for different business domains, such as a maintenance knowledge book and a cleaning knowledge book. These knowledge books originate from possible work order types formed within the professional domain. Based on the type of task unit, relevant key knowledge is selected and provided to the downstream model to enhance the model's in-depth understanding of the specific task, while reducing the amount of data the model needs to process, and consequently reducing the number of user tokens that the subsequent LLM needs to process. 4. Structured Extraction using Large Language Model (LLM): After obtaining text that has undergone preliminary classification by the topic classification module and keyword evaluation module, and knowledge base enhancement, the Large Language Model (LLM) is used to extract information in a structured manner. For example, from the sentence "The light in the conference room is broken," structured data such as {task type: repair, object: light, location: conference room} is extracted and generated to form standardized "work order information."

[0025] 5. Preview Generation and Manual Review: The system automatically generates a work order preview based on the structural work order information. This preview is then submitted to the administrator (or relevant person in charge) for final review. The administrator only needs to review and confirm, eliminating the need for manual copying, pasting, and filtering, greatly improving efficiency.

[0026] This solution effectively improves the accuracy of work order generation in complex scenarios by constructing a multi-level structure of "topic model or keyword measurement + knowledge enhancement + LLM extraction". In practice, this may involve fine-tuning or targeted training of the model, such as using techniques like SFT and RLHF, and implementing corresponding data cleaning and data construction. In a further embodiment, the user-input dialogue text is categorized by topic and evaluated for keywords, including: The user-input dialogue text is split into multiple sentence segments and word segments. In this embodiment, punctuation marks are typically used to segment sentences, and word segmentation algorithms are used to segment words. By evaluating the degree of association between the sentence segments and different preset topic knowledge bases, several candidate sentences and candidate preset topic knowledge bases corresponding to the candidate sentences are obtained based on the degree of association. In this embodiment, the association between sentence-level dialogue and different preset topic knowledge bases is to embed the content of the preset topic knowledge bases, then embed the split sentences, calculate the cosine similarity, filter the similarity through threshold scores, and select candidate preset topic knowledge base content. Candidate keywords are selected based on the importance of keywords in the word segments. In this embodiment, the importance of a keyword is determined by its commonness. The more common a word is, the lower its importance. The less common a word is, the higher its importance gradually becomes. Based on the candidate statements, their topics, and candidate keywords, multiple task units are obtained, wherein each task unit is associated with at least one preset topic knowledge base.

[0027] In a further embodiment, the matching of the task unit with knowledge in a preset topic knowledge base associated with the task unit to obtain first knowledge includes: Calculate the similarity between the task unit and the knowledge in the associated preset topic knowledge base, and obtain the knowledge in the preset topic knowledge base whose similarity is greater than the similarity threshold as the first knowledge.

[0028] In a further embodiment, candidate keywords are selected based on the importance of keywords in the word segmentation, including: Based on each keyword in the word segmentation and the preset word importance table, keywords with an importance greater than the preset importance threshold are selected as candidate keywords. The preset word importance table includes words and their importance. The importance is calculated based on the number of times the word appears in a large amount of corpus. The more times it appears, the lower the importance of the word. The fewer times it appears, the higher the importance of the word.

[0029] Specifically, keyword evaluation is usually an inverse function of assessing the word's commonness. For example: s = 1 / (x+1), where s is the word's importance score and x is the evaluation value for commonness. The more common a word is, the lower its importance; the less common a word is, the higher its importance gradually becomes. Common word calculation involves determining the commonness of words using a large corpus and can be viewed as a static table.

[0030] By processing the input text as described above, the computational load of subsequent LLM can be reduced. Assuming the original user input dialogue contains 3 topics plus approximately 30 tokens of irrelevant dialogue, and the pre-set topic knowledge base contains approximately 5k-10k tokens of related knowledge, the user input dialogue after topic modeling and keyword evaluation is reduced to only 3 topic dialogues plus a small number of keywords. The pre-set topic knowledge base, after matching, contains approximately 50-500 tokens of initial knowledge, significantly reducing the computational load of subsequent LLM.

[0031] In a further embodiment, based on the task unit and the first knowledge corresponding to the task unit, first structured information is generated, including: The task unit and the corresponding first knowledge are input into the large language model, and the large language model outputs the first structured information.

[0032] In a further embodiment, the large language model is obtained by fine-tuning or targeted training according to the business scenario.

[0033] This embodiment proposes a multi-level processing architecture. Its core is the introduction of an upstream "multi-topic classification model or keyword evaluation" specifically designed for the initial understanding, classification, and decomposition of complex multi-topic texts, thus changing the traditional model that handles all tasks with a single model.

[0034] Agent-Knowledge Base Collaboration: Combining the agent's initial decomposition capabilities with a domain knowledge base. The agent is responsible for "decomposition," while the knowledge base is responsible for "specialization," providing high-quality, highly relevant input for subsequent precise information extraction in LLM, reducing the computational power requirements of LLM, improving the accuracy of LLM inference, and reducing LLM illusions.

[0035] End-to-end automated process: It realizes a complete automated process from the original user dialogue input to the generation of a preview of the work order to be reviewed, transforming the human role from the heavy "copy and paste" and "filter and classify" to the easy "review and confirmation".

[0036] The invention will be further explained below with reference to a specific application scenario.

[0037] 1. Suppose a user sends multiple messages in WeChat Work: "How much electricity does C3703 have left? What's the password? Okay, across from the entrance of Building C2, from 2 AM to 2 PM on September 28th, a white car of a certain brand, license plate number Yue AXXX727, please help me, because I need video footage to go to the traffic police station this afternoon, okay?" 2. Receiving a message: The system receives this text input from WeChat Work; 3. Topic Model Decomposition: User input is synchronized to the model. The "topic" first identifies that this is a request containing multiple different topics: one is "water and electricity bill inquiry," another is "video surveillance retrieval," and there is also a part of the dialogue without a topic. The topic model and agent logically decompose this text into two independent tasks and filter out some useless dialogue.

[0038] 3. Knowledge base matching and LLM extraction: For the "water and electricity bill inquiry" task, the system calls upon highly relevant knowledge from the "bill bill inquiry knowledge book" to help the LLM understand terms such as "location" and "password." Subsequently, the LLM extracts structured information and generates work order information 1: {Task type: Water and electricity bill inquiry, account holder information: c3703, Task type-2: Balance inquiry}.

[0039] For the "video surveillance retrieval" task, the system may call another knowledge base (such as a "security knowledge book," not shown in the diagram but operating on the same principle) to help the LLM understand "location" and "time." Subsequently, the LLM extracts the information and generates work order information 2. {Task type: Security management, Task type-2: Video surveillance location information retrieval: opposite C2, time; other key features: white car of a certain brand, license plate number Yue AXXX727}.

[0040] 4. Generation and Approval: Based on these two pieces of structured information, the system generates two independent work order previews. The administrator will see two work orders awaiting approval in the system backend: one for utility bill inquiries and the other for security management - video surveillance retrieval. The administrator only needs to confirm each work order to complete its creation; no manual splitting and data entry is required.

[0041] Through this embodiment, the present invention successfully and accurately converts a complex multi-topic message into two independent, structured work orders, and reduces the speed of LLM inference while improving accuracy in specialized fields.

[0042] A second embodiment of the present invention provides an active work order decomposition system, such as... Figure 3 As shown, it includes: The acquisition module acquires the dialogue text input by the user; The first processing module performs topic classification and keyword evaluation on the dialogue text input by the user to obtain multiple task units; The first matching module matches each task unit with knowledge in multiple preset topic knowledge bases to obtain the first knowledge. The second processing module generates first structured information based on the task unit and the first knowledge corresponding to the task unit. The third processing module generates a work order preview for each task unit based on the first structured information of each task unit.

[0043] In a further embodiment, the first processing module performs topic classification and keyword evaluation on the user-input dialogue text, including: The user-input dialogue text is split into multiple sentence segments and word segments; By evaluating the degree of correlation between the sentence segments and different preset topic knowledge bases, several candidate sentences and candidate preset topic knowledge bases corresponding to the candidate sentences are obtained based on the degree of correlation. Candidate keywords are selected based on the importance of keywords in the word segments; Based on the candidate statements, their topics, and candidate keywords, multiple task units are obtained, wherein each task unit is associated with at least one preset topic knowledge base.

[0044] A third embodiment of the present invention provides an electronic device, including a processor and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, it implements the steps of the proactive work order decomposition method described in the first embodiment of the present invention.

[0045] The fourth embodiment of the present invention provides a computer storage medium storing a computer program. When the computer program is executed by a processor, it implements the steps of the proactive work order decomposition method described in the first embodiment of the present invention.

[0046] The same or similar labels correspond to the same or similar parts; The terms used to describe positional relationships in the accompanying drawings are for illustrative purposes only and should not be construed as limiting this patent. Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art can make other variations or modifications based on the above description. It is neither necessary nor possible to exhaustively describe all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A proactive work order decomposition method, characterized in that, Includes the following steps: Get the dialogue text entered by the user; The user-input dialogue text is categorized by topic and evaluated by keywords to obtain multiple task units; For each task unit, the knowledge in the task unit is matched with the knowledge in multiple preset topic knowledge bases to obtain the first knowledge; Based on the task unit and the first knowledge corresponding to that task unit, generate the first structured information; Based on the first structured information of each task unit, generate a work order preview corresponding to each task unit.

2. The proactive work order decomposition method according to claim 1, characterized in that, The user-input dialogue text is categorized by topic and evaluated for keywords, including: The user-input dialogue text is split into multiple sentence segments and word segments; By evaluating the degree of correlation between the sentence segments and different preset topic knowledge bases, several candidate sentences and candidate preset topic knowledge bases corresponding to the candidate sentences are obtained based on the degree of correlation. Candidate keywords are selected based on the importance of keywords in the word segments; Based on the candidate statements, their topics, and candidate keywords, multiple task units are obtained, wherein each task unit is associated with at least one preset topic knowledge base.

3. The proactive work order decomposition method according to claim 2, characterized in that, The process of matching the task unit with knowledge in a preset topic knowledge base associated with the task unit to obtain first knowledge includes: Calculate the similarity between the task unit and the knowledge in the associated preset topic knowledge base, and obtain the knowledge in the preset topic knowledge base whose similarity is greater than the similarity threshold as the first knowledge.

4. The proactive work order decomposition method according to claim 2, characterized in that, Based on the importance of keywords in the aforementioned word segments, candidate keywords are selected, including: Based on each keyword in the word segmentation and the preset word importance table, keywords with an importance greater than the preset importance threshold are selected as candidate keywords. The preset word importance table includes words and their importance. The importance is calculated based on the number of times the word appears in a large amount of corpus. The more times it appears, the lower the importance of the word. The fewer times it appears, the higher the importance of the word.

5. The proactive work order decomposition method according to claim 1, characterized in that, Based on the task unit and the corresponding first knowledge, first structured information is generated, including: The task unit and the corresponding first knowledge are input into the large language model, and the large language model outputs the first structured information.

6. The proactive work order decomposition method according to claim 5, characterized in that, The large language model is obtained by fine-tuning or targeted training based on the business scenario.

7. A proactive work order decomposition system, characterized in that, include: The acquisition module acquires the dialogue text input by the user; The first processing module performs topic classification and keyword evaluation on the dialogue text input by the user to obtain multiple task units, wherein each task unit is associated with at least one preset topic knowledge base. The first matching module matches each task unit with the knowledge in the preset topic knowledge base associated with that task unit to obtain the first knowledge. The second processing module generates first structured information based on the task unit and the first knowledge corresponding to the task unit. The third processing module generates a work order preview for each task unit based on the first structured information of each task unit.

8. The proactive work order decomposition system according to claim 7, characterized in that, The first processing module performs topic classification and keyword evaluation on the user-input dialogue text, including: The user-input dialogue text is split into multiple sentence segments and word segments; By evaluating the degree of correlation between the sentence segments and different preset topic knowledge bases, several candidate sentences and candidate preset topic knowledge bases corresponding to the candidate sentences are obtained based on the degree of correlation. Candidate keywords are selected based on the importance of keywords in the word segments; Based on the candidate statements, their topics, and candidate keywords, multiple task units are obtained, wherein each task unit is associated with at least one preset topic knowledge base.

9. An electronic device, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, it implements the steps of the proactive work order decomposition method according to any one of claims 1 to 6.

10. A computer storage medium, characterized in that, The computer storage medium stores a computer program, which, when executed by a processor, implements the steps of the proactive work order decomposition method according to any one of claims 1 to 6.

Images