Work order abstract generation method and device
By constructing a knowledge graph through intent recognition, entity extraction, and relation extraction, and combining it with parameter adjustment of a large language model and prompt word templates, the problem of missing key information in existing work order summary generation is solved, thereby improving the accuracy and efficiency of the summary.
Patent Information
- Application Number
- CN202511414069.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-29
- Publication Date
- 2026-02-03
AI Technical Summary
In existing work order summary generation methods, those based on regular expression matching often miss key information, resulting in low summary accuracy.
A knowledge graph is constructed by intent recognition, entity extraction, and relation extraction. The work order type is determined based on the knowledge graph, and the parameters of the large language model are adjusted. Prompt word templates are used to guide the generation of summaries.
It improves the accuracy and efficiency of work order summary generation, avoids the omission of key information, implements an adaptive summary generation strategy, and optimizes processing speed and quality.
Smart Images
Figure CN121457578A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of artificial intelligence, and in particular to a work order abstract generation method and device. BACKGROUND
[0002] In the field of operation and maintenance, customer service, etc., a work order system carries a large amount of information about device faults, user requests, processing processes, etc. This information usually exists in the form of unstructured or semi-structured text, and the content is lengthy and complex.
[0003] The existing method for generating an abstract of work order data generally uses a preset regular expression to match a key field to fill in a template. The hard matching method based on the expression or template often misses the key information in the work order data, resulting in a low accuracy of the finally generated work order abstract. SUMMARY
[0004] The present application provides a work order abstract generation method and device to improve the accuracy of work order abstract generation.
[0005] The present application provides a work order abstract generation method, comprising the following steps: Intention recognition is performed on the work order data, and based on the obtained intention information, a prompt word template for generating an abstract of the work order data is determined; Entity extraction and relationship extraction are performed on the work order data, and based on the obtained entity information and relationship information, a knowledge graph of the work order data is constructed; Based on the knowledge graph, the type of the work order data is determined, and based on the type of the work order data, the parameters of a large language model are adjusted to obtain a large language model after parameter adjustment; Based on the prompt word template, the large language model after parameter adjustment is guided to extract an abstract of the work order data to obtain abstract information of the work order data.
[0006] According to the work order abstract generation method provided by the present application, the type of the work order data is determined based on the knowledge graph, comprising: Based on the number of nodes and the relationship density of the knowledge graph, the type of the work order data is determined to be a complex work order type or a simple work order type.
[0007] According to the work order abstract generation method provided by the present application, the large language model is adjusted based on the type of the work order data to obtain a large language model after parameter adjustment, comprising: In the case where the type of the work order data is a complex work order type, the maximum generation length parameter and the beam search width parameter of the large language model are adjusted to a first preset range to obtain a large language model after parameter adjustment; In a case where the work order type is a simple work order type, a maximum generation length parameter and a beam search width parameter in the large language model are adjusted to a second preset range to obtain a large language model after parameter adjustment.
[0008] According to the work order abstract generation method provided in the application, the work order type of the work order data is determined based on the knowledge graph, and the work order type is determined as a fact statement type or a reasoning diagnosis type based on semantic content of entity information and relationship information in the knowledge graph. According to the work order abstract generation method provided in the application, the work order type of the work order data is determined based on the knowledge graph, and the work order type is determined as a fact statement type or a reasoning diagnosis type based on semantic content of entity information and relationship information in the knowledge graph.
[0009] According to the work order abstract generation method provided in the application, the work order type of the work order data is determined based on the knowledge graph, and the work order type is determined as a fact statement type or a reasoning diagnosis type based on semantic content of entity information and relationship information in the knowledge graph. In a case where the work order type is a fact statement type, a temperature coefficient of the large language model is set to be less than a preset temperature coefficient threshold to obtain a large language model after parameter adjustment. In a case where the work order type is a reasoning diagnosis type, a temperature coefficient of the large language model is set to be greater than the preset temperature coefficient threshold to obtain a large language model after parameter adjustment.
[0010] According to the work order abstract generation method provided in the application, the work order type of the work order data is determined based on the knowledge graph, and the work order type is determined as a fact statement type or a reasoning diagnosis type based on semantic content of entity information and relationship information in the knowledge graph. Based on the intent information, a prompt word template corresponding to the intent information is obtained from a prompt word template library. The prompt word template library is constructed based on a plurality of different intent information and prompt word templates corresponding to the different intent information.
[0011] According to the work order abstract generation method provided in the application, the work order data is subjected to entity extraction and relationship extraction, and the work order data is subjected to entity recognition based on a bidirectional encoder representation model (BERT) combined with a conditional random field model to determine entity information of the work order data. Based on a dependency syntax analysis method, a causal relationship and a time relationship between entities of the work order data are extracted as relationship information of the work order data. Based on a dependency syntax analysis method, a causal relationship and a time relationship between entities of the work order data are extracted as relationship information of the work order data.
[0012] The application further provides a work order abstract generation device comprising the following modules. An intent recognition module is configured to perform intent recognition on work order data and determine a prompt word template for abstract generation of the work order data based on obtained intent information. The knowledge graph construction module is configured to perform entity extraction and relationship extraction on the work order data, and construct a knowledge graph of the work order data based on the obtained entity information and relationship information. The parameter adjustment module is configured to determine a work order type of the work order data based on the knowledge graph, and perform parameter adjustment on the large language model based on the work order type to obtain a parameter-adjusted large language model. The summary generation module is configured to guide the parameter-adjusted large language model to perform summary extraction on the work order data based on the prompt word template to obtain summary information of the work order data.
[0013] The application further provides an electronic device including a memory, a processor, and a computer program stored in the memory and running on the processor, and the processor implements the work order summary generation method according to any one of the above when executing the program.
[0014] The application further provides a non-transitory computer-readable storage medium having a computer program stored thereon, and the computer program is executed by a processor to implement the work order summary generation method according to any one of the above.
[0015] The work order summary generation method and device provided by the application avoid missing of key information and ensure the accuracy of the summary by performing intent recognition on the work order data and determining the corresponding prompt word template. Moreover, the knowledge graph is constructed by performing entity and relationship extraction on the work order data, and the large language model is adjusted in a targeted manner based on the work order type determined based on the knowledge graph, so that the adaptive processing process of different generation strategies for different work orders is realized, thereby ensuring the summary quality while optimizing the processing speed, improving the efficiency and accuracy of the work order summary generation. BRIEF DESCRIPTION OF DRAWINGS
[0016] In order to more clearly illustrate the technical solutions in the application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are some embodiments of the application, and for those skilled in the art, other drawings can also be obtained without creative labor.
[0017] Figure 1 is a flowchart of the work order summary generation method provided by the application.
[0018] Figure 2 is a structural schematic diagram of the work order summary generation device provided by the application.
[0019] Figure 3 is a structural schematic diagram of the electronic device provided by the application. DETAILED DESCRIPTION
[0020] In order to make the objects, technical solutions and advantages of the present application clearer, the technical solutions of the present application will be described clearly and completely below in combination with the drawings in the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the protection scope of the present application.
[0021] Figure 1 is a flowchart of the work order summary generation method provided by the present application, as shown in Figure 1 Step 110, intent recognition is performed on work order data, and based on the obtained intent information, a prompt word template for generating a summary of the work order data is determined. Step 120, entity extraction and relationship extraction are performed on the work order data, and based on the obtained entity information and relationship information, a knowledge graph of the work order data is constructed. Step 130, based on the knowledge graph, the type of the work order data is determined, and based on the type of the work order, a large language model is adjusted in parameters to obtain a large language model adjusted in parameters. Step 140, based on the prompt word template, the large language model adjusted in parameters is guided to extract a summary of the work order data, and the summary information of the work order data is obtained.
[0022] The execution subject of the work order summary generation method provided by the present application can be an electronic device, a component in an electronic device, an integrated circuit, or a chip. The electronic device can be a mobile electronic device or a non-mobile electronic device. Exemplarily, the mobile electronic device can be a mobile phone, a tablet computer, a notebook computer, a palm computer, an ultra-mobile personal computer (UMPC), a netbook, or a personal digital assistant (PDA), etc., and the non-mobile electronic device can be a server, a network attached storage (NAS), or a personal computer (PC), etc., which are not limited by the present application.
[0023] The technical solutions of the present application will be described in detail below taking a computer executing the work order summary generation method provided by the present application as an example.
[0024] In step 110, intent recognition is performed on work order data, and based on the obtained intent information, a prompt word template for generating a summary of the work order data is determined.
[0025] The work order data can include various types of information obtained from various work order management systems. These information can be divided into structured data and unstructured data. The structured data is, for example, work order number, creation timestamp, processing status, network element ID, etc.; the unstructured data mainly refers to the text content filled by manual, such as problem description, processing process log, solution, etc.
[0026] It should be noted that the work order data can also include a variety of different modalities, such as text data parsed from log files by optical character recognition (OCR) technology. Before processing, the original work order data can be pre-processed, such as data cleaning (removing duplicate or format error data), format standardization (unifying fields from different sources to a standard model), and key information enhancement (such as automatically identifying and labeling time, device model, etc. in the text by tools).
[0027] Intention recognition refers to analyzing the text content in the work order data to understand the core purpose or category of the work order. For example, it can be judged whether a work order is a fault repair, a problem consultation or a function suggestion. Intention recognition can be achieved by various natural language processing techniques, such as rule-based methods based on keyword matching, or text classification models based on machine learning or deep learning.
[0028] Prompt template is a pre-designed text structure used to guide the subsequent large language model to generate summaries that meet specific formats and content. After identifying the intention of the work order, a prompt template matching the intention will be selected. For example, for a work order with a fault repair intention, the prompt template may require the model to summarize the fault phenomenon, the impact range and the solution. Selecting a prompt template corresponding to the intention can enhance the accuracy of the work order summary generation.
[0029] In step 120, entity extraction and relationship extraction are performed on the work order data, and a knowledge graph of the work order data is constructed based on the obtained entity information and relationship information.
[0030] Entity extraction is to identify meaningful information units from work order text. Extracted entities can be specific nouns such as servers, databases, network element names, or specific information categories such as problem types, solutions, error codes, etc.
[0031] Relationship extraction establishes connections between identified entities. These relationships can be diverse, such as causal relationships or temporal relationships between entities. Causal relationships can include, for example, power failure leading to business interruption, and temporal relationships can include restarting the routing service first and then the network recovering normally Through entity extraction and relation extraction, the unstructured text information in the work order can be converted into a structured knowledge graph. The knowledge graph is a graph-based data structure, where nodes represent entities and edges represent relationships between entities.
[0032] In step 130, based on the knowledge graph, the type of the work order data is determined, and based on the type of the work order, the large language model is adjusted in parameters to obtain the large language model after parameter adjustment.
[0033] The work order type is a further classification of the work order, and this classification is based on the structure or content features of the knowledge graph. For example, types can be divided according to the complexity or semantic connotation of the knowledge graph. Determining the type of the work order helps to optimize the summary generation strategy subsequently.
[0034] The large language model is the core tool for work order summary generation, which is a deep learning model with text understanding and generation capabilities. Optionally, a suitable base model can be selected according to application scenarios such as code understanding capabilities, Chinese semantic understanding accuracy, and performance indicators such as professional term recall rate and long text processing capabilities.
[0035] Parameter adjustment is to configure the behavior of the large language model according to the determined type of the work order. The generation behavior of the large language model is controlled by multiple parameters, and adjusting these parameters can make the generated summary more in line with expectations in terms of detail, creativity, etc. For example, for different types of work orders, the length, accuracy or efficiency of the generated summary can be adjusted. In this way, a large language model after parameter adjustment optimized for the current characteristics of the work order can be obtained.
[0036] In step 140, based on the prompt word template, the large language model after parameter adjustment is guided to extract the summary of the work order data, and the summary information of the work order data is obtained.
[0037] The prompt word template determined in the above steps and the original work order data are input into the large language model after parameter adjustment. The prompt word template provides a clear form requirement for the model, and the adjusted parameters ensure that the generation style of the model matches the type of the work order.
[0038] The large language model after adjusting the parameters understands, refines and reorganizes the content of the work order according to these inputs, and finally outputs a summary information.
[0039] Optionally, in order to continuously optimize the performance of the model, a human-machine collaborative optimization process can be designed, for example, users can modify the generated summary, collect these modifications as new training data, and continuously iterate the large language model through incremental fine-tuning.
[0040] The work order abstract generation method provided by the application can avoid missing of key information and ensure accuracy of the abstract. Moreover, the entity and relationship extraction is performed on the work order data to construct a knowledge graph, and the large language model is adjusted in parameters based on the work order type determined based on the knowledge graph, so that adaptive processing of different work orders is realized by using different generation strategies, thereby optimizing processing speed while ensuring abstract quality, and improving efficiency and accuracy of work order abstract generation.
[0041] In one embodiment, the work order type of the work order data is determined based on the knowledge graph, including: determining the work order type as a complex work order type or a simple work order type based on the number of nodes and the relationship density of the knowledge graph.
[0042] After the entity and relationship extraction on the work order data, a corresponding knowledge graph can be constructed. The number of nodes refers to the total number of entities contained in the knowledge graph, reflecting how many information points the work order involves. The relationship density measures the closeness of the connection between the nodes in the knowledge graph, which can be calculated by the ratio of the number of edges in the graph to the theoretically maximum possible number of edges, and it reflects the complexity of the association between information points in the work order.
[0043] One or more thresholds can be preset. When the number of nodes and the relationship density of the knowledge graph corresponding to a work order are greater than the preset threshold, it means that the work order contains many information points and the association between the information points is complex (for example, a chain fault involves multiple devices and multiple processing steps), and the work order is determined as a complex work order type. Conversely, if the number of nodes and the relationship density are less than or equal to the preset threshold, it means that the content of the work order is relatively direct and simple (for example, a request for password reset), and the work order is determined as a simple work order type.
[0044] Through this classification method based on the structural characteristics of the knowledge graph, the complexity of the work order can be quickly and objectively evaluated, providing a basis for subsequent adoption of different abstract generation strategies, making the entire process more efficient.
[0045] In one embodiment, the large language model is adjusted in parameters based on the work order type to obtain the large language model adjusted in parameters, including: in the case that the work order type is a complex work order type, adjusting the maximum generation length parameter and the beam search width parameter in the large language model to a first preset range to obtain the large language model adjusted in parameters; in the case that the work order type is a simple work order type, adjusting the maximum generation length parameter and the beam search width parameter in the large language model to a second preset range to obtain the large language model adjusted in parameters.
[0046] When a work order is determined to be a complex work order type, it generally means that the generated summary content needs to be longer and more complete in logic to cover all key information. Therefore, the relevant parameters need to be adjusted to the first preset range. Specifically, the maximum generation length parameter max_length needs to be set to a larger value to ensure that the summary will not be truncated prematurely due to length restrictions. At the same time, the beam search width parameter num_beams should also be set to a relatively large value, which can be set to 5. A larger beam search width means that the model will retain more candidate sequences at each generation step, which helps to generate more accurate summaries in complex scenarios.
[0047] When a work order is determined to be a simple work order type, its summary is usually short and direct. At this time, in order to pursue higher generation efficiency, the relevant parameters can be adjusted to the second preset range. Specifically, the maximum generation length parameter can be set to a smaller value. The beam search width parameter can be set to 1, which is called greedy decoding. In this mode, the model selects the word with the highest probability at each step, generating the fastest speed, which is sufficient to ensure the quality for simple work orders.
[0048] The work order summary generation method provided by the application dynamically adjusts the generation length and search strategy according to the complexity of the work order, achieving a balance between summary quality and generation efficiency. For complex work orders, the integrity and accuracy of the summary are prioritized; for simple work orders, the speed of processing is prioritized, thereby optimizing overall performance and resource utilization.
[0049] In one embodiment, based on the knowledge graph, the work order type of the work order data is determined, including: based on the semantic content of the entity information and the relationship information in the knowledge graph, determining the work order type to be a fact statement type or a reasoning diagnosis type.
[0050] If the knowledge graph of a work order is mainly composed of descriptive entities (such as device status, operation record) and temporal relationships, and lacks explicit causal chains, then its content is mainly objective record and description of facts. Such work orders are determined to be fact statement type. For example, user A reports an error at 3 pm, and the engineer restarts the server at 4 pm, and the service is restored.
[0051] If the knowledge graph of a work order contains a large number of causal relationships (such as because, cause), or the entities present a logical structure of problem, analysis, and conclusion, indicating that the content of the work order involves analysis of the problem and inference of the root cause, then such work orders are determined to be reasoning diagnosis type. For example, because the database connection pool is full, application A cannot log in, and the solution is to increase the size of the connection pool.
[0052] Through this semantic content-based classification method, the internal logic of the work order can be understood more deeply, not just the amount of information, but the nature of the information, which can be used to control the generation process of the large language model more finely later.
[0053] In one embodiment, based on the work order type, the large language model is parameter adjusted to obtain a parameter-adjusted large language model, including: in the case that the work order type is a fact statement type, setting the temperature coefficient of the large language model to be less than a preset temperature coefficient threshold to obtain the parameter-adjusted large language model; in the case that the work order type is a reasoning diagnosis type, setting the temperature coefficient of the large language model to be greater than the preset temperature coefficient threshold to obtain the parameter-adjusted large language model.
[0054] The temperature coefficient is a key parameter for controlling the randomness of the large language model when generating text.
[0055] When a work order is determined to be a fact statement type, the generated summary should be more rigorous to avoid unnecessary creation or guessing by the model. Therefore, the temperature coefficient needs to be set lower, for example, less than a preset temperature coefficient threshold, which can be less than 0.5. A lower temperature makes the model more inclined to choose the word with the highest probability when choosing the next word, and the output result will be more determined, conservative and close to the fact.
[0056] When the work order is determined to be a reasoning diagnosis type, the summary may need to summarize and reason scattered information, or even propose possible solutions. In this case, a reasonable reasoning process can help analysis. Therefore, the temperature coefficient can be set higher, for example, greater than the preset temperature coefficient threshold, which can be set to 0.7. A higher temperature will increase the possibility of choosing a low-probability word, making the output of the model more diversified and more reasoning, which helps to generate a more rich diagnostic summary.
[0057] The work order summary generation method provided by the application adjusts the output process of the model according to the semantic type of the work order, ensuring that the generated summary matches the nature of the work order content. For fact statements, the accuracy and reliability of the summary are guaranteed; for reasoning diagnosis, the model is guided to make reasonable induction and summary, thereby improving the quality and practicality of the summary.
[0058] In one embodiment, based on the obtained intent information, a prompt word template for generating a summary of the work order data is determined, including: based on the intent information, a prompt word template corresponding to the intent information is obtained from a prompt word template library; wherein the prompt word template library is constructed based on a plurality of different intent information and a prompt word template corresponding to each different intent information.
[0059] A prompt template library is pre-constructed. The construction process of the prompt template library can be based on analysis of a large number of historical work orders. An operation and maintenance expert or an algorithm engineer can design an abstract instruction, that is, a prompt template, according to different work order intent information. Each template is associated with one or more intents.
[0060] For example, the following template can be designed for the intent of a fault scenario: As a senior communication operation and maintenance expert, please generate a fault abstract of 150 characters or less according to the following work order information. The abstract needs to include: fault phenomenon-impact range (such as related devices or services)-temporary solution.
[0061] In actual operation, when intent recognition is completed for a new work order data, the intent information is used as an index to directly search and obtain the corresponding preset template in the prompt template library.
[0062] By using the prompt template library, the design work of the prompt is front-end and standardized, avoiding the complexity and uncertainty of dynamically generating the prompt each time. This way not only greatly improves the processing efficiency, but also ensures that the abstract structure of the work order of the same type of intent is consistent, improving the accuracy of subsequent abstract generation.
[0063] In an embodiment, entity extraction and relationship extraction are performed on the work order data, including: based on a bidirectional encoder representation model BERT combined with a conditional random field model, entity recognition is performed on the work order data to determine entity information of the work order data; based on a dependency syntax analysis method, a causal relationship and a time relationship between entities of the work order data are extracted as relationship information of the work order data.
[0064] In the entity extraction stage, a composite model combining a bidirectional encoder representation model BERT and a conditional random field (CRF) model is used. Specifically, first, the powerful context understanding ability of the BERT model is used to convert each word or character in the work order text into a vector representation rich in semantic information. BERT can capture the subtle differences of words in different contexts. Then, the vector sequence is input to the CRF layer. The CRF model can learn the constraint relationship between label sequences, so that when performing label prediction, it can output a globally optimal entity label sequence, effectively avoiding illogical labeling results. Optionally, BERT can be replaced by other advanced pre-training language models such as ELMo, RoBERTa, etc.
[0065] In the relationship extraction stage, when the entity information is determined, a dependency syntax analysis method is used to mine the relationship between entities. Dependency syntax analysis can analyze the grammatical dependency relationship between words in a sentence, such as subject-predicate, verb-object, and modifier-verb, to form a dependency relationship tree. By analyzing the tree, cause-effect relationships and time relationships can be effectively identified.
[0066] The work order summary generation device provided by the present application is described below. The work order summary generation device described below can be referred to in conjunction with the work order summary generation method described above.
[0067] The intent recognition module 210 is configured to perform intent recognition on the work order data and determine a prompt word template for generating a summary of the work order data based on the obtained intent information. The knowledge graph construction module 220 is configured to perform entity extraction and relationship extraction on the work order data and construct a knowledge graph of the work order data based on the obtained entity information and relationship information. The parameter adjustment module 230 is configured to determine a work order type of the work order data based on the knowledge graph and perform parameter adjustment on the large language model based on the work order type to obtain a parameter-adjusted large language model. The summary generation module 240 is configured to guide the parameter-adjusted large language model to perform summary extraction on the work order data based on the prompt word template to obtain summary information of the work order data.
[0068] The work order summary generation device provided by the present application avoids missing key information and ensures the accuracy of the summary by performing intent recognition on the work order data and determining the corresponding prompt word template. Furthermore, by performing entity and relationship extraction on the work order data to construct a knowledge graph and performing targeted parameter adjustment on the large language model based on the work order type determined from the knowledge graph, the adaptive processing process of different work orders using different generation strategies is realized, thereby optimizing the processing speed while ensuring the quality of the summary and improving the efficiency and accuracy of the work order summary generation.
[0069] In one embodiment, the knowledge graph construction module 220 is specifically configured to: determine a work order type of the work order data based on the knowledge graph, including: determine the work order type to be a complex work order type or a simple work order type based on the number of nodes and the relationship density of the knowledge graph.
[0070] In one embodiment, the parameter adjustment module 230 is specifically configured to: The large language model is adjusted in parameters based on the work order type to obtain an adjusted large language model, including: in a case where the work order type is a complex work order type, adjusting a maximum generation length parameter and a beam search width parameter in the large language model to a first preset range to obtain the adjusted large language model. In a case where the work order type is a simple work order type, the maximum generation length parameter and the beam search width parameter in the large language model are adjusted to a second preset range to obtain the adjusted large language model.
[0071] In an embodiment, the knowledge graph construction module 220 is specifically configured to: Determine the work order type of the work order data based on the knowledge graph, including: Determine, based on semantic content of entity information and relationship information in the knowledge graph, that the work order type is a fact statement type or a reasoning diagnosis type.
[0072] In an embodiment, the parameter adjustment module 230 is specifically configured to: Adjust the large language model in parameters based on the work order type to obtain an adjusted large language model, including: in a case where the work order type is a fact statement type, setting a temperature coefficient of the large language model to be less than a preset temperature coefficient threshold to obtain the adjusted large language model; In a case where the work order type is a reasoning diagnosis type, setting a temperature coefficient of the large language model to be greater than the preset temperature coefficient threshold to obtain the adjusted large language model.
[0073] In an embodiment, the intent recognition module 210 is specifically configured to: Determine, based on the obtained intent information, a prompt word template for generating an abstract of the work order data, including: Obtain, based on the intent information, a prompt word template corresponding to the intent information from a prompt word template library; The prompt word template library is constructed based on multiple different intent information and prompt word templates corresponding to the different intent information.
[0074] In an embodiment, the knowledge graph construction module 220 is further specifically configured to: Perform entity extraction and relationship extraction on the work order data, including: Perform entity recognition on the work order data based on a bidirectional encoder representation model BERT combined with a conditional random field model to determine entity information of the work order data; Extract causal relationships and temporal relationships between entities of the work order data as relationship information of the work order data based on a dependency syntax analysis method.
[0075] Figure 3 An example of a schematic diagram of a physical structure of an electronic device is shown in Figure 3 The electronic device can include a processor 310, a communications interface 320, a memory 330, and a communications bus 340, wherein the processor 310, the communications interface 320, and the memory 330 communicate with each other through the communications bus 340. The processor 310 can invoke a logical instruction in the memory 330 to execute a work order summary generation method, which includes: performing intent recognition on work order data, and determining a prompt word template for generating a summary of the work order data based on obtained intent information; performing entity extraction and relationship extraction on the work order data, and constructing a knowledge graph of the work order data based on obtained entity information and relationship information; determining a work order type of the work order data based on the knowledge graph, and performing parameter adjustment on a large language model based on the work order type to obtain a parameter-adjusted large language model; based on the prompt word template, guiding the parameter-adjusted large language model to perform summary extraction on the work order data to obtain summary information of the work order data.
[0076] In addition, the logical instructions in the memory 330 described above can be implemented in the form of a software function unit and sold or used as an independent product, which can be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the present application or parts of the present application that essentially contribute to the prior art or the parts of the technical solutions can be embodied in the form of a software product, which is stored in a storage medium and includes instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in the various embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes.
[0077] On the other hand, the present application also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium, and the computer program can be executed by a processor to enable a computer to execute the work order summary generation method provided by the above-mentioned methods, which includes: performing intent recognition on work order data, and determining a prompt word template for generating a summary of the work order data based on obtained intent information; entity extraction and relation extraction are performed on the work order data, and a knowledge graph of the work order data is constructed based on obtained entity information and relation information; based on the knowledge graph, a work order type of the work order data is determined, and a large language model is adjusted in parameters based on the work order type to obtain a large language model adjusted in parameters; based on the prompt word template, the large language model adjusted in parameters is guided to perform abstract extraction on the work order data to obtain abstract information of the work order data.
[0078] In another aspect, the present application also provides a non-transitory computer readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a work order abstract generation method provided by each of the above methods, and the method comprises: performing intent recognition on work order data, and determining a prompt word template for abstract generation of the work order data based on obtained intent information; entity extraction and relation extraction are performed on the work order data, and a knowledge graph of the work order data is constructed based on obtained entity information and relation information; based on the knowledge graph, a work order type of the work order data is determined, and a large language model is adjusted in parameters based on the work order type to obtain a large language model adjusted in parameters; based on the prompt word template, the large language model adjusted in parameters is guided to perform abstract extraction on the work order data to obtain abstract information of the work order data.
[0079] The device embodiments described above are only schematic, wherein the units shown as separate components can or can not be physically separate, and the components shown as units can or can not be physical units, i.e., can be located in one place, or can be distributed on multiple network units. Part or all of the modules can be selected according to actual needs to achieve the purpose of the present embodiment. Those skilled in the art can understand and implement without creative labor.
[0080] From the above description of the embodiments, those skilled in the art can clearly understand that the embodiments can be realized by means of software plus necessary general hardware platforms, and of course can also be realized by hardware. Based on such understanding, the above technical solutions can be embodied in the form of software products, which can be stored in a computer readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and include a number of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute the methods described in each embodiment or some parts of the embodiments.
[0081] Finally, it should be noted that the above examples are only used to illustrate the technical solutions of the present application, and are not intended to limit the same; although the present application has been described in detail with reference to the foregoing examples, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing examples, or make equivalent replacements for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.
Claims
1. A method for generating work order summaries, characterized in that, include: The intent of the work order data is identified, and based on the obtained intent information, a prompt word template for generating a summary of the work order data is determined. Entity and relationship extraction are performed on the work order data, and a knowledge graph of the work order data is constructed based on the obtained entity and relationship information. Based on the knowledge graph, the work order type of the work order data is determined, and based on the work order type, the parameters of the large language model are adjusted to obtain the large language model with adjusted parameters. Based on the prompt word template, the large language model with adjusted parameters is guided to extract a summary of the work order data, thereby obtaining the summary information of the work order data.
2. The work order summary generation method according to claim 1, characterized in that, The process of determining the work order type based on the knowledge graph includes: Based on the number of nodes and the relationship density of the knowledge graph, the work order type is determined to be either a complex work order type or a simple work order type.
3. The work order summary generation method according to claim 2, characterized in that, The step of adjusting the parameters of the large language model based on the work order type to obtain the parameter-adjusted large language model includes: When the work order type is a complex work order type, the maximum generation length parameter and the bundle search width parameter in the large language model are adjusted to a first preset range to obtain the large language model with adjusted parameters. When the work order type is a simple work order type, the maximum generation length parameter and the bundle search width parameter in the large language model are adjusted to a second preset range to obtain the large language model with adjusted parameters.
4. The work order summary generation method according to claim 1, characterized in that, The process of determining the work order type based on the knowledge graph includes: Based on the semantic content of entity and relation information in the knowledge graph, the work order type is determined to be either a fact statement type or a reasoning diagnosis type.
5. The work order summary generation method according to claim 4, characterized in that, The step of adjusting the parameters of the large language model based on the work order type to obtain the parameter-adjusted large language model includes: When the work order type is a fact statement type, the temperature coefficient of the large language model is set to be less than the preset temperature coefficient threshold to obtain the large language model with adjusted parameters. When the work order type is inference diagnosis type, the temperature coefficient of the large language model is set to be greater than the preset temperature coefficient threshold to obtain the large language model with adjusted parameters.
6. The work order summary generation method according to claim 1, characterized in that, The step of determining the prompt word template for summarizing the work order data based on the obtained intent information includes: Based on the intent information, obtain the prompt word template corresponding to the intent information from the prompt word template library; The prompt word template library is constructed based on multiple different intent information and the prompt word templates corresponding to each different intent information.
7. The work order summary generation method according to claim 1, characterized in that, The entity extraction and relation extraction of the work order data includes: Based on the bidirectional encoder representation model BERT combined with the conditional random field model, entity recognition is performed on the work order data to determine the entity information of the work order data. Based on dependency parsing, the causal and temporal relationships between entities in the work order data are extracted as relational information of the work order data.
8. A work order summary generation device, characterized in that, include: The intent recognition module is used to recognize the intent of the work order data and, based on the obtained intent information, determine the prompt word template for generating a summary of the work order data. The knowledge graph construction module is used to extract entities and relationships from the work order data, and to construct a knowledge graph of the work order data based on the obtained entity information and relationship information. The parameter adjustment module is used to determine the work order type of the work order data based on the knowledge graph, and adjust the parameters of the large language model based on the work order type to obtain the parameter-adjusted large language model. The summary generation module is used to guide the parameter-adjusted large language model to extract a summary of the work order data based on the prompt word template, so as to obtain the summary information of the work order data.
9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the work order summary generation method as described in any one of claims 1 to 7.
10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the work order summary generation method as described in any one of claims 1 to 7.