Knowledge mining method and device based on large model, equipment, medium and product
By generating and filtering triples for the target domain in a large language model and combining them with general knowledge to update model parameters, the problem of insufficient knowledge mining in specific domains by large language models is solved, and the performance of the model in specific domains is improved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA MOBILE GRP GUANGDONG CO LTD
- Filing Date
- 2026-01-26
- Publication Date
- 2026-05-12
AI Technical Summary
Large language models are insufficient in knowledge mining within specific domains, leading to a decrease in their effectiveness and credibility.
The first triplet is generated by driving the large language model with the pre-requirements of the target domain. The second triplet is obtained by expanding the head entity, and relation entity filtering and deduplication are performed. The model parameters are updated by combining general knowledge of the target domain.
It improves the effectiveness and accuracy of knowledge output of large language models in specific domains, and enhances the credibility of their application in specific domains.
Smart Images
Figure CN122021852A_ABST
Abstract
Description
Technical Field
[0001] This disclosure relates to the field of data processing technology, and in particular to a knowledge mining method, apparatus, device, medium and product based on a large model. Background Technology
[0002] With the rapid development of artificial intelligence, Large Language Models (LLMs) are widely used across various industries. Currently, in terms of knowledge mining within LLMs, there is a problem of insufficient knowledge mining in specific domains, affecting the performance of large models in practical applications. This reduces the effectiveness and reliability of large language models in specific domains. Summary of the Invention
[0003] This disclosure addresses some of the shortcomings mentioned in the background art by providing a knowledge mining method, apparatus, device, medium, and product based on a large model.
[0004] In a first aspect, embodiments of this disclosure provide a knowledge mining method based on a large model, comprising: Determine the target prerequisites corresponding to the target domain, and input the target prerequisites into the target large language model to obtain the first triplet; Based on the first head entity of the first triplet, knowledge expansion is performed to obtain the second triplet; Based on the first relation entity of the first triplet and the second relation of the second triplet, the entities are filtered to obtain the third triplet; wherein, the third relation entity of the third triplet meets the preset occurrence frequency condition; The third triplet is deduplicated to obtain a preset number of target triplets corresponding to the target domain; The parameters of the target large language model are updated based on the target triples and the general knowledge of the target domain to obtain the updated large language model.
[0005] In one embodiment of the first aspect, the step of extending knowledge based on the first head entity of the first triplet to obtain the second triplet includes: The first entity is retrieved through multiple rounds of iterative retrieval using an instruction retrieval model, resulting in a second entity with expanded knowledge. The second triplet is obtained by performing structural integrity processing based on the second head entity.
[0006] In one embodiment of the first aspect, the step of filtering entities based on the first relation entity of the first triplet and the second relation entity of the second triplet to obtain the third triplet includes: In the first triplet and the second triplet, a relation entity that satisfies the preset occurrence frequency condition is identified, and this relation entity is identified as the third relation entity; Determine the head entity corresponding to the third relation entity, and designate the head entity as the third head entity; Based on the third head entity and the third relation entity, a third tail entity that meets the preset related conditions with the third relation entity is generated, thereby obtaining a third triplet.
[0007] In one embodiment of the first aspect, the step of deduplicating the third triplet to obtain a preset number of target triplets corresponding to the target domain includes: Determine the set of duplicate triples in the third triplet, and retain any triplet in the set of duplicate triples to obtain the deduplicated triplet; If the number of triplets in the deduplicated triplet is greater than a preset number, the deduplicated triplet is optimized to obtain a preset number of fourth triplets.
[0008] Based on the general knowledge of the target domain, the fourth triple is retrieved and corrected using the target language big model to obtain the target triple.
[0009] In one embodiment of the first aspect, the fourth triplet is obtained by performing knowledge retrieval and correction based on the general knowledge of the target domain through the target language big model to obtain the target triplet, including: Based on the general knowledge of the target domain, knowledge retrieval is performed on the fourth triplet through the target language big model to obtain the fifth triplet corresponding to the first retrieval result and the sixth triplet corresponding to the second retrieval result; wherein, the first retrieval result is used to indicate the fourth triplet with correct knowledge and the second retrieval result is used to indicate the fourth triplet with incorrect knowledge. The sixth triplet is corrected to obtain the seventh triplet; The fifth triplet and the seventh triplet are identified as the target triplet.
[0010] In one embodiment of the first aspect, correcting the sixth triplet to obtain the seventh triplet includes: Obtain relevant knowledge and erroneous knowledge information about the sixth triplet; Based on the relevant knowledge, the erroneous knowledge information is corrected to obtain the corrected sixth triplet; The modified sixth triplet is determined as the seventh triplet.
[0011] In a second aspect, embodiments of this disclosure provide a knowledge mining apparatus based on a large model, comprising: The determination module is used to determine the target prerequisites corresponding to the target domain, and input the target prerequisites into the target large language model to obtain the first triplet; An extension module is used to extend knowledge based on the first head entity of the first triplet to obtain a second triplet. The filtering module is used to perform entity filtering based on the first relation entity of the first triplet and the second relation of the second triplet to obtain a third triplet; wherein the third relation entity of the third triplet meets a preset occurrence frequency condition. The deduplication module is used to deduplicatize the third triplet to obtain a preset number of target triplets corresponding to the target domain. The update module is used to update the parameters of the target large language model based on the target triples and the general knowledge of the target domain, so as to obtain the updated large language model.
[0012] In a third aspect, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement steps of a knowledge mining method based on a large model.
[0013] In a fourth aspect, a computer-readable storage medium is provided having a computer program / instructions stored thereon, which, when executed by a processor, implements the steps of a knowledge mining method based on a large model. In a fifth aspect, a computer program product is provided, including a computer program / instructions that, when executed by a processor, implement steps of a knowledge mining method based on a large model.
[0014] As will be described in detail below, a knowledge mining method, apparatus, device, medium, and product based on a large model according to embodiments of this disclosure are disclosed. By driving a large language model to generate a first triplet based on the pre-requirements of the target domain, and then expanding it based on the first head entity to obtain a second triplet, this process starts from the large model's own knowledge reserves, directionally mining initial knowledge associations in the target domain. This overcomes the dependence on external datasets and compensates for the insufficient breadth of knowledge mining in specific domains. Furthermore, by updating the model parameters by combining the target triplet with general knowledge of the target domain, the large language model can integrate core knowledge of the target domain, effectively improving the output effectiveness of the model when applied in specific domains, while ensuring the accuracy of knowledge, thereby enhancing the credibility of the large model in practical applications in specific domains.
[0015] The aforementioned technical process forms a closed loop of "targeted mining - precise screening - model update," improving the current situation of insufficient knowledge mining in specific domains from both the breadth and precision dimensions of knowledge mining. By integrating the core knowledge of the target domain after screening into the large language model, the model's knowledge output in specific domains is made more aligned with actual application needs. This avoids interference from irrelevant knowledge and strengthens the expression of key knowledge within the domain, ultimately significantly improving the application performance of the large language model in target domain scenarios and solving the problem of reduced model effectiveness and credibility caused by insufficient knowledge mining in specific domains. Attached Figure Description
[0016] Figure 1 A flowchart of a knowledge mining method based on a large model provided in this disclosure embodiment; Figure 2 A flowchart illustrating the determination of target triples in a knowledge mining method based on a large model, provided in this embodiment of the disclosure; Figure 3 A schematic diagram of a knowledge mining device based on a large model provided in this disclosure embodiment; Figure 4 This is a schematic diagram of an electronic device provided in an embodiment of the present disclosure. Detailed Implementation
[0017] The present disclosure will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the present disclosure and not intended to limit it. Furthermore, it should be noted that, for ease of description, only the parts relevant to the present disclosure are shown in the drawings, not the entire structure.
[0018] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.
[0019] In this document, the term "and / or" merely describes a relationship, indicating that three relationships can exist. For example, A and / or B can represent three cases: A alone, A and B simultaneously, and B alone. Furthermore, the term "at least one" in this document means any combination of at least two of any one or more elements. For example, including at least one of A, B, and C can mean including any one or more elements selected from the set consisting of A, B, and C.
[0020] Research has revealed that with the rapid development of artificial intelligence, large-scale models are widely used across various industries. Currently, in the area of knowledge mining within large language models, there is a deficiency in knowledge extraction within specific domains, affecting the performance of large-scale models in practical applications. This leads to a reduction in the effectiveness and reliability of large-scale models in specific domains.
[0021] Based on the above research, this disclosure provides a knowledge mining method based on a large model. It generates a first triplet by driving the large language model with the pre-requirements of the target domain, and then expands the second triplet based on the first entity. This process starts from the large model's own knowledge reserves, selectively mining initial knowledge associations in the target domain, breaking through the dependence on external datasets and compensating for the insufficient breadth of knowledge mining in specific domains. Furthermore, by combining the target triplet with general knowledge of the target domain to update the model parameters, the large language model can integrate core knowledge of the target domain, effectively improving the output effectiveness of the model in specific domain applications while ensuring the accuracy of knowledge, thereby enhancing the credibility of the large model in practical applications in specific domains.
[0022] The aforementioned technical process forms a closed loop of "targeted mining - precise screening - model update," improving the current situation of insufficient knowledge mining in specific domains from both the breadth and precision dimensions of knowledge mining. By integrating the core knowledge of the target domain after screening into the large language model, the model's knowledge output in specific domains is made more aligned with actual application needs. This avoids interference from irrelevant knowledge and strengthens the expression of key knowledge within the domain, ultimately significantly improving the application performance of the large language model in target domain scenarios and solving the problem of reduced model effectiveness and credibility caused by insufficient knowledge mining in specific domains.
[0023] To facilitate understanding of this embodiment, a detailed description of the knowledge mining method based on a large model disclosed in this disclosure will be provided first. The execution subject of the knowledge mining method based on a large model provided in this disclosure is generally an electronic device with a certain computing power. In some possible implementations, this knowledge mining method based on a large model can be implemented by a processor calling computer-readable instructions stored in memory.
[0024] Figure 1 The diagram shows a flowchart of a knowledge mining method based on a large model provided in this disclosure embodiment. The method includes steps S101 to S105, wherein: S101. Determine the target prerequisites corresponding to the target domain and input the target prerequisites into the target large language model to obtain the first triplet.
[0025] In the embodiments of this disclosure, firstly, the recognition domain corresponding to the target large language model can be determined, and the recognition domain can be identified as the target domain.
[0026] Then, based on the characteristics of the target domain, prerequisites related to that target domain can be generated. For example, if the target domain is communications, a series of prerequisites related to communications can be generated. These prerequisites can include questions, keywords, or phrases related to communication technologies, protocols, and equipment, to stimulate the large language model to generate knowledge related to the communications domain. That is, prerequisites such as "What are the key technologies of 5G communication?" and "What are the core components of a base station?"
[0027] Then, the above prerequisites can be used as input to the target large language model to drive the target large language model to generate the first triplet (i.e., the meta triplet).
[0028] S102. Based on the first head entity of the first triplet, knowledge expansion is performed to obtain the second triplet.
[0029] In embodiments of this disclosure, a head entity in the first triplet can be determined and designated as the first head entity.
[0030] Then, we can determine the relevant knowledge corresponding to the first entity, determine the relation triples related to the first entity, and then obtain the second triples based on the first entity.
[0031] S103. Filter the entities based on the first relation entity of the first triplet and the second relation of the second triplet to obtain the third triplet; wherein the third relation entity of the third triplet meets the preset occurrence frequency condition.
[0032] In the embodiments of this disclosure, a relational entity of a first triple can be determined and designated as a first relational entity; a relational entity of a second triple can be determined and designated as a second relational entity.
[0033] Then, the first and second relation entities that do not meet the preset frequency conditions can be filtered out, and the remaining first and second relation entities can be identified as the third relation entities.
[0034] Finally, based on the third relation entity, the corresponding head entity and tail entity are determined to obtain the third triplet.
[0035] S104. Perform deduplication on the third triplet to obtain a preset number of target triplets corresponding to the target domain.
[0036] In embodiments of this disclosure, the third head entity, the third relation entity, and the third tail entity in the third triplet can be determined.
[0037] Then, it can be determined that the third triplet in which the third head entity, the third relation entity, and the third tail entity are completely identical can be identified.
[0038] After identifying identical third triplets, any one of them can be retained to obtain a predetermined number of target triplets. For example, in the case of the target domain being the communications domain, duplicate <5G, core technology, Massive MIMO> are removed, and after multiple iterations, a predetermined number of target triplets for the communications domain are obtained.
[0039] S105. Update the parameters of the target large language model based on the target triples and general knowledge of the target domain to obtain the updated large language model.
[0040] In embodiments of this disclosure, the correctness of target triples can be judged based on general knowledge of the target domain, and erroneous target triples can be corrected to obtain correct triples (i.e., the seventh triple hereinafter). Then, the correct triples and general knowledge can be used to train the target large language model in combination to obtain an updated large language model.
[0041] Here, a strategy of using a mixture of knowledge-intensive supervised data and general-purpose data for training is adopted, which effectively avoids the catastrophic forgetting problem in large-scale language models (LLM).
[0042] In the embodiments of this disclosure, firstly, the target prerequisites corresponding to the target domain are determined and input into the target large language model to obtain a first triplet; secondly, knowledge expansion is performed based on the first head entity of the first triplet to obtain a second triplet; thirdly, entity filtering is performed based on the first relation entity of the first triplet and the second relation of the second triplet to obtain a third triplet; wherein the third relation entity of the third triplet meets a preset occurrence frequency condition; fourthly, deduplication is performed on the third triplet to obtain a preset number of target triples corresponding to the target domain; finally, the parameters of the target large language model are updated based on the target triples and the general knowledge of the target domain to obtain the updated large language model.
[0043] In the above implementation, the large language model generates the first triplet by driving the pre-requirements of the target domain, and then expands it based on the first-head entity to obtain the second triplet. This process starts from the large model's own knowledge reserves, selectively mining initial knowledge associations in the target domain, breaking through the dependence on external datasets and compensating for the insufficient breadth of knowledge mining in specific domains. Furthermore, by combining the target triplet with general knowledge of the target domain to update the model parameters, the large language model can integrate the core knowledge of the target domain, effectively improving the output effectiveness of the model when applied in specific domains, while ensuring the accuracy of knowledge, thereby enhancing the credibility of the large model in practical applications in specific domains.
[0044] The aforementioned technical process forms a closed loop of "targeted mining - precise screening - model update," improving the current situation of insufficient knowledge mining in specific domains from both the breadth and precision dimensions of knowledge mining. By integrating the core knowledge of the target domain after screening into the large language model, the model's knowledge output in specific domains is made more aligned with actual application needs. This avoids interference from irrelevant knowledge and strengthens the expression of key knowledge within the domain, ultimately significantly improving the application performance of the large language model in target domain scenarios and solving the problem of reduced model effectiveness and credibility caused by insufficient knowledge mining in specific domains.
[0045] In one optional implementation, knowledge expansion is performed based on the first head entity of the first triple to obtain the second triple, specifically including the following steps: First, the first entity is iteratively retrieved through multiple rounds using the instruction retrieval model to obtain the second entity after knowledge expansion. Then, structural integrity processing is performed based on the second head entity to obtain the second triplet.
[0046] In the embodiments of this disclosure, a knowledge exploration instruction template corresponding to the target domain can be preset based on the indication features of the target domain.
[0047] Here, the knowledge exploration instruction template has multiple expansion dimensions. These expansion dimensions include at least one of the following: attribute association, relationship mining, and application scenarios.
[0048] Next, the first entity can be substituted into the knowledge exploration instruction template to generate an exploration instruction. After determining the exploration instruction, multiple rounds of iterative retrieval can be performed based on the exploration instruction to generate the second triplet.
[0049] In each iteration, the knowledge exploration instruction template can be adjusted based on the generated knowledge relationships to further optimize the knowledge exploration process.
[0050] The above embodiments break through the limitations of single-dimensional knowledge exploration, enhance the breadth and depth of knowledge expansion in specific fields, make up for the shortcomings of knowledge exploration, make the generated second triplet knowledge associations richer and more accurate, provide high-quality data for subsequent knowledge processing, and effectively solve the problem of insufficient knowledge mining in specific fields.
[0051] In an optional implementation, the entities are filtered based on the first relation entity of the first triplet and the second relation of the second triplet to obtain the third triplet, specifically including the following steps: First, identify the relational entity that satisfies the preset occurrence frequency condition in the first triplet and the second triplet, and then identify the relational entity as the third relational entity. Secondly, determine the head entity corresponding to the third relation entity, and designate this head entity as the third head entity; Finally, based on the third head entity and the third relation entity, a third tail entity that meets the preset related conditions with the third relation entity is generated, thus obtaining the third triplet.
[0052] In the embodiments of this disclosure, a relational entity of a first triple can be determined and designated as a first relational entity; a relational entity of a second triple can be determined and designated as a second relational entity.
[0053] Then, the frequency of occurrence of the first relation entity and the second relation entity in the first triplet and the second triplet can be determined respectively.
[0054] Then, the first entity relationship and the second entity relationship, which appear at a frequency lower than the preset frequency threshold, can be identified as relationship entities that do not meet the preset frequency condition, and then the relationship entities that do not meet the preset frequency condition can be filtered.
[0055] Then, the remaining first and second relation entities after filtering can be identified as the third relation entity.
[0056] Then, the third head entity corresponding to the third relation entity can be determined.
[0057] Here, we can identify the third tail entity that meets the preset related conditions with the third head entity and the third relation entity, thus obtaining the third triplet.
[0058] Here, the preset relevance condition can be whether it conforms to the preset relevance with the third head entity and the third relation entity.
[0059] For example, by combining the third head entity "5G" with the third relational entity "key technology", multiple tail entities (Massive MIMO, ultra-dense networking, edge computing) can be generated. Then, the generated tail entities can be sorted by relevance, and the first preset number of tail entities can be selected as the third tail entities, forming a third triplet such as <5G, key technology, Massive MIMO>.
[0060] In the above embodiments, a heuristic large-scale model knowledge mining method based on knowledge triples is adopted to effectively discover and utilize the knowledge resources within the model. This method, through stages such as brainstorming knowledge retrieval in the communication domain, single-entity divergent knowledge exploration, high-frequency relation entity filtering and tail entity generation, and domain-specific knowledge deduplication, selectively mines domain-specific knowledge, thus making up for the lack of mining and localization of knowledge within the model itself in existing technologies.
[0061] In an optional implementation, the third triplet is deduplicated to obtain a preset number of target triplets corresponding to the target domains, specifically including the following steps: First, determine the set of duplicate triples in the third triplet, and retain any triplet in the set of duplicate triplets to obtain the deduplicated triplets; Secondly, if the number of triples in the deduplicated triplet is greater than the preset number, the deduplicated triplet is optimized to obtain the preset number of fourth triplets.
[0062] Finally, based on the general knowledge of the target domain, the fourth triplet is retrieved and corrected using the target language big model to obtain the target triplet.
[0063] In the embodiments of this disclosure, synonymous entities can be uniformly labeled according to domain standard terminology. For example, "core technology" and "key technology", "baseband unit" and "BBU".
[0064] Then, all third triples are converted into a unified structure of "<head entity, relation entity, tail entity>", and redundant punctuation, spaces, and semantically irrelevant modifiers are removed.
[0065] Then, semantically equivalent relations or entities are merged using a domain ontology library or semantic similarity algorithm. For example, "application scenario" and "applicable scenario" are grouped into the same relation entity, resulting in a set of repeated triples.
[0066] Then, for triples that are not completely repeated but semantically equivalent, identification can be performed by calculating semantic similarity (e.g., calculating the overall semantic similarity of triples based on the BERT model and setting the similarity threshold to 0.9).
[0067] Here, if the semantic similarity between two third triplets exceeds the threshold, it is determined to be semantic duplication. The triplet with more standardized expression and more complete information is selected as the retained item, and the other item is removed.
[0068] After identifying the fourth triplet, knowledge retrieval and correction can be performed on the fourth triplet using the target language big model based on general knowledge of the target domain, thus obtaining the target triplet.
[0069] Reference Figure 2 The diagram shown is a flowchart of a knowledge mining method based on a large model, provided in this embodiment of the present disclosure, for determining target triples, wherein: S21. Determine the target prerequisites corresponding to the target domain and input the target prerequisites into the target large language model to obtain the first triplet.
[0070] S22. Knowledge mining is performed on the first head entity of the first triplet using a large model to obtain the second triplet.
[0071] S23. Filter the relational entities in the first triplet and the second triplet that do not meet the preset occurrence frequency condition to obtain the filtered triplet.
[0072] S24. Based on the head entity and relation entity of the filtered triplet, process them to obtain the third tail entity, and then generate the third triplet.
[0073] S25. Perform deduplication on the third triplet to obtain a preset number of target triplets corresponding to the target domain.
[0074] In the above implementation, redundant third triples are removed to ensure the uniqueness and validity of the mined domain-specific knowledge, ultimately generating high-quality domain-specific fourth triples that meet the quantity requirements. This process is based on the full collection of third triples, uses multi-dimensional deduplication rules to achieve deduplication, and combines an iterative supplementation mechanism to achieve the predetermined number of fourth triples.
[0075] In an optional implementation, the fourth triple is obtained by knowledge retrieval and correction based on general knowledge of the target domain through a large target language model, specifically including the following conditions: First, based on the general knowledge of the target domain, knowledge retrieval is performed on the fourth triplet through the target language big model to obtain the fifth triplet corresponding to the first retrieval result and the sixth triplet corresponding to the second retrieval result; wherein, the first retrieval result is used to indicate the fourth triplet with correct knowledge and the second retrieval result is used to indicate the fourth triplet with incorrect knowledge. Secondly, the sixth triplet is corrected to obtain the seventh triplet; Finally, the fifth and seventh triplets were identified as the target triplets.
[0076] In embodiments of this disclosure, firstly, the correctness of the fourth triplet can be evaluated. Here, knowledge retrieval of the fourth triplet can be performed using the proxy capabilities of the target large language model through retrieval enhancement.
[0077] Then, the correctness of the fourth triplet can be determined by combining the knowledge retrieval results, and the fifth triplet corresponding to the first retrieval result and the sixth triplet corresponding to the second retrieval result can be obtained.
[0078] Here, the target language big data model can be used to extract the corresponding correct knowledge support corpus from the retrieval database, including authoritative definitions, technical parameters, industry data, etc.
[0079] For example, for <5G, base station coverage, 100 km>, extract the correct knowledge support corpus of "5G uses high-frequency electromagnetic waves, resulting in large signal propagation loss. The coverage range of macro base stations is usually 1-3 km, and the coverage range of micro base stations is about 100-300 meters".
[0080] Then, the sixth triplet can be corrected based on the correct knowledge-supported corpus to obtain the seventh triplet.
[0081] In the above implementation, the correctness verification and error correction of the fourth triplet knowledge are achieved by leveraging the general knowledge of the target domain and the retrieval enhancement capabilities of the large model, ultimately resulting in a high-quality target triplet.
[0082] In an optional implementation, the sixth triplet is corrected to obtain the seventh triplet, specifically including the following steps: First, acquire relevant and incorrect knowledge about the sixth triplet; Secondly, based on relevant knowledge, the erroneous knowledge information is corrected to obtain the corrected sixth triplet; Finally, the revised sixth triplet was determined to be the seventh triplet.
[0083] In the embodiments of this disclosure, the erroneous content of the sixth triple, the retrieved correct supporting corpus, and the triple reconstruction requirements can be integrated into a prompt word input model.
[0084] For example, the prompt is: "Given general knowledge in the field of communications: the coverage range of a 5G macro base station is approximately 1-3 kilometers. There is an incorrect triple <5G, base station coverage range, 100 kilometers>. Please reconstruct this triple based on the correct knowledge, ensuring that the head entity and relation entity remain unchanged, and the tail entity conforms to the general knowledge."
[0085] Here, after correcting the sixth triplet to obtain the corrected triplet, the corrected triplet can be compared with the correct knowledge in the search database to verify its accuracy again. If there is still a deviation, the prompt words are readjusted and iteratively corrected until the seventh triplet that conforms to general knowledge is generated.
[0086] In the above embodiments, this embodiment proposes a suitable knowledge editing method for specific, sophisticated domains lacking supervised training data (such as communication knowledge), overcoming the limitations of existing technologies in these domains. Through knowledge correctness judgment and correction, this application proposal can more accurately edit knowledge in specific domains.
[0087] Based on the same inventive concept, this disclosure also provides a knowledge mining device based on a large model. Since the principle of the device in this disclosure for solving the problem is similar to the knowledge mining method based on a large model described above in this disclosure, the implementation of the device can refer to the implementation of the method, and the repeated parts will not be described again.
[0088] Reference Figure 3 The diagram shown is a schematic of a knowledge mining device based on a large model provided in this embodiment of the present disclosure, including: a determination module 31, an expansion module 32, a filtering module 33, a deduplication module 34, and an update module 35; wherein: The determination module 31 is used to determine the target prerequisites corresponding to the target domain and input the target prerequisites into the target large language model to obtain the first triplet; Extension module 32 is used to extend knowledge based on the first head entity of the first triplet to obtain the second triplet; The filtering module 33 is used to perform entity filtering based on the first relation entity of the first triplet and the second relation of the second triplet to obtain a third triplet; wherein the third relation entity of the third triplet meets a preset occurrence frequency condition. The deduplication module 34 is used to perform deduplication processing on the third triplet to obtain a preset number of target triplets corresponding to the target domain. The update module 35 is used to update the parameters of the target large language model based on the target triples and the general knowledge of the target domain, so as to obtain the updated large language model.
[0089] This embodiment drives the large language model to generate the first triplet based on the pre-requirements of the target domain, and then expands it based on the first-head entity to obtain the second triplet. This process starts from the large model's own knowledge reserves, and selectively mines the initial knowledge associations of the target domain, breaking through the dependence on external datasets and making up for the lack of breadth in specific domain knowledge mining. Furthermore, by combining the target triplet with general knowledge of the target domain to update the model parameters, the large language model can integrate the core knowledge of the target domain, effectively improving the output effectiveness of the model when applied to specific domains, while ensuring the accuracy of knowledge, thereby enhancing the credibility of the large model in practical applications in specific domains.
[0090] The aforementioned technical process forms a closed loop of "targeted mining - precise screening - model update," improving the current situation of insufficient knowledge mining in specific domains from both the breadth and precision dimensions of knowledge mining. By integrating the core knowledge of the target domain after screening into the large language model, the model's knowledge output in specific domains is made more aligned with actual application needs. This avoids interference from irrelevant knowledge and strengthens the expression of key knowledge within the domain, ultimately significantly improving the application performance of the large language model in target domain scenarios and solving the problem of reduced model effectiveness and credibility caused by insufficient knowledge mining in specific domains.
[0091] The processing flow of each module in the device and the interaction flow between each module can be referred to the relevant descriptions in the above method embodiments, and will not be detailed here.
[0092] Corresponding to Figure 1 In addition to the knowledge mining methods based on large models, this disclosure also provides an electronic device 400, such as... Figure 4 The diagram shown is a structural schematic of an electronic device 400 provided in an embodiment of this disclosure, including: The system includes a processor 41, a memory 42, and a bus 43. The memory 42 stores execution instructions and includes main memory 421 and external memory 422. The main memory 421, also called internal memory, temporarily stores the computational data in the processor 41, as well as data exchanged with external memory such as a hard disk. The processor 41 exchanges data with the external memory 422 through the main memory 421. When the electronic device 400 is running, the processor 41 communicates with the memory 42 through the bus 43, causing the processor 41 to execute the following instructions: Determine the target prerequisites corresponding to the target domain, and input the target prerequisites into the target large language model to obtain the first triplet; Based on the first head entity of the first triplet, knowledge expansion is performed to obtain the second triplet; Based on the first relation entity of the first triplet and the second relation of the second triplet, the entities are filtered to obtain the third triplet; wherein, the third relation entity of the third triplet meets the preset occurrence frequency condition; The third triplet is deduplicated to obtain a preset number of target triplets corresponding to the target domain; The parameters of the target large language model are updated based on the target triples and the general knowledge of the target domain to obtain the updated large language model.
[0093] The basic principles of this disclosure have been described above with reference to specific embodiments. However, it should be noted that the advantages, benefits, and effects mentioned in this disclosure are merely examples and not limitations, and should not be considered as essential features of each embodiment of this disclosure. Furthermore, the specific details disclosed above are for illustrative and facilitative purposes only, and are not limitations. These details do not limit the scope of this disclosure to the necessity of employing the aforementioned specific details for implementation.
[0094] The block diagrams of devices, apparatuses, devices, and systems disclosed herein are merely illustrative examples and are not intended to require or imply that they must be connected, arranged, or configured in the manner shown in the block diagrams. As those skilled in the art will recognize, these devices, apparatuses, devices, and systems can be connected, arranged, and configured in any manner. Words such as “comprising,” “including,” “having,” etc., are open-ended terms meaning “including but not limited to,” and are used interchangeably with them. The terms “or” and “and” as used herein refer to the terms “and / or,” and are used interchangeably with them unless the context clearly indicates otherwise. The term “such as” as used herein refers to the phrase “such as but not limited to,” and is used interchangeably with it.
[0095] Additionally, as used herein, the "or" used in a list of items beginning with "at least one" indicates a separate list, such that a list of, for example, "at least one of A, B, or C" means A or B or C, or AB or AC or BC, or ABC (i.e., A and B and C). Furthermore, the word "exemplary" does not imply that the described example is preferred or better than other examples.
[0096] It should also be noted that in the systems and methods of this disclosure, the components or steps can be decomposed and / or recombined. These decompositions and / or recombinations should be considered as equivalent solutions to this disclosure.
[0097] Various changes, substitutions, and modifications can be made to the technology described herein without departing from the teachings defined by the appended claims. Furthermore, the scope of the claims of this disclosure is not limited to the specific aspects of the processes, machines, manufactures, events, means, methods, and actions described above. Currently existing or later-developed processes, machines, manufactures, events, means, methods, or actions that perform substantially the same function or achieve substantially the same result as the corresponding aspects described herein can be utilized. Therefore, the appended claims include such processes, machines, manufactures, events, means, methods, or actions within their scope.
[0098] The above description of the disclosed aspects is provided to enable any person skilled in the art to make or use this disclosure. Various modifications to these aspects will be readily apparent to those skilled in the art, and the general principles defined herein may be applied to other aspects without departing from the scope of this disclosure. Therefore, this disclosure is not intended to be limited to the aspects shown herein, but rather to be carried out within the widest scope consistent with the principles and novel features disclosed herein.
[0099] The above description has been given for purposes of illustration and description. Furthermore, this description is not intended to limit the embodiments of this disclosure to the forms disclosed herein. Although numerous exemplary aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, alterations, additions, and sub-combinations therein.
Claims
1. A knowledge mining method based on a large model, characterized in that, include: Determine the target prerequisites corresponding to the target domain, and input the target prerequisites into the target large language model to obtain the first triplet; Based on the first head entity of the first triplet, knowledge expansion is performed to obtain the second triplet; Based on the first relation entity of the first triplet and the second relation of the second triplet, the entities are filtered to obtain the third triplet; wherein, the third relation entity of the third triplet meets the preset occurrence frequency condition; The third triplet is deduplicated to obtain a preset number of target triplets corresponding to the target domain; The parameters of the target large language model are updated based on the target triples and the general knowledge of the target domain to obtain the updated large language model.
2. The method as described in claim 1, characterized in that, The second triplet is obtained by extending the knowledge based on the first head entity of the first triplet, including: The first entity is retrieved through multiple rounds of iterative retrieval using an instruction retrieval model, resulting in a second entity with expanded knowledge. The second triplet is obtained by performing structural integrity processing based on the second head entity.
3. The method as described in claim 1, characterized in that, The step of filtering entities based on the first relation entity of the first triplet and the second relation of the second triplet to obtain the third triplet includes: In the first triplet and the second triplet, a relation entity that satisfies the preset occurrence frequency condition is identified, and this relation entity is identified as the third relation entity; Determine the head entity corresponding to the third relation entity, and designate the head entity as the third head entity; Based on the third head entity and the third relation entity, a third tail entity that meets the preset related conditions with the third relation entity is generated, thereby obtaining a third triplet.
4. The method as described in claim 1, characterized in that, The process of deduplicating the third triplet to obtain a preset number of target triplets corresponding to the target domain includes: Determine the set of duplicate triples in the third triplet, and retain any triplet in the set of duplicate triples to obtain the deduplicated triplet; If the number of triplets in the deduplicated triplets is greater than a preset number, the deduplicated triplets are optimized to obtain a preset number of fourth triplets. Based on the general knowledge of the target domain, the fourth triple is retrieved and corrected using the target language big model to obtain the target triple.
5. The method as described in claim 4, characterized in that, The general knowledge based on the target domain is used to perform knowledge retrieval and correction on the fourth triplet through the target language big model to obtain the target triplet, including: Based on the general knowledge of the target domain, knowledge retrieval is performed on the fourth triplet through the target language big model to obtain the fifth triplet corresponding to the first retrieval result and the sixth triplet corresponding to the second retrieval result; wherein, the first retrieval result is used to indicate the fourth triplet with correct knowledge and the second retrieval result is used to indicate the fourth triplet with incorrect knowledge. The sixth triplet is corrected to obtain the seventh triplet; The fifth triplet and the seventh triplet are identified as the target triplet.
6. The method as described in claim 5, characterized in that, The process of correcting the sixth triplet to obtain the seventh triplet includes: Obtain relevant knowledge and erroneous knowledge information about the sixth triplet; Based on the relevant knowledge, the erroneous knowledge information is corrected to obtain the corrected sixth triplet; The modified sixth triplet is determined as the seventh triplet.
7. A knowledge mining device based on a large model, characterized in that, include: The determination module is used to determine the target prerequisites corresponding to the target domain, and input the target prerequisites into the target large language model to obtain the first triplet; An extension module is used to extend knowledge based on the first head entity of the first triplet to obtain a second triplet. The filtering module is used to perform entity filtering based on the first relation entity of the first triplet and the second relation of the second triplet to obtain a third triplet; wherein the third relation entity of the third triplet meets a preset occurrence frequency condition. The deduplication module is used to deduplicatize the third triplet to obtain a preset number of target triplets corresponding to the target domain. The update module is used to update the parameters of the target large language model based on the target triples and the general knowledge of the target domain, so as to obtain the updated large language model.
8. An electronic device comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method as described in any one of claims 1 to 6.
9. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method as described in any one of claims 1 to 6.
10. A computer program product, comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method as described in any one of claims 1 to 6.