An Evolvable Domain Knowledge Graph Design Method

Abstract

This invention discloses an evolvable domain knowledge graph design method, including generating a main knowledge base based on the knowledge graph; selecting at least one mutually exclusive definition entity and a regular entity; establishing sub-bases by branching from the mutually exclusive entities; if any unselected entity in the current main knowledge base conflicts with the new sub-base, the address of the unselected entity is included in the features of the new sub-base; when any user searches the main knowledge base and encounters an entity pointing to a sub-base, a user vote is triggered; if the proportion of users who approve of all the evidence in the sub-base is greater than or equal to a preset threshold, a sub-base merging operation with the main knowledge base is triggered; otherwise, the sub-base is deleted, and user voting is no longer triggered; if a user believes that some of the evidence in the sub-base is correct, a new sub-base can be created. This application enables the content in the main knowledge base to evolve in line with the times, dynamically modifying the content in the main knowledge base based on newly discovered evidence.

Description

Technical Field

[0001] This invention relates to the field of knowledge base management technology for large language models, and in particular to an evolvable domain knowledge graph design method. Background Technology

[0002] Due to the rapid development of large language models, various knowledge bases are needed to expand their application areas. Knowledge graph-based knowledge bases are a common type, frequently used for knowledge graph reasoning in RAGs. Currently, knowledge graphs used in scientific research cannot evolve automatically, cannot automatically update knowledge, and do not accept mutually exclusive knowledge. However, in actual scientific research, some newly discovered evidence is mutually exclusive, requiring falsification methods to determine which knowledge needs to be updated.

[0003] Currently, distributed knowledge bases based on multi-scale and knowledge graphs do not allow conflicts between knowledge points and assume that all knowledge is always true. During the RAG process, the large language model consistently seeks the most correct answer within the distributed knowledge base based on multi-scale and knowledge graphs through the knowledge graph. However, in practical applications, much scientific evidence (e.g., in geology) is directly conflicting, necessitating the introduction of evidence chain analysis around a central topic during knowledge retrieval. Based on a central knowledge base centered on research, methods for knowledge base merging and bifurcation are defined.

[0004] Existing decentralized knowledge bases based on blockchain cannot distinguish conflicting knowledge, nor can they determine whether new knowledge should be stored by adding content to the main knowledge base or creating a new sub-chain. Furthermore, due to their tamper-proof nature, they may result in too many sub-chains, meaning that sub-chains cannot be canceled or merged in a timely manner, ultimately leading to a knowledge base that cannot be dynamically expanded or contains useless knowledge. Summary of the Invention

[0005] To address all or some of the aforementioned problems, the present invention aims to provide an evolvable domain knowledge graph design method.

[0006] The present invention solves its problems through the following technical solution: An evolvable domain knowledge graph design method includes the following steps: S1. Generate a main knowledge base based on the knowledge graph, store defined entities and regular entities in the main knowledge base, and establish a searchable knowledge graph; S2. Traverse the newly discovered evidence, retrieve mutually exclusive definition entities and regular entities, and select at least one mutually exclusive definition entity and its related regular entity; S3. Generate sub-libraries by branching out from mutually exclusive entities; S4. Based on the knowledge in the current main knowledge base, determine whether all unselected entities and evidence entities in the new sub-base are mutually exclusive. If there is an unselected entity that conflicts with the new sub-base, then include the address of the unselected entity in the features of the new sub-base. S5. When any user searches the main knowledge base and encounters an entity that points to the sub-base, a user vote is triggered. If the proportion of users who approve all the evidence in the sub-base is greater than or equal to the preset proportion threshold Y, and the number of users who approve all the evidence in the sub-base exceeds the set first threshold, the sub-base and main knowledge base are merged. S6. If the proportion of users who approve all the evidence in the sub-library during the vote is less than the preset proportion threshold Y, and the total number of voting users exceeds the set second threshold, then the sub-library is deleted and user voting is no longer triggered; if users believe that some of the evidence in the sub-library is correct, while the rest is unreliable, then a new sub-library can be created and users are refused entry into the sub-library.

[0007] Optionally, methods for determining whether mutually exclusive knowledge exists in the knowledge base include: inputting the evidence chain in the knowledge graph into Credal Transformer to calculate the uncertainty u; when the uncertainty u is greater than 0.5 or a human-given threshold D, it is determined that mutually exclusive knowledge exists in the evidence chain.

[0008] Optionally, step S3 specifically includes: establishing a DAG graph, selecting the main knowledge base as the main node, defining sub-library characteristics based on mutually exclusive entity addresses specified by the user, establishing sub-library branches by specifying new sub-library names by the user, and establishing new branches from the main node in the DAG graph pointing to the sub-library.

[0009] Optionally, each unselected entity that conflicts with the new sub-library in step S4 contains information about the sub-library it points to, which can be queried during retrieval.

[0010] Optionally, in step S5, the merging operation between the sub-base and the main knowledge base includes: Suspend access to the main knowledge base, identify the knowledge in the main knowledge base that needs to be updated, deduplicate and clean the data, and delete invalid data; Search the main knowledge base for contradictory definitions and patterns between the main and sub-knowledge bases; The main knowledge base is backed up as a historical node. The defined entities and rule entities in the sub-base are used to update the main knowledge base, and the knowledge and data are also updated. Update the main knowledge base version number, keep the master node unchanged on the DAG graph, update the branch nodes, and re-enable each knowledge base.

[0011] In summary, the technical effects and advantages of this invention are as follows: This invention does not require all knowledge in the knowledge base to be correct, but only requires that it have a high degree of relevance when expanding the sub-base. That is, the expanded sub-base is required to have a high degree of relevance that can falsify a certain existing knowledge in the main knowledge base. When the large language model performs a retrieval, it will enter a certain sub-base (sub-chain) through existing evidence and perform hypothesis deduction. It enables the content in the main knowledge base to evolve with the times and dynamically modify the content in the main knowledge base based on newly discovered evidence. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a schematic diagram of an evolvable domain knowledge graph update process in one embodiment of the present invention; Figure 2 This is a schematic diagram of the process of discovering mutually exclusive knowledge entities and triggering voting in one embodiment of the present invention; Figure 3 This is a schematic diagram of the process of merging the main knowledge base and the sub-knowledge base in one embodiment of the present invention; Figure 4 This is a schematic diagram illustrating the process of determining whether a knowledge graph contains mutually exclusive knowledge, according to one embodiment of the present invention. Figure 5 This is a schematic diagram of the uncertainty u and Credal Transformer calculation process in one embodiment of the present invention; Figure 6 This is a schematic diagram of the main knowledge base and sub-knowledge base in one embodiment of the present invention. Detailed Implementation

[0014] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0015] In order to enable the content in the knowledge base to evolve with the times and dynamically modify the content of the knowledge base based on newly discovered evidence by researchers, a knowledge base that can model the coexistence of different viewpoints and can accept mutually exclusive knowledge is needed.

[0016] Specifically, this embodiment proposes an evolvable domain knowledge graph design method, such as... Figures 1-6 As shown, it includes the following steps: S1. Generate a main knowledge base based on a knowledge graph, store defined entities and regular entities in the main knowledge base, and establish a searchable knowledge graph.

[0017] To generate a knowledge graph-based master knowledge base using traditional methods, the research content of the master knowledge base is first determined. A searchable knowledge graph is then established by storing a sufficient number of defined entities and pattern entities in the master knowledge base. The knowledge in the master knowledge base is then traversed and retrieved to ensure it is self-consistent; in the initial state, no branching is required.

[0018] The basic unit of a knowledge graph is a triple consisting of "Entity-Relationship-Entity". Information extraction is a technique for automatically extracting structured information such as entities, relationships, and entity attributes from semi-structured and unstructured data. Key technologies include entity extraction, relation extraction, and attribute extraction. Entity extraction refers to automatically identifying named entities from text datasets. Current mainstream technologies are open-domain entity extraction. Relation extraction, in order to obtain semantic information, extracts the relationships between entities from related corpora. Only by linking entities through relationships can a network-like knowledge structure be formed. Its technological research has evolved from early "manually constructed grammar and semantic rules" (pattern matching) and "statistical machine learning" to a combination of "open-domain information extraction methods" and "closed-domain methods". Attribute extraction aims to collect attribute information of specific entities from different information sources. For example, for a public figure, information such as their nickname, birthday, nationality, and educational background can be obtained from publicly available online information. Data mining methods are used to directly extract the relationship patterns between entity attributes and attribute values ​​from text, thereby enabling the location of attribute names and attribute values ​​in the text.

[0019] S2. Based on the new findings, the researchers searched the main knowledge base, traversed the evidence of the new findings, retrieved mutually exclusive definition entities and regular entities, and selected at least one mutually exclusive definition entity and its related regular entity.

[0020] Optionally, a method for determining whether mutually exclusive knowledge exists in the knowledge base includes: inputting the evidence chain from the knowledge graph into the Credal Transformer to calculate the uncertainty u. When the uncertainty u is greater than 0.5 or a manually given threshold D, it is determined that mutually exclusive knowledge exists in the evidence chain. When mutually exclusive knowledge is found, manual intervention is required to determine whether the evidence chain needs to be forked, i.e., whether a sub-base should be established. The calculation process of the Credal Transformer is as follows: Figure 5 As shown, the specific calculation formula is expressed as follows: The final layer of the multi-head attention mechanism in large models can be represented as: The Q and K matrices are extracted from the bypass. The multi-head attention mechanism is a common-sense algorithm, and all symbolic representations are standard and accepted within the AI ​​industry. Furthermore, the following calculations are performed from the Q and K matrices: Then calculate: Overall strength of evidence: Total uncertainty: Where T is the length of the input sequence (token); u represents the total uncertainty; the greater the uncertainty, the greater the possibility that the data is mutually exclusive. If the uncertainty u is greater than 0.5 or a manually given threshold D, it is considered that there is a mutually exclusive evidence chain in the knowledge graph. The value of D can generally be set to 0.5 or other values, and the value of D can be adaptively set according to the actual application.

[0021] Repeat the above process to keep the main knowledge base running until its central issue is closed, at which point the process ends.

[0022] S3. Generate sub-libraries by branching out from mutually exclusive entities; Optionally, step S3 specifically includes: establishing a DAG graph, selecting the main knowledge base as the main node, defining sub-library characteristics based on mutually exclusive entity addresses specified by the user, establishing sub-library branches by specifying new sub-library names by the user, and establishing new branches from the main node in the DAG graph pointing to the sub-library.

[0023] A DAG (Directed Acyclic Graph) is a special data structure in graph theory. Its core feature is that all edges have a direction and there are no cycles in the graph.

[0024] S4. Based on the knowledge in the current main knowledge base, determine whether all unselected entities and evidence entities in the new sub-base are mutually exclusive. If an unselected entity conflicts with the new sub-base, then the address of the unselected entity is included in the features of the new sub-base. Each entity contains information about the sub-base it points to, which can be queried during retrieval. The Credal Transformer calculation in step S2 applies to all mutual exclusion determinations.

[0025] S5. When any user searches the main knowledge base and encounters an entity that points to the sub-base, a user vote is triggered. If the proportion of users who approve all the evidence in the sub-base is greater than or equal to the preset proportion threshold Y, and the number of users who approve all the evidence in the sub-base exceeds the set first threshold, the sub-base and main knowledge base are merged. Optionally, the sub-repository and main knowledge base merging operation includes: Suspend access to the main knowledge base, identify the knowledge in the main knowledge base that needs to be updated, deduplicate and clean the data, and delete invalid data. Invalid data refers to data that is outside the defined domain, such as data where the content ratio of a certain mineral exceeds the range of "0-1". Or it refers to data in an unknown format, such as year data containing letters or Chinese characters. Search the main knowledge base for contradictory definitions and patterns between the main and sub-knowledge bases; The main knowledge base is backed up as a historical node. The defined entities and rule entities in the sub-base are used to update the main knowledge base, and the knowledge and data are also updated. Update the main knowledge base version number, keep the master node unchanged on the DAG graph, update the branch nodes, and re-enable each knowledge base.

[0026] The threshold Y for determining the majority of users is set by the knowledge base builder, typically Y=0.9, meaning that if 90% of users believe the fork can be merged, then the content from the fork replaces the content in the main knowledge base. The first threshold for the number of users who agree with all evidence in the sub-base and the second threshold for the number of users who vote are specified by the knowledge base builder; it is not necessary for all users to have accessed the entity before a judgment can be made. The above process can be described using the following formula: Z = M / N; If (Z>Y), perform the merge; otherwise, abandon the sub-library. Where M represents the number of users who approve all evidence in a sub-database in a vote; N represents the total number of users who voted in a sub-database; and Z represents the percentage of users who approve all evidence in the sub-database in the current vote. However, if the total number of users who voted does not reach the second threshold, this percentage Z cannot be calculated, and voting needs to continue until the total number of users who voted exceeds the second threshold.

[0027] S6. If the proportion of users who approve all the evidence in the sub-library during the vote is less than the preset threshold Y, and the total number of voting users exceeds the set second threshold, then the sub-library is deleted, the corresponding branch in the DAG graph is deleted, and user voting is no longer triggered. If users believe that some of the evidence in the sub-library is correct, while the rest is unreliable, then a new sub-library can be created, and access to the current sub-library is refused. The steps for creating a new sub-library are the same as those in step S3 above. Here, a new sub-library can be created as long as any user believes that some of the evidence in the sub-library is correct. Since the information content of the new sub-library is less than that of the atomic library (the existing fork point), it will not cause infinite malicious forks.

[0028] Application examples: The paper "The Great Debate on Eclogites in the Western Gneiss Region of the Caledonian Orogen in Norway: The Debate on In-Crust Origin vs. Extraterrestrial Mantle Origin" presents a fundamental contradiction: the granitic rocks in the western gneiss region contain both eclogites originating from the mantle and gabbro intrusions that have metamorphosed into eclogites in situ; in addition, the discovery of coesite in the eclogites indicates that their formation pressure was far higher than that achievable through crustal thickening.

[0029] Following the acquisition of various geophysical and geochemical evidence, it was finally determined that the western gneiss region of Norway, which contains eclogites, is the continental basement of the Baltic Plate, which subducted into the mantle beneath the Laurentia continent during the Scandian orogeny at approximately 400 Ma.

[0030] Therefore, the relationship between the origin of the original crust and the origin of the extraterrestrial mantle should be modified in the knowledge base. In the main knowledge base, it should be modified to reflect the origin of continental crust subducting into the mantle.

[0031] The centralized distributed knowledge base proposed in this embodiment utilizes falsifiable logic to construct a main knowledge base (central knowledge base). A knowledge graph is built within the main knowledge base, marking two types of entities: definitions and rules. Each branching point must specify at least one definition entity and one related rule entity from the main knowledge base. When new knowledge is introduced at a branching point, an evidence chain must be specified, enabling the falsification of existing definitions or rules. Common sense and established research content should be directly merged into the main knowledge base after being judged by researchers, rather than being branched.

[0032] Scientific research often involves conflicting hypotheses that cannot be fully verified under current conditions. This necessitates a forked knowledge base, where each research node contributes a distributed sub-knowledge base. Each sub-knowledge base (sub-base or sub-chain) contains unverified hypotheses, enabling the large language model to perform centralized hypothesis reasoning around a specific specialized problem. The central knowledge base serves as the master knowledge base, upon which sub-bases can be built. When the content of a sub-base is proven valid, it is merged with the master knowledge base to form a new master knowledge base.

[0033] This invention addresses the challenge of modeling mutually exclusive knowledge in knowledge bases and establishes an evolvable distributed knowledge base by using falsifiable features to describe the relationship between the central and sub-bases. The key characteristic of this evolvable knowledge base is that when new knowledge is introduced at a bifurcation point, a chain of evidence must be specified that can falsify existing definitions or rules. In other words, bifurcation must originate from an entity and its relationships within the knowledge graph of the main knowledge base; these entities and relationships are typically referred to as rules. Common sense and established research content should be directly merged into the main knowledge base after being judged by researchers, rather than being bifurcerated.

[0034] This design method addresses the evolutionary requirements of a specific research project in scientific research by providing a dynamic approach to modifying the knowledge base. It allows for voting and modification of the content within the knowledge base, enabling the knowledge base to evolve dynamically over time.

[0035] This invention does not require all knowledge in the knowledge base to be correct, but only requires that it be highly relevant when expanding sub-bases. That is, the expanded sub-bases must be highly relevant enough to falsify some existing knowledge in the main knowledge base. When the large language model performs a retrieval, it will enter a sub-base (sub-chain) through existing evidence and perform hypothesis deduction. This enables the content in the main knowledge base to evolve with the times and dynamically modify the content in the main knowledge base based on newly discovered evidence.

[0036] This method records and displays core, mutually exclusive knowledge. After acquiring sufficient information, it can determine whether the knowledge needs to be evolved. In scientific research applications, this knowledge base is often used for discussion and modeling of a central issue. The evolution of existing knowledge is typically based on feedback from users with research privileges. This patent can assist researchers in discussing a central issue and help users organize and model existing scientific arguments.

[0037] Finally, it should be noted that: the accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments disclosed in this invention, and other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other. In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An evolvable domain knowledge graph design method, characterized in that, Includes the following steps: S1. Generate a main knowledge base based on the knowledge graph, store defined entities and regular entities in the main knowledge base, and establish a searchable knowledge graph; S2. Traverse the newly discovered evidence, retrieve mutually exclusive definition entities and regular entities, and select at least one mutually exclusive definition entity and its related regular entity; S3. Generate sub-libraries by branching out from mutually exclusive entities; S4. Based on the knowledge in the current main knowledge base, determine whether all unselected entities and evidence entities in the new sub-base are mutually exclusive. If there is an unselected entity that conflicts with the new sub-base, then include the address of the unselected entity in the features of the new sub-base. S5. When any user searches the main knowledge base and encounters an entity that points to the sub-base, a user vote is triggered. If the proportion of users who approve all the evidence in the sub-base is greater than or equal to the preset proportion threshold Y, and the number of users who approve all the evidence in the sub-base exceeds the set first threshold, the sub-base and main knowledge base are merged. S6. If the proportion of users who approve all the evidence in the sub-library during the vote is less than the preset proportion threshold Y, and the total number of voting users exceeds the set second threshold, then the sub-library will be deleted and user voting will no longer be triggered. If a user believes that some of the evidence in a sub-database is correct, while the rest is unreliable, they can create a new sub-database and refuse access to it.

2. The method for designing an evolvable domain knowledge graph according to claim 1, characterized in that, Methods for determining whether mutually exclusive knowledge exists in a knowledge base include: inputting the evidence chain in the knowledge graph into Credal Transformer to calculate the uncertainty u; when the uncertainty u is greater than 0.5 or a human-given threshold D, it is determined that mutually exclusive knowledge exists in the evidence chain.

3. The method for designing an evolvable domain knowledge graph according to claim 1, characterized in that, Step S3 specifically includes: establishing a DAG graph, selecting the main knowledge base as the master node, defining sub-library characteristics based on user-specified mutually exclusive entity addresses, establishing sub-library branches by specifying new sub-library names, and establishing new branches from the master node in the DAG graph pointing to the sub-library.

4. The method for designing an evolvable domain knowledge graph according to claim 1, characterized in that, In step S4, each unselected entity that conflicts with the new sub-library contains information about the sub-library it points to, which can be queried during retrieval.

5. The method for designing an evolvable domain knowledge graph according to claim 3, characterized in that, In step S5, the merging operation between the sub-base and the main knowledge base includes: Suspend access to the main knowledge base, identify the knowledge in the main knowledge base that needs to be updated, deduplicate and clean the data, and delete invalid data; Search the main knowledge base for contradictory definitions and patterns between the main and sub-knowledge bases; The main knowledge base is backed up as a historical node. The defined entities and rule entities in the sub-base are used to update the main knowledge base, and the knowledge and data are also updated. Update the main knowledge base version number, keep the master node unchanged on the DAG graph, update the branch nodes, and re-enable each knowledge base.

Images