Large-model multi-stage knowledge fusion method for intelligence industry research and judgment

Abstract

The invention discloses a large-model multi-stage knowledge fusion method for intelligence industry research and judgment, and the method comprises the following steps: determining an intelligence analysis range according to a user demand, and obtaining multi-source modal intelligence data; extracting an intelligence graph from the multi-source modal intelligence data; performing intelligence analysis by adopting a large language model based on the intelligence graph to generate an intelligence hypothesis; identifying an intelligence data gap based on the intelligence hypothesis, and obtaining intelligence data related to the gap; and repeating the process for one or more times to generate a complete information analysis result. According to the method, the reliability of the research and judgment conclusion is improved by identifying the evidence gap, actively discovering undercognition and directionally collecting the data, and the dependence on the qualification analysts is greatly reduced.

Description

Technical Field

[0001] This invention relates to a multi-stage knowledge fusion method for large-scale models used in intelligence industry analysis, belonging to the field of artificial intelligence technology. Background Technology

[0002] Intelligence analysis refers to the process of extracting valuable knowledge from massive amounts of multi-source information and forming predictions and assessments of specific events, trends, or threats. It has crucial applications in fields such as finance, security, technological competition, and business decision-making.

[0003] Traditional analytical methods rely heavily on analysts' experience, involving reading, analyzing, and logically reasoning through structured or semi-structured information such as text reports, news, and financial data. This approach suffers from inherent drawbacks, including inefficiency, limited information processing capacity, strong subjectivity, and difficulty in handling information overload.

[0004] With the development of information technology, especially the application of knowledge graphs and artificial intelligence, the level of automation in intelligence analysis has improved. Existing technical solutions typically focus on one or more of the following aspects: Information extraction and graph construction: Utilize natural language processing technology to automatically extract entities, relationships, and attributes from text data, construct a domain knowledge graph, and store intelligence elements in a structured form; Graph-based query and reasoning: using rule-based reasoning or simple graph algorithms to perform queries, path discovery, or association analysis on the graph; Multimodal information processing: Attempts were made to process non-textual data such as images and audio, but the strategy of processing each modality independently and then fusing them later was usually adopted, which failed to achieve deep cross-modal semantic alignment and fusion.

[0005] However, existing technologies still face many challenges: Insufficient multimodal fusion: Most existing systems are text-centric, which do not fully utilize the rich intelligence information contained in images and speech, and lack effective technical means to achieve deep semantic alignment and complementarity between features of different modalities, resulting in omission or misunderstanding of key information; Static analysis vs. passive response: Existing analysis systems are mostly based on existing, static datasets and lack the ability to proactively discover and collect data; Weak ability to handle complex and contradictory intelligence: Real intelligence environments are often filled with contradictory information. Traditional knowledge graphs usually use binary relationships, which makes it difficult to properly handle such contradictions. They often lose information by simply selecting and discarding information, and cannot reflect the dynamic evolution and uncertainty of events. Limited level of intelligence: Although machine learning models have been introduced, advanced analytical tasks such as deep reasoning, hypothesis generation, and critical thinking still mainly rely on manual labor, and the level of system automation and intelligence needs to be improved.

[0006] Therefore, it is necessary to conduct more in-depth research on intelligence analysis methods in order to solve the above problems. Summary of the Invention

[0007] Through in-depth research, this invention proposes a multi-stage knowledge fusion method for large-scale models used in intelligence industry analysis, comprising the following steps: S1. Determine the scope of intelligence analysis based on user needs and obtain multi-source modal intelligence data; S2. Extract intelligence maps from multi-source modal intelligence data; S3. Based on the intelligence graph, a large language model is used to conduct intelligence analysis and generate intelligence hypotheses; S4. Identify intelligence data gaps based on intelligence assumptions and obtain intelligence data related to the gaps; S5. Repeat S2-S4 once or multiple times to generate complete intelligence analysis results.

[0008] In a preferred embodiment, in S1, user needs are obtained based on a template, a structured query statement is obtained, and data is queried to obtain multi-source intelligence data.

[0009] In a preferred embodiment, in S1, after obtaining the structured query statement, a knowledge graph is used to expand the query statement.

[0010] In a preferred embodiment, the multi-source modal intelligence data includes text data, image data, and voice data.

[0011] In a preferred embodiment, S2 includes the following sub-steps: S21. Extract features from different modalities of data in each intelligence report; S22. To address the heterogeneity issue among multimodal features, a multi-head attention mechanism is used to recalibrate the feature vectors of each modality based on the eigenvectors of other modalities. The calibrated text feature vectors, image feature vectors, and speech feature vectors are then fused using a Transformer model to obtain a fused semantic vector. S23. Use a multi-task decoder to convert the fused semantic vector U into structured knowledge units; S24. Construct the obtained knowledge units into an intelligence graph, where entities are nodes, edges represent relationships between entities, and node attributes represent entity attributes.

[0012] In a preferred embodiment, when there is a conflict in the relationships between entities in different knowledge units, the node is marked as an entity with time.

[0013] In a preferred embodiment, each relation edge in the graph is assigned a weight value to characterize the authenticity of events between entities with contradictory relationships.

[0014] In a preferred embodiment, the weight is obtained by attaching source information to each knowledge unit, the source information including the source of the intelligence data corresponding to the knowledge unit, which is used to characterize the basic credibility of the intelligence data.

[0015] The present invention also provides an electronic device, comprising: At least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor to enable the at least one processor to perform the method described in any of the preceding claims.

[0016] The present invention also provides a computer-readable storage medium storing computer instructions, wherein the computer instructions are configured to cause the computer to perform the method according to any one of the preceding claims.

[0017] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention performs cross-modal feature calibration through a multi-head attention mechanism and then uses Transformer for deep fusion, thereby realizing the intrinsic correlation and unified representation of multimodal intelligence information; (2) Using large models to identify intelligence data gaps, we can proactively discover cognitive deficiencies and collect data in a targeted manner by identifying evidence gaps, thereby improving the reliability of the judgment conclusions and greatly reducing the reliance on senior analysts. Attached Figure Description

[0018] Figure 1 The diagram illustrates a flowchart of a large-scale, multi-stage knowledge fusion method for intelligence industry analysis, according to a preferred embodiment of the present invention. Detailed Implementation

[0019] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Through these descriptions, the features and advantages of the present invention will become clearer and more apparent.

[0020] The term “exemplary” as used herein means “serving as an example, embodiment, or illustration.” Any embodiment illustrated herein as “exemplary” is not necessarily to be construed as superior to or better than other embodiments. Although various aspects of embodiments are shown in the accompanying drawings, the drawings are not necessarily drawn to scale unless specifically indicated otherwise.

[0021] This invention provides a multi-stage knowledge fusion method for large-scale models used in intelligence industry analysis, such as... Figure 1 As shown, it includes the following steps: S1. Determine the scope of intelligence analysis based on user needs and obtain multi-source modal intelligence data; S2. Extract intelligence maps from multi-source modal intelligence data; S3. Based on the intelligence graph, a large language model is used to conduct intelligence analysis and generate intelligence hypotheses; S4. Identify intelligence data gaps based on intelligence assumptions and obtain intelligence data related to the gaps; S5. Repeat S2-S4 once or multiple times to generate complete intelligence analysis results.

[0022] In S1, user needs are obtained based on templates, structured query statements are obtained, and data queries are performed to obtain multi-source intelligence data.

[0023] Preferably, the template includes one or more of the following dimensions: Entities: The objects to be queried, such as companies, people, technologies, etc. Key questions: The questions that need to be answered, such as, **What is the next-generation technology roadmap for the product?** Time range: The time window of interest; Areas of interest: Areas of focus, such as AI algorithms, photovoltaic technology, etc.

[0024] In this invention, templates can transform vague user requirements into structured query statements, thereby enabling actionable task analysis and avoiding significant deviations between the analysis results and the requirements.

[0025] Preferably, in S1, after obtaining the structured query statement, a knowledge graph is used to expand the query statement.

[0026] The knowledge graph can be any existing graph, such as an industry knowledge graph, or it can be created by those skilled in the art according to actual needs. No restrictions are imposed in this invention.

[0027] Knowledge graphs can be used to expand query statements. For example, for a company entity, a knowledge graph can be used to expand to related entities such as its subsidiaries, key partners, core competitors, and major customers. For the key question of "market positioning", it can be expanded to "pricing strategy", "target customer group" and "channel strategy". In other words, knowledge graphs can effectively expand the scope of intelligence collection and improve the accuracy of intelligence analysis.

[0028] According to the present invention, the above method enables targeted and efficient multi-source data acquisition, avoids data overload, and improves information acquisition efficiency.

[0029] The method of obtaining the multi-source intelligence data is not limited in this invention, and it is usually done by web crawling or API querying.

[0030] The multi-source modal intelligence data includes text data, image data, and voice data.

[0031] In this invention, by fusing multi-source heterogeneous modal data such as text, images, and voice, intelligence elements can be extracted more fully and accurately, overcoming the limitations of single text intelligence.

[0032] S2 includes the following sub-steps: S21. Extract features from different modalities of data in each intelligence report.

[0033] Each intelligence data may contain a combination of multiple modalities. For example, a single intelligence report may contain text, images, video, or audio. Different modalities of data in each intelligence report are extracted separately.

[0034] For text data, a language encoder is used to obtain text feature vectors; For image data, a visual encoder is used to obtain image feature vectors; For speech data, an acoustic encoder is used to obtain speech feature vectors.

[0035] Traditional intelligence data consists only of text data. In this invention, image data and voice data are added to improve the accuracy of key information extraction.

[0036] In this invention, the specific structure of the language encoder, visual encoder, and acoustic encoder is not limited. Those skilled in the art can select a suitable encoder based on experience, such as BERT, BLIP, HuBERT, etc.

[0037] Furthermore, text feature vectors, image feature vectors, and speech feature vectors reside in different feature spaces, making direct comparison and application impossible.

[0038] In this invention, S22, to address the heterogeneity problem among multimodal features, a multi-head attention mechanism is used to recalibrate the feature vector of each modality based on the eigenvectors of other modalities, as shown below: t n =f(t,v,a); v n =f(v,t,a); a n =f(a,t,v); Where t represents the text feature vector, v represents the image feature vector, a represents the speech feature vector, and t n v represents the calibrated text feature vector. n Let a represent the calibrated image feature vector. n Let f represent the calibrated speech feature vector, and let f represent the multi-head attention mechanism.

[0039] Furthermore, the calibrated text feature vector, image feature vector, and speech feature vector are fused using a Transformer model to obtain a fused semantic vector U, denoted as: U = Transformer(Concat(t) n v n a n )) Here, Transformer represents the Transformer model, and Concat represents vector concatenation.

[0040] S23. Use a multi-task decoder to convert the fused semantic vector U into structured knowledge units.

[0041] The knowledge unit includes entities, relationships between entities, and attributes of entities.

[0042] The multi-task decoder includes an entity recognition head, a relationship classification head, and an attribute extraction head. The entity recognition head uses a sequence labeling architecture, taking a fused semantic vector U as input and outputting one or more entity labels. The relation classification head uses a pointer network architecture, taking the fused semantic vector U and the entity labels output by the entity recognition head as input, and outputting the relationship types between entities; The attribute extraction head is sequence-to-sequence price, which takes a fused semantic vector and entity label as input and outputs the entity's attributes.

[0043] S24. Construct the obtained knowledge units into an intelligence graph, where entities are nodes, edges represent relationships between entities, and node attributes represent entity attributes.

[0044] According to the present invention, in the process of constructing an intelligence graph, there may be conflicts in the relationships between entities in different knowledge units. Traditional intelligence graphs generally determine the truth or falsehood of conflicting relationships and select one. However, this approach may result in the loss of intelligence. For example, if Company A sues Company B first, and Company B then sues Company A, the method of selecting one will lead to discrepancies between the recorded intelligence graph and the actual situation.

[0045] In a preferred embodiment, when there is a conflict in the relationship between entities in different knowledge units, the node is marked as an entity with time, thereby preserving the conflicting relationship.

[0046] Traditional graph relationships are binary. In a preferred embodiment of the present invention, each relationship edge in the graph is assigned a weight value to characterize the authenticity of events between entities with contradictory relationships. While preserving contradictory relationships, the graph can more realistically and nuancedly reflect the dynamic, complex, and even contradictory intelligence world.

[0047] Preferably, the weight is obtained by attaching source information to each knowledge unit, the source information including the source of the intelligence data corresponding to the knowledge unit, which is used to characterize the basic credibility of the intelligence data.

[0048] For example, if the intelligence data comes from official media, its basic credibility is set to 0.9; if the intelligence data comes from self-media, its basic credibility is set to 0.2.

[0049] More preferably, the tracing information also includes the release time of the intelligence data, which is used to characterize the time reliability of the intelligence data; The more recently the intelligence data was released, the higher its time reliability. For example, the time reliability of intelligence data released one day ago is 0.6, and the time reliability of intelligence data released three years ago is 0.2.

[0050] In S3, a large language model is used to receive intelligence maps and user needs, and generate multiple intelligence hypotheses.

[0051] According to the present invention, the large language model can be any existing large language model, such as ChatGPT, Doubao, Deepseek, etc.

[0052] Preferably, those skilled in the art fine-tune the parameters of the large language model to provide more accurate assumptions.

[0053] Preferably, appropriate prompt words are designed for the large language model to guide it to output one or more hypothesis information.

[0054] For example: You are an intelligence analyst. Based on an intelligence map, please propose three different core hypotheses regarding the market adoption rate of technology X over the next six months. Summarize each hypothesis in one sentence and list the three strongest pieces of evidence in the map that support that hypothesis.

[0055] In S4, a large language model is used to receive intelligence maps and generate intelligence hypotheses, and to infer intelligence gaps in the intelligence hypotheses.

[0056] According to the present invention, the large language model in S4 may be the same as that in S3, or a different large language model may be used, and no limitation is made in the present invention.

[0057] Preferably, appropriate cue words are designed for the large language model to guide it in outputting the evidence gaps in each intelligence hypothesis.

[0058] For example: You are an intelligence analyst. Some people believe that "technology X will enter the mass production stage within the next 6 months". If this view is true, please analyze what key evidence is missing or what uncertainties exist in the intelligence map. For the missing key evidence or uncertainties, what intelligence gaps need to be filled.

[0059] Based on the intelligence gaps output by the large language model, the multi-source intelligence data corresponding to the gap intelligence is obtained in the same way as in S1.

[0060] In S5, the intelligence data corresponding to the gap intelligence is added to the original multi-source intelligence data. S2-S4 are repeated multiple times, and the intelligence hypothesis obtained in the last time is the complete intelligence analysis.

[0061] According to the present invention, through S5, the intelligence map can be dynamically supplemented in a targeted manner. This proactive process of identifying cognitive gaps, collecting data in a targeted manner, and self-improving continuously enhances the reliability of analytical conclusions, thereby constantly improving the accuracy of intelligence analysis. By embedding the deep reasoning and generation capabilities of a large language model into the analysis process, intelligence hypotheses are automatically generated from the map and evidence gaps are identified, significantly reducing reliance on senior analysts and improving analysis efficiency.

[0062] Various embodiments of the methods described above in this invention can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0063] The program code used to implement the methods of the present invention can be written in any combination of one or more programming languages. This program code can be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code can be executed entirely on the machine, partially on the machine, as a standalone software package partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0064] In the context of this invention, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. Machine-readable media can include, but are not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0065] To provide interaction with a user, the methods and apparatus described herein can be implemented on a computer having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0066] The methods and apparatus described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with embodiments of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication (e.g., a communication network) of any form or medium. Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0067] Computer systems can include clients and servers. Clients and servers are generally geographically separated and typically interact via communication networks. The client-server relationship is created by computer programs running on the respective computers and having a client-server relationship with each other. A server can be a cloud server, also known as a cloud computing server or cloud host, a hosting product within the cloud computing service ecosystem, addressing the shortcomings of traditional physical hosts and VPS (Virtual Private Server, or simply "VPS") services, such as high management difficulty and weak business scalability. Servers can also be servers for distributed systems or servers incorporating blockchain technology.

[0068] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this invention can be achieved, and this is not limited herein.

Claims

1. A multi-stage knowledge fusion method for large-scale models used in intelligence industry analysis, characterized in that, Includes the following steps: S1. Determine the scope of intelligence analysis based on user needs and obtain multi-source modal intelligence data; S2. Extract intelligence maps from multi-source modal intelligence data; S3. Based on the intelligence graph, a large language model is used to conduct intelligence analysis and generate intelligence hypotheses; S4. Identify intelligence data gaps based on intelligence assumptions and obtain intelligence data related to the gaps; S5. Repeat S2-S4 once or multiple times to generate complete intelligence analysis results.

2. The large-scale, multi-stage knowledge fusion method for intelligence industry analysis according to claim 1, characterized in that, In S1, user needs are obtained based on templates, structured query statements are obtained, and data queries are performed to obtain multi-source intelligence data.

3. The large-scale, multi-stage knowledge fusion method for intelligence industry analysis according to claim 1, characterized in that, In S1, after obtaining the structured query statement, a knowledge graph is used to expand the query statement.

4. The large-scale, multi-stage knowledge fusion method for intelligence industry analysis according to claim 1, characterized in that, The multi-source modal intelligence data includes text data, image data, and voice data.

5. The large-scale, multi-stage knowledge fusion method for intelligence industry analysis according to claim 1, characterized in that, S2 includes the following sub-steps: S21. Extract features from different modalities of data in each intelligence report; S22. To address the heterogeneity issue among multimodal features, a multi-head attention mechanism is used to recalibrate the feature vectors of each modality based on the eigenvectors of other modalities. The calibrated text feature vectors, image feature vectors, and speech feature vectors are then fused using a Transformer model to obtain a fused semantic vector. S23. Use a multi-task decoder to convert the fused semantic vector U into structured knowledge units; S24. Construct the obtained knowledge units into an intelligence graph, where entities are nodes, edges represent relationships between entities, and node attributes represent entity attributes.

6. The large-scale, multi-stage knowledge fusion method for intelligence industry analysis according to claim 5, characterized in that, When there is a conflict in the relationships between entities in different knowledge units, the node is marked as an entity with time.

7. The large-scale multi-stage knowledge fusion method for intelligence industry analysis according to claim 5, characterized in that, Each relation edge in the graph is assigned a weight value to characterize the authenticity of events between entities with contradictory relationships.

8. The large-scale, multi-stage knowledge fusion method for intelligence industry analysis according to claim 7, characterized in that, The weights are obtained by attaching source information to each knowledge unit. The source information includes the source of the intelligence data corresponding to the knowledge unit, which is used to characterize the basic credibility of the intelligence data.

9. An electronic device, comprising: At least one processor; And a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-8.

10. A computer-readable storage medium storing computer instructions, wherein, The computer instructions are used to cause the computer to perform the method according to any one of claims 1-8.

Images