Data analysis platform based on large model and knowledge graph

Abstract

The invention discloses a data analysis platform based on a large model and a knowledge graph, which is characterized in that information nodes where entities and relationships related to user questions are located are quickly positioned through the knowledge graph; and a data processing and semantic comprehension algorithm based on artificial intelligence and natural language processing technologies is introduced at the rear end to carry out semantic splicing on the plurality of matched related information node semantics and user consultation question semantics to form prompt information, and the prompt information is input into a large model to generate a consultation answer of the user. Therefore, the knowledge graph and the large model technology can be comprehensively utilized to automatically generate the matched consultation answer according to the consultation questioning information of the user, so that the intelligent level of the data analysis platform is improved, the questioning background of the user can be better understood, the high-quality answer is given, and the user experience is improved. And better service experience is provided for the user.

Description

Technical Field

[0001] The present application relates to the field of intelligent data analysis, and more specifically, to a data analysis platform based on large models and knowledge graphs. Background Art

[0002] With the rapid development of the Internet, users generate a large number of consultation requests every day. These requests may involve various topics, including product information, technical support, health consultation, etc. Traditional methods of generating answers to user consultations mainly include rule-based methods, template matching, and statistical language models. These methods often cannot capture the personalized needs of users well, resulting in the answers provided being relatively general and lacking in specificity.

[0003] Specifically, rule-based methods require manual writing of a large number of rules to cover various possible situations, which is not only time-consuming and labor-intensive, but also difficult to cover all situations. Although template matching is simple and easy to use, it has poor flexibility and may not correctly match unseen or slightly changed questions. Rule- and template-based methods require frequent manual updates of rule bases and template sets to adapt to new types of questions. This approach has a low degree of intelligence and is a difficult task in the face of a rapidly changing information environment. In addition, although traditional statistical language models (such as n-gram models) can be used to estimate the probability of sentences, they are weak in processing long-distance dependencies and understanding deep semantics, resulting in the generated answers that may not be natural enough or have low relevance.

[0004] Therefore, an optimized data analysis platform is desired, which can automatically generate personalized answers that meet user needs based on user consulting questions, and provide users with a better service experience. Summary of the invention

[0005] In order to solve the above technical problems, the present application is proposed. The embodiment of the present application provides a data analysis platform based on a large model and a knowledge graph, which quickly locates the information nodes where entities and relationships related to user questions are located through the knowledge graph, and introduces data processing and semantic understanding algorithms based on artificial intelligence and natural language processing technology at the back end to semantically splice the semantics of these matched multiple related information nodes and the semantics of user consultation questions into prompt information, and use this prompt information to input the large model to generate the user's consultation answer. In this way, the knowledge graph and large model technology can be comprehensively utilized to automatically generate matching consultation answers based on the user's consultation question information, thereby improving the intelligence level of the data analysis platform. In this way, the user's question background can be better understood, and high-quality answers can be given to provide users with a better service experience.

[0006] According to one aspect of the present application, a data analysis platform based on a large model and a knowledge graph is provided, which includes:

[0007] User consultation question collection module, used to obtain user consultation questions;

[0008] A knowledge graph related information node matching module is used to match related information nodes from a predetermined knowledge graph based on the user's inquiry question to obtain a set of related information nodes;

[0009] A node semantic embedding coding module, used for performing semantic embedding coding on each relevant information node in the set of relevant information nodes to obtain a set of semantic embedding coding features of the relevant information nodes;

[0010] A node semantic dynamic aggregation module is used to perform semantic feature distribution dynamic aggregation processing on the set of semantic embedding coding features of the relevant information nodes to obtain semantically significant representations of the relevant information nodes;

[0011] A consultation question semantic embedding coding module, used for performing semantic embedding coding on the user consultation question to obtain the user consultation question semantic embedding coding features;

[0012] The large model answer generation module is used to embed the user consultation question semantics into a coding vector and concatenate the relevant information node semantic saliency representation vector as a prompt semantics, and then use a large model-based question and answer generator to process the prompt semantics to obtain the user consultation answer.

[0013] Compared with the prior art, the present application provides a data analysis platform based on a large model and a knowledge graph, which quickly locates the information nodes where entities and relationships related to user questions are located through the knowledge graph, and introduces data processing and semantic understanding algorithms based on artificial intelligence and natural language processing technology at the back end to semantically splice the semantics of these matched multiple related information nodes and the semantics of user consultation questions into prompt information, and inputs this prompt information into the large model to generate the user's consultation answer. In this way, the knowledge graph and large model technology can be comprehensively utilized to automatically generate matching consultation answers based on the user's consultation question information, thereby improving the intelligence level of the data analysis platform. In this way, the user's question background can be better understood, and high-quality answers can be given to provide users with a better service experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] By describing the embodiments of the present application in more detail in conjunction with the accompanying drawings, the above and other purposes, features and advantages of the present application will become more apparent. The accompanying drawings are used to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the accompanying drawings, the same reference numerals generally represent the same components or steps.

[0015] Figure 1 A block diagram of a data analysis platform based on a large model and a knowledge graph according to an embodiment of the present application;

[0016] Figure 2 A schematic diagram of data flow of a data analysis platform based on a large model and a knowledge graph according to an embodiment of the present application;

[0017] Figure 3 It is a block diagram of a node semantic dynamic aggregation module in a data analysis platform based on a large model and a knowledge graph according to an embodiment of the present application. DETAILED DESCRIPTION

[0018] Below, the exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the exemplary embodiments described here.

[0019] As shown in this application and claims, unless the context clearly indicates an exception, the words "a", "an", "an" and / or "the" do not refer to the singular and may also include the plural. Generally speaking, the terms "include" and "comprise" only indicate the inclusion of the steps and elements that have been clearly identified, and these steps and elements do not constitute an exclusive list. The method or device may also include other steps or elements.

[0020] Although the present application makes various references to certain modules in the system according to the embodiments of the present application, any number of different modules can be used and run on the user terminal and / or server. The modules are only illustrative, and different aspects of the system and method can use different modules.

[0021] Flowcharts are used in the present application to illustrate the operations performed by the system according to the embodiments of the present application. It should be understood that the preceding or following operations are not necessarily performed accurately in order. On the contrary, various steps may be processed in reverse order or simultaneously as required. Meanwhile, other operations may also be added to these processes, or a certain step or several steps of operations may be removed from these processes.

[0022] Below, the exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the exemplary embodiments described here.

[0023] With the development of large language models and deep learning technology, machines can understand and generate human language more accurately. This is mainly due to the powerful feature extraction ability and self-learning mechanism of neural networks. Through a large amount of training data, the model can automatically learn complex patterns in the text, thereby improving the performance of tasks such as text classification, sentiment analysis, and machine translation. On the other hand, knowledge graphs, as a form of structured data storage, can clearly represent entities and their attributes, as well as the relationships between entities, providing a new perspective for information retrieval and recommendation systems. Knowledge graphs not only contain static factual information, but can also discover implicit knowledge through reasoning mechanisms, which is of great significance for improving the accuracy of information retrieval matching and the relevance of answer result generation.

[0024] However, using large-model deep learning or knowledge graphs alone has its limitations. Although large language models and deep learning techniques can capture complex semantic relationships, they may produce misleading results in the absence of clear prior knowledge; and although knowledge graphs can provide clear representations of semantic relationships, they are expensive to build and maintain when faced with large-scale unstructured data, and are also difficult to capture complex semantic expression information.

[0025] Based on this, in the technical solution of this application, a data analysis platform based on a big model and a knowledge graph is proposed, which can combine the advantages of the big model and the knowledge graph to build a more powerful intelligent question-answering system.

[0026] In the technical solution of this application, a data analysis platform based on a large model and a knowledge graph is proposed. Figure 1 It is a block diagram of a data analysis platform based on a big model and a knowledge graph according to an embodiment of the present application. Figure 2 Schematic diagram of data flow of a data analysis platform based on a large model and knowledge graph according to an embodiment of the present application. Figure 1 and Figure 2As shown, according to the embodiment of the present application, the data analysis platform 300 based on the big model and the knowledge graph includes: a user consultation question collection module 310, which is used to obtain user consultation questions; a knowledge graph related information node matching module 320, which is used to match the relevant information nodes from the predetermined knowledge graph based on the user consultation questions to obtain a set of relevant information nodes; a node semantic embedding coding module 330, which is used to perform semantic embedding coding on each relevant information node in the set of relevant information nodes to obtain a set of semantic embedding coding features of relevant information nodes; a node semantic dynamic aggregation module 340, which is used to perform semantic feature distribution dynamic aggregation processing on the set of semantic embedding coding features of the relevant information nodes to obtain semantically significant representations of relevant information nodes; a consultation question semantic embedding coding module 350, which is used to perform semantic embedding coding on the user consultation questions to obtain semantic embedding coding features of user consultation questions; a big model answer generation module 360, which is used to concatenate the user consultation question semantic embedding coding vector and the relevant information node semantically significant representation vector as prompt semantics, and then use a big model-based question and answer generator to process the prompt semantics to obtain a user consultation answer.

[0027] In particular, the user consultation question collection module 310 is used to obtain user consultation questions. It should be understood that due to the different habits and questioning methods of each user, the same question may obtain different analysis semantics, and the user consultation questions may involve professional domain knowledge. Traditional user consultation semantic understanding and answer generation solutions are difficult to apply these professional knowledge to generate high-quality answers that meet user consultation needs, affecting the user experience.

[0028] In particular, the knowledge graph related information node matching module 320 is used to match the related information nodes from the predetermined knowledge graph based on the user consultation question to obtain a set of related information nodes. That is, in the technical solution of the present application, in order to ensure that the system can accurately understand the user's question, and use the answer generator to generate answers that meet the user's consultation needs, and improve user satisfaction, in the technical solution of the present application, further based on the user consultation question, the relevant information nodes are matched from the predetermined knowledge graph to obtain a set of related information nodes. In this way, the constructed domain-specific knowledge graph can be used for searching to find the information nodes most relevant to the user's question and the connections between them, so that the answers generated subsequently can be ensured to be highly relevant. In particular, since the knowledge graph not only contains the information of the entity itself, but also the relationship between the entities. By matching the user's question based on the knowledge graph, the deep meaning behind the question can be excavated, so as to better understand the user's intention.

[0029] In particular, the node semantic embedding coding module 330 is used to perform semantic embedding coding on each relevant information node in the set of relevant information nodes to obtain a set of semantic embedding coding features of relevant information nodes. In a specific example of the present application, semantic embedding coding is performed on each relevant information node in the set of relevant information nodes to obtain a set of semantic embedding coding vectors of relevant information nodes as a set of semantic embedding coding features of relevant information nodes. It should be understood that semantic embedding coding is a technology that converts symbolic information (such as text, sentences, entities, etc.) into numerical vectors. In this way, the nodes in the knowledge graph can be converted into numerical representations so that the machine learning model can understand and process this information. Not only that, through the processing of semantic embedding coding, the semantic features on each relevant information node can also be learned, and this semantic feature can reflect the semantic relationship between words or entities. If two entities are semantically similar, their semantic embedding coding vectors of relevant information nodes will be close to each other in high-dimensional space. This helps to identify and strengthen the semantic expressions closely related to user questions in subsequent processing, and provides a basis for the generation of subsequent user consultation answers.

[0030] In particular, the node semantic dynamic aggregation module 340 is used to perform semantic feature distribution dynamic aggregation processing on the set of semantic embedding coding features of the relevant information nodes to obtain semantically significant representations of the relevant information nodes. It should be understood that each of the relevant information node semantic embedding coding vectors in the set of relevant information node semantic embedding coding vectors respectively contains the semantic features of each relevant information node related to the semantics of the user consultation question matched from the predetermined knowledge graph, and each relevant information node semantic feature is a semantic particle. When participating in the subsequent semantic feature aggregation process to obtain a more sufficient and accurate semantic aggregation representation of the relevant information node, these different relevant information node semantic particles will have different importance and contribution to the subsequent user consultation answer generation task. Based on this, in the technical solution of the present application, the set of semantic embedding coding features of the relevant information nodes is further subjected to semantic feature distribution dynamic aggregation processing to obtain semantically significant representations of the relevant information nodes. In particular, here, the process of semantic feature distribution dynamic aggregation processing is based on the interaction between the semantic particles of each relevant information node through a concept similar to a force field to model. This force field not only takes into account the spatial relationship between the semantics of each relevant information node (such as Euclidean distance), but also combines the properties of each relevant information node semantics itself (such as static energy factor) and the overall semantic structure of the entire relevant information node semantic set. In this way, the importance of the semantic features of each relevant information node can be dynamically adjusted during the semantic aggregation process, and the semantics of each relevant information node can be integrated into a more compact and representative relevant information node semantic aggregation representation according to a specific aggregation strategy, thereby obtaining a relevant information node semantic salient representation vector. Specifically, in a specific example of the present application, Figure 3 As shown, the node semantic dynamic aggregation module 340 includes: a static energy factor calculation unit 341, which is used to calculate the static energy factor of each relevant information node semantic embedding coding feature in the set of relevant information node semantic embedding coding features to obtain a set of relevant information node semantic static energy factors; a relevant information node semantic aggregation representation unit 342, which is used to perform feature significant aggregation based on the imitation field on the set of relevant information node semantic embedding coding features based on the set of relevant information node semantic static energy factors to obtain the relevant information node semantic significant representation.

[0031] Specifically, the static energy factor calculation unit 341 is used to calculate the static energy factor of each relevant information node semantic embedding coding feature in the set of relevant information node semantic embedding coding features to obtain a set of relevant information node semantic static energy factors. That is, for each relevant information node semantic embedding coding feature in the set of relevant information node semantic embedding coding features, its stability or importance in the feature space is evaluated based on a predefined metric function, and a numerical representation - static energy factor is assigned, so that the position of each relevant information node semantic embedding coding feature in the data distribution in which it is located can be quantified, providing a basis for the subsequent selection of the initial clustering center. Specifically, the specific process of calculating the static energy factor of each relevant information node semantic embedding coding feature in the set of relevant information node semantic embedding coding features includes: calculating the mean and variance of the relevant information node semantic embedding coding vector to obtain the relevant information node semantic mean and the relevant information node semantic variance; subtracting the relevant information node semantic embedding coding vector from the relevant information node semantic mean by position, and calculating the fourth power of each position of the subtracted feature vector to obtain the relevant information node semantic difference modulation vector; calculating the expected value of the relevant information node semantic difference modulation vector to obtain the relevant information node semantic expected value; dividing the relevant information node semantic expected value by the square of the relevant information node semantic variance, and inputting the divided value into the sigmoid function to obtain the relevant information node semantic static energy factor.

[0032] Specifically, the related information node semantic aggregation representation unit 342 is used to perform feature significant aggregation based on the imitation force field on the set of related information node semantic embedded coding features based on the set of related information node semantic static energy factors to obtain the related information node semantic significant representation. That is, first, the related information node semantic embedded coding vector corresponding to the largest related information node semantic static energy factor in the set of related information node semantic static energy factors is selected as the related information node semantic clustering initial center vector. Then, based on the spatial span between each related information node semantic embedded coding vector in the set of related information node semantic embedded coding vectors and the related information node semantic clustering initial center vector and the static energy factor of each related information node semantic embedded coding vector and the static energy factor of the related information node semantic clustering initial center vector, the dynamic aggregation energy factor of each related information node semantic embedded coding vector in the set of related information node semantic embedded coding vectors is calculated to obtain the set of related information node semantic dynamic aggregation energy factors; that is, based on the spatial span of each related information node semantic embedded coding feature vector relative to the related information node semantic clustering initial center vector, and combined with their respective static energy factors, the dynamic aggregation energy factor of each related information node semantic embedded coding feature vector is further derived. That is, by considering the feature space proximity of the semantic embedding coding feature vectors of each relevant information node relative to the core of the semantic feature distribution cluster of the relevant information node and the difference in the importance of the intrinsic attributes, its contribution to the semantically significant features of the relevant information node can be more accurately measured. Next, the set of the semantic dynamic aggregation energy factors of the relevant information nodes is input into the gated mask unit to obtain the set of semantic dynamic aggregation weight factors of the relevant information nodes; and the set of the semantic dynamic aggregation energy factors of the relevant information nodes is input into the gated mask unit to obtain the set of semantic dynamic aggregation weight factors of the relevant information nodes. Here, based on the gated mask mechanism, the obtained dynamic aggregation energy factors are nonlinearly transformed and screened to generate corresponding weights, so as to perform weighted aggregation on the set of semantic embedding coding vectors of the relevant information nodes, thereby enhancing the expression of important features and suppressing the influence of irrelevant or noise features, so as to obtain the semantically significant representation vector of the relevant information nodes.

[0033] Among them, based on the spatial span between each relevant information node semantic embedding coding vector in the set of relevant information node semantic embedding coding vectors and the initial center vector of the relevant information node semantic clustering, as well as the static energy factor of each relevant information node semantic embedding coding vector and the static energy factor of the initial center vector of the relevant information node semantic clustering, the specific process of calculating the dynamic aggregation energy factor of each relevant information node semantic embedding coding vector in the set of relevant information node semantic embedding coding vectors includes: multiplying the static energy factor of the relevant information node semantic embedding coding vector and the static energy factor of the initial center vector of the relevant information node semantic clustering with a first weighting parameter to obtain a first relevant information node semantic dynamic aggregation energy factor; multiplying the square of the spatial span between the relevant information node semantic embedding coding vector and the initial center vector of the relevant information node semantic clustering with a second weighting parameter to obtain a second relevant information node semantic dynamic aggregation energy factor; dividing the first relevant information node semantic dynamic aggregation energy factor by the second relevant information node semantic dynamic aggregation energy factor to obtain the relevant information node semantic dynamic aggregation energy factor.

[0034] More specifically, the specific process of inputting the set of related information node semantic dynamic aggregation energy factors into the gated mask unit includes: normalizing the set of related information node semantic dynamic aggregation energy factors to obtain a set of normalized related information node semantic dynamic aggregation energy factors; masking the set of normalized related information node semantic dynamic aggregation energy factors to obtain a set of related information node semantic dynamic aggregation weight factors. Wherein, normalizing the set of related information node semantic dynamic aggregation energy factors to obtain a set of normalized related information node semantic dynamic aggregation energy factors includes: calculating an exponential function value with a natural constant e as the base and the negative of each related information node semantic dynamic aggregation energy factor in the set of related information node semantic dynamic aggregation energy factors as the exponent to obtain a set of related information node semantic dynamic aggregation energy index factors; adding the set of related information node semantic dynamic aggregation energy index factors to the constant one by position, and then calculating the inverse of each related information node semantic dynamic aggregation energy index modulation factor in the set of related information node semantic dynamic aggregation energy index modulation factors to obtain the set of normalized related information node semantic dynamic aggregation energy factors. Among them, the set of normalized related information node semantic dynamic aggregation energy factors is masked to obtain the set of related information node semantic dynamic aggregation weight factors, including: in response to each normalized related information node semantic dynamic aggregation energy factor in the set of normalized related information node semantic dynamic aggregation energy factors being greater than a predetermined threshold, the normalized related information node semantic dynamic aggregation energy factors greater than the predetermined threshold are set to the original value, and the rest are set to zero, so as to obtain the set of related information node semantic dynamic aggregation weight factors.

[0035] In summary, in the above embodiment, the set of semantic embedding coding features of the relevant information nodes is subjected to dynamic aggregation processing of semantic feature distribution to obtain semantically significant representations of the relevant information nodes, including: performing dynamic aggregation processing of semantic feature distribution on the set of semantic embedding coding features of the relevant information nodes using the following semantic feature aggregation formula to obtain semantically significant representations of the relevant information nodes; wherein the semantic feature aggregation formula is:

[0036] X={x1,x2,...,x i ,...,x n}

[0037]

[0038] x c =x m

[0039]

[0040] w si =maak(w i )

[0041]

[0042] Where X is the set of semantic embedding coding vectors of the relevant information nodes, x1, x2, ..., x i ,...,x n are respectively the 1st, 2nd, ..., ith, ..., nth relevant information node semantic embedding coding vectors in the set of relevant information node semantic embedding coding vectors, is the eigenvalue of each position in the semantic embedding encoding vector of the i-th related information node, E(·) is the expected value, μ i and σ i 2 x i The mean and variance of , sigmoid(·) is the sigmoid function, is x i The corresponding related information node semantic static energy factor, arg max j (·) is the j value corresponding to the maximum value, m is the maximum matching value, x c is the initial center vector of semantic clustering of relevant information nodes, is x m The corresponding related information node semantic static energy factor, Count(x i →x m ) represents x i With x m The spatial span between them, γ and δ are weighting parameters, For x i The corresponding semantic dynamic aggregation energy factor of the relevant information node, w i is x i The corresponding normalized information node semantic dynamic aggregation energy factor, mask(·) is the masking process, θ is the predetermined threshold, w si is x i The corresponding dynamic aggregation weight factor of the semantics of the relevant information node, n is the number of vectors in the set of the semantic embedding encoding vectors of the relevant information node, x f is the semantically significant representation vector of the relevant information node.

[0043] In particular, the consultation question semantic embedding coding module 350 is used to perform semantic embedding coding on the user consultation question to obtain the user consultation question semantic embedding coding features. That is, in the technical solution of the present application, the user consultation question is semantically embedded to obtain the user consultation question semantic embedding coding vector as the user consultation question semantic embedding coding feature. Here, the user consultation question can be embedded and encoded by a transformer-based semantic embedding encoder.

[0044] In particular, the large model answer generation module 360 ​​is used to splice the user consultation question semantic embedding coding vector and the relevant information node semantic saliency representation vector as prompt semantics, and then use the large model-based question and answer generator to process the prompt semantics to obtain the user consultation answer. That is, in the technical solution of the present application, first, the user consultation question semantic embedding coding vector and the relevant information node semantic saliency representation vector are spliced ​​into a prompt vector as the prompt semantics; by splicing the semantic vector of the user question and the semantic saliency representation vector of the relevant information node, the semantic information of the user question and the semantic aggregation information of the relevant information node extracted from the knowledge graph can be integrated, thereby creating a complete context representation containing user intention semantics, professional background semantics and potential answer candidates. This helps the subsequent model to more comprehensively understand the background of the question and the possible answer direction, so as to ensure that the generated answer not only directly responds to the user's question, but also takes into account other important information related to the question, so that a richer and more detailed consultation answer can be generated. Then, the prompt semantics is input into the large model-based question and answer generator to obtain the user consultation answer. That is, the splicing information between the semantics of the user's question and the aggregation semantics of the nodes related to the knowledge graph is used as the prompt semantics, and the prompt semantics is input into the question-answer generator based on the big model, so as to use the big model technology to generate a big consultation answer that meets the user's consultation intention. In this way, the knowledge graph and big model technology can be used in combination to automatically generate matching consultation answers based on the user's consultation question information, thereby improving the intelligence level of the data analysis platform. In this way, the user's question background can be better understood, and high-quality answers can be given to provide users with a better service experience.

[0045] Preferably, the semantic embedding coding vector of the user consultation question and the semantic saliency representation vector of the related information node respectively represent the embedded coding semantic features of the user consultation question and the semantic force field distribution field aggregation features of the embedded coding semantics of the set of related information nodes, that is, the differences in the distribution of embedded coding semantic features under different semantics and the differences and correspondences in the distribution of force field features will cause different semantic aggregation modes during vector splicing, causing the prompt vector to have a complex semantic aggregation spatial structure. Therefore, when the prompt vector is input into a question and answer generator based on a large model to obtain user consultation answers, it is expected to improve the generative regression convergence and generalization effect of the boosted vector under the complex semantic aggregation spatial structure.

[0046] Therefore, when the applicant of the present application inputs the prompt vector into the question-answer generator based on the large model to obtain the user's inquiry answer, the prompt vector is optimized, and the optimization process includes:

[0047] The sum of the absolute values ​​of the eigenvalues ​​of the prompt vector is calculated to obtain a first prompt vector space structure value, and the square root of the sum of the squares of the eigenvalues ​​of the prompt vector is calculated to obtain a second prompt vector space structure value:

[0048]

[0049] where v i ∈V∈R 1×L , v i are the eigenvalues ​​of the prompt vector, α is the spatial structure value of the first prompt vector, and β is the spatial structure value of the second prompt vector;

[0050] multiplying each eigenvalue of the hint vector by the first hint vector space structure value and the second hint vector space structure value respectively to obtain a first hint vector structure reference value and a second hint vector structure reference value corresponding to each eigenvalue;

[0051] multiplying each eigenvalue of the hint vector by the length of the hint vector and the square root of the length respectively to obtain a first hint vector scale transformation value and a second hint vector scale transformation value corresponding to each eigenvalue;

[0052] Dividing the first hint vector structure reference value by the difference between the first hint vector space structure value and the first hint vector scale transformation value to obtain a first hint vector transformation adjustment value;

[0053] Dividing the second hint vector structure reference value by the difference between the second hint vector space structure value and the second hint vector scale transformation value to obtain a second hint vector transformation adjustment value; and

[0054] A weighted sum of the first hint vector transformation adjustment value and the second hint vector transformation adjustment value is calculated to obtain each eigenvalue of the optimized hint vector.

[0055] Here, the optimized hint vector V' is expressed as:

[0056] V'=V1⊕(ω⊙V2)

[0057] In v 1i ∈V1∈R 1×L , and v 2i ∈V2∈R 1×L In the case of:

[0058]

[0059] v 2i =(α×v i ) / (α-L×v i )

[0060] Among them, v 1i is the first hint vector transformation adjustment value, v 2i is the second prompt vector transformation adjustment value, L is the length of the prompt vector, V1 is the first prompt vector transformation adjustment vector composed of the first prompt vector transformation adjustment value, V2 is the second prompt vector transformation adjustment vector composed of the second prompt vector transformation adjustment value, ω is a weighted hyperparameter, ⊕ is positional addition, ⊙ is positional dot multiplication, and V' is the optimized prompt vector.

[0061] That is, for the spatial structure information of the feature set of the prompt vector in the high-dimensional space, the scale-based frame transformation of each eigenvalue of the prompt vector is performed by using the class norm space structured representation of the prompt vector as a reference window, and the spatial structure-based frame attention weight adjustment of each eigenvalue of the prompt vector is implemented to ensure the spatial transformation invariance of the prompt vector under the interaction of the feature space, thereby improving the convergence and generalization effect of the generative regression of the feature set of the prompt vector under the complex semantic space structure representation, and improving the semantic quality of the obtained user consultation answer. In this way, the knowledge graph and large model technology can be comprehensively utilized to automatically generate matching consultation answers based on the user's consultation question information, thereby improving the intelligence level of the data analysis platform and providing users with a better service experience.

[0062] As described above, the data analysis platform 300 based on the big model and knowledge graph according to the embodiment of the present application can be implemented in various wireless terminals, such as a server with a data analysis algorithm based on the big model and knowledge graph. In one possible implementation, the data analysis platform 300 based on the big model and knowledge graph according to the embodiment of the present application can be integrated into the wireless terminal as a software module and / or a hardware module. For example, the data analysis platform 300 based on the big model and knowledge graph can be a software module in the operating system of the wireless terminal, or it can be an application developed for the wireless terminal; of course, the data analysis platform 300 based on the big model and knowledge graph can also be one of the many hardware modules of the wireless terminal.

[0063] Alternatively, in another example, the data analysis platform 300 based on the big model and knowledge graph and the wireless terminal may also be separate devices, and the data analysis platform 300 based on the big model and knowledge graph may be connected to the wireless terminal via a wired and / or wireless network and transmit interactive information in accordance with an agreed data format.

[0064] The embodiments of the present disclosure have been described above, and the above description is exemplary, not exhaustive, and is not limited to the disclosed embodiments. Many modifications and changes will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the described embodiments. The selection of terms used herein is intended to best explain the principles of the embodiments, practical applications, or improvements to the technology in the market, or to enable other persons of ordinary skill in the art to understand the embodiments disclosed herein.

Claims

1. A data analysis platform based on a large model and knowledge graph, characterized in that: include: User consultation question collection module, used to obtain user consultation questions; A knowledge graph related information node matching module is used to match related information nodes from a predetermined knowledge graph based on the user's inquiry question to obtain a set of related information nodes; A node semantic embedding coding module, used for performing semantic embedding coding on each relevant information node in the set of relevant information nodes to obtain a set of semantic embedding coding features of the relevant information nodes; A node semantic dynamic aggregation module is used to perform semantic feature distribution dynamic aggregation processing on the set of semantic embedding coding features of the related information nodes to obtain semantically significant representations of the related information nodes, wherein the node semantic dynamic aggregation module includes: a static energy factor calculation unit, used to calculate the static energy factor of each relevant information node semantic embedding coding feature in the set of semantic embedding coding features of the related information nodes to obtain a set of semantic static energy factors of the related information nodes; a related information node semantic aggregation representation unit, used to perform feature significant aggregation based on the imitation force field on the set of semantic embedding coding features of the related information nodes based on the set of semantic static energy factors of the related information nodes to obtain semantically significant representations of the related information nodes; A consultation question semantic embedding coding module, used for performing semantic embedding coding on the user consultation question to obtain the user consultation question semantic embedding coding features; The large model answer generation module is used to embed the user consultation question semantics into a coding vector and concatenate the relevant information node semantic saliency representation vector as a prompt semantics, and then use a large model-based question and answer generator to process the prompt semantics to obtain the user consultation answer.

2. The data analysis platform based on large models and knowledge graphs according to claim 1, characterized in that: The node semantic embedding coding module is used to: perform semantic embedding coding on each relevant information node in the set of relevant information nodes to obtain a set of relevant information node semantic embedding coding vectors as a set of relevant information node semantic embedding coding features.

3. The data analysis platform based on large models and knowledge graphs according to claim 2 is characterized in that: The static energy factor calculation unit comprises: A related information node semantics calculation subunit, used for calculating the mean and variance of the related information node semantics embedding coding vector to obtain the related information node semantics mean and the related information node semantics variance; The related information node semantic difference modulation subunit is used to subtract the related information node semantic embedding coding vector from the related information node semantic mean by position, and calculate the fourth power of each position of the subtracted feature vector to obtain the related information node semantic difference modulation vector; A related information node semantic expected value calculation subunit, used to calculate the expected value of the semantic difference modulation vector of the related information node to obtain the related information node semantic expected value; The semantic static energy factor calculation subunit is used to divide the semantic expected value of the relevant information node by the square of the semantic variance of the relevant information node, and input the value obtained by the division into a sigmoid function to obtain the semantic static energy factor of the relevant information node.

4. The data analysis platform based on large models and knowledge graphs according to claim 3 is characterized in that: The relevant information node semantic aggregation representation unit includes: A clustering initial center vector selection subunit is used to select the relevant information node semantic embedding coding vector corresponding to the largest relevant information node semantic static energy factor in the set of relevant information node semantic static energy factors as the relevant information node semantic clustering initial center vector; A related information node semantic dynamic aggregation energy factor calculation subunit is used to calculate the dynamic aggregation energy factor of each related information node semantic embedding coding vector in the set of related information node semantic embedding coding vectors based on the spatial span between each related information node semantic embedding coding vector in the set of related information node semantic embedding coding vectors and the initial center vector of the related information node semantic clustering, as well as the static energy factor of each related information node semantic embedding coding vector and the static energy factor of the initial center vector of the related information node semantic clustering, so as to obtain a set of related information node semantic dynamic aggregation energy factors; A gated mask subunit, used for inputting the set of relevant information node semantic dynamic aggregation energy factors into a gated mask unit to obtain a set of relevant information node semantic dynamic aggregation weight factors; The related information node semantic salient representation subunit is used to calculate the weighted sum of the set of related information node semantic embedding coding vectors using the set of related information node semantic dynamic aggregation weight factors as weighted weights to obtain the related information node semantic salient representation vector as the related information node semantic salient representation.

5. The data analysis platform based on large models and knowledge graphs according to claim 4 is characterized in that: The relevant information node semantic dynamic aggregation energy factor calculation subunit is used to: Multiplying the static energy factor of the semantic embedding coding vector of the relevant information node and the static energy factor of the initial center vector of the semantic clustering of the relevant information node with a first weighting parameter to obtain a first relevant information node semantic dynamic aggregation energy factor; Multiplying the square of the spatial span between the relevant information node semantic embedding coding vector and the relevant information node semantic clustering initial center vector by a second weighting parameter to obtain a second relevant information node semantic dynamic aggregation energy factor; The first relevant information node semantic dynamic aggregation energy factor is divided by the second relevant information node semantic dynamic aggregation energy factor to obtain the relevant information node semantic dynamic aggregation energy factor.

6. The data analysis platform based on large models and knowledge graphs according to claim 5, characterized in that: The gated mask subunit comprises: A normalization secondary subunit, used for normalizing the set of semantic dynamic aggregation energy factors of the relevant information nodes to obtain a set of normalized semantic dynamic aggregation energy factors of the relevant information nodes; The masked secondary subunit is used to perform masking processing on the set of normalized semantic dynamic aggregation energy factors of the relevant information nodes to obtain the set of semantic dynamic aggregation weight factors of the relevant information nodes.

7. The data analysis platform based on large models and knowledge graphs according to claim 6, characterized in that: The normalized secondary subunit is used for: Calculate an exponential function value with the natural constant e as the base and the negative number of each relevant information node semantic dynamic aggregation energy factor in the set of relevant information node semantic dynamic aggregation energy factors as the exponent to obtain a set of relevant information node semantic dynamic aggregation energy exponential factors; After adding the set of related information node semantic dynamic aggregation energy index factors and constant one by position, the inverse of the semantic dynamic aggregation energy index modulation factor of each related information node in the set of related information node semantic dynamic aggregation energy index modulation factors is calculated to obtain the set of normalized related information node semantic dynamic aggregation energy factors.

8. The data analysis platform based on large models and knowledge graphs according to claim 7, characterized in that: The masked secondary subunit is used for: In response to each normalized relevant information node semantic dynamic aggregation energy factor in the set of normalized relevant information node semantic dynamic aggregation energy factors being greater than a predetermined threshold, the normalized relevant information node semantic dynamic aggregation energy factors greater than the predetermined threshold are set to the original value, and the rest are set to zero to obtain the set of relevant information node semantic dynamic aggregation weight factors.

9. The data analysis platform based on large models and knowledge graphs according to claim 8, characterized in that: The consultation question semantic embedding coding module is used to: perform semantic embedding coding on the user consultation question to obtain a user consultation question semantic embedding coding vector as the user consultation question semantic embedding coding feature.

10. The data analysis platform based on large models and knowledge graphs according to claim 9, characterized in that: The large model answer generation module is used to: The user consultation question semantic embedding coding vector and the relevant information node semantic saliency representation vector are concatenated into a prompt vector as the prompt semantics; The prompt semantics is input into the large model-based question and answer generator to obtain the user inquiry answer.

Images