Knowledge graph data intelligent retrieval method and system based on large language model

Abstract

The invention relates to the technical field of data analysis, in particular to a knowledge graph data intelligent retrieval method and system based on a large language model. Modeling is performed on emotion scores of public opinion comments, skewness, thick tail and multimodal characteristics are captured, comment data are divided into a positive category and a negative category, two types of emotion distribution are quantified, and event conflict factors not only reflect the opinion degree of positive and negative emotions in public opinions, but also provide a quantitative basis for fusion of external dynamic environment information; event conflict factors, event semantic dispersion, propagation structure change rate and context switching frequency are jointly combined into dynamic environment factors, and then the dynamic environment factors are converted into adjustment vectors with the same dimension as entity embedding, so that final query and entity matching can dynamically respond to external public opinion changes; retrieval and question and answer deviations caused by the fact that emotional polarization and abnormal emotions cannot be correctly captured in a traditional method are relieved, and the accuracy and robustness of the intelligent retrieval method during real-time public opinion information processing are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of data analysis, and particularly relates to an intelligent knowledge graph data retrieval method and system based on a large language model. Background Art

[0002] In current public opinion monitoring and intelligent question-and-answer systems, many methods mainly rely on text embeddings generated by large language models and semantic retrieval based on knowledge graphs to handle diverse information queries.

[0003] When facing a large amount of unstructured data on social media, the current methods often have the following problems: First, social media comment data has strong diversity and noise, and the emotional expressions often exhibit skewed, heavy-tailed, and multi-modal distribution characteristics. It is often difficult to accurately capture outliers and emotional polarization phenomena using traditional Gaussian mixture models or static threshold-based methods, resulting in inaccurate emotion classification. Second, existing methods lack effective measurement means for the opposing or conflicting information of emotions in public opinion comments, resulting in biases in judging the controversy or polarization degree of events. This will affect the response ability of the entire retrieval system to the dynamic public opinion environment, and thus reduce the accuracy and practicality of retrieval results. Summary of the Invention

[0004] Aiming at the above-mentioned drawbacks of the existing technology, the present invention provides an intelligent knowledge graph data retrieval method and system based on a large language model, which can effectively solve the problem in the existing technology that it is difficult to improve the accuracy of retrieval results by combining the actual emotional state of public opinion information.

[0005] To achieve the above object, the present invention is realized through the following technical solutions: The present invention provides an intelligent knowledge graph data retrieval method based on a large language model, which at least includes: Constructing a knowledge graph to generate semantic embedding vectors for the input query and entity nodes in the knowledge graph; Integrating time-dependent environmental features into entity representations so that the matching between the query and entity nodes dynamically reflects external environment and event changes, including: Constructing time-dependent environmental factors based on the change rate of event semantic dispersion, the change rate of event propagation structure, the context switching frequency, and the event conflict factor. Constructing the event conflict factor includes: Introducing a mixture distribution model to capture the positive and negative emotion distributions of emotions in public opinion comments, then: Establishing an emotional score set, based on the skewed, heavy-tailed, and multi-modal characteristics in the emotional scores, assuming that the emotional scores are mixed by multiple mixture distribution models: , represents the sentiment score The overall probability density function of the variable, represents the mixing weight of the th component, represents the distribution probability density function, based on the degrees of freedom of the component , scale parameter and mean to obtain; Calculate the posterior probability of the component based on the distribution probability density function, and determine whether the public opinion comment is a positive emotion distribution or a negative emotion distribution through the posterior probability; Map the environmental factor to an environmental embedding vector, and construct a dynamic similarity in combination with the semantic embedding vector to respond to external environment and event changes;

[0006] Furthermore, define the relational expression of the distribution probability density function as: represents the Gamma function, represents the normalization factor.

[0007] Furthermore, use the expectation maximization algorithm to estimate the parameters of the mixture distribution model, and the parameters include degrees of freedom, scale parameters and means, and iteratively optimize to obtain the optimal parameters.

[0008] Furthermore, the method for determining whether the public opinion comment is a positive emotion distribution or a negative emotion distribution through the posterior probability is: For the sentiment score of each public opinion comment and each component , calculate the posterior probability : represents the probability density function of the sentiment score , , and obtain the posterior probabilities of the two components; For each public opinion comment , according to its posterior probability judge whether it is classified as a positive emotion class or a negative emotion class; Combine kernel density estimation to obtain the positive emotion probability distribution function and the negative emotion probability distribution function.

[0009] Furthermore, the method for constructing the dynamic similarity in combination with the semantic embedding vector is: Define a mapping function to map environmental factors to environmental embedding vectors ; Define the dynamic similarity between a query and an entity node through cosine similarity .

[0010] Furthermore, the relational expression for calculating the multi-hop inference score is as follows: where represents the attenuation coefficient of the th hop, represents the adjusted matching degree of the user query at time with the th entity in the path , represents the number of hops of the path, represents the multi-hop inference score.

[0011] Furthermore, the construction method of the context switching frequency is as follows: Annotate sentiment labels for public opinion data, including positive, negative, and neutral; Expand the keyword set represents the similarity threshold, represents the initial keyword set, represents a keyword, represents calculating the semantic similarity between the keyword and the public opinion data; Identify the potential topics of public opinion data and analyze the context switching frequency based on topic changes : represents the probability distribution of the public opinion data belonging to the topic ; represents all public opinion data sets at time , and respectively represent and the probability distribution of each public opinion data in the public opinion data set at time represents the KL divergence.

[0012] Furthermore, the change rate of the event semantic divergence is calculated according to the following relational expression: represents the time window length, represents at time The semantic embedding vector, represents at the moment the semantic embedding vector.

[0013] A knowledge graph data intelligent retrieval system, characterized in that the system is implemented according to the knowledge graph data intelligent retrieval method based on a large language model described in any one of the above.

[0014] A computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the above method are implemented.

[0015] The technical solution provided by the present invention has the following beneficial effects compared with the known prior art: By using a mixture distribution model to model the sentiment scores of public opinion comments, more accurately capture skewness, heavy tails, and multi-modal characteristics, automatically classify the comment data into two categories: positive and negative, and quantify the difference between the two emotion distributions. Furthermore, the event conflict factor not only reflects the degree of opposition between positive and negative emotions in public opinion, but also provides a quantitative basis for the integration of external dynamic environment information; By combining sentiment analysis, similarity calculation, topics, and KL divergence to achieve precise tracking of topic and sentiment changes in public opinion data, and precisely define the context switching frequency; By combining the event conflict factor with the event semantic dispersion, propagation structure change rate, and context switching frequency into the dynamic environment factor, and then converting it into an adjustment vector with the same dimension as the entity embedding, the final query and entity matching can dynamically respond to external public opinion changes, alleviating the retrieval and question-and-answer biases caused by the traditional method's inability to correctly capture emotional polarization and abnormal emotions, and improving the accuracy, robustness, and adaptability of the intelligent retrieval method when processing real-time public opinion information. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention, and those of ordinary skill in the art can obtain other drawings without creative efforts based on these drawings.

[0017] Figure 1 It is a schematic diagram of the overall method of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0019] The present invention will be further described below in conjunction with embodiments.

[0020] Embodiment 1 (refer to Figure 1 ): An intelligent knowledge graph data retrieval method based on a large language model, at least including the following steps: Real-time capture of public opinion data from social media platforms (such as Douyin, Weibo, WeChat official accounts, etc.), including: text (news, comments, posts), pictures, videos, and related interaction data (likes, comments, forwards), and preprocess the public opinion information data, such as data cleaning through methods like denoising and standardization.

[0021] Extract event-related entities (such as events, people, locations, times, etc.) and their semantic relationships from the preprocessed public opinion data using LLM and NLP methods, then there are: Construct a knowledge graph , where represents all the extracted entity nodes, represents the edges between entities, and each edge is configured with weights to describe data credibility and relationship tightness; For the query input by the user and each entity node in the knowledge graph, generate embedding vectors respectively using the large language model: represents the semantic embedding vector obtained after the query is transformed by the large language model, and this vector captures the semantic information in the query, represents the semantic embedding vector obtained after the entity node is transformed by the large language model, and this vector represents the semantic content of the entity node , represents the process of LLM extracting embeddings; Integrate time-dependent environmental features into the entity representation so that the matching between the query and the entity node can dynamically reflect external environment and event changes, then there are: Define a time-dependent environmental factor : where, Represents the change rate of event semantic dispersion, which is used to measure the change of semantic embedding vectors over time within a time window. Represents the change rate of event propagation structure. Represents the event conflict factor. Represents the context switching frequency. Represents a fusion function that integrates multiple dynamic environment features into an overall environmental factor, such as weighting multiple dynamic environment features to output the environmental factor. .

[0022] Furthermore, the dynamic environment features involved in the above scheme are solved respectively, including: Change rate of event semantic dispersion Represents the length of the time window, describing how many past time steps, which can be minutes, hours or days, determined according to the scenario. Represents at the moment The semantic embedding vector, generated from the description or text of the public opinion event, reflecting the semantic state of the event at that moment. Represents at the moment The semantic embedding vector. Therefore, this change rate of event semantic dispersion reflects the "mutation" or "evolution" degree of the event topic or description content over time, capturing whether the event theme has changed significantly. If the change is large, it indicates that the semantic characteristics of the event have been updated. Change rate of event propagation structure Represents the function for extracting structural features from the propagation graph The structural features are such as the diameter of the graph (the maximum shortest path length between any two points in the propagation graph), the average shortest path (the average of the shortest path lengths between all node pairs), etc. , represents the propagation graph constructed at the moment containing the entity nodes extracted at that time and the relationships between them (edges). Therefore, this change rate of event propagation structure can quantify the change of the information diffusion pattern in the public opinion event over time. For example, when an event suddenly attracts a large amount of attention, it may lead to an increase in the connectivity of the propagation graph, a decrease in the diameter, or a decrease in the distance between nodes, and vice versa. By monitoring the dynamic change of the change rate of event propagation structure, the emergence, development or decline of public opinion hotspots can be captured in a timely manner, providing assistance for subsequent dynamic adjustment of the environment. In the public opinion data of social media, there are dynamic changes in emotions and topics, especially in informal language (such as slang, pinyin, emoticons, etc.) and complex emotional expressions (such as sarcasm, puns), sentiment analysis often faces accuracy problems. At the same time, social media topics and emotions change very quickly, often accompanied by emergencies or user interactions, resulting in frequent and complex switching of emotions and topics. In addition, emotions and topics in social media are often multi-layered and intertwined. Emotional fluctuations may directly affect topic changes, and vice versa. Therefore, sentiment analysis, topics, and KL divergence are combined to achieve context switching frequency. The construction steps are as follows: Label public opinion data with sentiment labels, including positive, negative, and neutral; Dynamically expand keyword sets based on sentiment tags and similarities , Represents the similarity threshold, which is used to filter out keywords with high similarity to the current public opinion data, helping to better describe the core topics in public opinion. represents the initial keyword set, Indicates keywords, Indicates the calculation of the semantic similarity between keywords and public opinion data (through cosine similarity); Use the LDA (Latent Dirichlet Allocation) model to identify potential topics in public opinion data and analyze the frequency of context switching based on topic changes : Represents public opinion data Belong to the topic The probability distribution of , describing the correlation between the two; Indicates at time All public opinion data sets under and Respectively and The probability distribution of each public opinion data and topic in the public opinion data set at the moment, It represents KL divergence, measures the difference between two probability distributions, quantifies the distribution change of each topic in the public opinion data at two time points, and reflects the drastic degree of topic change in public opinion.

[0023] Event conflict factor JSD stands for the symmetric Kullback-Leibler Divergence (KL divergence, measuring the difference between two probability distributions), which is used to measure the difference between two probability distributions. The value range is usually between 0 and 1. Indicates time Internal Mix The probability distribution function of positive emotions obtained by the distribution model division (describing the probability distribution of comments with an emotional score of at time in the positive emotion category), represents the probability distribution function of negative emotions obtained by the mixture distribution model within time (describing the probability distribution of comments with an emotional score of at time in the negative emotion category). The larger the JSD value, the higher the separation degree between the two, reflecting the more obvious emotional opposition and polarization in the comments, thus constituting the event conflict factor.

[0024] It should be noted that when constructing the event conflict factor above, considering the skewness, heavy tails, and multimodal phenomena in the emotional score data, the mixture distribution model is introduced to capture the degree of emotional opposition or polarization in public opinion comments, that is, the positive emotion distribution and the negative emotion distribution, to determine the event conflict factor. The specific steps are as follows: Within a preset time window, extract all public opinion comments from social media comments, and use an emotion analysis model (such as based on BERT or other pre-trained models) to calculate the continuous emotional score set of each public opinion comment represents the emotional score of the th public opinion comment, and its value usually ranges from -1 to 1. A higher value indicates more positive, and a lower value indicates more negative; Assume that the emotional score is a mixture of two mixture distribution models, where one distribution corresponds to positive emotions (positive sentiment), and the other corresponds to negative emotions (negative sentiment). Then there is:

[0025] represents the overall probability density function of the emotional score variable , represents the mixing weight of the th component, ; Similarly, represents the distribution probability density function with degrees of freedom of . For the th component, the expression is: Among them, represents the mean of the th component, describing the central tendency of emotions. The component with a higher mean represents positive emotions, while the component with a lower mean represents negative emotions. Represents the scale parameter of the th component, reflecting the dispersion degree of the sentiment score, Represents the th component's degree of freedom, controlling the tail thickness of the distribution, Represents the Gamma function, Represents the normalization factor; It should be noted that, A smaller allows for more extreme comments and is suitable for radical expressions in social media. The larger is, the closer it approaches the normal distribution. In this embodiment, the emotions of social media comments often do not have a symmetric distribution. For example, in the initial stage of an event, it triggers anger and the emotions are mostly concentrated in the negative range. However, in the later stage, with the spread of rumors and deepening of understanding, there are some positive tendencies, and the of the mixture distribution model can more accurately fit the central tendency of emotions under skewed conditions. Moreover, in public opinion events, extreme emotions such as "boycott" and "an eye for an eye" are common, and intense expressions will deviate significantly from the mainstream emotions. The degree of freedom can tolerate these extreme emotion values (thick tails), is insensitive to anomalies, and thus provides a more stable classification basis. Secondly, because multiple groups of people may have completely different emotions towards the same event, the constructed mixture

[0026] Among them, the expectation maximization (EM) algorithm is used to estimate the parameters of the mixture distribution model, and the optimal parameters (i.e., the updated , etc.) are obtained through iterative optimization, including: E-step: Calculate the posterior probability that each belongs to each component. At the same time, an auxiliary variable is introduced for updating in the M-step, then there is: Represents the square of the standardized deviation of the component; For the sentiment score of each public opinion comment and each component , calculate the posterior probability : M-step: According to the posterior probability and auxiliary variables (reflecting the contribution degree to the component ), update the parameters, then there is: represents the updated mixing weight, represents the total number of samples, that is, the number of all public opinion comments within a preset time window; represents the updated mean value; , by taking the square root of to obtain the updated ; Use numerical methods to solve this non-linear equation , then the updated represents the Digamma function; Thus, based on the above update formula, after completing the E-step and M-step, decide whether to stop the iteration according to the preset convergence criterion (such as the parameter change amplitude is lower than the preset threshold or the maximum number of iterations is reached), or return to the E-step to continue the next round of update. Furthermore, through the EM algorithm, parameter estimation is performed on the mixture distribution model to obtain the accurate parameters of each emotion component, which is convenient for accurately calculating the event conflict factor.

[0027] Because , the parameters of component 1 and component 2 after update can be obtained, that is, the posterior probabilities of the two components are obtained in the same way , for each public opinion comment , if there exists , then classify this public opinion comment as a positive emotion category, belonging to the set , otherwise it is a negative emotion category, belonging to the set ; Thus, perform kernel density estimation (KDE) on the sentiment scores of all public opinion comments in the set to obtain a smooth positive emotion probability distribution function ; Similarly, obtain the negative emotion probability distribution function , represents the kernel density estimation function.

[0028] Furthermore, in order to enable the environmental factor to directly affect the embedding representation, define the mapping function , map to a vector with the same dimension as the entity embedding, that is, the environmental embedding vector , then there is: A multi-layer neural network for environmental mapping tasks, consisting of a series of fully connected layers and non-linear activation functions (such as ReLU, Sigmoid, etc.). Denote The set of parameters, including the weight matrix of each layer And the bias vector , through non-linear transformation Can capture The non-linear and complex semantic information in, and "embed" it into a vector space consistent with the entity representation, thus facilitating subsequent vector addition and similarity calculation. Denote A real vector space of dimension, and add them to get the total embedding representation , which better reflects the real-time changes of the external environment and events.

[0029] In this way, the dynamic similarity between the query and the entity node can be defined using the entity embedding adjusted with the environment (A metric that measures the matching degree between the user query And the entity node in the knowledge graph At time :) Denote the semantic embedding vector of the query (converted by the large language model from the query ), And Denote the Euclidean norm (or modulus) of the respective vectors, which is used for normalization to ensure that the scale of the final similarity value is not affected by the vector length.

[0030] In summary, the dynamic similarity measures the matching degree between the user And the entity embedding adjusted with the environment Based on the cosine similarity method, which can capture the directional similarity between the two in the high-dimensional semantic space. Usually, the higher the value, the better the semantic matching. More importantly, by introducing Integrate the real-time external environment information directly into the entity representation, and can dynamically update the entity semantics during the emergence of public opinion hotspots and breaking news, so as to ensure that the query matching not only depends on the static text description, but also reflects the current environmental changes, improving the response accuracy of the user query and the real-time performance of the system.

[0031] Thus, the indirect relationship between the query and the event can be captured through multi-hop reasoning, including: Select a path containing multiple entity nodes from the knowledge graph One possible path can be: the topic involved in the query → event → related person → location, etc.; Obtain multi-hop reasoning scores based on dynamic similarity :

[0032] Among them, represents the attenuation coefficient of the th hop, represents at time , the user query and the th entity in the path after environment adjustment of the matching degree, represents the number of hops of the path. It should be noted that multi-hop reasoning can discover the indirect relationship established between the query and the event through intermediate nodes, so as to understand the information in the public opinion event; Furthermore, based on the multi-hop reasoning score combined with the data credibility (or other data information, such as context information, etc. can also be used), the final retrieval answer can be generated through the large language model , for example: Concatenate information such as multi-hop reasoning scores, data credibility, and context matching degree into a vector X, then obtain a hidden state through a feed-forward neural network, and then combine the hidden state with the prompt of the large language model to generate the final answer text that conforms to the current public opinion and query background. The specific implementation of this solution will not be elaborated here.

[0033] The present invention also provides a knowledge graph data intelligent retrieval system based on a large language model. This system is implemented according to the above method and will not be elaborated here.

[0034] A computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the above method are implemented.

[0035] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements will not cause the essence of the corresponding technical solutions to deviate from the protection scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for intelligent retrieval of knowledge graph data based on a large language model, comprising the following steps: Build a knowledge graph and generate semantic embedding vectors for the input query and entity nodes in the knowledge graph, which are characterized by: Incorporate time-dependent environmental features into entity representation to dynamically reflect changes in the external environment and events through matching between queries and entity nodes, including: The time-dependent environmental factors are constructed based on the change rate of event semantic discreteness, the change rate of event propagation structure, the frequency of context switching and event conflict factors. The event conflict factors include: Introducing Mix The distribution model captures the positive and negative sentiment distributions of public opinion comments, and we have: Establish a sentiment score set. Based on the skewness, fat tail and multi-peak characteristics in the sentiment score, assume that the sentiment score consists of multiple mixed Distribution Model Mixture: , Indicates sentiment score The overall probability density function of the variable, Indicates The mixing weights of the components, represents the probability density function of the distribution, according to the degrees of freedom of the components , scale parameters and mean get; The posterior probability of the component is calculated based on the distribution probability density function, and the posterior probability is used to determine whether the public opinion comment is a positive emotion distribution or a negative emotion distribution; Map environmental factors into environmental embedding vectors and combine them with semantic embedding vectors to construct dynamic similarity to respond to changes in external environment and events; The multi-hop reasoning score is calculated based on dynamic similarity and multi-hop reasoning, and the retrieval answer is output in combination with a large language model.

2. According to claim 1, a knowledge graph data intelligent retrieval method based on a large language model is characterized in that: The relationship defining the distribution probability density function is: represents the Gamma function, represents the normalization factor.

3. According to claim 2, a knowledge graph data intelligent retrieval method based on a large language model is characterized in that: Using the expectation maximization algorithm to The parameters of the distribution model, including degrees of freedom, scale parameter and mean, are estimated and iteratively optimized to obtain the optimal parameters.

4. According to claim 3, a knowledge graph data intelligent retrieval method based on a large language model is characterized in that: The method for judging whether a public opinion comment is a positive emotion distribution or a negative emotion distribution by using posterior probability is: Sentiment score for each public opinion comment And each component , calculate the posterior probability : Indicates sentiment score The probability density function of , and obtain the posterior probability of the two components ; For each public opinion comment , according to its posterior probability Classify the emotion into positive or negative categories; Combined with kernel density estimation, we obtain the probability distribution function of positive emotion and the probability distribution function of negative emotion.

5. According to claim 1, a knowledge graph data intelligent retrieval method based on a large language model is characterized in that: The method of combining semantic embedding vectors to construct dynamic similarity is: Define a mapping function to transform the environmental factors Mapping to environment embedding vector ; Defining the dynamic similarity between query and entity nodes through cosine similarity .

6. The method for intelligent retrieval of knowledge graph data based on a large language model according to claim 5, characterized in that: The relationship for calculating the multi-hop reasoning score is: in, Indicates The attenuation coefficient of the jump, Indicates at time Next, user query With the path Entities The environmentally adjusted match, Indicates the number of hops in the path. Represents the multi-hop reasoning score.

7. The method for intelligent retrieval of knowledge graph data based on a large language model according to claim 1, characterized in that: The context switching frequency The construction method is: Label public opinion data with sentiment labels, including positive, negative, and neutral; Expand keyword set represents the similarity threshold, represents the initial keyword set, Indicates keywords, Indicates the calculation of the semantic similarity between keywords and public opinion data; Identify potential topics in public opinion data and analyze the frequency of context switching based on topic changes : Represents public opinion data Belong to the topic The probability distribution of Indicates at time All public opinion data sets under and Respectively and The probability distribution of each public opinion data and topic in the public opinion data set at the moment, represents the KL divergence.

8. The method for intelligent retrieval of knowledge graph data based on a large language model according to claim 1, characterized in that: The change rate of the semantic discreteness of the event The calculation is performed according to the following relationship: represents the time window length, Indicates at time The semantic embedding vector of Indicates at time The semantic embedding vector of .

9. The knowledge graph data intelligent retrieval system is characterized by: The system is implemented according to the intelligent retrieval method for knowledge graph data based on a large language model as described in any one of claims 1-8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 8 are implemented.

Images