A Retrieval Enhancement Method and System Based on Structural Entropy Hierarchical Knowledge Tree

Abstract

This application discloses a retrieval enhancement method and system based on a structural entropy hierarchical knowledge tree. By combining the hierarchical clustering HCSE algorithm with RAG, a multi-level community structure with layer-by-layer refined information is constructed to improve the accuracy and generation efficiency of answers. The method specifically includes: using a large language model (LLM) to construct graphs for text datasets and image datasets to obtain corresponding graph structures, and using a hierarchical clustering HCSE algorithm based on structural entropy to directly generate a multi-level community structure from the obtained graph structure and construct a hierarchical knowledge tree; using a large language model (LLM) to analyze the node degree and community density of each community in the hierarchical knowledge tree to generate a corresponding community summary and sort it according to information priority; combining the community summary in the hierarchical knowledge tree, hierarchical retrieval is performed in the order of top-level community, middle-level community, and bottom-level community to generate local answers, and all local answers are integrated through Map-Reduce to generate the final answer.

Description

Technical Field

[0001] This application relates to the field of electrical data processing technology, and in particular to a retrieval enhancement method and system based on a hierarchical knowledge tree with structural entropy. Background Technology

[0002] In the field of modern medical technology, medical image processing technology helps doctors better understand patients' physiological states, facilitating the timely detection and diagnosis of various diseases, and possessing significant clinical value and practical significance. This is particularly prominent in the early diagnosis of cancer and patient consultation. Doctors and patients often face the challenge of accurately determining the stage and severity of cancer, which may lead to missing the optimal treatment window and seriously endangering the patient's life and health. Therefore, improving the quality of medical information extraction and generation is of paramount importance.

[0003] Currently, the Retrieval-Augmented Generation (RAG) method combines information retrieval and natural language generation to extract information from large-scale document sets and generate answers. It can quickly extract disease-related medical information to help doctors and patients accurately determine their conditions.

[0004] However, traditional RAG methods have significant limitations when faced with cross-topic, global problems. Especially when cross-domain summaries are required, existing methods struggle to generate high-quality answers effectively, leading to inefficient information integration and knowledge extraction. Summary of the Invention

[0005] To overcome the aforementioned technical deficiencies, this application provides a retrieval enhancement method and system based on a hierarchical knowledge tree with structural entropy.

[0006] In a first aspect, this application provides a retrieval enhancement method based on a hierarchical knowledge tree with structural entropy, comprising the following steps:

[0007] S1. Data Acquisition and Preprocessing:

[0008] With esophageal cancer as the focus of data collection, we collected raw text data and raw medical image data related to esophageal cancer, and preprocessed the collected data to obtain text datasets and image datasets.

[0009] S2. Hierarchical Knowledge Tree Construction:

[0010] We used the Large Language Model (LLM) to construct graphs from text and image datasets to obtain corresponding graph structures. Then, we used the Hierarchical Clustering (HCSE) algorithm based on structural entropy to directly generate multi-level community structures from the obtained graph structures and construct hierarchical knowledge trees.

[0011] S3, Community Summary Generation:

[0012] The Large Language Model (LLM) is used to analyze the node degree and community density of each community in the hierarchical knowledge tree, generate corresponding community summaries, and sort them according to information priority in order to quickly obtain important data.

[0013] S4. Query Processing and Answer Generation:

[0014] When a query request for esophageal cancer diagnosis is received, the system combines the community summaries in the hierarchical knowledge tree and performs a hierarchical search in the order of top-level community - middle-level community - bottom-level community to generate local answers. The Map-Reduce method is then used to integrate all the local answers to generate the final answer.

[0015] Optionally, in one possible implementation of the first aspect,

[0016] The graph structure built using the text dataset can be a document graph, in which the extracted entities are used as nodes, the relationships between entities are used as edges, and the covariates related to the diagnosis of esophageal cancer are used as node information for the entities.

[0017] The entities include symptoms, patient characteristics, and clinical trial information; the covariates include statements, attributes, and time spans. Statements describe specific statements or opinions related to the entity, attributes represent specific characteristics of the entity, and time spans represent the time during which the patient received treatment or the disease progressed.

[0018] Optionally, in one possible implementation of the first aspect,

[0019] The graph structure constructed using image data can be an interaction graph, which is constructed by jointly maximizing the structural entropy and minimizing the relative entropy between the generated weights and the original weight distribution. The generated weights are the weights of the interaction graph generated from the complete graph of the image, while the original weights are the weights of the interaction graph used in practice.

[0020] Optionally, in one possible implementation of the first aspect,

[0021] The multi-layered community structure includes a top-level community, an intermediate-level community, and a bottom-level community, with the intermediate-level community having at least two layers.

[0022] The top-level community represents the main topic in the field of esophageal cancer, generating a community with document-level granularity; the middle-level community is a sub-community with entity-level granularity generated by refining and recursively dividing the top-level community layer by layer using the HCSE algorithm; and the bottom-level community is a sub-community with feature-level granularity generated from the middle-level community using the HCSE algorithm.

[0023] Optionally, in one possible implementation of the first aspect,

[0024] A hierarchical knowledge tree includes a root node, intermediate nodes, and leaf nodes;

[0025] The root node represents the entire collection of medical documents, symbolizing the overall research field of esophageal cancer, and includes all relevant documents and information; the intermediate nodes are subdivisions of the topics under the root node, dividing them into various important topics and sub-topics; the leaf nodes correspond to specific documents, cases or related studies in the underlying community, providing detailed information and data.

[0026] Optionally, in one possible implementation of the first aspect,

[0027] The community summary is a general overview of the esophageal cancer diagnosis-related information contained in each community. The node degree is the number of edges connected to the node, and the community density is the strength of the relationship between entities. The information priority is positively correlated with the node degree and the community density.

[0028] Optionally, in one possible implementation of the first aspect,

[0029] The Map-Reduce method consists of two phases: Map and Reduce.

[0030] The MapReduce algorithm is used to integrate all the local answers, including:

[0031] In the Map phase, each local answer is scored for relevance and help, and a comprehensive score is obtained. The relevance score is used to evaluate the overlap between the keywords in the local answer and the query keywords, and the help score is used to evaluate the effectiveness in solving user problems or providing useful information.

[0032] In the Reduce phase, the optimal local answer is selected as the final answer based on the overall score.

[0033] Optionally, in one possible implementation of the first aspect,

[0034] The raw text data includes: electronic medical records (EMR), laboratory test results, medical literature, pathology reports, and patients' clinical symptoms and treatment records;

[0035] The raw medical imaging data includes CT images, MRI images, and endoscopic images.

[0036] Optionally, in one possible implementation of the first aspect,

[0037] Data preprocessing includes processing raw text data and raw image data;

[0038] The processing of raw text data specifically includes: cleaning the raw text data, and standardizing and normalizing the cleaned data to obtain a standardized and normalized text dataset;

[0039] The processing of the original image data specifically includes: adjusting the size, orientation, brightness and contrast of the original image data and adding noise to enhance the diversity of the image data; using correlation coefficients for feature selection; and using the selected features to combine features to obtain new features, thereby forming an image dataset.

[0040] The correlation coefficient is specifically a coefficient that characterizes the correlation between a patient's physiological characteristics and the incidence of esophageal cancer; the new feature obtained by combining features is a feature that can better reflect the patient's overall health status compared with the selected features.

[0041] Secondly, this application provides a retrieval enhancement system based on a hierarchical knowledge tree with structural entropy, characterized in that the retrieval enhancement system is used to execute the retrieval enhancement method described in any of the first aspects and their implementations above.

[0042] The technical solutions employed in this application embodiment achieve the following technical effects: Based on the Large Language Model (LLM) combined with the Hierarchical Clustering (HCSE) algorithm and MapReduce-style answer generation, it effectively handles global and local queries in large-scale datasets; through hierarchical knowledge trees, it can quickly locate relevant topics and extract key information, clearly organizing and displaying information, enabling users to quickly find the answers they need in complex information environments. Furthermore, during queries, users receive high-quality natural language answers when searching for information at different levels, whether it's a broad overview or specific details. The hierarchical knowledge tree design makes information organization and retrieval more efficient, improving the user experience. Through the effective integration of medical literature and clinical cases, the system can provide timely and accurate support for doctors and researchers. Attached Figure Description

[0043] The accompanying drawings exemplify embodiments and form part of the specification, serving together with the textual description to explain exemplary implementations of the embodiments. The illustrated embodiments are for illustrative purposes only and do not limit the scope of the claims. Throughout the drawings, the same reference numerals refer to similar but not necessarily identical elements.

[0044] Figure 1 This is a flowchart illustrating the retrieval enhancement method based on structural entropy hierarchical knowledge tree in this application;

[0045] Figure 2 This is a schematic diagram of the system structure of the retrieval enhancement system based on structural entropy hierarchical knowledge tree in this application;

[0046] Figure 3 This is a schematic diagram of the hardware structure of the computer device in this application. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application. All other embodiments obtained by those skilled in the art based on the embodiments in this application without inventive effort are within the scope of protection of this application.

[0048] It should be noted that the descriptions involving "first," "second," etc., in the embodiments of this application are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of that feature. Furthermore, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. If the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed in this application.

[0049] In the description of this application, it should be understood that the numerical labels before the steps do not indicate the order of the steps, but are only used to facilitate the description of this application and to distinguish each step, and therefore should not be construed as a limitation of this application.

[0050] First, a definition of the terminology used in this application is provided:

[0051] 1. Enhanced search function to generate RAG

[0052] RAG is a natural language processing technique that combines retrieval and generation. Its core idea is to first retrieve relevant information and then use a generative model to generate answers, thereby improving the quality and accuracy of text generation.

[0053] Specifically, RAG first retrieves information related to the input query from a large knowledge base or document collection. Given a query q, it retrieves relevant documents D from the knowledge base K using the retrieval model R.

[0054] D = R(q, K)

[0055] Where D is the set of retrieved documents, and R is the retrieval algorithm.

[0056] Using a search algorithm, the most relevant document fragments to the user's question are found. Next, the retrieved information will be combined with the user's input for use by a generative model (such as GPT, etc.) to generate a richer and context-related answer. The advantage of RAG is that it can dynamically retrieve the latest information and does not solely rely on the knowledge learned by the model during training. This method helps improve the accuracy and relevance of the generated content as it can utilize external information sources to supplement the model's knowledge. Additionally, RAG can generate diverse answers to adapt to different context requirements, thus performing well in areas such as question-answering systems, dialogue generation, and information summarization. This method effectively addresses the limitations that traditional generative models may encounter when dealing with specific knowledge.

[0057] 2. Structural Entropy

[0058] Structural entropy is a concept used to measure the complexity and information content of a system or structure, and is typically applied in fields such as network science and information theory. Intuitively, the structural entropy method encodes a tree structure by characterizing the uncertainty of the hierarchical topology. The structural entropy of a graph G is defined as the total minimum number of bits required to determine the encoding words of the nodes in G. Structural entropy has achieved success in fields such as information retrieval, traffic prediction, and reinforcement learning. By minimizing the structural entropy of a given graph G, the hierarchical clustering results of the vertices in G are retained by the associated encoding tree.

[0059] Encoding Tree: Let G = (V, E, W) be an undirected weighted graph, where V = {v1,..., v n} is the set of vertices, E is the set of edges, and W ∈ R n*n is the edge weight matrix.

[0060] The encoding tree T of graph G is defined as a hierarchical rooted tree as follows:

[0061] For each tree node α ∈ T, a vertex subset T α ∈ V is associated with it.

[0062] 1. The root node λ of the tree is associated with the vertex set V, i.e., T γ = V.

[0063] 2. For each α ∈ T, its direct successors are labeled as α^, increasing by i from left to right, and the direct predecessor is labeled as α - .

[0064] 3. For each α ∈ T with L direct successors, the vertex subset T α ^ They are disjoint and

[0065] 4. For each leaf node v∈T, Y v It contains only one vertex of V.

[0066] K-dimensional structural entropy: Given an arbitrary rooted coding tree T of a graph G, the structural entropy of G on T measures the remaining complexity of G after being reduced by T. For each non-root node α∈T, its assigned structural entropy is defined as:

[0067]

[0068] Where g α It is the tangent point, that is, at T. α Neutralization is not in T α The weights of the edges between nodes in the equation and V α and V C It is volume, i.e., T α The sum of the node degrees in G.

[0069] The structural entropy of G given T is defined as:

[0070]

[0071] To meet the requirements of downstream applications, the K-dimensional structural entropy of G is defined as:

[0072]

[0073] T takes values ​​in all coding trees whose height does not exceed K.

[0074] Structural entropy is introduced to quantify the internal information and organization of complex structures. It can be viewed as a measure of the relationships between the various components of a system.

[0075] 3. Hierarchical clustering algorithm

[0076] Hierarchical clustering algorithms are methods for organizing data into a hierarchical structure, often used to explore relationships between data. They are divided into two types: agglomerative (bottom-up) and divisive (top-down). Agglomerative clustering starts with each data point, treating each point as an independent cluster. It calculates the distance between all clusters, progressively merging the two closest clusters until all data points are merged into one cluster or a predetermined number of clusters is reached. The distance calculation during the merging process can be represented as:

[0077]

[0078] Where C i and C j Let a and b represent two clusters to be merged, where a and b represent cluster C. i and C j In the context of data points, d(a, b) represents the distance between data points a and b, typically measured using metrics such as Euclidean distance or Manhattan distance.

[0079] Split clusters, on the other hand, start with a single cluster and gradually divide it into smaller sub-clusters. The decision to split can be based on the internal compactness or external separation of the clusters.

[0080] Assume the internal distance of cluster C is D. in (C) and the external distance is D out (C) can be expressed by the following formula as the criterion for splitting:

[0081]

[0082] The main steps of the algorithm include calculating the distance matrix, initializing clusters, merging clusters, and generating a dendrogram. The distance matrix uses an appropriate distance metric to calculate the distance between each pair of data points. During the merging process, common merging strategies include single-link (shortest distance), full-link (longest distance), and average distance. The advantages of hierarchical clustering lie in its visualization capabilities and flexibility, allowing users to select the appropriate number of clusters based on the dendrogram. Hierarchical clustering is widely used in fields such as bioinformatics and text clustering, helping users gain a deeper understanding of the structure and relationships within data.

[0083] This application combines the hierarchical clustering algorithm HCSE with RAG to effectively construct multi-level community structures. This method refines information layer by layer from global to local, significantly improving the accuracy and efficiency of answer generation. The hierarchical clustering algorithm not only identifies potential relationships between data but also extracts key information at different levels, thus providing doctors with more accurate and comprehensive diagnostic support.

[0084] This innovative approach allows doctors to better understand patients' conditions, quickly access the latest cancer-related research findings and treatment options, and make more accurate judgments. This intelligent information processing method not only improves the efficiency of medical services but also provides more reliable assurance for early diagnosis and treatment, ultimately helping to improve patients' prognosis and quality of life.

[0085] The retrieval enhancement generation method based on structural entropy hierarchical knowledge tree in this application embodiment is described below with reference to the accompanying drawings:

[0086] Example 1

[0087] Taking the field of esophageal cancer medical research as an example, this application describes the retrieval enhancement generation method based on structural entropy hierarchical knowledge tree, specifically including:

[0088] S1. Data Acquisition and Preprocessing:

[0089] Using esophageal cancer as the focus of data collection, we collected raw text data and raw medical image data related to esophageal cancer, and preprocessed the collected data to obtain text datasets and image datasets.

[0090] Specifically, raw text data includes: electronic medical records (EMR), laboratory test results, medical literature, pathology reports, and patients' clinical symptoms and treatment records. Raw medical imaging data includes CT images, MRI images, and endoscopic images.

[0091] The above data preprocessing includes processing the raw text data and raw image data, specifically:

[0092] The processing of raw text data specifically includes: cleaning the raw text data, and standardizing and normalizing the cleaned data to obtain a standardized and normalized text dataset;

[0093] The processing of the original image data specifically includes: adjusting the size, orientation, brightness and contrast of the original image data and adding noise to enhance the diversity of the image data; using correlation coefficients for feature selection; and using the selected features to combine features to obtain new features, thereby forming an image dataset.

[0094] Specifically, the aforementioned correlation coefficient can be a coefficient characterizing the correlation between a patient's physiological characteristics and the incidence of esophageal cancer; the aforementioned new features can be features that better reflect the patient's overall health status compared to the selected features.

[0095] S2. Hierarchical Knowledge Tree Construction:

[0096] We used the Large Language Model (LLM) to construct graphs from text and image datasets to obtain corresponding graph structures. Then, we used the Hierarchical Clustering Sequencing (HCSE) algorithm based on structural entropy to directly generate multi-level community structures from the obtained graph structures and construct hierarchical knowledge trees.

[0097] The graph structure constructed using the text dataset can be a document graph. In the document graph, extracted entities are used as nodes, relationships between entities are used as edges, and covariates related to esophageal cancer diagnosis are used as node information for the entities. Specifically, entities include symptoms, patient characteristics, and clinical trial information; covariates include claims, attributes, and time spans, where claims describe specific statements or opinions related to the entity, attributes represent specific characteristics of the entity, and the time span represents the time the patient has received treatment or the disease has progressed.

[0098] Similarly, a graph structure built using image data can be an interaction graph. An interaction graph is constructed by jointly maximizing structural entropy and minimizing the relative entropy between the generated weights and the original weight distribution, where the generated weights are the weights used in the interaction graph from the complete graph of the image, and the original weights are the weights of the interaction graph used in practice.

[0099] The aforementioned multi-layered community structure includes a top-level community, a middle-level community, and a bottom-level community. The middle-level community has at least two layers. The top-level community represents the main topics in the field of esophageal cancer, generating a community with document-level granularity, such as the etiology, clinical symptoms, diagnostic methods, and treatment strategies of esophageal cancer. The middle-level community is a sub-community with entity-level granularity generated by refining and recursively dividing the top-level community layer by layer using the HCSE algorithm. The bottom-level community is a sub-community with feature-level granularity generated from the middle-level community using the HCSE algorithm.

[0100] The hierarchical knowledge tree described above includes a root node, intermediate nodes, and leaf nodes. The root node represents the entire collection of medical documents, symbolizing the overall research field of esophageal cancer and containing all relevant documents and information. The intermediate nodes are subdivisions of the topics under the root node, dividing them into various important topics and sub-topics. The leaf nodes correspond to specific documents, cases, or related studies in the underlying community, providing detailed information and data.

[0101] S3, Community Summary Generation:

[0102] By using the Large Language Model (LLM), the node degree and community density of each community in the hierarchical knowledge tree are analyzed to generate corresponding community summaries and sort them according to information priority, so as to quickly obtain important data.

[0103] Specifically, the community summary is a general overview of the esophageal cancer diagnosis-related information contained in each community. The node degree is the number of edges connected to the node, and the community density is the strength of the relationship between entities. The information priority is positively correlated with the node degree and the community density.

[0104] S4. Query Processing and Answer Generation:

[0105] When a query request for esophageal cancer diagnosis is received, the system combines the community summaries in the hierarchical knowledge tree and performs a hierarchical search in the order of top-level community - middle-level community - bottom-level community to generate local answers. The Map-Reduce method is then used to integrate all the local answers to generate the final answer.

[0106] Specifically, the Map-Reduce method includes two stages: Map and Reduce. Furthermore, the integration of all local answers through the Map-Reduce method includes: 1) In the Map stage, each local answer is scored for relevance and helpness, and a comprehensive score is obtained. The relevance score is used to evaluate the overlap between the keywords in the local answer and the query keywords, and the helpness score is used to evaluate the effectiveness in solving user problems or providing useful information; 2) In the Reduce stage, the best local answer is selected as the final answer based on the comprehensive score.

[0107] In this embodiment, the system effectively handles global and local queries in large-scale datasets by combining a Large Language Model (LLM) with the Hierarchical Clustering Sequence (HCSE) algorithm and MapReduce-style answer generation. Through a hierarchical knowledge tree, relevant topics can be quickly located and key information extracted, enabling clear organization and presentation of information. This allows users to quickly find the answers they need in complex information environments. Furthermore, during queries, users receive high-quality natural language responses regardless of whether they are seeking broad domain overviews or specific details. The hierarchical knowledge tree design makes information organization and retrieval more efficient, enhancing the user experience. Through the effective integration of medical literature and clinical cases, the system provides timely and accurate support for doctors and researchers.

[0108] Example 2

[0109] The technical solution for data acquisition and preprocessing in S1 above is described in detail below:

[0110] S1.1 Data Acquisition

[0111] Data collection is a crucial step. In this model training, esophageal cancer was selected as the data collection focus. Manual review ensured the collection of high-quality, diverse data to support the model's learning and inference capabilities. First, information on esophageal cancer cases was collected through relevant websites or partner hospitals, while protecting patient privacy. This process included selecting partner hospitals and research institutions with abundant medical data to ensure the authority and reliability of the data sources. Simultaneously, existing medical databases and online platforms, such as PubMed and ClinicalTrials.gov, were utilized to obtain publicly available case and research data. Second, the collection process ensured the coverage of multiple modalities of data, including but not limited to medical imaging (such as CT, MRI, and endoscopic images), electronic medical records (EMR), laboratory test results, pathology reports, and patients' clinical symptoms and treatment records. These different types of data will provide the model with a comprehensive perspective.

[0112] S1.2 Data Preprocessing

[0113] Data preprocessing is a crucial step in ensuring model performance, encompassing several important aspects: 1. Text data cleaning, 2. Data annotation and normalization, 3. Preprocessing of medical image data, and 4. Feature engineering. The specific processes are as follows:

[0114] 1. Text data cleaning:

[0115] First, input the original text dataset (such as electronic medical records and pathology reports). Duplicate records in this dataset can lead to overfitting and bias in the model. Deduplicating the data ensures that each sample is unique, thereby improving the model's training performance. Text data often contains spelling errors, inconsistent formatting, or other inaccurate entries. Missing values ​​also significantly impact model training. Missing data can be handled by deleting missing records, imputing them with the mean or median, or using more complex methods (such as interpolation or prediction models) to ensure dataset integrity. Additionally, cleaning up special characters, numbers, and extra spaces in the text reduces data noise and improves readability.

[0116] The cleaned dataset has higher data quality and consistency. This output will serve as input for subsequent standardization, normalization, and feature engineering.

[0117] 2. Data annotation and normalization:

[0118] Standardization and normalization are essential steps when processing data with different dimensions and distributions. The system takes a cleaned text dataset as input and performs standardization and normalization processes on it.

[0119] Data standardization transforms numerical data into a distribution with a mean of 0 and a standard deviation of 1. This process helps eliminate scale differences between features, making the model more efficient and stable during training. Normalization, on the other hand, scales feature values ​​to a specific range (e.g., 0 to 1). When data has different dimensions, normalization ensures all features are on the same scale, preventing some features from having an excessive impact on the model.

[0120] By standardizing and normalizing the dataset, scale differences between features are eliminated, providing a consistent input basis for feature engineering.

[0121] 3. Preprocessing of medical image data:

[0122] For medical imaging data, preprocessing is particularly important to ensure the quality and consistency of the input images.

[0123] This step takes raw medical image data as input and aims to remove irrelevant or redundant parts of the image, focusing on key areas to improve the model's ability to recognize important features. The image is then adjusted to a size suitable for the model's input. Standardizing image dimensions avoids errors caused by inconsistent sizes. Correcting the image shooting angle ensures the model can adapt to images from different orientations, improving its recognition capabilities. Finally, enhancing the diversity of image data through technical means (such as adjusting brightness and contrast, adding noise, etc.) improves the model's generalization ability.

[0124] Preprocessed medical images have higher consistency and quality. This image data will be used for subsequent model training to improve the model's ability to recognize key features.

[0125] 4. Feature Engineering:

[0126] Feature engineering is an important step performed after data preprocessing. It aims to extract useful information from the cleaned data to improve the model's performance and predictive ability. It typically consists of two main stages: feature selection and feature combination.

[0127] During the feature selection phase, a thorough analysis of the cleaned and processed dataset (including text and medical images) is required to identify which features significantly impact model performance. For example, in an esophageal cancer prediction model, correlation coefficient analysis can be used to select features. By calculating the correlation between various physiological characteristics and cancer incidence, a significant positive correlation (high correlation coefficient) was found between frequently consuming excessively hot food and smoking and esophageal cancer; therefore, these two features were selected as model inputs.

[0128] This process involves removing redundant or irrelevant features, ensuring that the final feature set maximizes the training efficiency and effectiveness of the model. By identifying features useful for training, resources and computing power can be concentrated, training time reduced, and model accuracy improved.

[0129] Feature combination involves combining existing features or creating new ones to enhance the expressive power of the data. In this model, a patient's age and BMI (Body Mass Index) can be combined into a new feature, the "Health Risk Index." This new feature may better reflect the patient's overall health status, thereby improving the model's predictive ability for esophageal cancer risk. In this way, the model can more effectively capture the complex relationships associated with esophageal cancer, not only improving its ability to capture complex patterns but also enabling it to better adapt to diverse data distributions.

[0130] Ultimately, these two steps will output an enhanced feature set to improve the data's expressive power and the model's predictive ability. These processed datasets will be used for model training and evaluation; the text dataset will be used to construct the document graph, while the medical image dataset will be used to generate the graph structure of medical images. These feature-engineered datasets will provide a solid foundation for subsequent hierarchical knowledge tree generation, analysis, and decision-making.

[0131] Through the above steps, data preprocessing can significantly improve the training efficiency and final performance of the model, providing a solid foundation for subsequent analysis and decision-making.

[0132] Example 3

[0133] The technical solution for constructing the hierarchical knowledge tree in S2 above will be described in detail below, including: construction of document graphs, generation of graph structures for medical images, hierarchical clustering and knowledge tree generation.

[0134] 2.1 Construction of the document diagram:

[0135] First, converting the preprocessed large-scale medical document collection into a graph structure for esophageal cancer diagnosis and information organization is a crucial step. The specific steps are as follows:

[0136] 2.1.1 Entity and Relationship Extraction:

[0137] First, large language models (LLM) are used to extract various entities from documents such as medical literature, case reports, and clinical studies, including symptoms, patient characteristics, and clinical trial information, and to construct their relationships.

[0138] Suppose we extract n entities from the text, these entities can be represented as a set:

[0139] E = {e1, e2, ..., e} n }; where each entity e i It may contain information such as age, gender, and medical history.

[0140] The relationship between entities can be represented by a binary relation R:

[0141] R = {e i e j ,r)|e i e j ∈E, r∈R}; where R is a set of relations, such as diagnosis, treatment suitability, etc.

[0142] These entities will serve as nodes in the graph G, where nodes represent entities and edges represent relationships, and can be represented in the following form:

[0143] G = (V, E); where V is the set of nodes, representing all extracted entities. E is the set of edges, representing the relationships between entities.

[0144] The relationships between them are represented as weighted edges in the graph structure. The weights can be set according to the strength of the relationship, and can be specifically expressed as follows:

[0145] w(e i e i ) = f(r); where f is a function representing the strength of the relation r.

[0146] 2.1.2 Extraction of Covariates:

[0147] During the graph structure construction process, the system not only extracts entities and relationships but also delves into covariates related to esophageal cancer diagnosis. These covariates provide richer and more multidimensional information for the nodes in the graph, specifically including:

[0148] Claims: Specific statements or opinions describing an entity. A claim can be represented as a function Claim(e) to describe statements related to entity e:

[0149] Claim(e)={claim1, claim2,...};

[0150] By analyzing these statements, doctors can access the latest research findings and clinical trial data regarding treatment.

[0151] Attributes: Used to represent the specific characteristics of an entity. An attribute can be represented by a vector containing the specific characteristics of the entity.

[0152] A(e) = [a1, a2, ..., a k ];

[0153] Where a k This represents patient characteristics, such as the patient's physical condition, cancer stage, and histological type. Detailed information about these attributes helps doctors consider more comprehensive factors when making diagnostic and treatment decisions.

[0154] Time span: Represents the time a patient has received treatment or the disease has progressed. A time span can be represented by a pair of points in time, defined as:

[0155] T(e)=(t start , t end );

[0156] Where t start Indicates the start time of treatment, t end Indicates the end time.

[0157] The effectiveness of treatment can be assessed by describing changes in efficacy over a time span:

[0158] ΔE=E(t end )-E(t start );

[0159] Here, E(t) represents the therapeutic effect at time t. For example, the therapeutic effect of a certain treatment may become apparent within 3 months, and this timeframe is crucial for doctors to assess the effectiveness of treatment.

[0160] 2.1.3 Data Integration and Analysis:

[0161] By extracting and integrating these covariates, the node information in the graph structure is enriched. This not only aids in the early diagnosis and treatment selection of esophageal cancer but also facilitates the organization and analysis of information. For example, researchers can use graph structures to analyze common characteristics among different esophageal cancer patients, identify potential therapeutic targets, or explore patterns of disease progression.

[0162] 2.2 Graph Structure Generation of Medical Images:

[0163] The input to this step is preprocessed medical image data, used to construct the interactive graph G of the images. The construction of the graph structure is guided by maximizing structural entropy. Structural entropy provides a quantitative measure of graph complexity, which has been validated in graph learning systems. This step uses structural entropy here as a guide for constructing the structure of the information processing system, rather than using entropy to measure the complexity of the network data. In this framework and other graph-based image segmentation methods, the algorithm is based on the graph generated from the image; therefore, the entropy of the graph structure measures the maximum information capacity of the model. Thus, the graph G is constructed by maximizing its structural entropy:

[0164] The system is constructed by jointly maximizing structural entropy and minimizing the relative entropy between the generated weights and the original weight distribution. Simultaneously, the graph should also satisfy constraints given by the computational system, such as the size of the edge set E.

[0165]

[0166] Where d represents the distance metric, λ is a preset coefficient, W is the weight used to generate the interaction graph from the complete graph, and W′ is the weight of the interaction graph G=(V,E,W′) used in practice, satisfying:

[0167]

[0168] To construct the graph, create an edge (u, v) for pixels u and v in the image whose horizontal and vertical distances are less than r. Then, for each vertex, retain the first k edges that connect to it with the smallest distance, and remove the rest.

[0169] 2.3 Hierarchical Clustering and Knowledge Tree Generation:

[0170] In this invention, HCSE (Hierarchical Clustering Algorithm Based on Structural Entropy) is used to directly generate a multi-level community structure, and based on this, a K-layer knowledge tree is constructed, specifically for the diagnosis and information organization of esophageal cancer in medicine. Each layer of this knowledge tree represents a progressive refinement from global topics to local details in the document set, helping doctors and researchers efficiently obtain relevant information.

[0171] The specific steps are as follows:

[0172] 2.3.1 Generation of K-level clustering trees

[0173] The HCSE algorithm automatically divides a collection of medical documents (including research papers, case reports, clinical guidelines, etc.) into a multi-level community structure by minimizing the structural entropy of the graph.

[0174] For each vertex v, define parameter d. v Let be the degree of vertex v, which is the number of edges directly connected to vertex v. Let be the total degree of the community containing vertex v, which is the sum of the degrees of all vertices within that community. Let m be the total number of edges in the graph.

[0175] if Then the structural entropy H of vertex v v The calculation formula is:

[0176]

[0177] For each community C, define parameter g. c V is the number of edges connecting community C to the outside; C Let V be the sum of the vertex degrees in community C; total Let be the sum of the degrees of all community vertices.

[0178] If V C If H > 0, then the structural entropy H of community C is... c The calculation formula is:

[0179]

[0180] These communities are generated through recursive clustering, and a K-layer cluster tree structure is directly constructed, where K ranges from 3 to 10 layers.

[0181] Top-level communities (Level 1): HCSE first performs global clustering on the entire medical dataset, generating a large community layer containing global themes. Each community represents a major theme in the field of esophageal cancer, such as the etiology, clinical symptoms, diagnostic methods, and treatment strategies of esophageal cancer.

[0182] Intermediate-layer communities (layers 2 to K-1): Next, the HCSE algorithm further refines each top-level community, recursively dividing it layer by layer to generate more granular sub-communities. These intermediate-layer communities capture the transition from broad subject categories to specific sub-topics in the dataset; for example, under diagnostic methods, they can be subdivided into endoscopy, imaging, and pathological diagnosis.

[0183] Bottom-level communities (Level K): Finally, at the Level K level (not too many, no more than 10 in this model), the finest-grained communities are generated using HCSE, with each community containing the most relevant documents or entities. These communities are the most localized and suitable for handling extremely specific problems or queries.

[0184] HCSE generates communities of different granularities:

[0185] Document level: the coarsest level, aggregating relevant literature or case reports.

[0186] Entity level: Medium granularity, aggregating related entities such as diseases, symptoms, and treatments.

[0187] Feature level: The finest granularity, aggregating specific data points or documents with similar features.

[0188] 2.3.2 Knowledge Tree Construction

[0189] After generating the K-layer community structure, the system organizes these communities into a hierarchical knowledge tree. Assuming the knowledge tree has K layers, a set of layers can be defined. L k:

[0190] L k ={n∈N|depth(n)=k};

[0191] Where depth(n) represents the depth of node n.

[0192] The root node of the knowledge tree represents the entire collection of medical documents, while the leaf nodes correspond to specific entities and documents within the underlying community. This structure provides a hierarchical semantic representation from global to local, suitable for information retrieval and generation tasks of varying granularities.

[0193] Root Node: The root node symbolizes the overall research field of esophageal cancer, containing all relevant documents and information. This layer establishes a basic framework for users, enabling them to clearly understand the broad background and developments in this field.

[0194] Intermediate nodes further subdivide the topics under the root node, dividing them into various important topics and subtopics. These nodes help users quickly navigate to research areas of interest, such as from the etiology of esophageal cancer to clinical symptoms, diagnostic methods, and treatment strategies. This hierarchical structure allows users to easily find relevant information, saving search time and effort.

[0195] Leaf nodes: Leaf nodes correspond to specific documents, case studies, or related research within the underlying community, providing detailed information and data. These nodes carry the most concrete content, allowing users to find specific research findings, clinical case analyses, or the latest treatment methods. The richness and specificity of information at this layer meet users' needs for detail.

[0196] This K-layer knowledge tree is capable of flexibly adapting to different types of query needs. Users can obtain global answers at a high level, helping them understand the overall trends and important findings in esophageal cancer research. At the same time, users can also delve into leaf nodes to obtain specific detailed information, thereby supporting more in-depth research and clinical decision-making.

[0197] For example, when doctors need to understand the latest research developments in esophageal cancer, they can obtain an overview of the overall topic from the top-level community. If they need to delve deeper into the effectiveness of a specific treatment, users can explore layer by layer to ultimately obtain specific clinical research and case data.

[0198] In this way, the system maintains both the comprehensiveness of global information and the accuracy of local details when processing large-scale medical document collections. This hierarchical knowledge structure not only improves the efficiency of information retrieval but also provides important support for the diagnosis and treatment of esophageal cancer, helping doctors make more accurate decisions.

[0199] Hierarchical clustering algorithms are used to construct knowledge trees, effectively organizing multi-level information within a dataset. This structure can handle both global topics and local details, allowing users to gradually delve from broad topics to specific instances. Utilizing hierarchical clustering, the system can layer documents according to the similarity of topics and content, thereby constructing a clear knowledge system.

[0200] In the medical field, the application of knowledge trees is particularly important. Doctors and researchers can quickly find research and cases related to specific diseases, treatments, or research findings, promoting academic exchange and collaboration. This structure not only improves the efficiency of information retrieval but also helps users find the specific information they need amidst complex medical knowledge, supporting clinical decision-making and research development.

[0201] Example 4

[0202] The technical solutions summarized and generated by the community in S3 will be explained in detail below:

[0203] For each community, the system generates community summaries using a Large Language Model (LLM) to provide crucial information for the diagnosis of esophageal cancer. These summaries contain key entities within the community and their relationships, helping doctors and researchers quickly access important data. The summarization process is as follows:

[0204] S3.1, Information Priority Sorting:

[0205] The system first identifies the most important entities and relationships within a community using metrics such as node degree and community density (a measure of the ratio of the number of edges within a community to the number of possible edges). Specific steps include:

[0206] Node degree analysis: The degree d(v) of node v is defined as the number of edges connected to that node, and can be expressed as:

[0207] d(v)=∑ u∈V A vu ;

[0208] Where A is the adjacency matrix, A vu This indicates the existence of an edge between node v and node u (1 if it exists, 0 otherwise).

[0209] The system calculates the degree of each entity (i.e., the number of connections to other nodes). Entities with higher degree values ​​typically represent the most critical patients, diseases, or treatments in that community. For example, in the context of esophageal cancer, key entities such as chemotherapy drugs or clinical trials might be prioritized.

[0210] Edge weight evaluation: The system also analyzes the relationships between entities, with edges having higher weights representing stronger connections. For an edge (u, v), its weight w(u, v) can be defined to represent the strength of the relationship between entities:

[0211] w(u, v) = f(relation);

[0212] Where f is a function that determines the weights based on the properties of the relationship (such as similarity, relevance, etc.). The method for calculating the weights is as follows:

[0213]

[0214] Where C(x) is the number of times node x appears, and C(u, v) is the number of times nodes u and v appear simultaneously.

[0215] For example, the strength of the relationship between a particular treatment method and patient response may be given higher weight.

[0216] Through this prioritization, the system is able to focus on processing the most representative and influential core elements in the community.

[0217] S3.2, Generate Community Summary:

[0218] Based on the identified key entities, relationships, and their covariates, the system integrates this data to form a comprehensive background.

[0219] For example, in a community about esophageal cancer, information such as key biomarkers, common clinical symptoms, and the latest treatments might be integrated. These summaries provide a concise overview of important information within the community and can be generated across different levels, progressively refining from higher-level community summaries to lower-level specific information, ensuring users can quickly access key information.

[0220] Community summaries, combining entity, relational, and covariate information, significantly improve the accuracy and comprehensiveness of answers. The key lies in using a Large Language Model (LLM) to integrate important information within the community, thus forming a comprehensive context. This context not only reflects the interrelationships between information but also provides more detailed context by leveraging covariates, enabling users to better understand the clinical significance and research value of the information.

[0221] In this way, users can gain deeper insights, such as the effectiveness of a particular biomarker in a specific population and the results of related clinical trials. This integration of information helps physicians make more informed choices in actual diagnosis and treatment, while also driving the development of related research.

[0222] Example 5

[0223] The technical solutions for query processing and answer generation in S5 will be explained in detail below:

[0224] When a user submits a query about the diagnosis of esophageal cancer, the system combines community summaries from the knowledge tree to gradually generate an answer. This process is completed through the following steps:

[0225] S5.1 Preliminary Search

[0226] The system first performs a preliminary search within the high-level communities of the knowledge tree to identify the global topics most relevant to the user's query. For example, if a user asks about the diagnostic methods for esophageal cancer, the system will automatically search for top-level topics related to esophageal cancer, including its diagnosis and clinical assessment. The goal of this stage is to quickly locate relevant areas, laying the foundation for subsequent in-depth searches.

[0227] S5.2 Hierarchical Search

[0228] Next, the system delves deeper into different levels of the knowledge tree, performing hierarchical searches to further refine information retrieval. The system first moves from the top-level community to the middle-level communities, such as imaging examinations and endoscopy, gradually narrowing the scope. During this process, the system will focus on different examination types and methods to ensure comprehensive coverage of all possible diagnostic techniques.

[0229] The system will then refine its search, delving into the underlying community to find specific examination methods, clinical guidelines, or case studies. For example, the system might search for information on the application of CT scans in esophageal cancer diagnosis, indications and procedures for endoscopy, and the latest biomarker detection technologies. This refinement process aims to ensure the system accesses the latest clinical research findings and practical experience, thereby providing users with detailed and reliable answers.

[0230] S5.3, Local Answer Generation

[0231] Within relevant communities, the system automatically generates multiple local answers. These local answers are generated based on summary information from each community and the generation capabilities of the Large Language Model (LLM), aiming to comprehensively cover the different aspects involved in the query.

[0232] Imaging Examinations: This section summarizes the applications of CT scans and MRI in the diagnosis of esophageal cancer. It discusses in detail the basic principles and applicable scenarios of these two imaging techniques, as well as their effectiveness and limitations in identifying esophageal cancer, helping users understand the advantages and disadvantages of different imaging examinations.

[0233] Endoscopy: This section provides information on the procedures, indications, and relevant cases of esophagoscopy. It details the specific operational process of esophagoscopy, including preparation, the procedure itself, possible complications, and when the procedure is recommended. Successful clinical cases are also presented to enhance user confidence in this examination.

[0234] These partial answers provide users with comprehensive, multi-dimensional information, helping them make more informed decisions regarding the diagnosis and treatment of esophageal cancer.

[0235] S5.4, Map-Reduce integration

[0236] The system integrates local answers using MapReduce. The first phase is the Map phase, where the system performs C operations for each community. k Generate local answer A k Each local answer is based on community-summarized information S. k Generate and score each answer, with each local answer A. k It will be assigned a relevance score R. k (Calculate the overlap between keywords in the local answer and the query keywords) and help score H k (Effectiveness in solving user problems or providing useful information, such as by rating the readability of the answer's structure, language, and expression):

[0237] R k =f r (A k )H k =f h (A k );

[0238] Where f h and f r It is a scoring function used to calculate relevance and helpfulness.

[0239] Next comes the Reduce phase, in which the system calculates the overall score M. k :

[0240] M k =R k +H k ;

[0241] Filter out the optimal local answer:

[0242] A * =Select(A1,A2,...,A K );

[0243] Ultimately, these partial answers are integrated into a comprehensive global answer. For example, the most important information is extracted from the partial answers of imaging examinations, endoscopy, and biomarkers to form a comprehensive response.

[0244] The MapReduce strategy effectively handles complex queries in answer generation, ensuring the system can respond to user needs quickly and accurately. This strategy ensures coverage of all aspects of the query by generating local answers from different communities in parallel. This parallel processing not only improves efficiency but also enhances the diversity and comprehensiveness of the answers.

[0245] During the generation process, the system scores and filters individual answers, comparing their relevance and quality to ultimately integrate a high-quality natural language answer. This integration process ensures the accuracy and clarity of the answer, enabling users to easily understand the information provided and thus supporting their decision-making process.

[0246] MapReduce-style workflow example:

[0247] Suppose a user asks about early diagnostic methods for esophageal cancer. The system's processing flow is as follows:

[0248] 1. Initial Search: First, the system searches for topics related to the diagnosis of esophageal cancer in the top-level communities of the knowledge tree. The goal of this stage is to obtain extensive background information and current research dynamics in the field, helping the system understand the user's basic needs.

[0249] 2. Hierarchical Search: Next, the system enters the intermediate-level community related to early diagnosis, further subdividing into specific examination methods. During this process, the system identifies various possible early diagnostic methods, including imaging examinations, endoscopy, and biomarker testing, ensuring coverage of all areas related to early diagnosis.

[0250] 3. Local Answer Generation: Within the segmented community, the system will generate multiple local answers, which will discuss in detail different checking methods. For example:

[0251] It describes its importance in the early detection of esophageal cancer, discusses the application of CT and endoscopic ultrasound, and its potential for early diagnosis.

[0252] 4. MapReduce Integration: The system integrates all local answers, filters out the most relevant information, and generates a comprehensive and specific answer, such as: Early diagnosis methods for esophageal cancer include endoscopy, endoscopic ultrasound, and CT scans.

[0253] In this way, the system can not only respond quickly to user queries, but also ensure that the information provided is accurate and comprehensive, greatly supporting medical decisions on the diagnosis and treatment of esophageal cancer.

[0254] Example 6

[0255] like Figure 2 As shown, the retrieval enhancement system 20 based on structural entropy hierarchical knowledge tree in this application includes:

[0256] Data acquisition and processing module 201, hierarchical knowledge tree construction module 202, community summary generation module 203, and query processing and answer generation module 204;

[0257] The data acquisition and processing module 201 is used to: collect raw text data and raw medical image data related to esophageal cancer, with esophageal cancer as the direction of data collection, and preprocess the collected data to obtain text datasets and image datasets.

[0258] The hierarchical knowledge tree construction module 202 is used to: construct graphs from text datasets and image datasets using the Large Language Model (LLM) to obtain corresponding graph structures, and use the hierarchical clustering algorithm based on structural entropy (HCSE) to directly generate multi-level community structures from the obtained graph structures and construct hierarchical knowledge trees.

[0259] The community summary generation module 203 is used to: analyze the node degree and community density of each community in the hierarchical knowledge tree using the Large Language Model (LLM) to generate corresponding community summaries and sort them according to information priority in order to quickly obtain important data;

[0260] The query processing and answer generation module 204 is used to: when a query request for esophageal cancer diagnosis is received, combine the community summary in the hierarchical knowledge tree, perform hierarchical retrieval in the order of top-level community - middle-level community - bottom-level community to generate local answers, and integrate all local answers through Map-Reduce to generate the final answer.

[0261] Optionally, the above-described retrieval enhancement system is also used to execute all the steps in the retrieval enhancement method based on the structural entropy hierarchical knowledge tree described in any of the embodiments 1 to 5 above. For a detailed description, please refer to the description of the relevant parts in the above embodiments, and this application will not repeat it here.

[0262] Example 7

[0263] Figure 3 This illustration schematically depicts a hardware architecture diagram of a computer device 10000 suitable for implementing a retrieval enhancement method according to an embodiment of this application. In some embodiments, the computer device 10000 may be a terminal device such as a smartphone, wearable device, tablet computer, personal computer, in-vehicle terminal, game console, virtual device, workbench, digital assistant, set-top box, robot, etc. In other embodiments, the computer device 10000 may be a rack server, blade server, tower server, or cabinet server (including standalone servers or server clusters composed of multiple servers), etc. Figure 3 As shown, the computer device 10000 includes, but is not limited to: a memory 10010, a processor 10020, and a network interface 10030 that can communicate and be linked with each other via a system bus. Wherein:

[0264] The memory 10010 includes at least one type of computer-readable storage medium, including flash memory, hard disk, multimedia card, card-type memory (e.g., SD or DX memory), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the memory 10010 may be an internal storage module of a computer device 10000, such as the hard disk or memory of the computer device 10000. In other embodiments, the memory 10010 may also be an external storage device of the computer device 10000, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the computer device 10000. Of course, the memory 10010 may also include both the internal storage module and the external storage device of the computer device 10000. In this embodiment, the memory 10010 is typically used to store the operating system and various application software installed on the computer device 10000, such as the program code for the retrieval enhancement method. Furthermore, the memory 10010 can also be used to temporarily store various types of data that have been output or will be output.

[0265] In some embodiments, processor 10020 may be a central processing unit (CPU), controller, microcontroller, microprocessor, or other chip. Processor 10020 is typically used to control the overall operation of computer device 10000, such as performing control and processing related to data interaction or communication with computer device 10000. In this embodiment, processor 10020 is used to run program code stored in memory 10010 or process data.

[0266] Network interface 10030 may include a wireless network interface or a wired network interface, which is typically used to establish a communication link between computer device 10000 and other computer devices. For example, network interface 10030 is used to connect computer device 10000 to an external terminal via a network, establishing a data transmission channel and communication link between computer device 10000 and the external terminal. The network may be an intranet, the Internet, Global System for Mobile Communication (GSM), Wideband Code Division Multiple Access (WCDMA), 4G network, 5G network, Bluetooth, Wi-Fi, or other wireless or wired networks.

[0267] It should be pointed out that, Figure 3 Only computer devices with components 10010-10030 are shown; however, it should be understood that it is not required to implement all of the shown components, and more or fewer components may be implemented instead.

[0268] In this embodiment, the retrieval enhancement method stored in memory 10010 can also be divided into one or more program modules and executed by one or more processors (such as processor 10020) to complete the embodiment of this application.

[0269] Example 8

[0270] This application also provides a computer-readable storage medium storing a computer program thereon, wherein the computer program, when executed by a processor, implements the steps of the retrieval enhancement method in the embodiments.

[0271] In this embodiment, the computer-readable storage medium includes flash memory, hard disk, multimedia card, card-type memory (e.g., SD or DX memory), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the computer-readable storage medium may be an internal storage unit of a computer device, such as the hard disk or memory of the computer device. In other embodiments, the computer-readable storage medium may also be an external storage device of the computer device, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., provided on the computer device. Of course, the computer-readable storage medium may include both the internal storage unit and the external storage device of the computer device. In this embodiment, the computer-readable storage medium is typically used to store the operating system and various application software installed on the computer device, such as the program code of the retrieval enhancement method in the embodiment. In addition, the computer-readable storage medium can also be used to temporarily store various types of data that have been output or will be output.

[0272] Example 9

[0273] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the methods described in the above embodiments.

[0274] Obviously, those skilled in the art should understand that the modules or steps of the embodiments of this application described above can be implemented using general-purpose computer devices. They can be centralized on a single computer device or distributed across a network of multiple computer devices. Optionally, they can be implemented using computer-executable program code, thereby storing them in a storage device for execution by a computer device. In some cases, the steps shown or described can be performed in a different order than those presented here, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, the embodiments of this application are not limited to any particular combination of hardware and software.

[0275] It should be noted that the above are merely preferred embodiments of this application and do not limit the scope of patent protection of this application. Any equivalent structural or procedural changes made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of this application.

Claims

1. A retrieval enhancement method based on a hierarchical knowledge tree with structural entropy, characterized in that, Includes the following steps: S1. Data Acquisition and Preprocessing: With esophageal cancer as the focus of data collection, we collected raw text data and raw medical image data related to esophageal cancer, and preprocessed the collected data to obtain text datasets and image datasets. S2. Hierarchical Knowledge Tree Construction: We used the Large Language Model (LLM) to construct graphs from text and image datasets to obtain corresponding graph structures. Then, we used the Hierarchical Clustering (HCSE) algorithm based on structural entropy to directly generate multi-level community structures from the obtained graph structures and construct hierarchical knowledge trees. The graph structure constructed using image data is called an interaction graph. The interaction graph is constructed by jointly maximizing the structural entropy and minimizing the relative entropy between the generated weights and the original weight distribution. The generated weights are the weights of the interaction graph generated from the complete graph of the image, and the original weights are the weights of the interaction graph used in practice. The multi-layered community structure includes a top-level community, an intermediate-level community, and a bottom-level community, with the intermediate-level community having at least two layers. The top-level community represents the main topic in the field of esophageal cancer, generating a community with document-level granularity; the middle-level community is a sub-community with entity-level granularity generated by refining and recursively dividing the top-level community layer by layer using the HCSE algorithm; and the bottom-level community is a sub-community with feature-level granularity generated from the middle-level community using the HCSE algorithm. A hierarchical knowledge tree includes a root node, intermediate nodes, and leaf nodes; The root node represents the entire medical document collection, symbolizing the overall research field of esophageal cancer, and includes all relevant documents and information; the intermediate nodes are subdivisions of the topics under the root node, dividing them into various important topics and sub-topics; the leaf nodes correspond to specific documents, cases or related studies in the underlying community, providing detailed information and data. S3, Community Summary Generation: The Large Language Model (LLM) is used to analyze the node degree and community density of each community in the hierarchical knowledge tree, generate corresponding community summaries, and sort them according to information priority in order to quickly obtain important data. S4. Query Processing and Answer Generation: When a query request for esophageal cancer diagnosis is received, the system combines the community summary in the hierarchical knowledge tree and performs a hierarchical search in the order of top-level community - middle-level community - bottom-level community to generate local answers. The Map-Reduce method is then used to integrate all local answers to generate the final answer. The Map-Reduce method consists of two phases: Map and Reduce. The MapReduce algorithm is used to integrate all the local answers, including: In the Map phase, each local answer is scored for relevance and help, and a comprehensive score is obtained. The relevance score is used to evaluate the overlap between the keywords in the local answer and the query keywords, and the help score is used to evaluate the effectiveness in solving user problems or providing useful information. In the Reduce phase, the optimal local answer is selected as the final answer based on the overall score.

2. The retrieval enhancement method according to claim 1, characterized in that, The graph structure constructed using the text dataset is a document graph. In the document graph, the extracted entities are used as nodes, the relationships between entities are used as edges, and the covariates related to the diagnosis of esophageal cancer are used as node information for the entities. The entities include symptoms, patient characteristics, and clinical trial information; the covariates include statements, attributes, and time spans. Statements describe specific statements or opinions related to the entity, attributes represent specific characteristics of the entity, and time spans represent the time during which the patient received treatment or the disease progressed.

3. The retrieval enhancement method according to claim 2, characterized in that, The community summary is a general overview of the esophageal cancer diagnosis-related information contained in each community. The node degree is the number of edges connected to the node, and the community density is the strength of the relationship between entities. The information priority is positively correlated with the node degree and the community density.

4. The retrieval enhancement method according to claim 1, characterized in that, The raw text data includes: electronic medical records (EMR), laboratory test results, medical literature, pathology reports, and patients' clinical symptoms and treatment records; The raw medical imaging data includes CT images, MRI images, and endoscopic images.

5. The retrieval enhancement method according to claim 1 or 4, characterized in that, Data preprocessing includes processing raw text data and raw image data; The processing of raw text data specifically includes: cleaning the raw text data, and standardizing and normalizing the cleaned data to obtain a standardized and normalized text dataset; The processing of the original image data specifically includes: adjusting the size, orientation, brightness and contrast of the original image data and adding noise to enhance the diversity of the image data; using correlation coefficients for feature selection; and using the selected features to combine features to obtain new features, thereby forming an image dataset. The correlation coefficient is specifically a coefficient that characterizes the correlation between a patient's physiological characteristics and the incidence of esophageal cancer; the new feature obtained by combining features is a feature that can better reflect the patient's overall health status compared with the selected features.

6. A retrieval enhancement system based on a hierarchical knowledge tree with structural entropy, characterized in that, The retrieval enhancement system is used to perform the retrieval enhancement method as described in any one of claims 1-5 above.

Images