Information interpretation method and device, equipment, storage medium and program product

Abstract

The invention discloses an information interpretation method and device, equipment, a storage medium and a program product, and relates to the technical field of finance, the method comprises the following steps: when it is detected that a user retrieves a question, retrieving a pre-constructed macroscopic knowledge graph library and a pre-constructed macroscopic text vector library to obtain corresponding cue words; and inputting the cue word into a pre-trained large language model to generate a corresponding answer. According to the method, the macroscopic viewpoint interpretation content required by the user is retrieved and generated from a large amount of macroscopic policy report information by means of a large language model through a graph retrieval enhancement technology, the user is helped to quickly extract key information and automatically generate viewpoint interpretation describing current macroeconomy, and the working efficiency is improved.

Description

Technical Field

[0001] The present application relates to the field of financial technology, and in particular to information interpretation methods, devices, equipment, storage media and program products. Background Art

[0002] In the current macro research field, researchers mainly query relevant information by reading a large number of macro policy research reports and other reports, and build a logic diagram of the current market macroeconomic status based on their own experience to assist researchers in predictive analysis. However, the current workflow has the following problems: the number of macro reports is large and complex, and manual information reading and retrieval is inefficient; due to the limitations of researchers' professional knowledge and experience, it may not be possible to fully describe the current macroeconomic status, thus affecting the accuracy of the analysis results.

[0003] With the development of large language model technology, it is now possible to use retrieval enhancement technology (RAG) to search and extract content from reports, which can improve researchers' information reading and retrieval efficiency. However, RAG is suitable for clear and specific information queries, and is not very effective for retrieving complex, extensive or implicit information. For example, when querying the views on credit in a single macro policy report, the full text of the report involves credit-related content in many places, and it is difficult for ordinary RAG technology to retrieve all the information and summarize it. For another example, when querying the changes in views on exchange rates in multiple macro policy reports, ordinary RAG technology does not work well in processing cross-document information aggregation. As for the interpretation of the views of macro policy reports, it is often necessary to summarize a large amount of information in the context of the report and draw conclusions based on the explicit and implicit correlation analysis between different macro elements. Summary of the invention

[0004] The main purpose of this application is to provide an information interpretation method, device, equipment, storage medium and program product, aiming to solve the technical problem of how to automatically and quickly extract key information from a large amount of macro information and accurately describe viewpoints.

[0005] To achieve the above purpose, the present application proposes an information interpretation method, which comprises:

[0006] When it is detected that the user searches for a question, the pre-built macro knowledge graph library and the pre-built macro text vector library are searched to obtain the corresponding prompt words;

[0007] The prompt word is input into a pre-trained large language model to generate a corresponding answer.

[0008] In one embodiment, when it is detected that the user searches for a question, the pre-built macro knowledge graph library and the pre-built macro text vector library are searched, and before the step of obtaining the corresponding prompt word, the step further includes:

[0009] Use a pre-trained large language model to process macro policy reports into knowledge graphs and build a macro knowledge graph library; and

[0010] The macro policy report is embedded and vectorized through a pre-trained vectorization model to build a macro text vector library.

[0011] In one embodiment, the step of processing the macro policy report into a knowledge graph by using a pre-trained large language model and constructing a macro knowledge graph library includes:

[0012] Segmenting the macro policy report to obtain a plurality of text units;

[0013] Leverage pre-trained large language models to extract macro elements, relations, and attributes from individual text units;

[0014] Construct a knowledge graph based on the macro elements, relationships and attributes;

[0015] Combining a community detection algorithm and a vector conversion algorithm to vectorize the structure of the knowledge graph, thereby obtaining a knowledge graph with a multi-level community structure;

[0016] The knowledge graph of the multi-level community structure is imported into a database to obtain a macro knowledge graph library.

[0017] In one embodiment, the step of communityizing and vectorizing the structure of the knowledge graph to obtain a knowledge graph with a multi-level community structure includes:

[0018] Communityize and vectorize the structure of the knowledge graph to generate multi-community modules;

[0019] The multi-community modules are hierarchically divided and summarized to construct a knowledge graph of a multi-level community structure.

[0020] In one embodiment, the step of embedding and vectorizing the macro policy report using a pre-trained vectorization model to construct a macro text vector library includes:

[0021] Segmenting the macro policy report to obtain a plurality of text units;

[0022] Use a pre-trained vector model to vectorize each text unit and generate a macro knowledge text vector library.

[0023] In one embodiment, when it is detected that the user searches for a question about a macro perspective, the step of searching a pre-built macro knowledge graph library and a pre-built macro text vector library to obtain corresponding prompt words includes:

[0024] If it is detected that the user's query question is specific opinion information, the query elements and information associated with the query elements are searched through the pre-built macro knowledge graph library;

[0025] The graph structure data corresponding to the query element and the document unstructured data corresponding to the associated information are used as a first retrieval candidate set;

[0026] Converting the query element and information associated with the query element into a vector form to obtain a query vector;

[0027] Searching for document fragments that match the query vector similarity in a pre-built macro text vector library to form a second retrieval candidate set;

[0028] The first search candidate set and the second search candidate set are converted into context environments of prompt words.

[0029] In one embodiment, when it is detected that the user searches for a question about a macro perspective, the step of searching the macro knowledge graph library and the macro text vector library to obtain corresponding prompt words also includes:

[0030] If it is detected that the question queried by the user is not specific opinion information, a multi-level community module summary report in the pre-built macro knowledge graph library is searched;

[0031] Querying the relevant content of the non-specific viewpoint information in the multi-level community module summary report, wherein the relevant content includes relevant community modules and relevant macro elements, relationships and attributes in the community;

[0032] The set of related contents is used as the context of the prompt word.

[0033] In one embodiment, the step of inputting the prompt word into the pre-trained large language model to generate a corresponding answer includes:

[0034] The context of the prompt word is combined with the user's query question and input into a pre-trained large language model to obtain an answer corresponding to the user's query question, and the answer is output to the user.

[0035] In addition, to achieve the above purpose, the present application also proposes an information interpretation device, which includes:

[0036] A retrieval module is used to search the pre-built macro knowledge graph library and the pre-built macro text vector library to obtain the corresponding prompt words when it is detected that the user searches for a question;

[0037] The generation module is used to input the prompt word into the pre-trained large language model to generate a corresponding answer.

[0038] In addition, to achieve the above-mentioned purpose, the present application also proposes an information interpretation device, which includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program is configured to implement the steps of the information interpretation method described above.

[0039] In addition, to achieve the above-mentioned purpose, the present application also proposes a storage medium, which is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, the steps of the information interpretation method described above are implemented.

[0040] In addition, to achieve the above-mentioned purpose, the present application also provides a computer program product, which includes a computer program, and when the computer program is executed by a processor, the steps of the information interpretation method described above are implemented.

[0041] The embodiment of the present application proposes an information interpretation method, device, equipment, storage medium and program product, the method comprising: when it is detected that the user searches for a question, searching the pre-built macro knowledge graph library and the pre-built macro text vector library to obtain the corresponding prompt words; inputting the prompt words into the pre-trained large language model to generate the corresponding answer. The method uses graph retrieval enhancement technology and the large language model to retrieve and generate the macro viewpoint interpretation content required by the user from a large amount of macro policy report information, helping users to quickly extract key information and automatically generate a viewpoint interpretation describing the current macro economy, thereby improving work efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.

[0043] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0044] Figure 1 A flowchart of the first embodiment of the method for interpreting information of this application is provided;

[0045] Figure 2 A flowchart of the second embodiment of the information interpretation method of this application is provided;

[0046] Figure 3 A flowchart of the third embodiment of the information interpretation method of this application is provided;

[0047] Figure 4 A brief flowchart of the information interpretation method provided in Example 1 and Example 2 of the present application;

[0048] Figure 5 A schematic diagram of the process of constructing a macro knowledge graph library using a pre-trained local large language model provided by the embodiment method of the present application;

[0049] Figure 6 A schematic diagram of the system architecture of the information interpretation method provided in the embodiment of the present application;

[0050] Figure 7 This is a schematic diagram of the module structure of the information interpretation device according to the embodiment of the present application;

[0051] Figure 8 Schematic diagram of the device structure of the hardware operating environment involved in the information interpretation method in the embodiment of the present application.

[0052] The purpose, features and advantages of this application will be further described in conjunction with the embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0053] It should be understood that the specific embodiments described herein are only used to explain the technical solutions of the present application and are not used to limit the present application.

[0054] In order to better understand the technical solution of the present application, a detailed description will be given below in conjunction with the accompanying drawings and specific implementation methods.

[0055] The main solution of the embodiment of the present application is: when it is detected that the user is searching for a question, the pre-built macro knowledge graph library and the pre-built macro text vector library are searched to obtain the corresponding prompt words; the prompt words are input into the pre-trained large language model to generate the corresponding answers.

[0056] In the current macro research field, researchers mainly query relevant information by reading a large number of macro policy research reports and other reports, and build a logic diagram of the current market's macroeconomic status based on their own experience and expertise to assist in predictive analysis. However, the existing workflow has the following problems: there are a large number of macro reports with complex content, and manual information reading and retrieval are inefficient. In addition, due to the limited expertise and experience of researchers, they may not be able to fully and accurately describe the current macroeconomic status, thus affecting the accuracy of the analysis results.

[0057] In recent years, with the development of large language model technology, especially the application of retrieval enhancement technology (RAG), researchers' information reading and retrieval efficiency has been improved to a certain extent. RAG technology can quickly locate and extract relevant information from a large number of documents by combining the capabilities of large-scale language models and search engines. However, RAG technology still has limitations when processing certain types of information queries: for macro information queries involving multiple aspects and extensive content, ordinary RAG technology is difficult to fully retrieve all relevant information. For example, when querying the views on credit in a single macro policy report, the full text of the report may involve credit-related content in many places, and ordinary RAG technology is difficult to retrieve all information and summarize it effectively. For queries that require information aggregation across multiple reports, ordinary RAG technology does not perform well. For example, when querying the changes in views on exchange rates in multiple macro policy reports, ordinary RAG technology is difficult to effectively integrate information from different reports and provide a coherent and comprehensive analysis. The views of macro policy reports often need to be combined with a large amount of information in the context for comprehensive analysis. Ordinary RAG technology is insufficient in extracting implicit information and analyzing the explicit and implicit associations between different macro elements.

[0058] This application provides a solution. When it is detected that a user is searching for a question, a pre-built macro knowledge graph library and a pre-built macro text vector library are searched to obtain the corresponding prompt words; the prompt words are input into the pre-trained large language model to generate the corresponding answers. This solution not only interprets, extracts and summarizes information with the help of a local large language model, but also realizes automatic and rapid interpretation of macro report viewpoints, greatly improving the work efficiency of researchers, and the operation of the local model better protects the security of private domain data. It is also based on graph retrieval enhancement technology to better summarize the viewpoints of the full report and extract implicit related information, improving the accuracy and completeness of information interpretation.

[0059] It should be noted that the execution subject of this embodiment can be a computing service device with data processing, network communication and program running functions, such as a tablet computer, a personal computer, a mobile phone, etc., or an electronic device capable of realizing the above functions, etc. The following takes a personal computer as an example to illustrate this embodiment and the following embodiments.

[0060] Based on this, the present application embodiment provides a method for interpreting information. Figure 1 , Figure 1 This is a flow chart of the first embodiment of the information interpretation method of this application.

[0061] In this embodiment, the information interpretation method includes steps S10 to S20:

[0062] Step S10, when it is detected that the user searches for a question, a pre-built macro knowledge graph library and a pre-built macro text vector library are searched to obtain corresponding prompt words;

[0063] It should be noted that the macro knowledge graph library is a structured database containing entities, relations, and events extracted from macro policy reports. The macro text vector library is a database or file collection that stores vector representations of macro policy reports.

[0064] When it is detected that a user is searching for a question, especially when it involves specific macro-viewpoint information, the retrieval method combining a pre-built macro knowledge graph and a pre-built macro text vector library can significantly improve the comprehensiveness and accuracy of the information.

[0065] In a feasible embodiment, step S30 may include steps A11 to A15:

[0066] Step A11: if it is detected that the user's query question is specific information from a macro perspective, the query element and information associated with the query element are searched through a pre-built macro knowledge graph library;

[0067] Specifically, when the system detects that the macro-opinion question queried by the user is specific opinion information, for example, "What is the latest policy opinion on domestic inflation?", the macro-knowledge graph library obtained in step S10 is first searched for nodes directly associated with the query element. For example, "domestic inflation" may be connected to multiple other nodes, such as "price level", "production level", "monetary policy", etc.

[0068] In addition, we should not only focus on directly related nodes, but also consider the relationship between these nodes and their relationship with other nodes. For example, "domestic inflation" is closely related to "price level", and "price level" may be affected by factors such as "international oil price". In this way, we can obtain more abundant and comprehensive relevant information.

[0069] Step A12, taking the graph structure data corresponding to the query element and the document unstructured data corresponding to the associated information as a retrieval candidate set;

[0070] Specifically, according to the key nodes and their associated nodes identified in the macro knowledge graph library, relevant document fragments are retrieved from the large amount of stored unstructured data. Among them, unstructured data includes but is not limited to policy reports, news reports, etc.; document fragments include but are not limited to detailed discussions and opinions on specific topics.

[0071] Then, the graph structure data where the query element is located and the document unstructured data where the associated information is located are taken as the retrieved candidate set, which is taken as the first retrieval candidate set.

[0072] Step A13, converting the query element and the information associated with the query element into a vector form to obtain a query vector;

[0073] Specifically, query elements such as "domestic inflation" and their associated information such as "price level", "production level", "monetary policy", etc. are extracted from the macro knowledge graph. The extracted query elements and associated information are preprocessed, including word segmentation, stop word removal, stemming, etc., to ensure the cleanliness and standardization of text data.

[0074] Use the selected vectorization model to convert the preprocessed text data into vector form. For example, if you use the BERT (Bidirectional Encoder Representations from Transformers) model, you can input the text into the model and get the vector representation of each word or sentence.

[0075] For multiple related information, their vectors can be weighted averaged or concatenated to form a comprehensive query vector. The generated vectors are normalized to ensure that the lengths of the vectors are consistent, which is convenient for subsequent similarity calculations.

[0076] Step A14, searching the macro text vector library for document segments that match the query vector similarity to form a second retrieval candidate set;

[0077] Specifically, firstly, a pre-built macro text vector library is loaded, wherein the macro text vector library contains vector representations of a large number of text data such as policy reports, news reports, etc. A similarity calculation method is used to calculate the similarity between the query vector and each document fragment vector in the text vector library.

[0078] Then, based on the calculated similarity values, the top N document fragments with the highest similarity to the query vector are selected. N can be adjusted according to actual needs, and 10-50 fragments are usually selected as candidate sets.

[0079] Finally, the selected document fragments are extracted from the text vector library to form the second retrieval candidate set. These document fragments usually contain detailed information and opinions related to the query topic.

[0080] Step A15: convert the first search candidate set and the second search candidate set into a context environment of a prompt word.

[0081] Specifically, the first search candidate set and the second search candidate set formed by step A13 and step A14 are integrated and converted into the context of prompt. For example, domestic inflation is also related to factors such as price level and production level. Therefore, when searching the macro knowledge graph, it is necessary not only to search for the domestic inflation factor node, but also to search for the price level, production level and other factor nodes, and search for the original document fragments where these factor nodes are located.

[0082] Through the above steps, by combining the macro knowledge graph and the macro text vector library, it is possible to understand and respond to user queries from a broader perspective and provide more accurate and in-depth answers.

[0083] Step S20, input the prompt word into the pre-trained large language model to generate a corresponding answer.

[0084] Finally, the prompt words obtained through the above steps are combined with the macro-perspective questions queried by the user and input into the pre-trained large language model to obtain the answers corresponding to the questions queried by the user (especially those involving macro-perspectives), and the answers are output to the user.

[0085] Through the above-mentioned embodiment method, not only can the automatic and rapid interpretation of macro report viewpoints be achieved by using the local large language model for information retrieval, extraction and summarization, thereby greatly improving the work efficiency of researchers, but the operation of the local model can better protect the security of private domain data. In addition, compared with the use of traditional RAG technology for report analysis, the technical method based on graph retrieval enhancement (GraphRAG) can better summarize the viewpoints of the full report and extract implicit related information, thereby improving the accuracy and completeness of information retrieval.

[0086] Based on the first embodiment of the present application, in the second embodiment of the present application, the same or similar contents as those in the above-mentioned embodiment 1 can be referred to the above introduction, and will not be repeated in the following. Figure 2 , step S10 further includes steps B11 to B13:

[0087] Step B11, if it is detected that the question queried by the user is not specific opinion information, searching for a multi-level community module summary report in the macro knowledge graph library;

[0088] When it is detected that the user's query question is non-specific opinion information, such as "What specific measures have been implemented in our country's latest monetary policy?", understand the user's query intention and extract key elements such as "latest monetary policy" and "specific measures".

[0089] Conduct a global search in the macro knowledge graph and find content related to the "latest monetary policy" from different community summary reports, such as "credit", "interest rate", "exchange rate", "liquidity", etc.

[0090] Step B12, querying the multi-level community module summary report for relevant content of the non-specific opinion information;

[0091] Specifically, we continue to search downward according to the community structure to find specific measure nodes related to "monetary policy". We aggregate all relevant element nodes and document fragments to form a comprehensive retrieval candidate set.

[0092] Step B13: taking the set of related contents as the context of the prompt word.

[0093] Finally, the comprehensive retrieval candidate set is integrated and converted into a natural language context to generate a prompt.

[0094] For example, monetary policy includes many specific measures. First, we need to search out the community structure related to monetary policy, and then continue to find related communities and elements such as credit, interest rates, exchange rates, and liquidity based on the community structure. By aggregating all relevant element nodes and document fragments, we will ultimately use them as the context environment.

[0095] Through the above-mentioned embodiment method, it is possible to effectively handle the problem of users querying macro-viewpoints, especially the query of non-specific viewpoint information. This method not only improves the comprehensiveness and accuracy of information, but also improves the efficiency of query and user experience, and provides users with more in-depth and multi-angle answers.

[0096] Based on the first embodiment of the present application, in the third embodiment of the present application, the same or similar contents as those in the above-mentioned embodiment 1 can be referred to the above introduction, and will not be repeated in the following. Figure 3 , before step 10, it also includes steps S01 to S02:

[0097] Step S01, using a pre-trained large language model to process the macro policy report into a knowledge graph and construct a macro knowledge graph library;

[0098] It should be noted that the Large Language Model (LLM) is a deep learning model that has strong natural language understanding and generation capabilities through pre-training on large-scale text data. Macroeconomic policy reports usually include descriptions of current economic conditions, forecasts of future economic trends, policy recommendations, and economic indicator analysis.

[0099] It is understandable that since querying a large amount of macro information requires a lot of labor costs, performing step S01 can quickly locate and retrieve key information in the macro policy report, avoiding the inefficient process of manually reading and screening each article one by one.

[0100] In a feasible implementation, step S01 may include steps A011 to A015:

[0101] Step A011, segmenting the macro policy report to obtain multiple text units;

[0102] It can be understood that executing step A011 can split the long macro policy report into multiple smaller text units to facilitate subsequent processing and analysis.

[0103] Specifically, after obtaining the macro policy report, the macro policy report is segmented into paragraphs, and each paragraph is regarded as a text unit. Then each paragraph is further segmented into sentences, and each sentence is regarded as a text unit. Alternatively, the macro policy report is segmented according to the subject or chapter title given in the macro policy report, and each subject or chapter is regarded as a text unit.

[0104] Step A012, extracting macro elements, relations and attributes from each text unit using a pre-trained large language model;

[0105] It can be understood that step A012 is executed to extract key macro elements, relationships and attributes from the segmented text units to provide basic data for constructing a knowledge graph.

[0106] Specifically, a pre-trained large language model is used to identify entities in each text unit, such as enterprises, economic indicators, etc., and to identify relationships between entities, such as causal relationships, influence relationships, time relationships, etc., and to extract attribute information of entities, such as value, time, location, etc.

[0107] Step A013, constructing a knowledge graph based on the macro elements, relationships and attributes;

[0108] It can be understood that step A013 is executed to organize the extracted macro elements, relationships and attributes into a knowledge graph to form a structured data representation.

[0109] Specifically, the extracted macro indicators and entities are used as nodes in the graph; the relationships between entities are used as edges in the graph, and the corresponding nodes are connected using edges; and the extracted attribute information is attached to the corresponding nodes or edges.

[0110] Step A014, communityizing and vectorizing the structure of the knowledge graph to obtain a knowledge graph with a multi-level community structure;

[0111] Optionally, in this embodiment, the structure of the knowledge graph is communityized and vectorized by combining a community detection algorithm and a vector conversion algorithm to generate a knowledge graph with a hierarchical community structure.

[0112] It should be noted that the community detection algorithm is used to identify the community structure in the network, that is, to divide the nodes in the network into several subgroups with dense internal connections and sparse external connections. The community detection algorithm includes but is not limited to the Louvain (modular) algorithm, LPA (Label Propagation Algorithm), etc. The vector conversion algorithm is used to convert the nodes in the graph into vector form to facilitate subsequent machine learning and data analysis tasks. The vector conversion algorithm includes but is not limited to Node2Vec, DeepWalk, etc.

[0113] In another feasible implementation manner, step A014 may further include steps a1 to a2:

[0114] Step a1: Communityize and vectorize the structure of the knowledge graph to build a multi-community module;

[0115] Specifically, first select an adapted community detection algorithm to perform community detection on the knowledge graph. The specific operations include: initializing each knowledge graph data node as a separate community, and checking each node one by one, moving it to the adjacent community to maximize the increase in modularity. Repeat the above steps until the modularity no longer increases. Treat each community as a super node, build a new network, and repeat the above process until the modularity no longer increases. Finally, the community detection results can be verified by visualization and other methods.

[0116] After completing the community detection, a vectorization algorithm is applied to the graph data after community detection. In this embodiment, the Node2Vec algorithm is applied to convert the nodes of the graph data into vector form. The specific operations include: defining a random walk strategy, including return parameters and gradient parameters. Random walks are performed from each node to generate a node sequence. Finally, the semantic processing model is used to convert the node sequence into a vector.

[0117] Through the above steps, the community structure in the graph is identified, and the nodes are divided into multiple subgroups with dense internal connections and sparse external connections. In addition, the nodes in the graph are converted into vector form to retain the structure and semantic information of the nodes.

[0118] Step a2, hierarchically divide and summarize the multi-community modules to construct a knowledge graph of a multi-level community structure.

[0119] Specifically, the community detection algorithm is applied again to each community to identify subcommunities. The size, density and other features of each community are calculated. The text content within the community is summarized to extract key information. LDA is used to extract the topics of the community. For example, the topics of the "economic growth" community may include "GDP growth", "employment rate changes", etc.

[0120] Through the above steps, a macro knowledge graph with a multi-level community structure is constructed. The multi-level community structure allows users to quickly locate specific communities or sub-communities of interest, thereby finding the required information more quickly. In addition, this multi-community module and multi-level structured knowledge graph fully and comprehensively explores the complex connections and implicit relationships in macro policy reports, which helps to understand the overall relationship between macro elements and indicators.

[0121] Step A015, importing the knowledge graph of the multi-level community structure into a database to obtain a macro knowledge graph library.

[0122] It should be noted that the databases used in this embodiment include but are not limited to: Neo4j, which is a powerful database that supports complex graph query and analysis; ArangoDB, which supports multi-model data storage, including graph data; JanusGraph, which is highly scalable and suitable for large-scale graph data storage.

[0123] According to the attributes and types of the knowledge graph of the multi-level community structure, the corresponding database is selected, and the corresponding knowledge graph data is stored in the database to obtain a macro knowledge graph library.

[0124] Through the above steps, the knowledge graph of multi-level community structure can be successfully imported into the database to build an efficient macro knowledge graph library, providing strong support for macroeconomic analysis, policy making and market forecasting.

[0125] Step S02, embedding and vectorizing the macro policy report through a pre-trained vectorization model to build a macro text vector library.

[0126] It should be noted that a vectorization model is a model that converts text data into numerical vectors, and is usually used in natural language processing (NLP) tasks. These models convert words or sentences in the text into high-dimensional vectors by learning the semantics and contextual relationships in the text, so that similar texts are closer in the vector space. Embedding vectorization refers to the process of converting text data into numerical vectors.

[0127] It is understandable that since macro policy reports usually contain a large amount of text data, and these data are highly complex and diverse, therefore, while executing step S01, step S02 is executed to efficiently convert the text data in the macro policy report into numerical vectors and construct a macro text vector library. This not only improves the efficiency and accuracy of data processing, but also enhances the ability of data analysis, supports a variety of advanced applications, and optimizes query and retrieval.

[0128] In a feasible implementation manner, step S02 may further include steps B021 to B022:

[0129] Step B021, segmenting the macro policy report to obtain multiple text units;

[0130] Similarly, the macro policy report is segmented into paragraphs, and each paragraph is regarded as a text unit. Then each paragraph is further segmented into sentences, and each sentence is regarded as a text unit. Alternatively, the macro policy report is segmented according to the subject or chapter title given in the macro policy report, and each topic or chapter is regarded as a text unit. According to the above steps, multiple text units are obtained.

[0131] Step B022, use the pre-trained vector model to vectorize each text unit to generate a macro knowledge text vector library.

[0132] Specifically, after obtaining multiple text units, a vector model, such as BGE (BAAI General Embedding, general embedding model), is used to vectorize each text unit to obtain a macro knowledge text vector library.

[0133] Through the above steps, the macro policy report is embedded in the vectorization process and the corresponding vector library is constructed, which not only improves the efficiency and quality of retrieval, but also compares with the traditional keyword matching method. The deep learning-based vectorization model can better capture the semantic relationship between texts, making the system more accurate and intelligent when processing complex queries.

[0134] For example, to help understand the implementation process of the information interpretation method obtained by combining the above-mentioned embodiments 1, 2 and 3, please refer to Figure 4 , Figure 4 A brief flowchart of an information interpretation method is provided, specifically:

[0135] The method of this embodiment first uses a pre-trained local large language model (LLM) to process the macro policy report into a knowledge graph, and constructs a macro knowledge graph library that can be queried (Graph Querying).

[0136] At the same time, the pre-trained local vectorization model is used to embed and vectorize the macro policy reports and build a macro text vector library that can be queried (Embedding Querying).

[0137] When searching for macro-viewpoint questions from users, the text vector library and the knowledge graph library are searched at the same time, and the retrieved relevant candidate set data are integrated and converted into prompt content.

[0138] Finally, the above merged content is used as the context of the user's question and input into the pre-trained local large language model to generate an answer.

[0139] Based on the above, the method of this embodiment is described in more detail:

[0140] Specifically, first, a flowchart of using a pre-trained local large language model to build a macro knowledge graph library is shown in the figure below. Figure 5 As shown. The process of building a macro knowledge graph library using a pre-trained local large language model mainly includes: first, segmenting the macro policy report into multiple text units; then, using the pre-trained local large language model to build a knowledge graph, extracting macro elements, relationships, and attributes of macro elements from each text unit; and using the community detection algorithm to build a multi-community module, and using the Node2Vec algorithm to vectorize the graph structure; then, hierarchically divide and summarize the community modules to build a multi-level community structure graph. This multi-community module, multi-level structure knowledge graph fully and comprehensively explores the complex connections and implicit relationships in the macro policy report, which helps to understand the overall connection between macro elements and indicators. Finally, based on the above content, a macro knowledge graph library is constructed.

[0141] At the same time, the process of building a macro knowledge vector library using a pre-trained vectorization model mainly includes: segmenting the macro policy report into multiple text units; using the pre-trained vector model to vectorize each text unit to generate a macro knowledge text vector library.

[0142] Then, when searching for macro viewpoints in user queries, it can generally be divided into two situations. On the one hand, when querying specific viewpoint information, first search the information related to the query element through the macro knowledge graph, and then use the graph structure data where the query element is located and the unstructured data of the document where the related information is located as the retrieved candidate set, and integrate them into the prompt context environment. On the other hand, when querying non-specific viewpoints, it is necessary to reason about the entire document. First, it is necessary to globally search the multi-level community report generated by the macro knowledge graph, find relevant content from different community reports, aggregate them, and use the collection as the prompt context environment. The prompt context environment finally generated by the above method will be used as the input of the next large language model and combined with the user's query question.

[0143] After the prompt is generated, a suitable large language model is selected, and the answer corresponding to the end user's question is obtained through the understanding and generation capabilities of the large language model, and then output to the user.

[0144] Through the above steps, the entire process from user input to answer output is completed, realizing a technical solution for interpreting the views of macroeconomic policy reports based on terminal graph retrieval enhancement.

[0145] For example, to help understand the implementation process of the information interpretation method obtained by combining the above-mentioned embodiments 1, 2 and 3, please refer to Figure 6 , Figure 6 A system architecture diagram of an information interpretation method is provided, specifically:

[0146] The system includes a text input module, a graph construction module, a text processing module, a vector embedding module, a retrieval processing module and a text generation module.

[0147] In this embodiment, the document is first processed, and graph construction and vector embedding are performed respectively; then the input content is searched in the two modules respectively, and the search results and input are merged and input into the text generation module to finally obtain the result.

[0148] The above system architecture is conducive to graph retrieval enhancement technology, and with the help of local large language models, it is possible to retrieve and generate the macro-view interpretation content required by researchers from a large amount of macro policy report information, aiming to help researchers quickly extract key information and automatically generate interpretations of views describing the current macro-economy, thereby improving the work efficiency of researchers.

[0149] It should be noted that the above examples are only used to understand the present application and do not constitute a limitation on the information interpretation method of the present application. More simple transformations based on this technical concept are all within the scope of protection of the present application.

[0150] This application also provides an information interpretation device, please refer to Figure 7 , the information interpretation device comprises:

[0151] The retrieval module 10 is used to search the pre-built macro knowledge graph library and the pre-built macro text vector library to obtain the corresponding prompt words when it is detected that the user searches for a question;

[0152] The generation module 20 is used to input the prompt word into a pre-trained large language model to generate a corresponding answer.

[0153] The information interpretation device provided by the present application adopts the information interpretation method in the above embodiment, which can solve the technical problem of how to automatically and quickly extract key information from a large amount of macro information and accurately describe the viewpoint. Compared with the prior art, the beneficial effects of the information interpretation device provided by the present application are the same as the beneficial effects of the information interpretation method provided by the above embodiment, and the other technical features in the information interpretation device are the same as the features disclosed in the above embodiment method, which will not be repeated here.

[0154] The present application provides an information interpretation device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the information interpretation method in the above-mentioned embodiment one.

[0155] Reference below Figure 8 , which shows a schematic diagram of the structure of an information interpretation device suitable for implementing the embodiment of the present application. The information interpretation device in the embodiment of the present application may include but is not limited to mobile terminals such as mobile phones, laptop computers, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Descriptions), PMPs (Portable Media Players), etc., and fixed terminals such as digital TVs, desktop computers, etc. Figure 8 The information interpretation device shown is merely an example and should not bring any limitation to the functions and scope of use of the embodiments of the present application.

[0156] like Figure 8As shown, the information interpretation device may include a processing device 1001 (e.g., a central processing unit, a graphics processor, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM: Read Only Memory) 1002 or a program loaded from a storage device 1003 to a random access memory (RAM: Random Access Memory) 1004. In RAM1004, various programs and data required for the operation of the information interpretation device are also stored. The processing device 1001, ROM1002, and RAM1004 are connected to each other through a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Generally, the following systems can be connected to the I / O interface 1006: an input device 1007 including, for example, a touch screen, a touch pad, a keyboard, a mouse, an image sensor, a microphone, an accelerometer, a gyroscope, etc.; an output device 1008 including, for example, a liquid crystal display (LCD: Liquid Crystal Display), a speaker, a vibrator, etc.; a storage device 1003 including, for example, a magnetic tape, a hard disk, etc.; and a communication device 1009. The communication device 1009 can allow the information interpretation device to communicate with other devices wirelessly or by wire to exchange data. Although the figure shows an information interpretation device with various systems, it should be understood that it is not required to implement or have all the systems shown. More or fewer systems can be implemented or have alternatively.

[0157] In particular, according to the embodiments disclosed in the present application, the process described above with reference to the flowchart can be implemented as a computer software program. For example, the embodiments disclosed in the present application include a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program includes a program code for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from a network through a communication device, or installed from a storage device 1003, or installed from a ROM 1002. When the computer program is executed by the processing device 1001, the above-mentioned functions defined in the method of the embodiment disclosed in the present application are executed.

[0158] The information interpretation device provided by the present application adopts the information interpretation method in the above embodiment to solve the technical problem of information interpretation. Compared with the prior art, the beneficial effects of the information interpretation device provided by the present application are the same as the beneficial effects of the information interpretation method provided by the above embodiment, and the other technical features in the information interpretation device are the same as the features disclosed in the method of the previous embodiment, which will not be repeated here.

[0159] It should be understood that the various parts disclosed in this application can be implemented by hardware, software, firmware or a combination thereof. In the description of the above embodiments, specific features, structures, materials or characteristics can be combined in any one or more embodiments or examples in a suitable manner.

[0160] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

[0161] The present application provides a computer-readable storage medium having computer-readable program instructions (ie, computer programs) stored thereon, and the computer-readable program instructions are used to execute the information interpretation method in the above-mentioned embodiment.

[0162] The computer-readable storage medium provided in the present application may be, for example, a USB flash drive, but is not limited to electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, systems or devices, or any combination of the above. More specific examples of computer-readable storage media may include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in combination with an instruction execution system, system or device. The program code contained on the computer-readable storage medium may be transmitted using any appropriate medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination of the above.

[0163] The computer-readable storage medium may be included in the information interpretation device; or may exist independently without being assembled into the information interpretation device.

[0164] The computer-readable storage medium carries one or more programs. When the one or more programs are executed by the information interpretation device, the information interpretation device: when detecting that the user searches for a question, searches a pre-built macro knowledge graph library and a pre-built macro text vector library to obtain corresponding prompt words; inputs the prompt words into a pre-trained large language model to generate corresponding answers.

[0165] Computer program code for performing the operations of the present application may be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, C++, and conventional procedural programming languages ​​such as "C" or similar programming languages. The program code may be executed entirely on the user's computer, partially on the user's computer, as a separate software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0166] The flow chart and block diagram in the accompanying drawings illustrate the possible architecture, function and operation of the system, method and computer program product according to various embodiments of the present application. In this regard, each square box in the flow chart or block diagram can represent a module, a program segment or a part of a code, and the module, the program segment or a part of the code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the square box can also occur in a sequence different from that marked in the accompanying drawings. For example, two square boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each square box in the block diagram and / or flow chart, and the combination of the square boxes in the block diagram and / or flow chart can be implemented with a dedicated hardware-based system that performs a specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0167] The modules involved in the embodiments described in this application may be implemented by software or hardware, wherein the name of the module does not constitute a limitation on the unit itself in some cases.

[0168] The readable storage medium provided by the present application is a computer-readable storage medium, which stores computer-readable program instructions (i.e., computer programs) for executing the above-mentioned information interpretation method, and can solve the technical problem of how to automatically and quickly extract key information from a large amount of macro information and accurately describe the viewpoints. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided by the present application are the same as the beneficial effects of the information interpretation method provided by the above-mentioned embodiment, and will not be repeated here.

[0169] The present application also provides a computer program product, including a computer program, which implements the steps of the above-mentioned information interpretation method when executed by a processor.

[0170] The computer program product provided by this application can solve the technical problem of information interpretation. Compared with the prior art, the beneficial effects of the computer program product provided by this application are the same as the beneficial effects of the information interpretation method provided by the above embodiment, which will not be repeated here.

[0171] The above descriptions are only some embodiments of the present application, and are not intended to limit the patent scope of the present application. All equivalent structural changes made using the contents of the present application specification and drawings under the technical concept of the present application, or direct / indirect applications in other related technical fields are included in the patent protection scope of the present application.

Claims

1. A method for interpreting information, characterized in that: The method comprises: When it is detected that the user searches for a question, the pre-built macro knowledge graph library and the pre-built macro text vector library are searched to obtain the corresponding prompt words; The prompt word is input into a pre-trained large language model to generate a corresponding answer.

2. The method according to claim 1, characterized in that When it is detected that the user searches for a question, the pre-built macro knowledge graph library and the pre-built macro text vector library are searched to obtain the corresponding prompt words, and before the step, the method further includes: Use a pre-trained large language model to process macro policy reports into knowledge graphs and build a macro knowledge graph library; and The macro policy report is embedded and vectorized through a pre-trained vectorization model to build a macro text vector library.

3. The method according to claim 2, characterized in that The steps of processing the macro policy report into a knowledge graph by using a pre-trained large language model and constructing a macro knowledge graph library include: Segmenting the macro policy report to obtain a plurality of text units; Leverage pre-trained large language models to extract macro elements, relations, and attributes from individual text units; Construct a knowledge graph based on the macro elements, relationships and attributes; The structure of the knowledge graph is communityized and vectorized to obtain a knowledge graph with a multi-level community structure; The knowledge graph of the multi-level community structure is imported into a database to obtain a macro knowledge graph library.

4. The method according to claim 3, characterized in that The step of communityizing and vectorizing the structure of the knowledge graph to obtain a knowledge graph with a multi-level community structure includes: Communityizing and vectorizing the structure of the knowledge graph to generate a multi-community module; The multi-community modules are hierarchically divided and summarized to construct a knowledge graph of a multi-level community structure.

5. The method according to claim 2, characterized in that The step of embedding and vectorizing the macro policy report through a pre-trained vectorization model to construct a macro text vector library comprises: Segmenting the macro policy report to obtain a plurality of text units; Use a pre-trained vector model to vectorize each text unit and generate a macro knowledge text vector library.

6. The method according to any one of claims 2 to 5, characterized in that When it is detected that the user searches for a question, the step of searching the pre-built macro knowledge graph library and the pre-built macro text vector library to obtain the corresponding prompt words includes: If it is detected that the user's query question is specific opinion information, the query elements and information associated with the query elements are searched through the pre-built macro knowledge graph library; The graph structure data corresponding to the query element and the document unstructured data corresponding to the associated information are used as a first retrieval candidate set; Converting the query element and information associated with the query element into a vector form to obtain a query vector; Searching for document fragments that match the query vector similarity in a pre-built macro text vector library to form a second retrieval candidate set; The first search candidate set and the second search candidate set are converted into context environments of prompt words.

7. The method according to any one of claims 2 to 5, characterized in that The step of searching the pre-built macro knowledge graph library and the pre-built macro text vector library to obtain the corresponding prompt words when detecting that the user searches for a question also includes: If it is detected that the question queried by the user is not specific opinion information, a multi-level community module summary report in the pre-built macro knowledge graph library is searched; Querying the relevant content of the non-specific viewpoint information in the multi-level community module summary report, wherein the relevant content includes relevant community modules and relevant macro elements, relationships and attributes in the community; The set of related contents is used as the context of the prompt word.

8. The method according to any one of claims 6 to 7, characterized in that The step of inputting the prompt word into a pre-trained large language model to generate a corresponding answer includes: The context of the prompt word is combined with the user's query question and input into a pre-trained large language model to obtain an answer corresponding to the user's query question, and the answer is output to the user.

9. An information interpretation device, characterized in that: The device comprises: A retrieval module is used to search the pre-built macro knowledge graph library and the pre-built macro text vector library to obtain the corresponding prompt words when it is detected that the user searches for a question; The generation module is used to input the prompt word into the pre-trained large language model to generate a corresponding answer.

10. An information interpretation device, characterized in that: The device comprises: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program is configured to implement the steps of the information interpretation method according to any one of claims 1 to 8.

11. A storage medium, characterized in that: The storage medium is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, the steps of the information interpretation method according to any one of claims 1 to 8 are implemented.

12. A computer program product, characterized in that The computer program product comprises a computer program, and when the computer program is executed by a processor, the steps of the information interpretation method according to any one of claims 1 to 7 are implemented.

Images