Information generation method and device, electronic equipment and storage medium

Abstract

The invention relates to the technical field of data processing, and provides an information generation method and device, electronic equipment and a storage medium, and the method comprises the steps: retrieving candidate data related to a query request of a user from a multi-source heterogeneous data source, and distributing an authority level for a data source corresponding to each candidate data; determining the data source with the highest authority level from the data sources corresponding to the candidate data with the conflicts, and initiating a verification request to the data source with the highest authority level, so that the data source with the highest authority level returns the latest version data for the candidate data with the conflicts; updating candidate data with conflicts according to the latest version data; and generating business cooperation information of the query request based on the updated candidate data and the candidate data without conflicts. According to the method, the conflict data is updated through the data source with the highest authority level, so that logic contradictions and numerical illusion caused by data timeliness lag or incapability of judging authenticity of the data are avoided.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to an information generation method, apparatus, electronic device, and storage medium. Background Technology

[0002] In enterprise-level business collaboration scenarios, it is often necessary to integrate structured data generated by internal business systems and unstructured information from external sources to support business decisions.

[0003] In existing technologies, retrieval-enhanced generation methods are commonly used to generate business information. These methods retrieve relevant fragments from external knowledge bases and input them into a large language model for answering questions. However, when data from different sources (such as internal systems and external news) contradict each other in terms of values ​​or status, this method usually can only simply present them side by side or select them randomly. This results in information that is outdated, logically contradictory, or numerically inaccurate, making it difficult to meet the requirements of business decision-making for information accuracy and consistency. Summary of the Invention

[0004] This invention provides an information generation method, apparatus, electronic device, and storage medium to address the deficiencies in the prior art.

[0005] This invention provides an information generation method, comprising the following steps: Retrieve candidate data related to the user's query request from multiple heterogeneous data sources, and assign an authority level to the data source corresponding to each candidate data; If a conflict is detected between at least two candidate data, the data source with the highest authority level is determined from the data sources corresponding to the conflicting candidate data, and a verification request is sent to the data source with the highest authority level so that the data source with the highest authority level returns the latest version of the data for the conflicting candidate data. Update conflicting candidate data based on the latest version data; Based on the updated candidate data and the candidate data that do not conflict, the business collaboration information for the query request is generated.

[0006] According to an information generation method provided by the present invention, the conflict detection step between candidate data includes: Extract attribute values ​​related to the query request from each candidate data; If the attribute values ​​corresponding to the same attribute of any two candidate data are inconsistent, it is determined that there is a conflict between the two candidate data.

[0007] According to an information generation method provided by the present invention, after extracting attribute values ​​related to the query request from each candidate data, the method further includes: When the attribute value is text, the attribute values ​​corresponding to the same attribute of any two candidate data are converted into semantic vectors; Calculate the similarity between the semantic vectors corresponding to the same attribute of any two candidate data; If the similarity is lower than a first threshold, it is determined that the attribute values ​​corresponding to the same attribute of any two candidate data are inconsistent.

[0008] According to an information generation method provided by the present invention, after extracting attribute values ​​related to the query request from each candidate data, the method further includes: When the attribute value is numerical, calculate the degree of difference between the numerical values ​​corresponding to the same attribute of any two candidate data. If the difference exceeds the second threshold, it is determined that the attribute values ​​corresponding to the same attribute of any two candidate data are inconsistent.

[0009] According to an information generation method provided by the present invention, the step of assigning authority levels to the data sources corresponding to each candidate data includes: Assign an initial authority level to each data source based on its type; Based on the data sensitivity of each candidate data, the initial authority level of the data source corresponding to each candidate data is adjusted to obtain the authority level of the data source corresponding to each candidate data. The data sensitivity of each candidate data is determined based on the business category to which the attribute contained in the corresponding candidate data belongs.

[0010] According to an information generation method provided by the present invention, adjusting the initial authority level of the data source corresponding to each candidate data based on the data sensitivity of each candidate data includes: If the data sensitivity of any candidate data exceeds a preset sensitivity threshold, the initial authority level of the data source corresponding to that candidate data is increased.

[0011] According to an information generation method provided by the present invention, retrieving candidate data related to a user's query request from multi-source heterogeneous data sources includes: The query request is semantically parsed to extract the temporal semantic features from the query request; If the temporal semantic features contain keywords indicating the acquisition of the latest data, then candidate data with the latest timestamp will be retrieved first, based on the timestamp carried by each candidate data.

[0012] According to an information generation method provided by the present invention, retrieving candidate data related to a user's query request from multi-source heterogeneous data sources includes: The query request is subjected to intent recognition to determine the type of the query request; If the query request is a numeric query, the query request is converted into a structured query statement, and candidate data is retrieved from the multi-source heterogeneous data source based on the structured query statement.

[0013] According to an information generation method provided by the present invention, the step of converting the query request into a structured query statement includes: Extract the triple of entity, attribute, and operation from the query request; Generate a structured query statement for the multi-source heterogeneous data source based on the triples.

[0014] According to an information generation method provided by the present invention, the step of performing intent recognition on the query request and determining the type of the query request includes: If the type of the query request is a text query, the query request is classified based on a pre-trained classification model to obtain a first classification result; Feature words are extracted from the query request, and the query request is classified according to the matching results of the feature words and the preset feature word library to obtain a second classification result; If the first classification result is consistent with the second classification result, the first classification result or the second classification result shall be used as the type of the query request.

[0015] According to an information generation method provided by the present invention, generating the business collaboration information of the query request includes: Based on the updated candidate data and the candidate data that did not conflict, first information is generated through a prediction model; Based on the updated candidate data and the candidate data that did not conflict, second information is generated through logical calculation; If the deviation between the first information and the second information exceeds a preset deviation threshold, the second information is used as the business collaboration information.

[0016] According to an information generation method provided by the present invention, retrieving candidate data related to a user's query request from multi-source heterogeneous data sources includes: Based on the user's permission information, determine the range of data sources that the user can access; Retrieve candidate data related to the user's query request from the range of the data source.

[0017] The present invention also provides an information generation device, comprising the following modules: The allocation module is used to retrieve candidate data related to the user's query request from multi-source heterogeneous data sources and assign an authority level to the data source corresponding to each candidate data. The determination module is used to determine the data source with the highest authority level from the data sources corresponding to the conflicting candidate data when a conflict is detected between at least two candidate data, and to initiate a verification request to the data source with the highest authority level so that the data source with the highest authority level returns the latest version data for the conflicting candidate data. The update module is used to update conflicting candidate data based on the latest version data; The generation module is used to generate the business collaboration information for the query request based on the updated candidate data and the candidate data that does not conflict.

[0018] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the information generation method as described above.

[0019] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the information generation method as described above.

[0020] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the information generation method as described above.

[0021] The information generation method, apparatus, electronic device, and storage medium provided by this invention retrieve data from multiple heterogeneous data sources and assign authority levels. When data conflicts are detected, a real-time verification request is initiated to the data source with the highest authority level to obtain the latest version of the data, and the conflicting data is updated accordingly. Finally, business collaboration information is generated based on the updated data. After obtaining candidate data, this invention first assigns authority levels to the data sources. Then, when inconsistencies and conflicts are found among multiple data sources, it does not simply discard or splice the data, but uses the authority level as a judgment criterion to actively trigger the verification process to the highest-level data source to obtain the latest version of the data. This not only breaks down data silos within enterprises but also effectively solves the problem of fragmented decision-making basis caused by the separation of internal and external information. At the same time, by verifying and updating conflicting data to high-authority data sources, it avoids logical contradictions and numerical illusions caused by data timeliness delays or the inability to determine the authenticity of data. This significantly improves the accuracy, consistency, and credibility of the generated business collaboration information, meeting the high standards required for decision-making data in enterprise-level scenarios. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0023] Figure 1 This is a flowchart illustrating the information generation method provided by the present invention.

[0024] Figure 2 This is a schematic diagram of the conflict resolution process prioritizing sovereignty provided by the present invention.

[0025] Figure 3 This is a schematic diagram of the intelligent semantic segmentation workflow provided by the present invention.

[0026] Figure 4 This is a flowchart illustrating another information generation method provided by the present invention.

[0027] Figure 5 This is a schematic diagram of the information generation device provided by the present invention.

[0028] Figure 6 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0030] As enterprises deepen their digital transformation, the acquisition, integration, and collaborative processing of cross-domain heterogeneous data have become crucial for improving enterprise decision-making efficiency. Early enterprise systems commonly adopted a federated search architecture based on inverted indexes and Boolean logic matching, using statistical relevance algorithms to rank search results. With the development of deep learning technology, semantic retrieval based on vector representation has gradually become mainstream. Pre-trained language models such as BERT are used to map text into high-dimensional dense vectors, and semantic matching is achieved through cosine similarity. In recent years, generative large language models (LLMs) based on the Transformer architecture have been widely introduced. Some systems combine LLMs with retrieval-enhanced generation, attempting to introduce external knowledge into the generation process to solve business problems.

[0031] However, despite the progress made in text generation, these technologies still face significant technical bottlenecks when directly applied to complex enterprise-level business collaboration scenarios. On one hand, current systems often lack the ability to semantically process structured data, typically flattening all input data into unstructured text blocks. This leads to a loss of capability when faced with precise query requests involving full data traversal, numerical comparison, or aggregation operations, and makes it difficult to accurately identify temporal semantics, easily causing logical confusion or statistical function failure. On the other hand, and more critically, these technologies suffer from severe fragmentation and conflict issues in cross-source heterogeneous data fusion. Most systems merely use physical splicing to display internal and external information, lacking deep modeling of semantic relationships between entities. Especially when internal private data and external public intelligence contradict each other in terms of status or value—for example, the internal system shows a project as "in progress" while external reports show it as "suspended"—the technologies lack effective conflict resolution mechanisms and cannot dynamically judge and verify based on the authority of the data source. This results in logically contradictory generated content, severely impacting the accuracy and credibility of business collaboration information.

[0032] To address this issue, the present invention provides an information generation method that aims to retrieve candidate data from multiple heterogeneous data sources and assign them authority levels. When a data conflict is detected, a verification request is initiated to the data source with the highest authority level using the principle of sovereignty priority to obtain the latest version of the data. The conflicting data is then updated accordingly. This method generates accurate and logically consistent business collaboration information based on the updated candidate data and the candidate data that has not conflicted, thereby effectively solving the problems of multi-source data conflict and timeliness verification, and improving the reliability of decision-making basis.

[0033] All actions involving the acquisition of signal information or data in this invention are carried out in compliance with the relevant data protection laws and policies of the country where the device is located, and with the authorization granted by the owner of the device.

[0034] in, Figure 1 This is a flowchart illustrating the information generation method provided by the present invention, as shown below. Figure 1 As shown, the method includes steps 110, 120, 130 and 140.

[0035] Step 110: Retrieve candidate data related to the user's query request from multi-source heterogeneous data sources, and assign an authority level to the data source corresponding to each candidate data.

[0036] Here, a user's query request can be understood as natural language text, keyword combinations, or structured instructions entered by the user to obtain specific business information. This query request represents the user's business intent, such as querying project progress, financial indicators, personnel scheduling, or market intelligence. As an optional embodiment, the user's text or voice input can be acquired through the input interface of the terminal device (such as a keyboard, microphone, etc.) and converted into a computer-processable query request. For example, a user's query request could be "query the latest progress of project A" or "statistics on contracts with sales exceeding 1 million last quarter."

[0037] Multi-source heterogeneous data sources refer to a collection of data with diverse origins and varying data structures. Specifically, data sources can include internal business systems such as enterprise resource planning, customer relationship management, and office automation, as well as external public information sources such as industry news websites, public opinion platforms, and public report repositories. The diverse data structures are reflected in the differences in data format and semantic system, encompassing both structured data stored in relational databases, such as financial statements and inventory quantities, and unstructured or semi-structured text data, such as project weekly reports, news reports, and scanned copies of contracts.

[0038] Candidate data refers to data fragments or records retrieved from the aforementioned multi-source heterogeneous data sources based on the user's query request, and which are semantically or logically related to the query request. For example, a query for "Project A" might retrieve structured records with "Project Status: In Progress" from Enterprise Resource Planning, while simultaneously retrieving unstructured text reports from Internet news stating "Project A is suspected to have been suspended." These retrieved raw data constitute candidate data.

[0039] Based on the retrieved candidate data, this embodiment introduces the concept of authority level. Authority level characterizes the credibility, accuracy, and priority of the data source in business decisions. Because data from different sources varies in timeliness, authenticity, and compliance, simple data concatenation cannot meet the stringent accuracy requirements of enterprise applications. Therefore, this embodiment assigns a quantified authority level to the data source corresponding to each candidate data point.

[0040] As an optional implementation, the assignment of authority levels can be based on a pre-defined sovereignty-aware federated knowledge network architecture. This architecture divides data sources into different tiers; for example, the internal core database has the highest authority level because it represents the most authentic and approved operational data within the enterprise; subsidiary databases have a medium authority level; and internet sources have a lower authority level because external information may be outdated or misleading.

[0041] Step 120: If a conflict is detected between at least two candidate data, determine the data source with the highest authority level from the data sources corresponding to the conflicting candidate data, and send a verification request to the data source with the highest authority level so that the data source with the highest authority level returns the latest version of the data for the conflicting candidate data.

[0042] After obtaining multiple candidate data points, due to the diversity and timeliness of data sources, different data sources may have inconsistent descriptions of the same business entity or attribute, resulting in conflicts. These conflicts can manifest as semantic contradictions, such as the internal system displaying "in progress" while the external system displays "paused," or numerical differences, such as an internal record showing an amount of 1.5 million while an external report shows 2 million.

[0043] Furthermore, considering that simply listing conflicting information could lead to difficulties in user decision-making or even be misleading, this embodiment does not directly expose the conflict to the user when a conflict is detected between at least two candidate data points. Instead, it initiates a conflict resolution mechanism. The core of this mechanism lies in identifying which data source has the highest authority level from among the multiple data sources involved in the conflict, based on the authority levels assigned in the aforementioned steps.

[0044] After identifying the data source with the highest authority, this embodiment considers that even authoritative data sources may not have the latest real-time data in their retrieval cache due to system synchronization delays, batch processing update cycles, etc. Directly using this data could lead to outdated authoritative data overwriting newer external information, resulting in new erroneous judgments. Therefore, this embodiment does not directly use the cached data of this data source at the time of retrieval, but instead initiates a verification request. This verification request is a dynamic, real-time interactive process, that is, sending an instruction to the highest authoritative data source to confirm the current real-time state. This ensures that the data used to resolve conflicts is the latest and most accurate record from the authoritative source at that moment, preventing the use of outdated high-authority data due to cache delays or untimely data synchronization.

[0045] By initiating a verification request, this embodiment can obtain the latest version of data from the data source with the highest authority. The latest version of data refers to the most recent and definitive state or value recorded by the data source at the current moment. For example, when a project status conflict is detected, a real-time query interface call is initiated to the internal project management system to retrieve the project's real-time status field in the database.

[0046] Step 130: Update the candidate data that conflict with the latest version data.

[0047] After obtaining the latest version of data from the most authoritative data source, the conflicting parts of the candidate data are updated based on this. The update operation can be understood as a cleaning or overwriting process, that is, replacing candidate data that is inconsistent with the latest version, has a lower authority level, or is outdated with the verified latest version of data with the highest credibility. This ensures that the underlying data used to generate information in the future is logically consistent and up-to-date.

[0048] For example, if the verification results confirm that the project status is "in progress", the conflicting data "paused" originally from external news will be marked as invalid, or its status will be updated to be based on internal authoritative data, thereby eliminating the contradiction at the data level.

[0049] Step 140: Based on the updated candidate data and the candidate data that did not conflict, generate the business collaboration information for the query request.

[0050] After obtaining the updated candidate data, the updated candidate data and the candidate data without conflicts are combined to generate the final business collaboration information. Business collaboration information refers to structured or semi-structured information content directly fed back to users to assist in business decision-making. Business collaboration information is no longer fragmented data pieces, but a complete answer after logical integration, reasoning, and verification. For example, a natural language description can be generated: "As of now, Project A is in progress (data source: internal project management system, updated at 10:00), with a total contract amount of 1.5 million yuan."

[0051] As an optional embodiment, updated candidate data and non-conflicting candidate data can be input into the LLM, which will then generate business collaboration information for the query request.

[0052] The information generation method provided in this embodiment retrieves data from multiple heterogeneous data sources and assigns them authority levels. When a data conflict is detected, a real-time verification request is initiated to the data source with the highest authority level to obtain the latest version of the data. The conflicting data is then updated accordingly, and finally, business collaboration information is generated based on the updated data. After obtaining candidate data, this embodiment first assigns authority levels to the data sources. Then, when inconsistencies and conflicts are found among multiple data sources, it does not simply discard or splice the data. Instead, it uses the authority level as a judgment criterion to actively trigger the verification process to the highest-level data source to obtain the latest version of the data. This not only breaks down data silos within the enterprise but also effectively solves the problem of fragmented decision-making basis caused by the separation of internal and external information. At the same time, by verifying and updating conflicting data from high-authority data sources, it avoids logical contradictions and numerical illusions caused by data timeliness delays or the inability to determine the authenticity of data. This significantly improves the accuracy, consistency, and credibility of the generated business collaboration information, meeting the high standards required for decision-making data in enterprise-level scenarios.

[0053] Given that the candidate data originates from multiple heterogeneous data sources, their data formats (such as structured database records, semi-structured logs, and unstructured news texts) and semantic systems may differ significantly. Directly comparing data fragments as a whole often fails to accurately pinpoint contradictions and may even lead to misjudgments due to format noise. To accurately identify the true differences in data at both the logical and numerical levels, it is necessary to map heterogeneous data to a unified semantic space for fine-grained comparison. Therefore, the conflict detection steps between candidate data in this embodiment include: Extract attribute values ​​relevant to the query request from each candidate data; If the attribute values ​​corresponding to the same attribute of any two candidate data are inconsistent, it is determined that there is a conflict between the two candidate data.

[0054] Specifically, query requests typically focus on specific attributes, such as "project status" or "contract amount." By extracting attribute values ​​relevant to the query request, unstructured text or complex database records can be transformed into specific numerical values ​​or status descriptions on a single dimension, thereby filtering out interference from irrelevant information.

[0055] Here, attribute values ​​can be understood as the specific content, status, or quantified value described for a particular business attribute in the candidate data. For example, for the query request "Current status of project A", the attribute is "Project Status", and the corresponding attribute value might be "In Progress" or "Paused".

[0056] As an optional implementation, natural language processing technology can be used to scan each candidate data, identify the entities contained therein and the descriptive fragments corresponding to the attributes of interest in the query request, and extract them as independent attribute values. For example, from an unstructured news report "Open source framework A is currently suspended from maintenance," the entity "open source framework A" can be extracted, with the attribute "status" and the attribute value "suspended"; from records in an enterprise's internal system, the entity "project A" can be extracted, with the attribute "status" and the attribute value "in progress."

[0057] After extracting attribute values ​​from each candidate data set, considering that only descriptions of the same attribute are comparable, it is necessary to align and compare the extracted attribute values ​​to identify potential contradictions. Based on this, this embodiment determines a conflict between any two candidate data sets if the attribute values ​​corresponding to the same attribute are inconsistent.

[0058] This embodiment ensures the correctness of the logical premise of the comparison by locking the same attribute; by judging the inconsistency of attribute values, it can keenly capture the abnormal situation of semantically opposite numerical deviation, thereby determining the existence of conflict.

[0059] Here, "same attribute" refers to a dimension described by two candidate data points that has the same business meaning. For example, both data sources might describe "sales revenue in the first quarter of 2024." Inconsistent attribute values ​​mean that two attribute values ​​are semantically contradictory or differ beyond the permissible error range in numerical statistics.

[0060] As an optional implementation, any two candidate data can be mapped to the same comparison dimension. For example, suppose candidate data A comes from an internal system and contains the attribute "contract amount" with a value of "1.5 million"; candidate data B comes from an external public report and contains the same attribute "contract amount" with a value of "2 million". A comparison reveals a significant discrepancy between "1.5 million" and "2 million", thus determining a conflict between candidate data A and candidate data B regarding the "contract amount" attribute. Similarly, for the attribute "project progress", if candidate data A has the attribute value "in progress" while candidate data B has the attribute value "paused", their semantics are completely opposite, again indicating a conflict between the two candidate data.

[0061] Considering that in actual business collaboration scenarios, attribute values ​​extracted from multi-source heterogeneous data sources often exist in the form of unstructured text, and given the ambiguity, synonymy, and diversity of expression in natural language (e.g., "project start" and "project begin" express the same meaning, while "project paused" and "project terminated" are semantically similar despite their different literal meanings), using only traditional string matching methods for comparison can easily misjudge attribute values ​​with different expressions but identical meanings as conflicts, or fail to identify attribute values ​​with similar literal meanings but opposite meanings. To achieve a deeper semantic understanding and accurate determination of the logical consistency of text-based data, this embodiment, after extracting attribute values ​​related to the query request from each candidate data, also includes: When the attribute value is text, convert the attribute values ​​corresponding to the same attribute of any two candidate data into semantic vectors; Calculate the similarity between the semantic vectors corresponding to the same attribute of any two candidate data; If the similarity is below the first threshold, it is determined that the attribute values ​​corresponding to the same attribute of any two candidate data are inconsistent.

[0062] Typically, computers cannot directly understand the deep meaning of human language and need to map abstract text characters into high-dimensional numerical values ​​that computers can compute. By converting them into semantic vectors, the semantic information of the text can be encoded into mathematical representations, making semantically similar words and phrases closer together in the vector space. Here, a semantic vector refers to a high-dimensional real-number vector containing semantic features generated after encoding text through a natural language processing model.

[0063] As an alternative implementation, semantic vectors can be obtained using a pre-trained language model. Specifically, the extracted attribute values ​​are input into the pre-trained language model. Through learning from massive amounts of corpus data, the model has mastered the contextual relationships between words, thereby outputting semantic vectors of fixed dimensions representing the two attribute values. For example, attribute value A can be converted into semantic vector V. A Convert attribute value B into semantic vector V B .

[0064] After converting attribute values ​​into semantic vectors, considering that geometric distance or angle in vector space can reflect the degree of semantic similarity, this embodiment calculates the similarity between the semantic vectors corresponding to the same attribute of any two candidate data in order to quantitatively evaluate the semantic consistency between two attribute values. By calculating the similarity, semantic differences that are difficult to judge intuitively can be transformed into a specific value between 0 and 1, thereby providing a quantitative basis for determining conflicts.

[0065] As an alternative implementation, a cosine similarity algorithm can be used to calculate the similarity between two semantic vectors. Cosine similarity determines the similarity of two vectors by measuring the cosine of the angle between them in a multidimensional space.

[0066] After obtaining the similarity score, considering that this score only represents the semantic closeness, a clear standard is needed to define the boundary between "consistency" and "conflict." Based on this, this embodiment flexibly controls the system's tolerance for semantic differences by setting a first threshold. When the similarity score is below the first threshold, it means that the angle between the two attribute values ​​in the vector space is large, indicating a significant semantic distance. This suggests that the content described by the two attributes has an essential difference or logical contradiction, and therefore they are judged as inconsistent. Here, the first threshold is a pre-set critical value used to determine whether there is a semantic conflict.

[0067] As an optional implementation, assume the first threshold is set to 0.85. If attribute value A is "in progress" and attribute value B is "in execution," the calculated cosine similarity is 0.95, which is greater than 0.85. Therefore, they are considered to be consistent and there is no conflict. If attribute value A is "in progress" and attribute value C is "paused," since their semantics are opposite, the calculated cosine similarity may be 0.3, which is lower than 0.85. Based on this, it is determined that the attribute values ​​corresponding to the same attribute of these two candidate data are inconsistent, i.e., there is a conflict. Through this dynamic determination mechanism based on semantic vectors, logical contradictions in cross-domain entity associations can be effectively identified, making up for the shortcomings of traditional string matching methods.

[0068] Considering that enterprise business data includes not only textual descriptions but also a large amount of numerical data such as contract amounts, completion percentages, and timestamps, conflicts in this type of data often manifest not as semantic contradictions but as quantitative numerical discrepancies. Directly applying textual semantic similarity calculations might overlook the mathematical meaning and precision requirements of the numerical values ​​themselves; for example, "1 million" and "1.01 million" may be semantically similar but numerically inconsistent. To accurately identify contradictions in numerical data, this embodiment extracts attribute values ​​relevant to the query request from each candidate data set, and then further includes: When the attribute value is numerical, calculate the degree of difference between the numerical values ​​corresponding to the same attribute of any two candidate data. If the difference exceeds the second threshold, it is determined that the attribute values ​​corresponding to the same attribute of any two candidate data are inconsistent.

[0069] Specifically, if the attribute value is numerical, then by calculating the degree of difference between the numerical values ​​corresponding to the same attribute of any two candidate data, the relationship between the two values ​​can be transformed into a relative and measurable index, thereby eliminating the influence of units and facilitating the establishment of a unified judgment standard. Here, the degree of difference refers to an index that measures the relative difference between two values. The degree of difference can be determined by calculating the relative error.

[0070] After obtaining the degree of difference between the values, considering that in actual business operations, a certain tolerance range is often allowed for indicators of different precision—for example, financial statements may require accuracy to the minute, while project progress estimates may allow for an error of 5%—it is not possible to simply require the values ​​to be completely equal. Based on this, this embodiment sets a second threshold to introduce a flexible fault-tolerance mechanism for numerical comparison.

[0071] When the difference exceeds the second threshold, it indicates that the numerical deviation between the two data sources has exceeded the acceptable range for business purposes and should be considered a substantial conflict. Here, the second threshold is a pre-set critical percentage used to determine whether the numerical difference constitutes a conflict.

[0072] As an optional embodiment, assume the second threshold is set to 5%. If the "contract amount" in candidate data A is 1 million yuan and the "contract amount" in candidate data B is 1.02 million yuan, the difference calculated according to the above formula is: |100-102| / 102≈1.96%. Since 1.96% is less than 5%, the system can determine that the two are consistent within the allowable error range. However, if the "contract amount" in candidate data C is 1.5 million yuan, the difference is: |100-150| / 150≈33.3%. Since 33.3% is significantly greater than 5%, the system determines that the attribute values ​​corresponding to the same attribute of the two candidate data are inconsistent, that is, there is a numerical conflict. Through this difference-based judgment mechanism, false alarms caused by small data fluctuations or differences in statistical methods can be effectively prevented, while accurately capturing significant data contradictions.

[0073] Given the diverse types of enterprise data sources and the varying reliability of different data sources across different business scenarios, simple static grading is insufficient to adapt to complex and ever-changing business needs. For example, internal systems are generally more reliable than external news, but when specific sensitive business matters are involved, the authority requirements for data sources are higher. To construct a dynamic authority assessment system that reflects both the differences in data sources and the sensitivity of business needs, this embodiment assigns authority levels to the data sources corresponding to each candidate data, including: Assign an initial authority level to each data source based on its type; Based on the data sensitivity of each candidate data, the initial authority level of the data source corresponding to each candidate data is adjusted to obtain the authority level of the data source corresponding to each candidate data. The data sensitivity of each candidate data is determined based on the business category to which the attribute contained in the corresponding candidate data belongs.

[0074] Considering that the type of data source is the most fundamental and direct basis for judging its data quality, assigning an initial authority level can establish a baseline trust hierarchy architecture and quickly distinguish core data from reference data. Therefore, this embodiment assigns an initial authority level to each data source based on its type. The data source type refers to the physical environment and management attributes of the data generation, storage, or publication, such as whether it belongs to a controlled internal network or the external open internet. The initial authority level refers to a preset basic trust score based on the data source's physical attributes and management affiliation.

[0075] As an optional implementation, a three-tier data architecture can be constructed and levels assigned accordingly. This three-tier data architecture includes an internal core database, subsidiary databases, and internet information sources. The internal core database, as the single source of factual information for the enterprise's operations, is assigned the highest initial authority level; the subsidiary databases, being internal data but at a lower level, are assigned a medium initial authority level; and the internet information sources, as external reference information and subject to uncertainty, are assigned the lowest initial authority level.

[0076] After assigning initial authority levels to each data source, considering that data within the same data source may contain business information of varying sensitivity (e.g., employee names are general information, while salary data is highly sensitive), treating all data equally might introduce risks due to insufficient authority when processing highly sensitive data. Therefore, it is necessary to fine-tune the authority levels based on the specific business attributes of the data. Based on this, this embodiment adjusts the initial authority levels of the data sources corresponding to each candidate data point according to their data sensitivity, thus obtaining the authority levels of the data sources corresponding to each candidate data point.

[0077] By introducing a data sensitivity dimension, more granular access control and trust management can be achieved. For highly sensitive data, the source is given higher authority, or the weight of internal sources is increased while the weight of external sources is decreased. This allows for more prudent conflict resolution, ensuring the security and accuracy of core business data.

[0078] The data sensitivity of each candidate data point is determined based on the business category to which the attributes contained in the corresponding candidate data belong. Here, data sensitivity refers to the degree to which a data breach or error impacts the company's business. Business category refers to the classification of data attributes according to the company's management functions.

[0079] As an optional embodiment, a mapping relationship between business categories and sensitivity can be preset, thereby determining the data sensitivity of candidate data based on this mapping relationship. Specifically, this mapping relationship can be as follows: financial attributes are considered high sensitivity, corresponding to a high sensitivity level; personnel attributes are considered high sensitivity, corresponding to a high sensitivity level; project attributes are considered medium sensitivity, corresponding to a medium sensitivity level; and public attributes are considered low sensitivity, corresponding to a low sensitivity level.

[0080] For example, for candidate data containing "financial attributes," if it originates from an "internal core database," its authority weight can be further enhanced, giving it an overwhelming advantage when it conflicts with external financial reports. Conversely, if highly sensitive data comes from a low-authority "internet source," its effective authority level may be reduced, or it may even be directly marked as untrustworthy in conflict detection. This dynamic adjustment mechanism ensures that the final authority level reflects both the origin of the data source and the current business importance of the data.

[0081] Considering that not all data requires equal adjustment when adjusting the initial authority level, the initial authority level may be sufficient to reflect the credibility of data with moderate sensitivity. However, for highly sensitive data, relying solely on the initial authority level may not fully reflect its highest judgment weight or priority of acceptance in conflict resolution. To further strengthen the protection and authority confirmation of key data, this embodiment adjusts the initial authority level of the data source corresponding to each candidate data according to the data sensitivity of each candidate data, including: If the data sensitivity of any candidate data exceeds a preset sensitivity threshold, the initial authority level of the data source corresponding to any candidate data will be increased.

[0082] Specifically, when data sensitivity is high, it means that the accuracy of the candidate data has a significant impact on business operations and a low tolerance for error. By increasing the authority level of the corresponding data source when the sensitivity exceeds a preset sensitivity threshold, the weight of the candidate data in the conflict detection and resolution process can be increased. This ensures that when facing high-risk decisions, there is a greater tendency to accept data marked as highly sensitive and from reliable sources, or that when conflicts occur between data sources, the data source carrying the highly sensitive data is prioritized for verification.

[0083] Here, the preset sensitivity threshold refers to a critical reference value set to define highly sensitive data. Data with a sensitivity higher than this threshold means that the data belongs to the company's core secrets or key business indicators.

[0084] As an optional implementation, assume that data sensitivity is quantified into a score from 1 to 10, and the preset sensitivity threshold is set to 8. If a candidate data has a data sensitivity score of 9, and the data originates from the "internal core database," with an initial authority level of 90, then an enhancement mechanism is triggered to temporarily raise the authority level of the data source for that data to 95, or to attach a "high priority" label to it.

[0085] Conversely, if the data sensitivity is lower than the preset sensitivity threshold, the initial authority level remains unchanged, and the data is processed according to the normal procedure.

[0086] Considering that user queries often implicitly contain specific requirements for data timeliness—for example, when a user asks for "latest updates," their intention is to obtain the current status, not historical records—if the retrieval process ignores this timeliness requirement and only returns all relevant data based on keyword matching, it may result in a large amount of outdated information being mixed in, increasing the burden of subsequent conflict detection and even leading to misjudgments. To accurately respond to users' timeliness needs, this embodiment retrieves candidate data related to the user's query request from multi-source heterogeneous data sources, including: Perform semantic parsing on the query request to extract the temporal semantic features from the query request; When the temporal semantic features contain keywords indicating the acquisition of the latest data, the candidate data with the latest timestamp is retrieved first, based on the timestamp carried by each candidate data.

[0087] Given the diverse ways time can be expressed in natural language, such as "latest," "current," "last month," and "2024," direct keyword matching cannot cover all scenarios. Semantic parsing allows us to abstract standardized temporal semantic features from natural language, thereby understanding the specific constraints users place on the time dimension.

[0088] Here, temporal semantic features refer to semantic tags or structured information extracted from query requests to describe the time attributes or timeliness requirements of data. For example, temporal semantic features may include timestamps, relative temporal positions, business cycle identifiers, etc.

[0089] As an optional implementation, a preset time-series keyword library or time-series recognition model can be used to identify keywords such as latest, current, now, real-time, and recently updated that indicate the latest data, as well as keywords such as history, past, formerly, and before that indicate historical data. When a user initiates a request to query the latest status of item A, the word "latest" contained within it is identified as belonging to the words in the preset time-series keyword library used to indicate the latest status, thereby determining that the time-series semantic features corresponding to the query request contain keywords indicating the latest data.

[0090] After extracting temporal semantic features, considering that users explicitly request the latest data, the retrieval strategy should shift from comprehensive coverage to timeliness priority to reduce interference from invalid data. Utilizing the timestamp information inherent in the data for sorting and filtering allows for the direct filtering out of outdated data during the retrieval stage, ensuring that the returned candidate data best meets user needs in terms of timeliness, thereby improving the accuracy of retrieval results and the efficiency of subsequent processing.

[0091] Based on this, in this embodiment, when the temporal semantic features contain keywords indicating the retrieval of the latest data, the candidate data with the latest timestamp is retrieved first, according to the timestamp carried by each candidate data. Here, a timestamp refers to the specific point in time information attached to a data record, used to identify the data's generation, update, or the business cycle to which it belongs. Furthermore, keywords indicating the retrieval of the latest data include, but are not limited to, latest, current, and now.

[0092] When the temporal semantic features contain keywords indicating the acquisition of the latest data, it indicates that the user intends to obtain the data status at the current moment. In this case, based on the timestamps carried by each candidate data, the relevant data are sorted in descending order of timestamps, and the data record at the top of the sort is selected as the search result. This ensures that the returned data is up-to-date in terms of timeliness and avoids temporal logic errors caused by referencing historical data.

[0093] Considering that enterprise-level business data is mostly stored in structured form in relational databases or business systems, traditional semantic retrieval often flattens all inputs into unstructured text blocks. When users make requests involving precise numerical comparisons or aggregation operations, fuzzy matching based on vector similarity struggles to perform deterministic logical operations, easily leading to information omissions or statistical errors. To compensate for the lack of semantic processing of structured data, this embodiment retrieves candidate data related to the user's query request from multi-source heterogeneous data sources, including: Perform intent recognition on the query request to determine the type of query request; When the query request type is a numeric query type, the query request is converted into a structured query statement, and candidate data is retrieved from multi-source heterogeneous data sources based on the structured query statement.

[0094] Considering that not all query requests require complex logical calculations, intent recognition can categorize queries into different types, such as text-based queries for query progress and numerical queries for query counts, thus providing a basis for adopting different retrieval strategies subsequently.

[0095] Here, the query request type refers to the result of categorizing the query content based on user intent. Common types include, but are not limited to, numerical query types and text progress query types.

[0096] As an optional implementation, a first classification result can be obtained by analyzing the overall semantics of the query statement using a deep learning model; simultaneously, a second classification result can be obtained by scanning the query for specific feature words using semantic feature engineering. The type of query request is then determined by combining the first and second classification results.

[0097] After determining the type of query request, considering that directly retrieving text fragments cannot meet the needs of precise calculation for numerical query types, it is necessary to utilize the computing power of the database itself. Since Structured Query Language (SQL) is the standard language for interacting with relational databases, capable of precisely performing logical operations such as numerical comparison, range filtering, summation, and counting, this embodiment converts the query request into a structured query statement when the query type is numerical. Based on the structured query statement, candidate data is retrieved from multi-source heterogeneous data sources, transforming ambiguous questions into precise calculation instructions. The underlying computing engine of the database directly obtains accurate results, fundamentally avoiding the illusions that generative models may produce when processing numerical values. Here, a structured query statement refers to instruction code that conforms to specific database syntax rules and is used to perform data queries and operations, such as Structured Query Language (SQL).

[0098] Given the significant gap in syntactic structure and expressive logic between natural language and structured query statements, direct end-to-end translation can easily lead to unexecutable statements or incorrect results due to semantic ambiguity. To improve the accuracy and robustness of the translation, it is necessary to first decompose the natural language into intermediate structured elements and then assemble them. Based on this, the conversion of query requests into structured query statements in this embodiment includes: Extract the triples of entity, attribute, and operation from the query request; Generate structured query statements for multi-source heterogeneous data sources based on triples.

[0099] Here, an entity refers to the business object that the query points to, such as a project or contract; an attribute refers to the specific characteristics or indicators of an entity, such as amount or status; and an operation refers to the logical judgment or calculation performed on the attribute, such as greater than, equal to, or sum.

[0100] As an optional implementation, by performing word segmentation and dependency parsing on the query request, the subject-verb-object structure and modification relationship in the sentence are identified, and then the nouns representing business objects are located as entities, the nouns representing data features are located as attributes, and the verbs or adjectives representing comparison or aggregation logic are located as operations, thereby extracting triples of entities, attributes and operations from the query request.

[0101] For example, for the query request "Statistics on orders with sales exceeding 5 million in zone H", the extracted triple is: entity = "order", attribute = "sales", operation = ">5 million AND zone = zone H".

[0102] After successfully extracting the core triples, considering that these elements are currently at the business semantic level, they need to be translated into instruction code that the database can understand. Based on this, this embodiment generates structured query statements for multi-source heterogeneous data sources based on the triples, thereby accurately translating the user's natural language intent into the database's execution logic and keeping the error rate of numerical calculations at a low level.

[0103] As an optional implementation, a mapping knowledge base of "business semantics-database schema" can be maintained. This knowledge base stores the correspondence between business entities and database table names, and business attributes and database field names. First, based on this mapping knowledge base, the entities in the extracted triples are mapped to the target query table, the attributes are mapped to the query fields, and the operations are mapped to the query conditions. Then, the mapped table names, field names, and conditions are concatenated according to the syntax rules of the structured query language to generate a structured query statement for multi-source heterogeneous data sources.

[0104] in, Figure 2 This is a schematic diagram of the conflict resolution process prioritizing sovereignty provided by the present invention, as shown below. Figure 2 As shown, the process includes: First, responding to query requests and performing multi-source data acquisition operations. The data sources cover a three-tier architecture, including the internal core database as the highest authority, the subsidiary database as the intermediate authority, and the internet information source as the lowest authority.

[0105] Then, based on the data source type and data sensitivity, dynamic confidence weights are calculated to assign an initial authority level to each candidate data.

[0106] Next, the process moves to the dual-threshold conflict detection stage. Specifically, semantic conflict detection is performed using semantic vector similarity, and numerical conflict detection is performed using difference calculation. These two detection methods are used to identify conflicts and determine whether data conflicts exist.

[0107] Subsequently, if no conflict exists, the process proceeds directly to the information generation stage. If a conflict exists, a secondary confirmation mechanism is triggered, which involves sending a verification request to the data source with the highest authority to obtain the latest version of the data.

[0108] After obtaining the latest version of the data, a confidence reassessment is performed, the data status is updated based on the latest version of the data, and the knowledge graph is updated to correct conflicting data and generate results with conflict resolution tags.

[0109] Finally, based on the conflict-resolved and updated data, LLM is used to generate business collaboration information with traceability tags. The generated information is not only logically consistent, but also includes timestamped data source tags, supporting penetrating traceability verification.

[0110] Considering the limitations of single intent recognition methods, while deep learning-based models have strong generalization capabilities, they may exhibit biases when handling specific industry terminology. Rule-based keyword matching, although accurate, struggles to cover complex natural language expressions. To improve the accuracy and robustness of intent recognition, this embodiment performs intent recognition on query requests to determine the type of query request, including: When the query request is a text query, the query request is classified based on a pre-trained classification model to obtain the first classification result; Extract feature words from the query request, and classify the query request based on the matching results of the feature words with the preset feature word library to obtain the second classification result; If the first classification result and the second classification result are the same, the first classification result or the second classification result will be used as the type of query request.

[0111] Specifically, the pre-trained classification model can capture the deep semantics and contextual information of query requests, demonstrating good recognition capabilities for queries with varied sentence structures and ambiguous expressions. Through the pre-trained classification model, query requests can be mapped to a high-dimensional semantic space, and the probability of belonging to a preset category is predicted based on the learned probability distribution. The category with the highest probability is selected as the prediction result, yielding the first classification result. Here, the first classification result refers to the predicted label about the type of query request obtained through model inference. Since the first classification result is based on semantic probability inference trained on a large-scale corpus, it provides a generalized judgment of query intent from the perspectives of deep semantic understanding and contextual association. The pre-trained model can be a BERT model based on the Transformer architecture, fine-tuned for a specific business domain.

[0112] Meanwhile, to compensate for potential black-box uncertainties in the model, and since business queries often contain highly indicative keywords, explicit feature word matching can quickly and accurately pinpoint user intent. Therefore, this embodiment also extracts feature words from the query request and classifies the query request based on the matching results of the feature words with a preset feature word library, obtaining a second classification result. Here, feature words refer to words in the query request that have clear business meaning or intent. The preset feature word library is a collection containing a large number of industry terms and their corresponding intent types. Since the second classification result is based on precise matching of specific business keywords, it provides accurate positioning of query intent from the perspectives of explicit features and business rules.

[0113] After obtaining two classification results, considering that the credibility is highest when the judgment results of different mechanisms corroborate each other, this embodiment, when the first classification result and the second classification result are consistent, uses the first classification result or the second classification result as the type of query request, thereby significantly reducing the misclassification rate of a single path.

[0114] For example, if the classification model identifies the query as a "numerical query," and the feature word matching also finds the keyword "how much," then the query request type is determined to be a "numerical query type." If the two are inconsistent, manual review or a rejection mechanism can be triggered to prevent erroneous execution.

[0115] in, Figure 3 This is a schematic diagram of the intelligent semantic segmentation workflow provided by the present invention, such as... Figure 3 As shown, the system first receives user query requests and performs query preprocessing. This preprocessing can involve operations such as word segmentation, stop word removal, part-of-speech tagging, and error correction to clean up noisy data and extract key information.

[0116] Next, a chatter detection function is performed to determine if a request is indeed a chatter question. If it is, a rejection mechanism is triggered, a fallback response is generated, and the process ends, avoiding wasting computational resources on invalid requests.

[0117] If the question is not casual conversation, the process proceeds to the dual-path intent recognition stage, which involves launching two recognition paths in parallel. Specifically, one path is based on a domain-fine-tuned binary classifier model for prediction. This model is trained using the company's historical query logs for domain adaptation, calculating the probabilities of progress-related and question-related categories respectively. The other path is based on semantic feature engineering matching. This engineering maintains a dynamically configurable lexicon, including a progress-related feature lexicon (containing temporal semantic words such as "progress," "status," "latest," and "current") and a question-related feature lexicon (containing quantitative semantic words such as "how much," "amount," "statistics," and "percentage"). Progress-related feature matching and question-related feature matching are used to assist in the judgment.

[0118] Subsequently, the confidence scores of the two paths are evaluated, and it is determined whether the confidence score is ≥0.85. If the confidence score is ≥0.85, the query request type is further determined. If it is a text query, it is routed to the text processing path; if it is a numerical query, it is routed to the numerical processing path, and the intent recognition process ends.

[0119] If the confidence level is <0.85, a dynamic routing strategy is activated, triggering a manual review mechanism. Manual confirmation by business administrators not only ensures a single classification accuracy >90%, but also accumulates high-quality labeled data. The results of this manual confirmation will be used to update the training dataset and fed back into the incremental learning framework.

[0120] Finally, based on the incremental learning framework, misclassification cases are automatically collected, and the model is regularly fine-tuned and updated to optimize the binary classifier model. At the same time, business administrators can update the feature word library through a visual interface, thereby achieving adaptability to business scenarios.

[0121] While LLM excels in natural language generation, it can still produce illusions when handling rigorous numerical calculations or logical reasoning, generating seemingly reasonable but actually erroneous content. To ensure the absolute reliability of the generated business collaboration information in terms of logic and numerical values, the business collaboration information for generating query requests in this embodiment includes: Based on the updated candidate data and the candidate data that did not conflict, the first information is generated through the prediction model; Based on the updated candidate data and the candidate data that did not conflict, a second piece of information is generated through logical calculation; If the deviation between the first information and the second information exceeds a preset deviation threshold, the second information is used as the business collaboration information.

[0122] Leveraging the powerful language organization capabilities of predictive models (such as LLM), fragmented data snippets can be integrated into fluent and natural business responses, providing a superior user experience. Based on this, this embodiment inputs updated candidate data and non-conflicting candidate data into the predictive model to generate first information. This first information refers to natural language text or descriptive content containing preliminary calculation results, generated by the predictive model based on the input data.

[0123] Meanwhile, for content involving numerical aggregation, comparison, or deterministic logical judgment, model-based generation may lead to probabilistic biases or logical errors due to the deterministic and zero-error characteristics of logical calculations. Logical calculations can yield a precise and reliable reference answer, serving as a benchmark for verifying model generation. Therefore, this embodiment also generates second information based on updated candidate data and conflict-free candidate data through logical calculations. This second information refers to the precise numerical value or logical conclusion obtained after calculating the data using deterministic algorithms or logical rules.

[0124] After obtaining the first and second information, considering that a significant difference between them usually indicates a decrease in the credibility of the model's generated results, although the text generated by the model may be more in line with user reading habits, its accuracy in key numerical values ​​may not be as good as that of logical calculations. Therefore, in this embodiment, when the deviation between the first and second information exceeds a preset deviation threshold, the second information, derived from deterministic logical calculations, possesses mathematical precision and logical uniqueness, exhibiting higher accuracy and credibility compared to model predictions based on probability generation. Thus, this embodiment uses the second information as business collaboration information, prioritizing the adoption of logical calculation results to correct the model-generated content. Here, the preset deviation threshold refers to the maximum allowable difference between the model-generated results and the logical calculation results, set to tolerate a certain degree of natural fluctuation in model generation while ensuring data accuracy.

[0125] Given that enterprise data often involves sensitive information such as trade secrets and personnel privacy, employees at different levels or in different departments should have different data access permissions. Indiscriminately scanning all data sources during retrieval not only wastes computing resources but also poses a serious risk of data leakage. To ensure the compliance and security of data retrieval, this embodiment retrieves candidate data related to the user's query request from multi-source heterogeneous data sources, including: Based on the user's permission information, determine the range of data sources that the user can access; Retrieve candidate data relevant to the user's query request from the range of data sources.

[0126] Considering that access control is the basic security barrier of enterprise information systems, if access boundaries are not defined in advance, any retrieval operation may become a potential data leakage path. This embodiment needs to determine the range of data sources that users can access. This range clearly defines the boundaries of data sources that the current user has legal read permissions in this session, thereby blocking illegal access from the source and ensuring that users can only see the data they are authorized to view, thus realizing the principle of least privilege.

[0127] Specifically, this embodiment determines the range of data sources that a user can access based on the user's permission information. Here, the user's permission information refers to the user's identity identifier and the associated access control policy. The data source range refers to the set of data sources that the user is allowed to query.

[0128] As an optional implementation, when a user initiates a query request, their identity tag is read. Based on preset permission rules, the range of data sources that the user can access is determined.

[0129] After determining the scope of secure data sources, and considering the need to further improve retrieval efficiency and prevent out-of-bounds retrieval, this embodiment retrieves candidate data related to the user's query request from the scope of the data sources. This strictly limits the retrieval operation to a compliant range, ensuring data security and improving retrieval speed by narrowing the search space.

[0130] in, Figure 4 This is a flowchart illustrating another information generation method provided by the present invention, as shown below. Figure 4 As shown, the process includes: First, users upload historical documents to the database, where data is sliced ​​and vectorized. Based on this, a temporal knowledge graph is constructed, preserving the entity-attribute-value structure, and a timestamp fingerprint is implanted into each data entry to ensure the semantic integrity and temporal traceability of the structured data. Simultaneously, a sovereign-aware federated knowledge network is established, integrating the internal core database, subsidiary databases, and internet information sources to form a multi-level data architecture.

[0131] Then, when a user submits a query request, semantic decomposition is initiated, using a dual-path intent recognition engine to identify the type of query request. If it is identified as a numerical query, a key element extractor performs key element extraction, identifying the "entity + numerical attribute + comparison operation" triple and converting it into a structured query statement. Next, a numerical inference engine directly performs aggregation calculations in the database to obtain accurate results. To prevent the large model from generating erroneous values, an anti-illusion verification is performed, comparing the logical calculation results with the model prediction results to ensure the accuracy of the values.

[0132] If the problem is identified as a text query, the system first checks if there are any data conflicts among the retrieved candidate data. If so, a secondary verification mechanism is triggered. Based on the principle of sovereignty priority, a verification request is sent to the data source with the highest authority level to obtain the latest version of the data to resolve the conflict. If not, or after the conflict has been resolved, time window pruning is performed using timestamp fingerprints to filter out candidate data that meets the timeliness requirements.

[0133] Finally, the processing results for the question and progress categories are input into the large model. The large model integrates the verified accurate values, the status information after conflict resolution, and the relevant context to generate structured business collaboration information with data traceability tags, and finally feeds it back to the user.

[0134] The information generation apparatus provided by the present invention will be described below. The information generation apparatus described below and the information generation method described above can be referred to in correspondence.

[0135] Based on any of the above embodiments Figure 5 This is a schematic diagram of the information generation device provided by the present invention, as shown below. Figure 5 As shown, the device includes: The allocation module 510 is used to retrieve candidate data related to the user's query request from multi-source heterogeneous data sources and assign an authority level to the data source corresponding to each candidate data. The determination module 520 is used to determine the data source with the highest authority level from the data sources corresponding to the conflicting candidate data when a conflict is detected between at least two candidate data, and to initiate a verification request to the data source with the highest authority level so that the data source with the highest authority level returns the latest version of the data for the conflicting candidate data. Update module 530 is used to update conflicting candidate data based on the latest version of the data; The generation module 540 is used to generate business collaboration information for the query request based on the updated candidate data and the candidate data that has not conflicted.

[0136] Figure 6 This is a schematic diagram of the structure of the electronic device provided by the present invention, such as... Figure 6 As shown, the electronic device may include a processor 610, a communications interface 620, a memory 630, and a communication bus 640. The processor 610, communications interface 620, and memory 630 communicate with each other via the communication bus 640. The processor 610 can call logical instructions stored in the memory 630 to execute an information generation method.

[0137] Furthermore, the logical instructions in the aforementioned memory 630 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0138] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer is able to execute the information generation methods provided by the above methods.

[0139] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the information generation methods provided by the methods described above.

[0140] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0141] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0142] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An information generation method, characterized in that, include: Retrieve candidate data related to the user's query request from multiple heterogeneous data sources, and assign an authority level to the data source corresponding to each candidate data; If a conflict is detected between at least two candidate data, the data source with the highest authority level is determined from the data sources corresponding to the conflicting candidate data, and a verification request is sent to the data source with the highest authority level so that the data source with the highest authority level returns the latest version of the data for the conflicting candidate data. Update the conflicting candidate data based on the latest version data; Based on the updated candidate data and the candidate data that do not conflict, the business collaboration information for the query request is generated.

2. The information generation method according to claim 1, characterized in that, The conflict detection steps between candidate data include: Extract attribute values ​​related to the query request from each candidate data; If the attribute values ​​corresponding to the same attribute of any two candidate data are inconsistent, it is determined that there is a conflict between the two candidate data.

3. The information generation method according to claim 2, characterized in that, After extracting attribute values ​​related to the query request from each candidate data, the process further includes: When the attribute value is text, the attribute values ​​corresponding to the same attribute of any two candidate data are converted into semantic vectors; Calculate the similarity between the semantic vectors corresponding to the same attribute of any two candidate data; If the similarity is lower than a first threshold, it is determined that the attribute values ​​corresponding to the same attribute of any two candidate data are inconsistent.

4. The information generation method according to claim 2, characterized in that, After extracting attribute values ​​related to the query request from each candidate data, the process further includes: When the attribute value is numerical, calculate the degree of difference between the numerical values ​​corresponding to the same attribute of any two candidate data. If the difference exceeds the second threshold, it is determined that the attribute values ​​corresponding to the same attribute of any two candidate data are inconsistent.

5. The information generation method according to any one of claims 1 to 4, characterized in that, Assigning authority levels to the data sources corresponding to each candidate data includes: Assign an initial authority level to each data source based on its type; Based on the data sensitivity of each candidate data, the initial authority level of the data source corresponding to each candidate data is adjusted to obtain the authority level of the data source corresponding to each candidate data. The data sensitivity of each candidate data is determined based on the business category to which the attribute contained in the corresponding candidate data belongs.

6. The information generation method according to claim 5, characterized in that, The adjustment of the initial authority level of the data source corresponding to each candidate data based on the data sensitivity of each candidate data includes: If the data sensitivity of any candidate data exceeds a preset sensitivity threshold, the initial authority level of the data source corresponding to that candidate data is increased.

7. The information generation method according to any one of claims 1 to 4, characterized in that, The step of retrieving candidate data related to the user's query request from multi-source heterogeneous data sources includes: The query request is semantically parsed to extract the temporal semantic features from the query request; If the temporal semantic features contain keywords indicating the acquisition of the latest data, then candidate data with the latest timestamp will be retrieved first, based on the timestamp carried by each candidate data.

8. The information generation method according to any one of claims 1 to 4, characterized in that, The step of retrieving candidate data related to the user's query request from multi-source heterogeneous data sources includes: The query request is subjected to intent recognition to determine the type of the query request; If the query request is a numeric query, the query request is converted into a structured query statement, and candidate data is retrieved from the multi-source heterogeneous data source based on the structured query statement.

9. The information generation method according to claim 8, characterized in that, The step of converting the query request into a structured query statement includes: Extract the triple of entity, attribute, and operation from the query request; Generate a structured query statement for the multi-source heterogeneous data source based on the triples.

10. The information generation method according to claim 9, characterized in that, The step of performing intent recognition on the query request and determining the type of the query request includes: If the type of the query request is a text query, the query request is classified based on a pre-trained classification model to obtain a first classification result; Feature words are extracted from the query request, and the query request is classified according to the matching results of the feature words and the preset feature word library to obtain a second classification result; If the first classification result is consistent with the second classification result, the first classification result or the second classification result shall be used as the type of the query request.

11. The information generation method according to any one of claims 1 to 4, characterized in that, The business collaboration information that generates the query request includes: Based on the updated candidate data and the candidate data that did not conflict, first information is generated through a prediction model; Based on the updated candidate data and the candidate data that did not conflict, second information is generated through logical calculation; If the deviation between the first information and the second information exceeds a preset deviation threshold, the second information is used as the business collaboration information.

12. The information generation method according to any one of claims 1 to 4, characterized in that, The step of retrieving candidate data related to the user's query request from multi-source heterogeneous data sources includes: Based on the user's permission information, determine the range of data sources that the user can access; Retrieve candidate data related to the user's query request from the range of the data source.

13. An information generation device, characterized in that, include: The allocation module is used to retrieve candidate data related to the user's query request from multi-source heterogeneous data sources and assign an authority level to the data source corresponding to each candidate data. The determination module is used to determine the data source with the highest authority level from the data sources corresponding to the conflicting candidate data when a conflict is detected between at least two candidate data, and to initiate a verification request to the data source with the highest authority level so that the data source with the highest authority level returns the latest version data for the conflicting candidate data. The update module is used to update conflicting candidate data based on the latest version data; The generation module is used to generate the business collaboration information for the query request based on the updated candidate data and the candidate data that does not conflict.

14. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the information generation method as described in any one of claims 1 to 12.

15. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the information generation method as described in any one of claims 1 to 12.

Images