Work summary generation method and device, electronic equipment and storage medium

Abstract

The invention relates to the technical field of data processing, and provides a work summary generation method and device, electronic equipment and a storage medium, and the method comprises the steps: obtaining a work knowledge graph, and constructing nodes and edges in the work knowledge graph based on work data; determining a first work sub-graph in a first time range from the work knowledge graph, and dividing the first work sub-graph into project type units based on project identifiers and work types of nodes in the first work sub-graph; generating a work task chain of a project corresponding to the project type unit based on the work timestamps of the nodes in the project type unit and edges between the nodes, and generating summary information of the project type unit based on the work timestamps of the nodes in the project type unit, the work content description information and the related personnel information; and generating a work summary corresponding to the first time range based on the work task chain and the summary information. According to the invention, the logicality and traceability of the automatically generated work summary are ensured.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method, apparatus, electronic device, and storage medium for generating work summaries. Background Technology

[0002] Currently, most work summaries are written manually. When writing work summaries, employees often need to manually recall, search, and summarize data from heterogeneous data sources because their work results are often scattered across different systems or platforms. This method is not only inefficient but also prone to omissions of key information due to memory lapses.

[0003] While there are solutions for generating work summaries through template-based filling systems or text generation models, these solutions cannot effectively integrate diverse work data and lack a deep understanding of the work content and timeline logic. As a result, the work summaries obtained are often superficial and lack factual support.

[0004] Therefore, how to automate and intelligently generate work summaries is an urgent problem to be solved in this field. Summary of the Invention

[0005] This invention provides a method, apparatus, electronic device, and storage medium for generating work summaries, in order to address the shortcomings of insufficient automation and intelligence in the generation of work summaries in related technologies.

[0006] This invention provides a method for generating work summaries, comprising: Obtain a job knowledge graph, wherein the nodes and edges in the job knowledge graph are constructed based on job data; From the work knowledge graph, a first work subgraph within a first time range is determined, and based on the project identifier and work type of the nodes in the first work subgraph, the first work subgraph is divided into project type units; Based on the work timestamps of the nodes in the project type unit and the edges between the nodes, a work task chain corresponding to the project of the project type unit is generated. Based on the work timestamps of the nodes in the project type unit, the work content description information and the relevant personnel information, a summary information of the project type unit is generated. Based on the work task chain and the summary information, a work summary corresponding to the first time range is generated.

[0007] According to a method for generating a work summary provided by the present invention, the step of obtaining a work knowledge graph includes: Transform the working data from multiple working data sources into intermediate representations; Based on the data type of the work data corresponding to the intermediate representation, the node attributes of the intermediate representation are extracted, and the nodes corresponding to the intermediate representation are established based on the node attributes; the node attributes include work timestamp, work content description information, project identifier, work type and related personnel information; Based on the node attributes of the nodes, construct the edges between the nodes.

[0008] According to a work summary generation method provided by the present invention, the step of extracting node attributes of the intermediate representation based on the data type of the work data corresponding to the intermediate representation includes: Based on the data type of the working data corresponding to the intermediate representation, determine the associated intermediate representation of the working data associated with the intermediate representation; Based on the intermediate representation and the associated intermediate representation, extract the node attributes of the intermediate representation.

[0009] According to a work summary generation method provided by the present invention, the step of constructing edges between nodes based on the node attributes of the nodes includes: Based on the project identifier and / or relevant personnel information of the nodes, establish temporal edges between the nodes; Based on at least one of the following: the co-existing entities in the work data corresponding to the nodes, the semantic similarity between the work content description information of the nodes, and the causal relationship between the node attributes of the nodes, an association edge is established between the nodes.

[0010] According to a work summary generation method provided by the present invention, the step of obtaining a work knowledge graph further includes: Based on the semantic vectors of the work content description information of the nodes, the nodes are clustered to obtain multiple node clusters; Based on a large language model, different nodes within the same node cluster are merged and / or their work content description information is disambiguated.

[0011] According to a method for generating a work summary provided by the present invention, determining a first work subgraph within a first time range from the work knowledge graph includes: From the work knowledge graph, nodes whose work timestamps belong to a first time range and whose relevant personnel information includes the target personnel are selected as target nodes; Based on the target node and its associated nodes in the working knowledge graph, the first working subgraph is obtained; the associated nodes are nodes in the working knowledge graph that are connected to the target node through edges.

[0012] A method for generating a work summary according to the present invention further includes: Based on the work summary corresponding to the first time range, a new report node is added to the work knowledge graph, and edges are established between the report node and each node in the first work subgraph.

[0013] A method for generating a work summary according to the present invention further includes: The report nodes within the second time range are obtained from the working knowledge graph, and the second working subgraph within the second time range is determined based on the connection edges of the report nodes within the second time range. Based on the second working subgraph, a work summary corresponding to the second time range is generated.

[0014] According to a method for generating a work summary provided by the present invention, the step of generating a work summary corresponding to the second time range based on the second work subgraph includes: Based on the density of nodes in the second working subgraph, the second working subgraph is divided into multiple working theme units; A work summary for the second time period is generated based on at least one of the summary information of the multiple work topic units, the project information of the second work subgraph, and the indicator information of the second work subgraph.

[0015] A method for generating a work summary according to the present invention further includes: Obtain the user's question and determine the query complexity of the user's question; When the query complexity is complex, the user question is broken down into multiple sub-queries; Extract the user's question or the query intent and query entity of the multiple subqueries; Based on the query intent, the query entity, and the working knowledge graph, a response text for the user's question is generated.

[0016] The present invention also provides a work summary generation device, comprising: The graph maintenance unit is used to acquire a working knowledge graph, wherein the nodes and edges in the working knowledge graph are constructed based on the working data. The subgraph analysis unit is used to determine the first working subgraph within a first time range from the working knowledge graph, and to divide the first working subgraph into project type units based on the project identifier and working type of the nodes in the first working subgraph. The project analysis unit is used to generate a work task chain for the project corresponding to the project type unit based on the work timestamps of the nodes in the project type unit and the edges between the nodes, and to generate summary information for the project type unit based on the work timestamps, work content description information and relevant personnel information of the nodes in the project type unit. The summary generation unit is used to generate a work summary corresponding to the first time range based on the work task chain and the summary information.

[0017] 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 work summary generation method as described above.

[0018] 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 work summary generation method as described above.

[0019] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the work summary generation method as described above.

[0020] The present invention provides a method, apparatus, electronic device, and storage medium for generating work summaries. Based on work data, a dynamic work knowledge graph is constructed. This allows for the determination of a first work subgraph based on the work knowledge graph. Based on the node attributes of each node in the first work subgraph and the edges between nodes, a work summary for a first time range is generated. On the basis of achieving fully automated work summary generation, deep analysis and reasoning of the temporal logic of work content are achieved through reasoning about the work task chain of the project and extracting summary information for project type units. This ensures the logicality and traceability of the work summary generated. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in this invention or related technologies, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is one of the flowcharts illustrating the work summary generation method provided by this invention.

[0023] Figure 2 This is the second flowchart of the work summary generation method provided by the present invention.

[0024] Figure 3 This is a schematic diagram of the work summary generation device provided by the present invention.

[0025] Figure 4 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0026] 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.

[0027] 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 relevant device.

[0028] Currently, most work summaries are written manually. Typically, the process of writing a complete work summary involves reviewing past work activities, extracting key data, and organizing the text logically.

[0029] With the widespread adoption of digital office work, employees' work outputs are often scattered across various heterogeneous systems and platforms. For example, code commit records can be stored in code repositories such as Git, task progress data can be stored in project management tools such as Jira, and communication and collaboration data can be found in systems and platforms such as email, instant messaging records, and document collaboration platforms.

[0030] Therefore, in the process of manually writing work summaries, it is necessary to manually recall, search, and summarize data from the aforementioned various heterogeneous systems and platforms. This method is not only tedious and inefficient, but also prone to omission of important work content and details due to memory lapses when writing longer-term work summaries such as quarterly or annual work summaries, resulting in the omission of key achievements or a disordered timeline.

[0031] Among related technologies, there are also solutions for generating work summaries through template filling systems, log aggregation tools, or text generation models.

[0032] Template-based data entry systems, or data entry systems based on templates, often result in work summaries that are merely information piles, lacking in-depth analysis and structured presentation of the internal logic, chronological evolution, and overall project progress. While these systems can solve the problem of inconsistent formats, they are superficial and fail to reflect the true value and relevance of the work, leading to the continued need for manual completion of most tasks.

[0033] Log aggregation tools can aggregate activity logs from a single data source according to a timeline, such as code commit records from a code repository. However, log aggregation tools cannot integrate diverse data such as experimental conclusions and meeting minutes, and they lack the ability to mine semantic correlations and causal relationships between work events. Therefore, they cannot generate work summaries that conform to the inherent logic of the work content.

[0034] While text generation models can produce fluent work summaries based on prompts, these summaries heavily rely on materials provided by personnel. Furthermore, because text generation models lack an understanding of internal company knowledge, interpersonal relationships, and the long-term evolution of projects, the generated work summaries are prone to creating illusions detached from reality, or containing empty content lacking concrete factual and data support.

[0035] In addition, among related technologies, there are also work summary generation schemes based on static knowledge bases. Specifically, when a user inputs a specific report title, the report title goes through a pipeline of steps, including extracting keywords, matching or generating an outline, and retrieving materials to fill in paragraphs, and finally assembling a complete report.

[0036] However, the above-mentioned solutions still have significant limitations in practical applications and are insufficient to meet the needs for intelligent and in-depth work summary generation. Specifically, their main shortcomings are reflected in the following aspects: First, the data is disconnected from the context, heavily relying on manual title creation. The above solution is a typical title-driven approach, its operation depending on the user-inputted report title as the trigger. However, in practice, it is often difficult for personnel to pre-conceive precise titles. This solution cannot automatically discover and define the themes, cycles, and key points for summarizing from the raw work data accumulated daily, lacking perception-induction capabilities, resulting in the core task of title creation still requiring manual completion.

[0037] Second, the knowledge representation is static and discrete, lacking semantic connections. This scheme sets up a knowledge base and a knowledge graph. The knowledge base is essentially a collection of documents and materials used as a reference when generating paragraphs, while the knowledge graph is a template library for report outlines used as a reference when generating structured outlines. Neither the knowledge base nor the knowledge graph in this scheme performs fine-grained modeling of the basic entities of the work content. Therefore, it cannot capture the dynamic semantic relationships such as temporal, causal, and attributive relationships that may exist between basic entities. This results in the scheme's understanding of the work content remaining only at the document level, unable to support deep relational analysis and logical reasoning.

[0038] Third, the generation process is mechanical, lacking chronological logic and interpretability. Because this solution employs a linear, pipeline-divide-and-conquer strategy of keywords-outline-paragraphs, its report generation is essentially based on material retrieval and text synthesis using user-input titles. The resulting work summaries often fail to reflect the evolution of work tasks over time, nor do they explain why multiple tasks are organized together. The solution's content organization relies excessively on a shallow understanding of title keywords using pre-set outline templates or models, resulting in work summary texts with weak logic and poor traceability.

[0039] Therefore, how to overcome the shortcomings of data and scenario separation, static knowledge representation, and lack of temporal logic in the related technologies, and realize an automated and intelligent method for generating work summaries, remains an urgent problem to be solved in this field.

[0040] Figure 1 This is one of the flowcharts illustrating the work summary generation method provided by this invention, such as... Figure 1 As shown, the method includes: Step 110: Obtain the work knowledge graph. The nodes and edges in the work knowledge graph are constructed based on the work data.

[0041] Specifically, work data refers to the data required to generate a work summary. Work data can be understood as data generated by personnel during the work process that can reflect the work content or progress. For example, it can include code commit records, experiment records, document reports, meeting minutes, emails, task items, etc. Work data can be obtained from one or more data sources, such as code repositories, project management tools, emails, instant messaging records, document collaboration platforms, etc.

[0042] In this embodiment of the invention, information reflecting the work content can be extracted from the work data, and the extracted information reflecting the work content can be instantiated into a tuple, thereby constructing nodes based on the tuple. Here, a node can be understood as a work content node; each node represents an actual work content being executed, and the node attributes are the aforementioned tuple. That is, the work content represented by the node can be specifically reflected through the node attributes.

[0043] Here, the information reflecting the work content extracted from the work data, namely the node attributes, can include a work timestamp, work content description information, project identifier, work type, and relevant personnel information. Furthermore, the above information can be represented in the form of a tuple. Optionally, in addition to including a work timestamp, work content description information, project identifier, work type, and relevant personnel information, node attributes can also include at least one of the following: node identifier, data type, metadata, and links pointing to the work data.

[0044] Among them, the work timestamp can be recorded as Timestamp, which is used to record the precise time point of the work; the work content description information can be recorded as Content, which is a summary description of the work content reflected by the work data, such as "fixed the concurrency vulnerability of the user login module"; the project identifier can be recorded as Project, which is used to mark the specific project to which the work belongs; the work type can be recorded as Type, which is used to indicate the type of work, such as "code development", "documentation writing", "meeting discussion" or "testing and verification"; and the relevant personnel information can be recorded as Contributors, which can be a list of relevant personnel who participated in the work.

[0045] In addition, node identifiers can be unique identifiers for nodes; data types can be the data types of the working data corresponding to the node, such as experimental records, document reports, meeting minutes, emails, task items, etc.; metadata refers to the structured attributes related to the data types of the working data corresponding to the node, such as the commit hash value of code commit records, and the hypothesis and result of experimental records.

[0046] Furthermore, after constructing nodes based on working data, edges between nodes can be built based on the node attributes of each constructed node. Here, an edge between nodes is an edge connecting two nodes. The existence of an edge between any two nodes means that these two nodes have a semantic relationship. Specifically, the semantic relationship here can include at least one of the following: temporal sequence relationship, causal influence relationship, project affiliation relationship, semantic reference relationship, and collaborative work relationship.

[0047] Furthermore, the attributes of the edges between nodes can include semantic relationship type, as well as weight and basis. The weight reflects the strength or confidence of the semantic relationship represented by the edge, while the basis is the foundation upon which the semantic relationship is constructed, such as co-occurring text or rule matching.

[0048] It is understandable that after constructing the nodes and the edges between them, a knowledge graph containing the nodes and the edges between them can be obtained, which is denoted as the working knowledge graph in this embodiment of the invention. The working knowledge graph is the knowledge graph obtained by structuring the working data. The working knowledge graph can reflect the key information in the working data and the relationships between different working data.

[0049] In this embodiment of the invention, acquiring a work knowledge graph can refer to the process of constructing a work knowledge graph, or it can refer to the process of dynamically updating an existing work knowledge graph over time based on newly added work data. That is, a work knowledge graph is a network structure that dynamically evolves by continuously parsing work data over time.

[0050] Step 120: Determine the first working subgraph within the first time range from the working knowledge graph, and divide the first working subgraph into project type units based on the project identifier and working type of the nodes in the first working subgraph.

[0051] Specifically, the first time frame is the time frame corresponding to the work summary to be generated. The first time frame can be a day, a week, a month, a quarter, a year, or other time granularities.

[0052] After determining the first time range, the first working subgraph within the first time range can be determined from the working knowledge graph based on the working timestamps in the node attributes of each node. Specifically, the first time range can be used as a filtering condition to filter nodes in the working knowledge graph whose working timestamps are within the first time range. Then, by using the edges of the filtered nodes in the working knowledge graph, the nodes in the working knowledge graph that are related to the filtered nodes can be obtained. Thus, the first working subgraph can be composed of the filtered nodes and the edges between the nodes, or the filtered nodes, the edges between the nodes, and the nodes that are related.

[0053] It is understandable that the first working subgraph is part or all of the working knowledge graph. Compared to the working knowledge graph, the first working subgraph focuses more on the working content within the first time frame.

[0054] After obtaining the first working subgraph, the nodes in the first working subgraph can be divided into projects and work types based on the project identifier and work type in the node attributes, thereby obtaining at least one project type unit. Here, a project type unit only includes nodes that belong to the same project and the same work type; that is, a project type unit only includes nodes that have the same project identifier and the same work type.

[0055] By dividing the first work subgraph into at least one project type unit, a refined division of work content based on project affiliation and work type is achieved.

[0056] Step 130: Based on the work timestamps of the nodes in the project type unit and the edges between the nodes, generate the work task chain of the project corresponding to the project type unit. Based on the work timestamps, work content description information and relevant personnel information of the nodes in the project type unit, generate the summary information of the project type unit.

[0057] Specifically, based on dividing the first working subgraph into at least one project type unit, for all project type units belonging to the same project, the logical chain of work tasks performed by the project can be reconstructed based on the work timestamps of each node in the project type unit and the edges between the nodes, thus obtaining the work task chain of the project. For example, based on the work timestamps of each node in each project type unit belonging to the same project and the edges between the nodes representing causal relationships such as "caused" and "based on", a complete work task chain from "writing the design plan in the morning" to "reviewing the plan with colleagues in the afternoon" and then to "submitting the preliminary implementation code in the evening" can be identified.

[0058] In addition, for each project type unit, key information can be extracted based on the work timestamps, work content descriptions, and relevant personnel information of each node in the project type unit, thereby obtaining summary information that reflects the specific work content of the project type unit.

[0059] For example, for each project type unit, the work content descriptions and relevant personnel information of each node can be arranged in the order of the work timestamps of each node, thus obtaining a chronologically arranged list containing the work timestamps, work content descriptions, and relevant personnel information of each node in the project type unit. Subsequently, this list, along with the corresponding project and work type for this project type unit, can be input into a large language model to generate a structured paragraph containing three aspects: "Main Achievements," "Key Progress," and "Problems Encountered," serving as a summary of the project type unit.

[0060] Step 140: Based on the work task chain and summary information, generate a work summary corresponding to the first time range.

[0061] Specifically, after obtaining the task chain of each project in the first working subgraph and the summary information of each project type unit in the first working subgraph, a work summary corresponding to the first time range can be generated based on the task chain and summary information.

[0062] For example, the task chain and summary information can be input into a large language model. Based on a pre-set work summary template, the large language model can organize the task chain and summary information into a complete work summary with a standardized format and fluent language. This allows the output of the work summary corresponding to the first time range from the large language model.

[0063] In the method provided in this embodiment of the invention, a dynamic work knowledge graph is constructed based on work data. This allows for the determination of a first work subgraph based on the work knowledge graph. Based on the node attributes of each node in the first work subgraph and the edges between nodes, a work summary for a first time range is generated. On the basis of achieving fully automated work summary generation, deep analysis and reasoning of the temporal logic of the work content are achieved through reasoning about the work task chain of the project and extracting summary information for the project type unit. This ensures the logic and traceability of the work summary generated thereby.

[0064] Based on the above embodiments, step 110, obtaining the working knowledge graph, includes: Transform the working data from multiple working data sources into intermediate representations; Based on the data type of the work data corresponding to the intermediate representation, the node attributes of the intermediate representation are extracted, and the nodes corresponding to the intermediate representation are established based on the node attributes; the node attributes include work timestamp, work content description information, project identifier, work type and related personnel information; Based on the node attributes, construct edges between nodes.

[0065] Specifically, a working data source refers to a data source that stores working data, such as a code repository, project management tool, experimental platform, document system, calendar, report, etc. Various working data sources can be accessed uniformly through the adapter pattern to obtain the working data from each source.

[0066] Furthermore, since the working data from different data sources may belong to different data types, in order to facilitate unified processing of the working data from different data sources, the acquired working data can be converted into an intermediate representation. Here, the intermediate representations of the working data from different data sources reside in the same feature space.

[0067] Building upon this foundation, for each intermediate representation of work data, a large-scale language model fine-tuned for that data type, or a rule engine configured for that data type, can be invoked to extract and structure the node attributes of the intermediate representation. For example, node attributes might include a work timestamp, work content description, project identifier, work type, and relevant personnel information. Alternatively, node attributes might include at least one of the following: node identifier, data type, metadata, and links pointing to the work data.

[0068] After obtaining the node attributes of the intermediate representation of each working data, the node corresponding to that working data, that is, the node corresponding to the intermediate representation, can be constructed based on this.

[0069] After the nodes are constructed, edges between them can be built based on the node attributes of each node. For example, the execution order of the work corresponding to each node can be inferred based on the work timestamp of each node, thereby establishing edges between nodes that reflect the temporal sequence of work execution; or, based on the semantic relevance between the work content description information of each node, it can be determined whether there is a semantic reference relationship between the nodes, thereby establishing edges between nodes that reflect the semantic reference relationship.

[0070] Based on any of the above embodiments, in step 110, extracting the node attributes of the intermediate representation based on the data type of the working data corresponding to the intermediate representation includes: Based on the data type of the working data corresponding to the intermediate representation, determine the associated intermediate representation of the working data associated with the intermediate representation; Based on intermediate representations and associated intermediate representations, extract the node attributes of the intermediate representations.

[0071] Specifically, the information contained in a single working data point may be insufficient, making it difficult to extract node attributes from the intermediate representation corresponding to that working data. To address this issue, for any intermediate representation of working data, we can determine the working data associated with that working data based on its data type, and denote the intermediate representation of the associated working data as the associated intermediate representation. Here, the associated intermediate representation can provide contextual information for extracting node attributes from the intermediate representation.

[0072] It's understandable that different data types of work data can have different associated work data. Therefore, it's possible to pre-define the types of work data associated with different data types. For example, code commit records can be associated with file lists and task items.

[0073] Therefore, node attributes can be extracted based on the intermediate representation of the working data and the associated intermediate representation of the working data associated with it.

[0074] For example, when the work data is code commit records, the following prompt can be set: "You are a technical expert. Please extract the work content nodes according to {schema} based on the following code commit information {commit_msg}, the list of changed files {files}, and the associated task {linked_issue}." Here, {commit_msg} corresponds to the intermediate representation of the code commit record, {files} is the associated intermediate representation of the file list linked to the code commit record, {linked_issue} is the associated intermediate representation of the task item linked to the code commit record, and {schema} is the node attribute extraction strategy set for the code commit record. Based on this, the above prompt can be input into a large language model that has been fine-tuned for the code commit records, thereby obtaining the node attributes of the code commit records output by the large language model, which can include the work content description information of the code commit records.

[0075] In the method provided in this embodiment of the invention, the node attributes of the intermediate representation are extracted based on the associated intermediate representation of the working data associated with the intermediate representation. This can effectively overcome the problem of insufficient intermediate representation information of the original working data, provide rich contextual information for node attribute extraction, and thereby improve the reliability and accuracy of node attribute extraction.

[0076] Based on any of the above embodiments, step 110, constructing edges between nodes based on node attributes, includes: Establish temporal edges between nodes based on node project identifiers and / or related personnel information; Based on at least one of the following: the co-existing entities in the work data corresponding to the nodes, the semantic similarity between the work content description information of the nodes, and the causal relationship between the node attributes of the nodes, establish the association edges between the nodes.

[0077] Specifically, after the nodes are constructed, temporal edges between nodes can be established based on the project identifier and / or relevant personnel information in the node attributes of each constructed node.

[0078] For example, for nodes with the same project identifier, a time-series chain can be naturally formed based on the node's working timestamp. Thus, for nodes with the same project identifier, a time-series edge can be created between two nodes with adjacent working timestamps based on the working timestamps in chronological order. In this way, a link can be realized to connect nodes belonging to the same project in a time sequence through the time-series edge.

[0079] For example, for nodes with the same related personnel information, a time-series chain can be naturally formed based on the node's work timestamp. Thus, for nodes with the same related personnel information, a time-series edge can be created between two nodes with adjacent work timestamps based on the order of their work timestamps from front to back. In this way, a link can be realized through the time-series edge to connect nodes belonging to the same work team in a time sequence.

[0080] Furthermore, after the nodes are constructed, the associated edges between nodes can be established based on at least one of the following: the common entities in the work data corresponding to the nodes, the semantic similarity between the work content description information of the nodes, and the causal relationship between the node attributes of the nodes.

[0081] This can be achieved by establishing association edges based on co-occurring entities in the work data corresponding to nodes. Co-occurring entities refer to entities that co-occur in the same data of the work data corresponding to two nodes, such as entities that appear in the same document or the same meeting minutes of the work data corresponding to two nodes. The association edges established in this way can reflect collaborative work relationships or semantic reference relationships.

[0082] Furthermore, semantic similarity can be used to establish association edges reflecting the semantic reference relationship between nodes based on the semantic similarity between the node's job description information. For example, BERT (Bidirectional Encoder Representations from Transformers) or similar language models can be used to calculate the semantic features of the job description information of two nodes, and then the similarity between the two semantic features can be calculated, thus obtaining the semantic similarity between the node's job description information. If the semantic similarity is greater than a preset threshold, an association edge can be established between the two nodes. Further, the value of the semantic similarity can be used as the weight of the association edge between these two nodes.

[0083] Furthermore, association edges reflecting the causal influence between nodes can be established based on the causal relationships between node attributes. For example, a finely tuned causal relationship classification model can be used to determine whether a causal relationship exists between two nodes based on their node attributes. If it does, an association edge reflecting the causal influence between the two nodes can be established.

[0084] Based on any of the above embodiments, step 110, obtaining the working knowledge graph, further includes: Based on the semantic vectors describing the work content of nodes, nodes are clustered to obtain multiple node clusters; Based on a large language model, different nodes within the same node cluster are merged and / or their work content description information is disambiguated.

[0085] Specifically, after creating nodes based on the node attributes of the intermediate representation of each work data, nodes can be merged, and / or the work content description information in the node attributes can be disambiguated.

[0086] Furthermore, after establishing all nodes, semantic vectors can be extracted from the job description information in the node attributes of all nodes; that is, semantic vectors of the job description information of all nodes can be obtained. For example, the semantic vectors of the job description information can be extracted using the Sentence-BERT model or other language models.

[0087] Building upon this, node clustering can be performed based on the semantic vectors of the job description information of all nodes. This clustering can be achieved using clustering algorithms such as K-Means and DBSCAN (Density-Based Spatial Clustering of Applications with Noise). This results in multiple node clusters, each containing multiple nodes with similar semantic vectors of their job description information.

[0088] It is understandable that any two nodes within the same node cluster have very similar job descriptions. Based on this, there are two possibilities for these two nodes: one is that their job descriptions are accurate and their job content is identical, in which case they can be merged into one node; the other is that their job content is different, but their job descriptions are similar due to ambiguity or errors, in which case the job descriptions can be adjusted to eliminate ambiguity.

[0089] Therefore, for each node cluster, a large language model can be used to determine whether to merge nodes within the cluster or disambiguate the descriptions of the nodes' work content. Based on this, either merging of nodes or disambiguation of the node's work content descriptions can be performed.

[0090] In the method provided in the embodiments of the present invention, node clustering is used to perform node merging or disambiguation processing within the node cluster, thereby ensuring the reliability and rationality of the working indication map.

[0091] Based on any of the above embodiments, step 120, determining the first working subgraph within a first time range from the working knowledge graph, includes: From the work knowledge graph, nodes whose work timestamps belong to the first time range and whose relevant personnel information includes the target personnel are selected as target nodes; Based on the target node and its associated nodes in the working knowledge graph, a first working subgraph is obtained; the associated nodes are nodes in the working knowledge graph that are connected to the target node through edges.

[0092] Specifically, after determining the first time frame, a node filtering action can be performed on the work knowledge graph. The criteria for node filtering can include the time frame (the first time frame) and the personnel whose work needs to be summarized (the target personnel). During the node filtering process, all nodes in the work knowledge graph can be traversed, retaining only those nodes whose work timestamps fall within the aforementioned first time frame and whose relevant personnel information includes the target personnel. The nodes obtained through this filtering are denoted as target nodes in this embodiment of the invention. Target nodes reflect the work traces of the target user within the first time frame.

[0093] After obtaining the target node, the knowledge graph can be extended outwards based on the target node, i.e., relationship capture can be performed. Here, relationship capture refers to obtaining other nodes in the knowledge graph that are connected to the target node through semantic relationships such as causal influence, semantic reference, and temporal sequence. In this embodiment of the invention, these are referred to as the target node's associated nodes. For example, associated nodes connected to the target node through edges reflecting causal influence can reflect which tasks triggered the target node's work; another example is that associated nodes connected to the target node through edges reflecting semantic reference can reflect the documents or code applied to the target node's work; yet another example is that associated nodes connected to the target node through edges reflecting temporal sequence can reflect the natural chronological order between tasks. This embodiment of the invention does not specifically limit this.

[0094] Therefore, the target node and its associated nodes connected by edges can be integrated, thereby obtaining a first working subgraph with rich content and complete logic corresponding to the first time range from the working knowledge graph, laying a data foundation for subsequent automated analysis of the working content within the first time range.

[0095] For example, when analyzing the first working subgraph later, all nodes in the first working subgraph can be sorted according to the order in which the work was discovered, so as to facilitate the analysis of the timeline of the first working subgraph.

[0096] Based on any of the above embodiments, the work summary generation method further includes: Based on the work summary corresponding to the first time range, a new report node is added to the work knowledge graph, and edges are established between the report node and each node in the first work subgraph.

[0097] Specifically, after generating the work summary corresponding to the first time frame, a node corresponding to the work summary can be constructed in the work knowledge graph based on the generated work summary; this node is denoted as a report node. Here, the node attributes of the report node can be extracted based on the work summary, and the work summary itself can also be stored as part of the node attributes.

[0098] Furthermore, for the newly added report node, all nodes referenced in the first working subgraph when generating the working summary can be connected to the report node through edges, where the edges can reflect the derived relationships.

[0099] Therefore, by constructing report nodes based on work summaries, the summary knowledge reflected in the work summaries can be persistently stored in the work knowledge graph. By establishing edges between report nodes and nodes in the first work subgraph, it can be ensured that every description in the work summary can be traced back to the original work data, thus achieving transparency and verifiability of the information reflected in the work summary.

[0100] Based on any of the above embodiments, the work summary generation method further includes: Obtain the report nodes within the second time range from the working knowledge graph, and determine the second working subgraph within the second time range based on the connection edges of the report nodes within the second time range; Based on the second working subgraph, a work summary corresponding to the second time range is generated.

[0101] Specifically, the second time range is the time range corresponding to the work summary to be generated. Furthermore, compared to the first time range, the second time range has a larger time granularity. For example, if the time granularity of the first time range is one day, then the time granularity of the second time range can be one week, one month, etc. This embodiment of the invention does not specifically limit this.

[0102] In this embodiment of the invention, for a second time range with a larger time granularity than the first time range, when generating a work summary for the second time range, the work summary for the first time range, which has a smaller granularity and is covered by the second time range, can be referenced, and a higher-level work summary can be generated based on this. For example, when a weekly report needs to be generated, the report node for the daily report can be obtained from the work knowledge graph; when a monthly report needs to be generated, the report node for the weekly report can be obtained from the work knowledge graph; when a quarterly report needs to be generated, the report node for the monthly report can be obtained from the work knowledge graph; and when an annual report needs to be generated, the report node for the quarterly report can be obtained from the work knowledge graph.

[0103] Therefore, we can first obtain report nodes within the second time frame from the work knowledge graph. These report nodes are the report nodes for the work summaries corresponding to the first time frame, which has a smaller time granularity within the second time frame. For example, for a second time frame of one week, we can obtain the report nodes corresponding to the daily work summaries within that week from the work knowledge graph. These report nodes can be regarded as the core raw materials for generating the work summaries corresponding to the second time frame.

[0104] Based on this, the nodes used to generate the work summary corresponding to the report node can be traced in the work knowledge graph based on the connection edges of the report nodes in the second time range obtained therefrom. The resulting work subgraph includes the report nodes in the second time range and the nodes connected to the report nodes in the second time range, which is referred to here as the second work subgraph.

[0105] For example, after obtaining the report nodes within the second time range, nodes with derivative relationships to each report node can be traversed in the work knowledge graph. This allows the collection of all nodes that have been summarized within the second time range to generate the aforementioned report nodes. These nodes can serve as auxiliary raw materials for generating the work summary corresponding to the second time range, providing fine-grained information for generating the content of the work summary corresponding to the second time range and ensuring its traceability.

[0106] After obtaining the second working subgraph, a work summary corresponding to the second time range can be generated based on the report nodes contained in the second working subgraph, the nodes directly generated based on the work data, and the edges between the above nodes.

[0107] In the method provided in the embodiments of the present invention, a second working subgraph is constructed by applying report nodes in the second time range of the working knowledge graph, thereby generating a working summary corresponding to the second time range. This realizes a bottom-up, hierarchical, multi-granularity working summary generation system, which can effectively improve the reliability of generating working summaries with larger time granularity.

[0108] Based on any of the above embodiments, a work summary corresponding to the second time range is generated based on the second working subgraph, including: Based on the density of nodes in the second working subgraph, the second working subgraph is divided into multiple working theme units; A work summary for the second time frame is generated based on at least one of the following: summary information from multiple work topic units, project information from the second work subplot, and indicator information from the second work subplot.

[0109] Specifically, after obtaining the second working subgraph, it can be divided into multiple working theme units based on the density of the nodes within it. During this process, for the second working subgraph that carries the complete working memory within the second time frame, a community detection algorithm can be run to calculate the density of the nodes. Here, the community detection algorithm can be a modularity-based optimization algorithm such as the Louvain algorithm or the Leiden algorithm, or a label propagation-based algorithm such as LPA (Label Propagation Algorithm). The density of the nodes obtained reflects whether nodes share the same project label, whether there are frequent collaborative relationships, and whether there is a high degree of semantic relevance. The community detection algorithm can divide the second working subgraph into multiple independent communities based on the density of the nodes. In this embodiment of the invention, the resulting communities are denoted as working theme units.

[0110] Understandably, each resulting work theme unit is essentially a naturally clustered set of nodes, with all nodes within that set representing work related to a single work theme. For example, nodes within a work theme unit might represent Monday's design document, Tuesday's review meeting minutes, related code commits from Wednesday to Thursday, and Friday's test report—all revolving around the theme of "functional module development." This approach of dividing work into theme units effectively addresses the problem of overlooking logical connections across time dimensions during manual review.

[0111] After obtaining multiple work topic units, a work summary within a second time frame can be generated based on at least one of the summary information of the multiple work topic units, the project information of the second work subplot, and the indicator information of the second work subplot.

[0112] The summary information of a work topic unit can be extracted based on all nodes within that unit. For example, the work content descriptions of all nodes within a work topic unit can be input into a large language model, which will then generate a highly condensed topic overview as the summary information for that work topic unit. The resulting summary information can include key information about work progress, value output, and bottlenecks. The project information in the second working subplot is derived from horizontal data integration across the subplot, focusing on the project dimension. This information can encompass the input details for each project within the subplot, including, for example, work hours or task items. Furthermore, the project information can list all thematic units involved in each project. Therefore, the project information in the second working subplot clearly demonstrates the distribution of effort and contribution of an individual or team across different projects.

[0113] The metrics for the second work subplot are global quantitative data calculated specifically for that subplot. These metrics may include core theme focus, cross-project collaboration index, and percentage of unplanned work. Specifically, core theme focus refers to the proportion of time allocated to the work theme unit with the longest time commitment; cross-project collaboration index refers to the number of work items that the target personnel collaborate on with different personnel; and percentage of unplanned work refers to the proportion of work content not associated with any predetermined project or task.

[0114] At least one of the following can be input into a large language model: summary information of multiple working topic units, project information of the second working subplot, and indicator information of the second working subplot. The large language model can generate a structurally complete and linguistically professional work summary based on a preset template, thereby obtaining the work summary corresponding to the second time range output by the large language model.

[0115] Optionally, based on the work summary corresponding to the second time range, a new report node can be added to the work knowledge graph, and edges can be established between the report node and each node in the second work subgraph, thereby completing the knowledge loop.

[0116] Based on any of the above embodiments, the work summary generation method further includes: Obtain the user's question and determine the query complexity of the user's question; When the query complexity is complex, the user's question is broken down into multiple subqueries; Extract the user's question or the query intent and query entities of multiple subqueries; Based on the query intent and query entities, as well as the working knowledge graph, generate the answer text for the user's question.

[0117] Specifically, in this embodiment of the invention, the working knowledge graph can not only be used to generate work summaries, but also serve as a dynamic and interactive organizational intelligence hub to engage in deep question-and-answer sessions with users based on working memory.

[0118] First, user questions can be obtained. Here, user questions refer to work-related inquiries raised by users; user questions can also be understood as work-related queries. User questions can be natural language text directly entered by the user, or text obtained by transcribing the user's speech. This embodiment of the invention does not specifically limit the types of questions.

[0119] After obtaining the user's question, its query complexity can be determined. This can be achieved using a large language model. Specifically, the large language model can be used to classify the user question using zero-shot or few-shot methods, thereby determining whether the user question contains implicit multiple issues, requires multi-step reasoning, or involves comparing and synthesizing multiple independent facts. Understandably, if the user question involves only a clear topic and requires only a single search or simple statistics to obtain the answer, the query complexity can be determined to be simple. If the user question needs to be broken down into several simpler sub-questions to be answered accurately, the query complexity can be determined to be complex.

[0120] For example, user questions can be populated into pre-designed prompts and input into a large language model, which then determines the query complexity. Here, the prompts could take the following form: Please determine whether the following job query needs to be broken down into multiple simpler sub-questions for an accurate answer. Criteria for no breakdown: The question addresses only a single, well-defined topic, and the answer can be derived through a single fact search or simple statistics.

[0121] Query: "{User Query}"; Please output only one JSON object: {"needs_decomposition": true / false, "reason": "A one-sentence explanation of the reason"}.

[0122] Here, "{User query}" represents the user's query; "needs_decomposition" being "true" indicates that decomposition is needed, meaning the query complexity is complex; "needs_decomposition" being "false" indicates that decomposition is not needed, meaning the query complexity is simple. "reason" is the reason why the large language model determines the query complexity of the user's query.

[0123] For example, when the user question is "What did Zhang San do last week?", the query complexity output by the large language model can be simple because it is a "single-topic fact query". As another example, when the user question is "What is the reason for the delay of project A, and what impact does this have on project B?", the query complexity output by the large language model can be complex because it "includes two logical steps: causal reasoning and impact analysis".

[0124] For queries with high complexity, the user problem can be broken down into multiple subqueries. These subqueries are sub-problems of the user problem, and multiple subqueries can constitute the user problem. Furthermore, the subqueries are logically connected and can be retrieved independently.

[0125] For example, complex user questions can be populated into pre-designed prompts and input into a large language model, which then breaks down the user question. Here, the prompts can take the following form: Original query: "{original query}" Task: Please break this down into 2-4 simpler subqueries, ensuring coverage of all aspects of the original problem. The subqueries should satisfy: 1. Each subquery focuses on a core intent (such as query status, tracing the reason, or comparing data).

[0126] 2. Information dependencies may exist between subqueries; please indicate them.

[0127] 3. Subqueries should contain the key entity words required for graph retrieval.

[0128] Output in JSON format: { "sub_queries": [ {"id": 1, "query": "...", "depends_on": [ ]}, {"id": 2, "query": "...", "depends_on": [1]} ] } Here, "{original query}" represents the user's question. "query" represents a subquery, and "depends_on" indicates the information dependencies between this subquery and other subqueries.

[0129] For example, to address the user question "Compare Zhang San and Li Si's contributions to project A and analyze the reasons for Zhang San's slow progress," we can break it down into three subqueries: Subquery 1: List all work items, types, and times for Zhang San in project A; Subquery 2: List all work items, types, and times for Li Si in project A; Subquery 3: Find the work blocking records or decision nodes related to Zhang San's progress delay in project A. Subquery 3 depends on the result of Subquery 1.

[0130] Based on this, the query intent and query entities of the user's question or multiple subqueries can be extracted. Specifically, for cases with simple query complexity, the query intent and query entities of the user's question can be extracted directly; for cases with complex query complexity, the query intent and query entities of each subquery of the user's question can be extracted.

[0131] Here, the extraction of query intent and query entities can be achieved using a large language model. For example, query intent and query entities can be extracted using the following prompts: Analyze the following job query to extract its intent and all entities involved.

[0132] Query: "{query text}" Output in JSON format: { "entities": { "intent": ["Progress Inquiry", …], "persons": ["Zhang San", ...], "projects": ["Project A", ...], "time_ranges": ["Last week", "Q1 of 2024", ...], "work_types": ["code submission", "design review", ...], "keywords": ["latency", "bottleneck", ...] } } Here, "{query text}" represents the user's question or subquery. "intent" represents the query intent, and "persons", "projects", "time_ranges", "work_types", and "keywords" are all query entities, namely, people, projects, time ranges, work types, and keywords, respectively.

[0133] After obtaining the query intent and query entity, relevant information can be retrieved from the working knowledge graph based on the query intent and query entity, and a response text for the user's question can be generated. Here, the retrieval and response text generation based on the working knowledge graph can be implemented using Graph Retrieval-Augmented Generation (GraphRAG) technology.

[0134] Furthermore, the user question and the extracted query entities can be input into the graph retrieval augmented generative model. The graph retrieval augmented generative model concatenates the user question and query entities and encodes them into a vector. It then searches the vector indexes of all nodes in the working knowledge graph and returns the K most relevant nodes as seed nodes. Here, K is an integer greater than 1.

[0135] Subsequently, a Cypher query can be dynamically generated and executed based on all seed nodes, the extracted query entities, and the query intent. During this process, the working knowledge graph can be expanded in a targeted manner, starting from the seed nodes and extending along the edges. Specifically, for queries with attribution intent, backtracking can be prioritized along edges representing causal relationships; for queries with statistical intent, all nodes under specific types, personnel, or projects can be retrieved. Thus, by executing a Cypher query, a semantically rich evidence subgraph containing relevant nodes and edges can be obtained.

[0136] Next, the answer text to the user question can be generated based on the user question and the structured information of the evidence subgraph. Here, the structured information of the evidence subgraph can be a textual description of the evidence subgraph or a small-scale graph structure.

[0137] The method provided in this embodiment of the invention implements question answering based on a work-related knowledge graph. In this process, by determining the query complexity of the user's question and decomposing the user's question into subqueries in complex cases, the reliability and accuracy of the generated response text can be improved, ensuring the rationality and reliability of the questions based on the work-related knowledge graph.

[0138] Based on any of the above embodiments Figure 2 This is the second flowchart of the work summary generation method provided by the present invention, as shown below. Figure 2 As shown, the method includes: First, construct a dynamic work knowledge graph. This process may include the following steps: Work data access and parsing: The adapter pattern allows for unified access to diverse and heterogeneous work data sources, such as code repositories, project management tools, experimental platforms, document systems, calendars, and reports, thereby obtaining raw work data. Furthermore, the adapter can parse the obtained work data into an intermediate representation.

[0139] Node construction and edge construction: For each intermediate representation of working data, a large language model fine-tuned for that data type, or a rule engine configured for that data type, can be invoked to extract and structure the node attributes of the intermediate representation. Based on these attributes, the corresponding node for that working data—that is, the node corresponding to that intermediate representation—can be constructed.

[0140] After the nodes are constructed, edges between them can be built based on the node attributes of each node. In essence, once the nodes and the edges between them are constructed, a working knowledge graph is formed.

[0141] Knowledge graph of dynamic maintenance work: Once the working knowledge graph is created, it can be saved. Furthermore, the node and edge construction processes are triggered whenever new working data is generated, thus enabling real-time, incremental updates to the working knowledge graph. Additionally, the working knowledge graph can be automatically updated even when original working data is manually updated or deleted.

[0142] After creating a work knowledge graph, work summaries can be generated based on the work knowledge graph, and intelligent interactions can be performed based on the work knowledge graph.

[0143] The automatic generation of multi-granularity summaries may include the following steps: Periodic / Event Triggered: Work summary generation can be automated either on a periodic or event-triggered basis. For example, a built-in scheduler can automatically generate work summaries at fixed times, such as every Friday at 6:00 PM or the last day of each month. Alternatively, work summaries can be automatically generated when the work knowledge graph detects a specific event, such as when a project is marked as completed. Furthermore, multi-turn dialogues between users can be supported for work summary generation.

[0144] Working subgraph extraction and cell discovery: Assuming the first time frame is one day, i.e., a daily report needs to be generated, nodes whose work timestamps belong to the first time frame and whose relevant personnel information includes the target personnel can be selected from the work knowledge graph as target nodes. Based on the target nodes and their associated nodes in the work knowledge graph, the first work subgraph is obtained. This first work subgraph can also be understood as the work subgraph for the current day.

[0145] Based on this, the first working subgraph can be divided into project type units according to the project identifier and working type of the nodes in the first working subgraph.

[0146] Assuming the second time frame is one week, i.e., the scenario requiring weekly report generation, we retrieve the report nodes within the second time frame from the work knowledge graph. Based on the connecting edges of these report nodes, we determine the second working subgraph within that second time frame. This second working subgraph can also be understood as the working subgraph for this week. Furthermore, the second time frame can also be one month, one quarter, or one year, etc.

[0147] Based on this, a community discovery algorithm can be run on the second working subgraph, thereby dividing the second working subgraph into multiple working topic units.

[0148] Temporal clustering and information fusion: For each project type unit in the first working subgraph, the work task chain of the project can be reconstructed based on the work timestamps of the nodes and the edges between the nodes that reflect the causal relationship. Furthermore, for each project type unit in the first working subgraph, key information can be extracted from the project type unit based on the work timestamps, work content description information and relevant personnel information of each node in the project type unit, thereby obtaining summary information that reflects the specific work content of the project type unit.

[0149] For each working topic unit in the second working subgraph, a topic-level deep summary can be generated, thus obtaining summary information for each working topic unit. Furthermore, information can be integrated from a project perspective for the second working subgraph, thereby obtaining project information for the second working subgraph. Additionally, global metrics can be calculated for the second working subgraph, thus obtaining metric information for the second working subgraph.

[0150] Work summary generation: For the first time frame, a work summary corresponding to the first time frame can be generated based on the work task chain and summary information.

[0151] For the second time frame, a work summary for the second time frame can be generated based on at least one of the following: summary information of multiple work thematic units, project information of the second work subplot, and indicator information of the second work subplot.

[0152] Report node construction and linking: After generating the work summary, a corresponding node, denoted as a report node, can be constructed in the work knowledge graph based on the generated work summary. Here, the node attributes of the report node can be extracted based on the work summary, and the work summary itself can also be stored as part of the node attributes.

[0153] Furthermore, for the newly added report node, all nodes in the work subgraph corresponding to the time range of the report node that are referenced when generating the work summary can be connected to the report node through edges. These edges can reflect the derived relationships.

[0154] In addition, intelligent interaction with the graph can be performed. This process may include the following steps: User acquisition issues: It can retrieve user questions. Here, user questions refer to work-related questions raised by users; user questions can also be understood as job inquiries.

[0155] Query complexity assessment and subquery decomposition: After obtaining the user's query, the query complexity can be determined. If the query complexity is complex, the user's query can be broken down into multiple subqueries. If the query complexity is simple, then no decomposition is necessary.

[0156] Query intent and query entity extraction: For cases with simple query complexity, the query intent and query entities of the user's question can be directly extracted; for cases with complex query complexity, the query intent and query entities of each subquery of the user's question can be extracted.

[0157] Enhanced graph retrieval generation: The user question and the extracted query entities can be input into a graph retrieval-enhanced generative model. The model concatenates the user question and query entities into a vector, which is then searched through the vector indexes of all nodes in the working knowledge graph. The K most relevant nodes are returned as seed nodes. Subsequently, based on all seed nodes, the extracted query entities, and the query intent, a Cypher query can be dynamically generated and executed to obtain an evidence subgraph. Finally, based on the user question and the structured information of the evidence subgraph, a response text can be generated for the user question.

[0158] Output response text: After receiving the answer text for the user's question, the answer text can be output to enable interaction with the user.

[0159] In the method provided in the embodiments of the present invention, key information is extracted from massive, heterogeneous, and diverse work data in an automated, structured, and intelligent manner to construct a work knowledge graph with temporal correlation. Based on the work knowledge graph, multi-granularity work summary reports such as daily, weekly, and monthly reports are intelligently generated. At the same time, intelligent question answering and progress tracking services based on the work knowledge graph are provided.

[0160] The work summary generation device provided by the present invention is described below. The work summary generation device described below and the work summary generation method described above can be referred to in correspondence.

[0161] Figure 3 This is a schematic diagram of the work summary generation device provided by the present invention, as shown below. Figure 3 As shown, the device includes: Graph maintenance unit 310 is used to acquire a working knowledge graph, wherein the nodes and edges in the working knowledge graph are constructed based on working data; The subgraph analysis unit 320 is used to determine a first working subgraph within a first time range from the working knowledge graph, and to divide the first working subgraph into project type units based on the project identifier and working type of the nodes in the first working subgraph. Project analysis unit 330 is used to generate a work task chain for the project corresponding to the project type unit based on the work timestamps of the nodes in the project type unit and the edges between the nodes, and to generate summary information of the project type unit based on the work timestamps, work content description information and relevant personnel information of the nodes in the project type unit. The summary generation unit 340 is used to generate a work summary corresponding to the first time range based on the work task chain and the summary information.

[0162] In the apparatus provided in this embodiment of the invention, a dynamic work knowledge graph is constructed based on work data. This allows for the determination of a first work subgraph based on the work knowledge graph. Based on the node attributes of each node in the first work subgraph and the edges between nodes, a work summary for a first time range is generated. On the basis of achieving fully automated work summary generation, deep analysis and reasoning of the temporal logic of the work content are achieved through reasoning about the work task chain of the project and extracting summary information for the project type unit. This ensures the logic and traceability of the work summary generated thereby.

[0163] Based on any of the above embodiments, the map maintenance unit is specifically used for: Transform the working data from multiple working data sources into intermediate representations; Based on the data type of the work data corresponding to the intermediate representation, the node attributes of the intermediate representation are extracted, and the nodes corresponding to the intermediate representation are established based on the node attributes; the node attributes include work timestamp, work content description information, project identifier, work type and related personnel information; Based on the node attributes of the nodes, construct the edges between the nodes.

[0164] Based on any of the above embodiments, the map maintenance unit is specifically used for: Based on the data type of the working data corresponding to the intermediate representation, determine the associated intermediate representation of the working data associated with the intermediate representation; Based on the intermediate representation and the associated intermediate representation, extract the node attributes of the intermediate representation.

[0165] Based on any of the above embodiments, the map maintenance unit is specifically used for: Based on the project identifier and / or relevant personnel information of the nodes, establish temporal edges between the nodes; Based on at least one of the following: the co-existing entities in the work data corresponding to the nodes, the semantic similarity between the work content description information of the nodes, and the causal relationship between the node attributes of the nodes, an association edge is established between the nodes.

[0166] Based on any of the above embodiments, the map maintenance unit is further configured to: Based on the semantic vectors of the work content description information of the nodes, the nodes are clustered to obtain multiple node clusters; Based on a large language model, different nodes within the same node cluster are merged and / or their work content description information is disambiguated.

[0167] Based on any of the above embodiments, the subgraph analysis unit is specifically used for: From the work knowledge graph, nodes whose work timestamps belong to a first time range and whose relevant personnel information includes the target personnel are selected as target nodes; Based on the target node and its associated nodes in the working knowledge graph, the first working subgraph is obtained; the associated nodes are nodes in the working knowledge graph that are connected to the target node through edges.

[0168] Based on any of the above embodiments, the map maintenance unit is further configured to: Based on the work summary corresponding to the first time range, a new report node is added to the work knowledge graph, and edges are established between the report node and each node in the first work subgraph.

[0169] Based on any of the above embodiments, the subgraph analysis unit is further configured to: The report nodes within the second time range are obtained from the working knowledge graph, and the second working subgraph within the second time range is determined based on the connection edges of the report nodes within the second time range. The summary generation unit is also used for: Based on the second working subgraph, a work summary corresponding to the second time range is generated.

[0170] Based on any of the above embodiments, the summary generation unit is specifically used for: Based on the density of nodes in the second working subgraph, the second working subgraph is divided into multiple working theme units; A work summary for the second time period is generated based on at least one of the summary information of the multiple work topic units, the project information of the second work subgraph, and the indicator information of the second work subgraph.

[0171] Based on any of the above embodiments, the device further includes a question-and-answer unit, used for: Obtain the user's question and determine the query complexity of the user's question; When the query complexity is complex, the user question is broken down into multiple sub-queries; Extract the user's question or the query intent and query entity of the multiple subqueries; Based on the query intent, the query entity, and the working knowledge graph, a response text for the user's question is generated.

[0172] Figure 4 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 4 As shown, the electronic device may include: a processor 410, a communications interface 420, a memory 430, and a communication bus 440, wherein the processor 410, the communications interface 420, and the memory 430 communicate with each other via the communication bus 440. The processor 410 can call logical instructions in the memory 430 to execute a work summary generation method, which includes: Obtain a job knowledge graph, wherein the nodes and edges in the job knowledge graph are constructed based on job data; From the work knowledge graph, a first work subgraph within a first time range is determined, and based on the project identifier and work type of the nodes in the first work subgraph, the first work subgraph is divided into project type units; Based on the work timestamps of the nodes in the project type unit and the edges between the nodes, a work task chain corresponding to the project of the project type unit is generated. Based on the work timestamps of the nodes in the project type unit, the work content description information and the relevant personnel information, a summary information of the project type unit is generated. Based on the work task chain and the summary information, a work summary corresponding to the first time range is generated.

[0173] Furthermore, the logical instructions in the aforementioned memory 430 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, or the part that contributes to related technologies, or a portion 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.

[0174] On the other hand, the present invention also provides a computer program product, the computer program product comprising a computer program that can be stored on a non-transitory computer-readable storage medium, wherein when the computer program is executed by a processor, the computer is able to execute the work summary generation method provided by the above methods, the method comprising: Obtain a job knowledge graph, wherein the nodes and edges in the job knowledge graph are constructed based on job data; From the work knowledge graph, a first work subgraph within a first time range is determined, and based on the project identifier and work type of the nodes in the first work subgraph, the first work subgraph is divided into project type units; Based on the work timestamps of the nodes in the project type unit and the edges between the nodes, a work task chain corresponding to the project of the project type unit is generated. Based on the work timestamps of the nodes in the project type unit, the work content description information and the relevant personnel information, a summary information of the project type unit is generated. Based on the work task chain and the summary information, a work summary corresponding to the first time range is generated.

[0175] 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 work summary generation method provided by the above methods, the method comprising: Obtain a job knowledge graph, wherein the nodes and edges in the job knowledge graph are constructed based on job data; From the work knowledge graph, a first work subgraph within a first time range is determined, and based on the project identifier and work type of the nodes in the first work subgraph, the first work subgraph is divided into project type units; Based on the work timestamps of the nodes in the project type unit and the edges between the nodes, a work task chain corresponding to the project of the project type unit is generated. Based on the work timestamps of the nodes in the project type unit, the work content description information and the relevant personnel information, a summary information of the project type unit is generated. Based on the work task chain and the summary information, a work summary corresponding to the first time range is generated.

[0176] 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.

[0177] 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 parts that contribute to the related technology, can be embodied in the form of software products. 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.

[0178] 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. A method for generating work summaries, characterized in that, include: Obtain a job knowledge graph, wherein the nodes and edges in the job knowledge graph are constructed based on job data; From the work knowledge graph, a first work subgraph within a first time range is determined, and based on the project identifier and work type of the nodes in the first work subgraph, the first work subgraph is divided into project type units; Based on the work timestamps of the nodes in the project type unit and the edges between the nodes, a work task chain corresponding to the project of the project type unit is generated. Based on the work timestamps of the nodes in the project type unit, the work content description information and the relevant personnel information, a summary information of the project type unit is generated. Based on the work task chain and the summary information, a work summary corresponding to the first time range is generated.

2. The method for generating work summaries according to claim 1, characterized in that, The acquisition of the working knowledge graph includes: Transform the working data from multiple working data sources into intermediate representations; Based on the data type of the work data corresponding to the intermediate representation, the node attributes of the intermediate representation are extracted, and the nodes corresponding to the intermediate representation are established based on the node attributes; the node attributes include work timestamp, work content description information, project identifier, work type and related personnel information; Based on the node attributes of the nodes, construct the edges between the nodes.

3. The method for generating work summaries according to claim 2, characterized in that, The step of extracting node attributes of the intermediate representation based on the data type of the working data corresponding to the intermediate representation includes: Based on the data type of the working data corresponding to the intermediate representation, determine the associated intermediate representation of the working data associated with the intermediate representation; Based on the intermediate representation and the associated intermediate representation, extract the node attributes of the intermediate representation.

4. The method for generating work summaries according to claim 2, characterized in that, The process of constructing edges between nodes based on their node attributes includes: Based on the project identifier and / or relevant personnel information of the nodes, establish temporal edges between the nodes; Based on at least one of the following: the co-existing entities in the work data corresponding to the nodes, the semantic similarity between the work content description information of the nodes, and the causal relationship between the node attributes of the nodes, an association edge is established between the nodes.

5. The method for generating work summaries according to claim 2, characterized in that, The acquisition of the work knowledge graph also includes: Based on the semantic vectors of the work content description information of the nodes, the nodes are clustered to obtain multiple node clusters; Based on a large language model, node merging and / or disambiguation of work content description information are performed on different nodes within the same node cluster.

6. The method for generating work summaries according to claim 1, characterized in that, Determining the first working subgraph within a first time range from the working knowledge graph includes: From the work knowledge graph, nodes whose work timestamps belong to a first time range and whose relevant personnel information includes the target personnel are selected as target nodes; Based on the target node and its associated nodes in the working knowledge graph, the first working subgraph is obtained; the associated nodes are nodes in the working knowledge graph that are connected to the target node through edges.

7. The method for generating a work summary according to any one of claims 1 to 6, characterized in that, Also includes: Based on the work summary corresponding to the first time range, a new report node is added to the work knowledge graph, and edges are established between the report node and each node in the first work subgraph.

8. The method for generating work summaries according to claim 7, characterized in that, Also includes: The report nodes within the second time range are obtained from the working knowledge graph, and the second working subgraph within the second time range is determined based on the connection edges of the report nodes within the second time range. Based on the second working subgraph, a work summary corresponding to the second time range is generated.

9. The method for generating work summaries according to claim 8, characterized in that, The step of generating a work summary corresponding to the second time range based on the second work subgraph includes: Based on the density of nodes in the second working subgraph, the second working subgraph is divided into multiple working theme units; A work summary for the second time period is generated based on at least one of the summary information of the multiple work topic units, the project information of the second work subgraph, and the indicator information of the second work subgraph.

10. The method for generating a work summary according to any one of claims 1 to 6, characterized in that, Also includes: Obtain the user's question and determine the query complexity of the user's question; When the query complexity is complex, the user question is broken down into multiple sub-queries; Extract the user's question or the query intent and query entity of the multiple subqueries; Based on the query intent, the query entity, and the working knowledge graph, a response text for the user's question is generated.

11. A work summary generation device, characterized in that, include: The graph maintenance unit is used to acquire a working knowledge graph, wherein the nodes and edges in the working knowledge graph are constructed based on the working data. The subgraph analysis unit is used to determine the first working subgraph within a first time range from the working knowledge graph, and to divide the first working subgraph into project type units based on the project identifier and working type of the nodes in the first working subgraph. The project analysis unit is used to generate a work task chain for the project corresponding to the project type unit based on the work timestamps of the nodes in the project type unit and the edges between the nodes, and to generate summary information for the project type unit based on the work timestamps, work content description information and relevant personnel information of the nodes in the project type unit. The summary generation unit is used to generate a work summary corresponding to the first time range based on the work task chain and the summary information.

12. 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 work summary generation method as described in any one of claims 1 to 10.

13. 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 work summary generation method as described in any one of claims 1 to 10.

14. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the work summary generation method as described in any one of claims 1 to 10.

Images