A method and apparatus for automatic generation of intelligent report descriptions based on hierarchical attention pointer generation networks.

Abstract

This invention provides a method and apparatus for automatically generating intelligent report descriptions based on a hierarchical attention pointer generation network. By acquiring a multi-row structured data table, adding contribution values, and determining one or more required data rows based on a leaf node contribution model, the data rows are flattened to form single-row data. A description of the data table is then generated based on the hierarchical attention pointer generation network model. This solves the problem of poor matching between the output of general descriptions and the input data of the data table in existing technologies. It directly generates more human-like general descriptions from logically complex multi-row structured data, concisely, efficiently, and professionally expressing the information described in the data table.

Description

Technical Field

[0001] This invention relates to the field of intelligent semantic analysis, and more specifically, to a method and apparatus for automatically generating intelligent report descriptions based on a hierarchical attention pointer generation network. Background Technology

[0002] Business intelligence (BI) analysts often need to convey the core content of data to users using natural language descriptions based on data insights. However, in some scenarios, such as anomaly detection and automated daily / weekly report generation, due to factors like rapid data updates and a large amount of repetitive work, it is necessary to automate the manual generation of these descriptions. Conveying important information to users accurately, concisely, and vividly presents a significant challenge.

[0003] Most mainstream business intelligence tools currently offer some form of Natural Language Generation (NLG) functionality. However, the mainstream NLG capabilities in the BI field are primarily rule-based and template-based. Template-based methods suffer from overly formatted content and difficulties in business expansion. Existing deep learning-based models rely heavily on data, requiring minimal human intervention and easily generating rich and fluent text descriptions. However, users often cannot directly control content generation, making it difficult to ensure accurate matching between the output and input data. Furthermore, specifically in the field of BI report description generation, the aforementioned deep learning solutions still suffer from a lack of flexibility, vividness, conciseness, and professionalism.

[0004] Therefore, there is a need for an intelligent method for automatically generating report descriptions based on hierarchical attention pointer generation networks, which can concisely, vividly, and professionally express the information that BI data tables want to describe, and can flexibly adapt to multiple business scenarios. Summary of the Invention

[0005] In view of this, the purpose of the present invention is to provide a method and apparatus for automatically generating intelligent report descriptions based on a hierarchical attention pointer generation network, so as to solve the problem of poor matching between the general description of intelligent reports and the report input data in the prior art, and directly generate more human-like general descriptions from logically complex multi-line structured data.

[0006] To solve the above-mentioned technical problems, the following solution is proposed:

[0007] A method for automatically generating intelligent report descriptions based on a hierarchical attention pointer generation network includes:

[0008] Add contribution values ​​to the acquired structured data table, which includes multiple data rows;

[0009] Based on the leaf node contribution model, one or more data rows are determined in the structured data table according to the contribution value;

[0010] The determined data rows are flattened to form single-row data;

[0011] The single row of data is input into the hierarchical attention pointer generation network model to generate a description of the data table.

[0012] Preferably, adding contribution values ​​to the acquired structured data table specifically includes:

[0013] Add a contribution value to one or more data rows, wherein the contribution value occupies one or more columns in the data table.

[0014] Preferably, the leaf node contribution model, which determines one or more data rows in the structured data table based on the contribution value, specifically includes:

[0015] Determine whether a data row in the structured data table is a leaf node or a non-leaf node;

[0016] Based on the contribution value of the specific data row, extract the data row containing the leaf node.

[0017] Preferably, the determined data rows are flattened to form single-row data, specifically including:

[0018] One or more of the determined data rows are arranged into a single row of data in a preset order, wherein the data rows are considered as paragraphs of the single row of data.

[0019] Preferably, the step of inputting the single row of data into the hierarchical attention pointer generation network model to generate a description of the data table specifically includes:

[0020] The single-line data is input into the hierarchical attention pointer generation network model to obtain descriptive frame words and structured extracted information;

[0021] The description of the data table is generated using the descriptive framework terms and the structured extracted information.

[0022] Preferably, obtaining descriptive framework terms and structured extraction information specifically includes:

[0023] The pointer probability distribution of specific anchor segments for each paragraph is obtained using the first-layer attention model. The pointer probability distribution of specific anchor segments is used to indicate the anchor segment distribution of the descriptive frame words and structured extracted information.

[0024] The positional probability distribution of specific information for each specific anchor segment is obtained using a second-layer attention model. This positional probability distribution of specific information is used to indicate the positional distribution of the descriptive frame words and structured extracted information.

[0025] Based on the pointer probability distribution of the specific anchor segment and the position probability distribution of the specific information, the output probability distribution of the descriptive frame words and structured extracted information is obtained;

[0026] The descriptive frame terms and structured extracted information are determined based on the output probability distribution.

[0027] An apparatus for automatically generating intelligent report descriptions based on a hierarchical attention pointer generation network, comprising:

[0028] A data processing unit is used to add contribution values ​​to an acquired structured data table, which includes multiple data rows.

[0029] A data filtering unit is used to determine one or more data rows in the structured data table based on the leaf node contribution model and the contribution value.

[0030] The data conversion unit is used to flatten the determined data rows to form single-row data.

[0031] The data description unit is used to input the single row of data into the hierarchical attention pointer generation network model to generate a description of the data table.

[0032] Preferably, the data processing unit specifically includes:

[0033] Add a contribution value to one or more data rows, wherein the contribution value occupies one or more columns in the data table.

[0034] Preferably, the data filtering unit specifically includes:

[0035] Determine whether a data row in the structured data table is a leaf node or a non-leaf node;

[0036] Based on the contribution value of the specific data row, extract the data row containing the leaf node.

[0037] Preferably, the data conversion unit specifically includes:

[0038] One or more of the determined data rows are arranged into a single row of data in a preset order, wherein the data rows are considered as paragraphs of the single row of data.

[0039] Preferably, the data description unit specifically includes:

[0040] The information extraction unit is used to input the single-line data into the hierarchical attention pointer generation network model to obtain descriptive frame words and structured extracted information;

[0041] The description generation unit is used to generate a description of the data table using the descriptive framework terms and the structured extracted information.

[0042] Preferably, the information extraction unit specifically includes:

[0043] The first probability processing unit is used to obtain the pointer probability distribution of specific anchor segments for each paragraph using the first-layer attention model. The pointer probability distribution of specific anchor segments is used to indicate the anchor segment distribution of the descriptive frame words and structured extracted information.

[0044] The second probability processing unit is used to obtain the position probability distribution of specific information for each specific anchor segment using the second-layer attention model. The position probability distribution of the specific information is used to indicate the position distribution of the descriptive frame words and the structured extracted information.

[0045] The third probability processing unit obtains the output probability distribution of the descriptive frame words and the structured extracted information based on the pointer probability distribution of the specific anchor segment and the position probability distribution of the specific information; the information determination unit is used to determine the descriptive frame words and the structured extracted information according to the output probability distribution.

[0046] An apparatus for automatically generating intelligent report descriptions based on a hierarchical attention pointer generation network includes a memory, a processor, and a computer program stored in the memory and executable on the processor. The processor executes the computer program to implement the method steps for automatically generating intelligent report descriptions based on a hierarchical attention pointer generation network.

[0047] A computer-readable storage medium storing a computer program, characterized in that, when executed by a processor, the computer program implements a method for automatically generating intelligent report descriptions based on a hierarchical attention pointer generation network.

[0048] As can be seen from the above technical solutions, the method for automatically generating intelligent report descriptions based on hierarchical attention pointer generation networks provided in this application obtains a multi-row structured data table, adds contribution values, determines one or more required data rows based on the leaf node contribution model, flattens the data rows to form single-row data, and generates a description of the data table based on the hierarchical attention pointer generation network model. This solves the problem of poor matching between the output of the general description and the input data of the data table in the prior art, and directly generates a more human-like general description from logically complex multi-row structured data, expressing the information described in the data table concisely, efficiently, and professionally. Attached Figure Description

[0049] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0050] Figure 1 This is one of the structured data tables in the intelligent report description automatic generation method based on hierarchical attention pointer generation network of the present invention.

[0051] Figure 2 This is the second structured data table in the intelligent report description automatic generation method based on hierarchical attention pointer generation network of the present invention.

[0052] Figure 3 This is a flowchart of the intelligent report description automatic generation method based on hierarchical attention pointer generation network of the present invention.

[0053] Figure 4 This is the third structured data table in the intelligent report description automatic generation method based on hierarchical attention pointer generation network of the present invention.

[0054] Figure 5 This is a flowchart of the flattening transformation of the intelligent report description automatic generation method based on hierarchical attention pointer generation network of the present invention.

[0055] Figure 6 This is a flowchart of the first embodiment of the automatic generation method for intelligent report descriptions of the present invention, which generates a data table description.

[0056] Figure 7 This is a flowchart of the second embodiment of the automatic generation method for generating data table descriptions based on hierarchical attention pointer generation network of the present invention.

[0057] Figure 8This is one of the structural schematic diagrams of the intelligent report description automatic generation device based on hierarchical attention pointer generation network of the present invention.

[0058] Figure 9 The second schematic diagram of the intelligent report description automatic generation device based on hierarchical attention pointer generation network of the present invention. Detailed Implementation

[0059] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0060] The method for automatically generating intelligent report descriptions based on hierarchical attention pointer generation networks provided by this invention is applicable to the field of intelligent semantic analysis, specifically involving generating more human-like generalized descriptions from logically complex multi-line structured data.

[0061] Business intelligence (BI), also known as business wisdom or business intelligence, refers to the use of modern data warehousing technology, online analytical processing technology, data mining, and data visualization technology to analyze data in order to realize business value.

[0062] Business intelligence is the process of collecting, managing, and analyzing business information, and it generally consists of data warehousing, online analytical processing (OLAP), data mining, data backup, and recovery. The implementation of business intelligence involves software, hardware, consulting services, and applications, and its basic architecture includes three parts: data warehousing, OLAP, and data mining.

[0063] Business intelligence is a solution that focuses on extracting and cleaning useful data from various enterprise operating systems to ensure its accuracy. This data is then extracted, transformed, and loaded—the ETL process—and merged into an enterprise-level data warehouse, providing a global view of the enterprise's data. Based on this, appropriate query and analysis tools, data mining tools (Big Data Mirror), and OLAP tools are used to analyze and process the data (at which point the information becomes knowledge to support decision-making). Finally, this knowledge is presented to managers to support their decision-making process.

[0064] Pointer-Generator Networks (PGNs) are a type of generative algorithm that combines pointing and generation methods, integrating extractive and abstract summarization. They allow word copying via pointing and can also generate words from a fixed vocabulary, thus addressing the issues of out-of-vocabulary words and duplicate generation to some extent.

[0065] Attention mechanism: Originating from research on human vision. In cognitive science, due to information processing bottlenecks, humans selectively focus on a portion of all information while ignoring other visible information; this mechanism is commonly referred to as the attention mechanism. The attention mechanism mainly has two aspects: deciding which part of the input to focus on; and allocating limited information processing resources to the important parts.

[0066] Drill up, which means going back up from the current data to the previous level of data.

[0067] Drill down: Expand the data to the next level from the current data.

[0068] Up-drilling and down-drilling are collectively referred to as drilling and extraction.

[0069] Slice: Displays data at the same level.

[0070] Business intelligence analysts often need to convey the core content of data to users using natural language descriptions based on data insights. However, in some scenarios, such as anomaly detection and automated daily / weekly report generation, due to factors like rapid data updates and a large amount of repetitive work, it's necessary to automate the manual generation of these descriptions. Conveying important information to users accurately, concisely, and vividly presents a significant challenge.

[0071] For example, regarding such Figure 1 The formatted data table shown might lead us to the following summary: "Last month, existing user retention still accounted for the majority of total DAU, while new user growth continued to decline year-on-year."

[0072] Such general descriptions are difficult to generate automatically using ordinary template techniques, especially when dealing with complex drill-down scenarios. Even more complex is the fact that... Figure 2 As shown, drill-down scenarios include not only DAU scenarios but also GMV scenarios. If templates are used, they need to be finely customized for different scenarios, which is very time-consuming and labor-intensive.

[0073] If we use the deep learning model from this solution, we only need to... Figure 1 The structured data table shown is used as training sample data block 1, and will be as follows: Figure 2The structured data table shown is used as training sample data block 2. The two are paired, and then the model can automatically learn the rules from the training data and generate descriptions based on the new data. This also helps analysts simplify the template configuration work to some extent.

[0074] Current research focuses on data-to-text conversion, attempting to use deep learning models to transform structured data into textual descriptions. This relies heavily on high-quality training data. However, often the text and structured data in the training data are not perfectly aligned, posing a significant challenge to structured data-to-text generation.

[0075] The paper "Operation-guided Neural Networks for High Fidelity Data-To-Text Generation" proposes a computation-guided neural text generation model. Specifically, this model pre-executes several predefined computations. During text generation, the decoder dynamically selects the information source from the input data and the results using a threshold mechanism. Furthermore, accurately conveying numerical information and optimizing language expression based on this understanding remains a significant challenge.

[0076] The paper "Learning Latent Semantic Annotations for Grounding Natural Language to Structured Data" mentions that the model uses different probabilistic models to model the correspondence between different types of variables, such as string variables, attribute variables, and numerical variables. Text that cannot be mapped to the input data is uniformly labeled with a specific empty tag. While acquiring diverse lexical expressions, the alignment results also provide rich rules and templates for natural language generation, combining interpretability and controllability.

[0077] Overall, the deep learning-based models mentioned above are primarily data-driven, requiring minimal human intervention and easily generating rich and fluent text descriptions. However, users often cannot directly control content generation, making it difficult to ensure an accurate match between the output and the input data. Furthermore, specifically in the field of BI report description generation, the aforementioned deep learning solutions still suffer from the following issues:

[0078] Inflexible: Poor scalability, requiring significant customization for different scenarios. For example, in drill-down scenarios, the business logic is complex, and the drill-down logic varies across different scenarios.

[0079] Lacks vividness: The content is formatted, the sentences are stiff, and it doesn't resemble natural language descriptions; it lacks vividness.

[0080] Not concise enough: There are too many objective factors described, especially numerical information, which makes it difficult to reveal the most critical factors and ignore the key trend factors hidden behind the numbers.

[0081] Insufficient professionalism: For data with complex business logic, the previous models could only handle some simple data logic, such as judging the win or loss relationship based on the score of a basketball game. They were not up to the task of business intelligence analysis scenarios with diverse business modalities.

[0082] In response to the above issues, such as Figure 3 As shown, the present invention provides a method for automatically generating intelligent report descriptions based on a hierarchical attention pointer generation network, specifically including:

[0083] Step 101: Add contribution values ​​to the acquired structured data table, which contains multiple data rows.

[0084] In most business intelligence (BI) analytics scenarios, describing drill-down results with hierarchical relationships is a major challenge for Natural Language Logic (NLG). Statistics show that over 60% of the analytical content in a BI report is related to drill-down logic. Data insight teams define the breadth and depth of drill-down based on certain logic, storing relevant data in different rows of the insight offline table. Previously, template-based solutions required the NLG model to search layer by layer through all insight data tables according to the drill-down nodes, revealing all relevant data information. Determining which nodes' information is more critical might require expert experience to define thresholds, but this definition method varies greatly across different businesses, and manually defining thresholds is not objective enough.

[0085] Step 102: Based on the leaf node contribution model, determine one or more data rows in the structured data table according to the contribution value.

[0086] Specifically, determine whether a data row in a structured data table is a leaf node or a non-leaf node; extract the data row containing the leaf node based on the contribution value of a specific data row.

[0087] Data annotation in BI is more objective and easier than template organization in BI. In a sense, the model also helps BI business abstract templates. Here, we propose a model based on the contribution of leaf nodes.

[0088] like Figure 4As shown, when a user asks, "How was today's DAU performance?" the returned analysis report should describe, in natural language, the prominent or abnormal information discovered during the drill-down process, layer by layer. In the original insight data, new users and returning users are two parallel nodes. The child and grandchild nodes under the returning user node, such as the historical recall user and returning user node, actually have an inclusion relationship. However, in the NLG model, it is difficult to distinguish the hierarchical relationship between historical recall users and new users, and between historical recall users and returning users. To solve the problem of the model's difficulty in understanding the hierarchical relationships in the insight data, we added a contribution column to each row of data.

[0089] Therefore, this invention only needs to focus on the leaf node data when analyzing data. For example, new users account for 15% of the total, and returning users account for 85%. Specifically, within the returning user leaf node, historically recalled users account for 5% of the total, and the retention rate of returning users from last month accounts for 57%. By focusing only on leaf nodes, historically recalled users and new users are both leaf nodes in the current scenario, and they are parallel. This ensures that returning user and last month's returning user retention information are not duplicated. On the other hand, based on the contribution-based approach, we need the model to understand the meaning of the current contribution value, and then the model ultimately determines which node information corresponds to which contribution value in the current scenario. A higher contribution value does not necessarily mean it will be described; the specific details need to be learned based on the distribution of the training samples. In the diagram, the rows marked with gray boxes are all leaf nodes; the white rows are not leaf nodes and are not considered here.

[0090] from Figure 4 As can be seen, there are four leaf points: new users, last month's new user retention, last month's old user retention, and historical recalled users. Of course, contribution can have more than one column. Contribution can be calculated for both absolute values ​​and year-on-year comparisons. This example only shows the contribution for absolute values. Finally, a multi-row structured information related to drill-down can be flattened after adding a contribution column and then used as a single row of input.

[0091] Step 103: Flatten the determined data rows to form single-row data.

[0092] Specifically, one or more defined data rows are arranged into a single row of data in a preset order, where each data row is a paragraph of the single row of data.

[0093] We have prepared a structured business intelligence data table and want to describe the information in the table using natural language. The ultimate goal of our model is to achieve end-to-end functionality for converting structured data into a generalized natural language description. The input may be one or more lines of structured data. Describing the data line by line would be overly formatted and make it difficult to simultaneously summarize and learn from multiple lines of information, failing to grasp the overall trend and obtain a comprehensive description. Therefore, as... Figure 5 As shown, in the data preprocessing stage, we need to flatten the multi-line formatted data.

[0094] After flattening, this data is mapped to corresponding natural language descriptions. As shown in the diagram, some information is framework-level information, which functions similarly to templates (including the black parts in the text). It primarily describes the structured information, and this part depends mainly on the training corpus. The language probability model learned from the training corpus generates the most suitable word for the current position. Another part of the information needs to be extracted from the input table (the blue parts in the text). This mainly relies on the Pointer mechanism in the model; that is, this part of the information is not directly derived from the language probability model, but rather the model learns which coordinate in the structured input should be retrieved for the current position. The red parts in the text require further descriptive learning based on an understanding of numerical information; here, they are not distinguished from the black parts and are collectively referred to as framework-level descriptive vocabulary. The training corpus preparation issue will be described in detail later.

[0095] Step 104: Input the single row of data into the hierarchical attention pointer to generate the network model and generate a description of the data table.

[0096] Specifically, a single row of data is input into a hierarchical attention pointer generation network model to obtain descriptive frame terms and structured extraction information; the descriptive frame terms and the structured extraction information are then used to generate a description of the data table.

[0097] Specifically, obtaining descriptive framework terms and structured extraction information includes:

[0098] The pointer probability distribution of specific anchor segments for each paragraph is obtained using the first-layer attention model. The pointer probability distribution of specific anchor segments is used to indicate the anchor segment distribution of descriptive frame words and structured extracted information.

[0099] The second-layer attention model is used to obtain the positional probability distribution of specific information for each specific anchor segment. The positional probability distribution of specific information is used to indicate the positional distribution of descriptive frame words and structured extracted information.

[0100] Based on the pointer probability distribution of specific anchor segments and the position probability distribution of specific information, the output probability distribution of descriptive frame words and structured extracted information is obtained.

[0101] Descriptive frame words and structured extracted information are determined based on the output probability distribution.

[0102] When describing structured information, as mentioned in Section 1, one part of the information is the framework description of the text, and the other part is extracted from the structured information. Typically, we can obtain a description similar to the following using the PGN model:

[0103] Key Account (KA) merchants (Level 100): Daily GMV in March was 2.02 million RMB, a year-on-year decrease of 5.8%, which slowed down by 6.1 percentage points compared to January and February. Meanwhile, the transaction growth rates of other merchant levels all increased significantly.

[0104] Average store traffic: The average daily number of members visiting the store in March was 7,132, a year-on-year decrease of 17.7%, continuing the negative growth trend.

[0105] Average store conversion rate: The average store conversion rate in March was 3.1%, a year-on-year increase of 27.3%.

[0106] Average transaction value per store: The average transaction value per store in March was $16, a year-on-year decrease of 9.9%, which is basically stable.

[0107] However, as mentioned in the section on drawbacks above, this type of information listing is not vivid or concise enough, and users still find it difficult to directly grasp the core information of the entire text. Therefore, we hope to further summarize it as follows: "Merchant stratification: The transaction growth rate of top-tier (KA) merchants (level 100) continues to slow down, and the average store traffic growth rate has significantly declined; large KA merchants (mostly standard categories) generally experienced negative transaction growth, which requires attention." As you can see, this is a natural language expression, not a list of information. It doesn't deliberately reveal dazzling numerical information, but rather highly abstractly expresses the important trends and core concerns contained within the numerical information using generalized natural language.

[0108] There are two approaches to moving from listing basic information to providing a general description.

[0109] First embodiment

[0110] like Figure 6 As shown, we additionally train a deep learning model. The encoder's input is a basic information description, and the decoder's output is a summary. In other words, to obtain a summary of structured information, we need two deep learning models: 1. A PGN model that transforms structured information into a basic information description. 2. A model that summarizes the basic information description. This approach is feasible.

[0111] Second Embodiment

[0112] This paper proposes an anchor-hierarchical attention pointer network model, which is directly trained end-to-end. The encoder input is flattened structured data, and the encoder output is directly a summarization model. From this... Figure 7 As shown in the algorithm framework diagram, the square portion in the flattened data represents an anchor. Without additional information, it can be a placeholder; in drill-down scenarios, it can be a placeholder with contribution information, depending on the scenario requirements. The entire diagram actually uses a PGN model similar to a hierarchical attention mechanism. The entire descriptive language still consists of descriptive frame words and structured extracted information. However, the combination of frame words and extracted words here is more flexible; for example, the extracted information in the same sentence can even come from different lines in the original structured information. Here, the first-layer attention mechanism is mainly used to learn from which anchor segment the current output word should start extracting information, ideally obtaining a pointer probability distribution for an anchor segment. The second-layer attention mechanism focuses on learning which words to extract within a specified segment after the specified anchor segment. This part involves the PGN problem, which is learning the output probability distribution of frame words and extracted information; we call this distribution the local merge distribution. In the merging stage, the second-layer local merge distribution is multiplied by the first-layer anchor segment pointer probability distribution to obtain the multiply distribution. In this way, during the final general description decoding stage, each step automatically determines whether to output framework descriptions or extractable information from a specific location within a particular anchor segment. This results in an end-to-end model.

[0113] For understanding numerical information, since numerical types are also extracted as words, while ensuring the accuracy of the numbers, numerical types can be pre-marked to distinguish them from non-numerical types. Then, during the decoding stage, if a numerical type mark is encountered, it can be paired with certain threshold rules to generate different descriptions.

[0114] For the above solutions, we all need to construct training samples that transform structured data into descriptions. In the actual business intelligence field, readily available data for this is difficult to find. Therefore, we adopt a semi-automated generation solution combining expert experience. There are two forms: 1. A semi-automated sample generation solution based on rules and manual labeling: Descriptive terms for the data, such as "significant growth" and "stable," are manually selected and labeled. The full description generated from the template is manually labeled, removing duplicate or less informative items / dimensions. 2. Expert labeling or user drag-and-drop: Data accumulation is achieved by using user drag-and-drop template splicing. BI labeling data is more objective and easier than BI template organization; in a sense, the model also helps BI business abstract templates. Next, we need to construct samples that transform basic information descriptions into general descriptions. This part requires some expert experience for labeling. This part of the training data can directly use general descriptions transformed from structured information to expert experience labeling.

[0115] To address the aforementioned drawbacks, the present invention solves them through the following methods:

[0116] To address the issues of inaccuracy and lack of professionalism, a pointer-based network approach is used to extract crucial information directly from structured data, ensuring accurate data presentation. Simultaneously, by incorporating numerical type labeling and threshold rules, the system aims to understand numerical values ​​to a certain extent and generate corresponding descriptions.

[0117] To address the issues of insufficient flexibility and poor scalability, a scheme based on leaf node contribution is proposed, which effectively solves the most common and complex drill-down logic processing in the field of business intelligence.

[0118] To address the issues of insufficient vividness and conciseness, an end-to-end anchor hierarchical attention mechanism model is proposed. This model directly generates more human-like general descriptions from complex multi-line structured data, avoiding redundant data descriptions and directly reaching the key trend factors hidden behind the numerical values.

[0119] The present invention achieves the following technical effects:

[0120] We propose an anchor-based hierarchical attention pointer network model that combines multi-layer attention mechanisms with pointers, which can directly generate more human-like generalized descriptions from logically complex multi-line structured data.

[0121] A scheme based on the contribution of leaf nodes is proposed, which focuses on solving the complex business intelligence drill-down logic processing.

[0122] A pointer scheme combining numerical labeling and threshold rules is proposed, which can accurately reveal numerical information and also provide a certain degree of comprehensible description of the numerical information.

[0123] Based on the same concept as the intelligent report description automatic generation method based on hierarchical attention pointer generation network provided above, this invention also provides an apparatus for intelligent report description automatic generation based on hierarchical attention pointer generation network, such as... Figure 8 As shown, the device includes: a data processing unit 100, a data filtering unit 200, a data conversion unit 300, and a data description unit 400.

[0124] The data processing unit 100 is used to add contribution values ​​to the acquired structured data table, which includes multiple data rows.

[0125] The data filtering unit 200 is used to determine one or more data rows in the structured data table based on the contribution value of the leaf node contribution model.

[0126] The data conversion unit 300 is used to flatten the determined data rows to form single-row data.

[0127] The data description unit 400 is used to input a single row of data into the hierarchical attention pointer to generate a network model and generate a description of the data table.

[0128] Based on the same concept as the intelligent report description automatic generation method based on hierarchical attention pointer generation network provided above, this invention also provides an apparatus for intelligent report description automatic generation based on hierarchical attention pointer generation network, such as... Figure 9 As shown, the device includes: a memory 101, a processor 102, and a computer program stored in the memory and executable on the processor 102. When the processor 102 executes the computer program, it implements method steps for sign blur detection.

[0129] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. The terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0130] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For apparatus embodiments, since they are fundamentally similar to method embodiments, the descriptions are relatively simple; relevant parts can be referred to the descriptions in the method embodiments.

[0131] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0132] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0133] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0134] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0135] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of methods, apparatus, and computer program products according to various embodiments of the present invention. In this regard, each block in the flowcharts and block diagrams may represent a module, program segment, or portion of code, containing one or more computer-executable instructions for implementing logical functions. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. Furthermore, it should be noted that each block or combination of blocks in the block diagrams and flowcharts may be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

Claims

1. A method for automatically generating intelligent report descriptions based on a hierarchical attention pointer generation network, characterized in that, The method includes: Add contribution values ​​to the acquired structured data table, which includes multiple data rows; Based on the leaf node contribution model, one or more data rows containing leaf nodes are determined in the structured data table according to the contribution value. The data rows containing one or more leaf nodes are flattened to form single rows of data. The single-line data is input into a hierarchical attention pointer generation network model to generate a description of the data table. The hierarchical attention pointer generation network model is used to obtain descriptive frame terms and structured extracted information based on the input single-line data. The descriptive frame terms and the structured extracted information are determined according to an output probability distribution. This output probability distribution is obtained based on the pointer probability distribution of specific anchor segments in each paragraph and the position probability distribution of specific information in each specific anchor segment. The pointer probability distribution of specific anchor segments in each paragraph is obtained using a first-layer attention model, and the position probability distribution of specific information in each specific anchor segment is obtained using a second-layer attention model. The pointer probability distribution of specific anchor segments indicates the anchor segment distribution of the descriptive frame terms and the structured extracted information, and the position probability distribution of specific information indicates the position distribution of the descriptive frame terms and the structured extracted information.

2. The method according to claim 1, characterized in that, The step of adding contribution values ​​to the acquired structured data table specifically includes: Add a contribution value to one or more data rows, wherein the contribution value occupies one or more columns in the data table.

3. The method according to claim 2, characterized in that, The leaf node contribution model determines one or more data rows in the structured data table based on the contribution value, specifically including: Determine whether a data row in the structured data table is a leaf node or a non-leaf node; Based on the contribution value of the data rows in the structured data table, extract the data rows containing the leaf nodes.

4. The method according to claim 1, characterized in that, The data rows containing one or more leaf nodes are flattened to form single rows of data, specifically including: The data rows containing one or more leaf nodes are arranged into a single row of data in a preset order, wherein the data rows are used as paragraphs of the single row of data.

5. The method according to claim 4, characterized in that, The step of inputting the single row of data into the hierarchical attention pointer generation network model to generate a description of the data table specifically includes: The single-line data is input into the hierarchical attention pointer generation network model to obtain descriptive frame words and structured extracted information; The description of the data table is generated using the descriptive framework terms and the structured extracted information.

6. A device for automatically generating intelligent report descriptions based on a hierarchical attention pointer generation network, characterized in that, The device includes: A data processing unit is used to add contribution values ​​to an acquired structured data table, which includes multiple data rows. A data filtering unit is used to determine the data rows containing one or more leaf nodes in the structured data table based on the leaf node contribution model and the contribution value. The data conversion unit is used to flatten the data rows containing the one or more leaf nodes to form a single row of data. A data description unit is used to input the single-line data into a hierarchical attention pointer generation network model to generate a description of the data table. The hierarchical attention pointer generation network model is used to obtain descriptive frame terms and structured extracted information based on the input single-line data. The descriptive frame terms and the structured extracted information are determined according to an output probability distribution. The output probability distribution is obtained based on the pointer probability distribution of specific anchor segments in each paragraph and the position probability distribution of specific information in each specific anchor segment. The pointer probability distribution of specific anchor segments in each paragraph is obtained using a first-layer attention model, and the position probability distribution of specific information in each specific anchor segment is obtained using a second-layer attention model. The pointer probability distribution of specific anchor segments is used to indicate the anchor segment distribution of the descriptive frame terms and the structured extracted information, and the position probability distribution of specific information is used to indicate the position distribution of the descriptive frame terms and the structured extracted information.

7. The apparatus according to claim 6, characterized in that, The data processing unit specifically includes: Add a contribution value to one or more data rows, wherein the contribution value occupies one or more columns in the data table.

8. The apparatus according to claim 7, characterized in that, The data filtering unit specifically includes: Determine whether a data row in the structured data table is a leaf node or a non-leaf node; Based on the contribution value of the data rows in the structured data table, extract the data rows containing the leaf nodes.

9. The apparatus according to claim 6, characterized in that, The data conversion unit specifically includes: The data rows containing one or more leaf nodes are arranged into a single row of data in a preset order, wherein the data rows are used as paragraphs of the single row of data.

10. The apparatus according to claim 9, characterized in that, The data description unit specifically includes: The information extraction unit is used to input the single-line data into the hierarchical attention pointer generation network model to obtain the descriptive frame words and the structured extracted information; The description generation unit is used to generate a description of the data table using the descriptive framework terms and the structured extracted information.

11. An apparatus for automatically generating intelligent report descriptions based on a hierarchical attention pointer generation network, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1-5.

12. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1-5.

Images