Text summarization method and apparatus

Abstract

This disclosure provides a text summarization method and apparatus, including receiving a user question and a data source, obtaining raw query results related to the user question from the data source, configuring at least one target low-rank adaptive label for the user question, obtaining customized low-rank adaptive weights corresponding to the user question based on the target low-rank adaptive label, generating a text summarization model based on the customized low-rank adaptive weights, and inputting the user question and the raw query results into the text summarization model to obtain a text summary. This disclosure's embodiments dynamically combine multiple low-rank adaptive modules to update and obtain a large multi-domain model, meeting diverse text summarization needs. Simultaneously, it performs multi-dimensional segmentation and intelligent compression on NL2SQL-generated data, inputting it into the optimized model, overcoming input length and memory limitations, and achieving efficient and accurate processing of text summarization tasks.

Description

Technical Field

[0001] This disclosure relates to the field of information processing technology, and in particular to a text summarization method and apparatus. Background Technology

[0002] This section is intended to provide background or context for the embodiments of this disclosure as set forth in the claims. The description herein is not intended to be a prior art simply because it is included in this section.

[0003] Traditional data-driven text summarization solutions typically employ the following process: The user's natural language request is converted into executable SQL using an NL2SQL (Natural Language to SQL) model. After establishing a database connection, a cursor is created and the SQL statement is executed to retrieve the full dataset, which is then formatted into a structured format. This structured format is then converted into natural language easily understood by a large language model. Based on this model, prompts for the data requirements are constructed, and the large language model generates and outputs a text summary based on these prompts.

[0004] When faced with text summarization tasks across multiple categories and domains, a single general-purpose large language model is insufficient to simultaneously meet diverse needs. Deploying a dedicated large model for each specific category leads to redundant investment and significant waste of computing resources, substantially increasing hardware costs and computational power consumption, making it uneconomical and impractical for engineering applications. Furthermore, limited by the input length and GPU memory constraints of large language models, when the amount of data returned by NL2SQL queries is too large, or the total length of the input text exceeds the model's processing capacity, data cannot be completely input into the model, resulting in the inability to generate a valid text summary. In severe cases, this can trigger an OutOfMemoryError (OOM) error, causing system crashes. Summary of the Invention

[0005] In view of this, the purpose of this disclosure is to propose a text summarization method and apparatus, which at least to some extent solves one of the technical problems in the related art.

[0006] To achieve the above objectives, the first aspect of this exemplary embodiment provides a text summarization method, comprising:

[0007] Receive user questions and data sources; obtain raw query results related to user questions from data sources; configure at least one target low-rank adaptive label for user questions; obtain customized low-rank adaptive weights corresponding to user questions based on target low-rank adaptive labels; generate a text summary model based on customized low-rank adaptive weights; input user questions and raw query results into the text summary model to obtain a text summary.

[0008] Based on the same inventive concept, a second aspect of the exemplary embodiments of this disclosure provides a text summarization apparatus, comprising: The information receiving module is configured to receive user questions and data sources; the data source query module is configured to obtain raw query results related to user questions from the data source; the tag acquisition module obtains multiple low-rank adaptive tags carried by user questions; the large model acquisition module is configured to obtain customized low-rank adaptive weights corresponding to user questions based on target low-rank adaptive tags; generate a text summary model based on customized low-rank adaptive weights; and the text summary module is configured to input user questions and raw query results into the text summary model to obtain a text summary.

[0009] Based on the same inventive concept, a third aspect of the exemplary embodiments of this disclosure 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 method as described in the first aspect.

[0010] Based on the same inventive concept, a fourth aspect of the exemplary embodiments of this disclosure provides a non-transitory computer-readable storage medium storing computer instructions for causing a computer to perform the method as described in the first aspect.

[0011] Based on the same inventive concept, a fifth aspect of the exemplary embodiments of this disclosure provides a computer program product including computer program instructions that, when run on a computer, cause the computer to perform the method as described in the first aspect.

[0012] As can be seen from the above, the embodiments of this disclosure, on the one hand, generate a large model that meets the needs of multiple domains through the dynamic combination of multiple low-rank adaptive modules, which can meet the needs of various types of text summarization and provide customized optimal model support for text summarization in various task scenarios; on the other hand, by performing multi-dimensional segmentation and intelligent compression on the data generated by Natural Language to Structured Query Language (NL2SQL), and inputting the processed data into the updated large language model that uses dynamically combined low-rank adaptive modules, the limitations of input length and memory capacity of existing large language models are overcome, and efficient and accurate processing of text summarization tasks can be achieved. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in this disclosure 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 only embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1A and Figure 1BA schematic diagram illustrating the original database query and text summary results according to an embodiment of the present disclosure is shown; Figure 2 A schematic diagram of the text summarization device module structure according to an embodiment of the present disclosure is shown; Figure 3 A schematic flowchart of a text summarization method according to an embodiment of the present disclosure is shown; Figure 4 A schematic diagram of the compression process for excessive data according to an embodiment of the present disclosure is shown; Figure 5 A schematic diagram of the hardware structure of a text summary electronic device according to an embodiment of the present disclosure is shown. Detailed Implementation

[0015] It is understood that before using the technical solutions disclosed in the various embodiments of this application, users should be informed of the types, scope of use, and usage scenarios of the personal information involved in this application in an appropriate manner in accordance with relevant laws and regulations, and user authorization should be obtained.

[0016] For example, upon receiving a user's active request, a prompt message is sent to the user to explicitly inform them that the requested operation will require the acquisition and use of the user's personal information. This allows the user to independently choose whether to provide personal information to the software or hardware, such as the electronic device, application, server, or storage medium performing the operations of this application's technical solution, based on the prompt message.

[0017] As an optional but non-limiting implementation, in response to a user's active request, sending a prompt message to the user can be done via a pop-up window, where the prompt message can be presented in text format. Furthermore, the pop-up window can also include a selection control allowing the user to choose "agree" or "disagree" to provide personal information to the electronic device.

[0018] It is understood that the above notification and user authorization process are merely illustrative and do not constitute a limitation on the implementation of this application. Other methods that comply with relevant laws and regulations may also be applied to the implementation of this application.

[0019] It is understood that the data involved in this technical solution (including but not limited to the data itself, the acquisition or use of the data) shall comply with the requirements of relevant laws, regulations and related provisions.

[0020] To make the objectives, technical solutions, and advantages of this disclosure clearer, the principles and spirit of this disclosure will be described below with reference to several exemplary embodiments. It should be understood that these embodiments are provided merely to enable those skilled in the art to better understand and implement this disclosure, and are not intended to limit the scope of this disclosure in any way. Rather, these embodiments are provided to make this disclosure more thorough and complete, and to fully convey the scope of this disclosure to those skilled in the art.

[0021] In this article, it is important to understand that any number of elements in the accompanying figures is for illustrative purposes and not for limitation, and any naming is for distinction only and has no limiting meaning.

[0022] It should be noted that, unless otherwise defined, the technical or scientific terms used in the embodiments of this disclosure should have the ordinary meaning understood by one of ordinary skill in the art to which this disclosure pertains. The terms "first," "second," and similar words used in the embodiments of this disclosure do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly. The article "a" or "an" preceding an element does not exclude the existence of multiple such elements.

[0023] The principles and spirit of this disclosure will be explained in detail below with reference to several representative embodiments. Invention Overview As described in the background section, traditional data-based text summarization solutions face text summarization tasks across multiple categories and fields. A single general-purpose large language model is insufficient to simultaneously meet diverse needs. On the other hand, deploying dedicated large models for each sub-category problem results in redundant investment and significant waste of computing resources, significantly increasing hardware costs and computing power consumption, and lacks economic efficiency and feasibility in engineering applications. Furthermore, due to the limitations of the input length and memory capacity of large language models, traditional data-based text summarization solutions may fail to input complete data into the model when the amount of data returned by the NL2SQL query is too large or the total length of the input text exceeds the model's processing range. Consequently, effective text summaries cannot be generated, and in severe cases, an OutOfMemoryError (OOM) may be triggered, leading to system crashes.

[0025] To better understand the technical problem addressed by this disclosure, let's assume the user's question is "How to obtain TV price trends for the next 4 months and the next 12 months". Figure 1A The original query results related to the user's question, obtained from the data source, are shown. Figure 1A As can be seen, the original query results contain 39 pages of information, which is a very large amount of data. If the original query results are directly input into the existing large model for text summarization, the input length limit and memory capacity constraint of the large language model will result in the data not being fully input into the model, thus failing to generate a valid text summary, and may even lead to a memory overflow error, causing the system to crash.

[0026] In addition, if the original query result data volume is not very large, it can be input into the existing large model used for text summarization. However, since the existing large model is a single general-purpose large language model, it is difficult to meet the diverse needs at the same time when facing text summarization tasks of multiple categories and multiple fields. The solution of independently deploying multiple dedicated single large models for each sub-category problem will result in the repeated investment of computing resources and a great waste, significantly increasing hardware costs and computing power consumption, and lacking economic efficiency and feasibility in engineering applications.

[0027] To address the aforementioned issues, this disclosure provides a text summarization method, specifically including: receiving a user question and a data source; obtaining original query results related to the user question from the data source; obtaining multiple low-rank adaptive tags carried by the user question; obtaining multiple low-rank adaptive modules adapted to the multiple low-rank adaptive tags from a low-rank adaptive weight storage; generating customized low-rank adaptive weights for the user question based on the adapted multiple low-rank adaptive modules; combining the customized low-rank adaptive weights into a large model to obtain an updated large model; determining whether the data volume of the original query results exceeds a set threshold; and, in response to the data volume of the original query results not exceeding the set threshold, inputting the user question and the original query results into the updated large model to obtain a text summary. Optionally, the text summarization method of this disclosure further includes: in response to the data volume of the original query results exceeding the predetermined threshold, performing pattern linking compression on the original query results to obtain compressed query results; performing data segmentation and keyword compression on the compressed query results to obtain processed query results; and inputting the user question and the processed query results into the updated large model to obtain the text summary. The technical solution proposed in this disclosure generates a large, updated model that satisfies multiple domains by constructing a dynamic combination of multiple low-rank adaptive modules, thereby meeting the needs of various types of text summarization. Its multi-module collaborative processing architecture enables a deep understanding of database query results, providing customized optimal model support for text summarization in various task scenarios. Preferably, this method further performs multi-dimensional segmentation and intelligent compression on the large amount of query data generated by Natural Language to Structured Query Language (NL2SQL), and inputs the processed data into an updated large language model using dynamically combined low-rank adaptive modules. This overcomes the input length limitations and memory capacity constraints of existing large language models, enabling efficient and accurate processing of text summarization tasks.

[0028] Figure 1B This shows the text summary that is expected based on the user's question and the original query results. Figure 1B The text summary performs semantic analysis on the original query results based on time recognition capabilities, trend analysis capabilities, and financial calculation capabilities.

[0029] After introducing the basic principles of this disclosure, various non-limiting embodiments of this disclosure will be described in detail below.

[0030] Exemplary methods and apparatus specifically refer to Figure 2 , Figure 2 A block diagram schematically illustrates an apparatus 200 for summarizing text according to an embodiment of the present invention. According to... Figure 2The device 200 includes: an information receiving module 202, a data source query module 204, a tag acquisition module 206, a large model acquisition module 208, and a text summarization module 210.

[0031] Information receiving module 202 is configured to receive user questions and data sources; The data source query module 204 is configured to obtain raw query results related to the user's question from the data source. The tag acquisition module 206 acquires multiple low-rank adaptive tags carried by the user's question; The large model acquisition module 208 is configured to obtain customized low-rank adaptive weights corresponding to user questions based on the target low-rank adaptive labels; and generate a text summary model based on the customized low-rank adaptive weights. The text summary module 210 is configured to input user questions and raw query results into the text summary model to obtain a text summary.

[0032] For ease of description, the above apparatus is described in terms of its functions, divided into various modules. Of course, in implementing this disclosure, the functions of each module can be implemented in one or more software and / or hardware.

[0033] The operating modes of each module of device 200 can be combined with subsequent... Figure 3 This can be explained using the following method.

[0034] specifically refer to Figure 3 , Figure 3 An illustration of a text summarization method 300 according to an embodiment of the present disclosure is shown, specifically including: Step 301: The information receiving module 202 receives user questions and data sources. Specifically, methods for receiving user questions and data sources can include input through interactive entry points (e.g., text input boxes, conversational interactions, or template selection), API interface input (e.g., RESTful API, GraphQL precise query interfaces, or WebSocket real-time dialogue scenarios), or batch import of uploaded Excel / CSV / TXT files. User questions can be in text, JSON, or XML format, while data sources can be in database table structure, CSV, Excel, Parquet, or JSON format. It is important to note that when receiving user questions and data sources, NL2SQL must know the table structure, fields, data types, and relationships to generate correct SQL. Data sources can be accessed through structured or semi / unstructured data sources.

[0035] Step 302: The data source query module 204 obtains the original query results related to the user's question from the data source. Specifically, obtaining the query results can include obtaining the original query results by interpreting the SQL message through a large language model. Specifically, this can be achieved by using a large language model to read the user's question, connect to the database for parsing, and finally return the original query results.

[0036] It is important to emphasize that the above embodiments are merely one possible implementation method for obtaining raw query results related to the user's question. Those skilled in the art should understand that the methods for obtaining raw query results are not limited to the above embodiments. In fact, those skilled in the art can implement this step using various methods such as large language models and database analysis. The key is to accurately understand the user's question and generate accurate raw query results.

[0037] Step 303: The label acquisition module 206 configures at least one target low-rank adaptive (LORA) label for the user question. Specifically, this step can be implemented by analysis through a large language model or by comparison and lookup in a database.

[0038] Step 304: The large model acquisition module 208 obtains the customized low-rank adaptive weights corresponding to the user problem based on the target low-rank adaptive label.

[0039] Specifically, this method can obtain multiple low-rank adaptive modules adapted to multiple low-rank adaptive labels from an external low-rank adaptive weight storage, and then generate customized low-rank adaptive weights for the user's problem based on the adapted multiple low-rank adaptive modules. Specific customization methods and embodiments are described below.

[0040] In step 305, the large model acquisition module 208 further generates a text summary model based on customized low-rank adaptive weights. The technical solution in step 305 differs from traditional low-rank adaptive model weight loading methods. In existing technologies, only one type of low-rank adaptive weight can be loaded each time a large model requiring low-rank adaptive configuration is started. However, the architecture built using the method of this application can load multiple low-rank adaptive weights at once, and can dynamically route or concatenate different low-rank adaptive weights according to different problems, enabling a single model to have multiple text semantic understanding capabilities. Specific combination methods and embodiments are described below.

[0041] Step 306: The text summary module 210 inputs the user's question and the original query results into the text summary model to obtain a text summary. This text summary, while retaining core data conclusions, refines, logically organizes, and paraphrases the query results in natural language, while simultaneously outputting a structured and easily understandable concise answer that aligns with the user's question intent. Optionally, the original query results input into the text summary model include a compression step. This compression step determines whether to compress the original query results by statistically analyzing the number of rows, fields, and total data bytes of the original query results and comparing these values ​​with preset row count thresholds and data volume thresholds. Specific compression steps are detailed below.

[0042] The method disclosed herein can load multiple low-rank adaptive algorithms at once, and dynamically route or concatenate different low-rank adaptive weights according to different problems, so that a model has multiple text semantic understanding capabilities. At the same time, loading the SQL generation part of the low-rank adaptive algorithm can enhance the original model's capabilities in professional vertical domains and improve accuracy.

[0043] In one possible implementation, the low-rank adaptive weights are stored in a low-rank adaptive weight storage. The low-rank adaptive weight storage for obtaining low-rank adaptive modules includes a plurality of low-rank adaptive labels and low-rank adaptive modules adapted to each low-rank adaptive label; the low-rank adaptive module includes low-rank adaptive weights; wherein, the low-rank adaptive weights are used to indicate the specialized processing capability of the large model in the domain specified by the low-rank adaptive label adapted to the low-rank adaptive module.

[0044] In one possible implementation, Figure 3 In step 303, at least one target low-rank adaptive label is configured for the user question, including: First, the user's question is semantically deconstructed and categorized to resolve it into at least one minimum business unit. Specifically, this step can use Intent Recognition Technology to semantically deconstruct the input question and use a Multi-level Question Classifier to categorize it, resolving the original question request to the minimum business unit.

[0045] Then, a target low-rank adaptive tag is configured for each minimum business unit. This step yields multiple low-rank adaptive tags carried by the user's question. This step can employ text matching to assign low-rank adaptive tags. Specifically, the keywords of the minimum business unit are matched against different low-rank adaptive tags, and the low-rank adaptive tag with the highest matching degree is selected as the assigned low-rank adaptive tag for that minimum business unit.

[0046] The intent recognition technology described above first preprocesses the user's input text or speech through word segmentation, noise reduction, and vectorization, transforming unstructured language into feature vectors that the model can recognize. Then, through text classification, sequence labeling, or deep learning models, semantic features and key information are extracted and matched with pre-defined intent labels. The system combines contextual information and slot filling techniques to pinpoint the user's core needs and constraints, eliminate ambiguous expressions in the input text, and ultimately determine the user's true intent and output the corresponding classification result.

[0047] The multi-level question classifier described above employs a hierarchical architecture to classify user questions progressively. First, the first-layer model performs coarse-grained segmentation of the question, determining its domain or broad category. Then, the second-layer model performs fine-grained identification, further distinguishing specific intents and subcategories. The question text is preprocessed with word segmentation and vectorization to extract semantic features, which are then sequentially processed through multiple classification models. Each layer focuses on the current level's label, gradually narrowing the matching range while simultaneously eliminating ambiguity by incorporating context and key information. This structure reduces the pressure of single-step classification and improves classification accuracy and efficiency in complex scenarios.

[0048] In one possible implementation, generating customized low-rank adaptive weights for a user problem based on multiple adapted low-rank adaptive modules includes, in response to more than one target low-rank adaptive label configured for the user problem, weighting and summing the multiple target low-rank adaptive labels to obtain the customized low-rank adaptive weights corresponding to the user problem.

[0049] In one possible implementation, generating a text summary model based on customized low-rank adaptive weights includes: integrating the customized low-rank adaptive weights into a base model; keeping the original pre-trained weights frozen in the base model; and, based on the integrated customized low-rank adaptive weights, performing feature enhancement and forward propagation calculations to generate the text summary model.

[0050] Specifically, in the technical solution disclosed herein, low-rank decomposition of model weights can be used to perform a weighted summation of multiple target low-rank adaptive labels. Specifically, low-rank decomposition of model weights refers to the forward propagation process of the entire neural network layer with low-rank adaptive plugin in the model. The pre-trained weight matrix is ​​defined as... 0∈ This represents the weights of a pre-trained fully connected or attention layer in a large language model, where R represents the real number field, the input dimension is d, and the output dimension is k. Two decomposition matrices, A and B, are added in the low-rank adaptive module to learn the weight increments. W; where the condition matrix is ​​satisfied. ∈ This indicates that matrix A has an input dimension of d and an output dimension of r. ∈ Let represent matrix A, where the input dimension is r and the output dimension is k. And satisfy . ( , The forward propagation process of the entire neural network layer with low-rank adaptive plugin in the large model can be expressed as Equation 1: (1) During the training process, the original weights 0 is frozen and does not participate in training updates; only the weights of A and B are updated. The low-rank adaptive multi-weight matrix merging is performed based on formulas (2) and (3): (2) (3) in , It is the low-rank matrix of the i-th low-rank adaptive plugin, ˆ ,ˆ It is the low-rank matrix of the new low-rank adaptive plugin. These are the weight coefficients for each low-rank adaptive plugin, calculated based on the matching degree between the business unit and the low-rank adaptive domain. The higher the matching degree, the larger the coefficient. Alternatively, they can be set according to the importance of each dimension in the user's problem, such as a primary intention coefficient of 0.7 and a secondary intention coefficient of 0.3. The final sum is guaranteed to be 1 to avoid weight amplification that could lead to model instability. After fusion, only one set of low-rank adaptive weights is obtained, which is then linearly combined with the base model, rather than loading multiple sets of low-rank adaptive weights simultaneously.

[0051] Specifically, This is used to quantify the fit between the i-th low-rank adaptive plugin and the current user problem. The higher the fit, the better the corresponding... The larger the value, the larger the low-rank matrix of this low-rank adaptive plugin. , In the merged new matrix 、ˆ The higher the weight of the component, the better; conversely, the lower the fit, the better. The smaller the value, the lower its influence weight. This weighted summation method allows the fusion of multiple low-rank adaptive modules tailored to different semantic dimensions of the user's problem, generating a unique, customized low-rank matrix. 、ˆ This leads to customized low-rank adaptive weights. According to formula (1), the merged new low-rank adaptive plugin still follows the low-rank decomposition principle, and its weight increment... By substituting the forward propagation formula, the large model can accurately meet the needs of the current user problem, while maintaining the advantages of a concise model parameter set and efficient inference, and avoiding redundancy and conflicts caused by parallel operation of multiple low-rank adaptive modules.

[0052] specifically refer to Figure 4 In one possible implementation, the method may include a compression step for the original query results. To address the compression problem of the original query results, the method disclosed herein can integrate the capabilities of the NL2SQL model with the reasoning mechanism of a large language model to construct a data compression strategy based on semantic understanding. By sending a query request to the SQL large model and employing the thought chain (COT) guidance technology, the model is driven to generate a concise SELECT statement, thereby filtering out the necessary fields in the data. The core implementation steps are: first, adding a small number of example hint models; then, based on the semantic similarity between keywords and column descriptions, retrieving the top k columns for each keyword; then, enabling the value retriever to adopt a two-stage retrieval strategy based on Locality Sensitive Hash (LSH) to identify similar values ​​in the database; finally, based on the SELECT results in the generated SQL statement, the original data is filtered by field, retaining only the key data fields, achieving efficient data compression, effectively reducing the data transmission and model processing burden, and improving the processing efficiency and accuracy of text summarization tasks. Specifically, the compression step includes: Step 401: In response to the original query result data volume exceeding a predetermined threshold, the original query result is compressed using schema-linking to obtain a compressed query result. Step 401 addresses the technical problem in large-scale database scenarios where traditional database description schema files may become too large, exceeding the context length limit of the large language model, or cause information redundancy that prevents the model from accurately extracting effective information. The method disclosed in this paper employs schema-link data compression technology. This technology leverages the large language model's processing capabilities for Structured Query Language (SQL), achieving lightweight database description and data processing through a closed-loop process of generating, validating, executing SQL statements, and iteratively optimizing.

[0053] Specifically, the schema link compression in step 401 may include extracting data tables and key fields related to the current problem, and eliminating irrelevant tables and fields; simplifying table relationships into key foreign key links, omitting redundant constraints and comments; using concise identifiers to represent field types and meanings, reducing the length of natural language descriptions; and finally forming a lightweight schema structure, which is then concatenated with the query results into the input model.

[0054] It is worth noting that step 401 can also be implemented when the data table field length is relatively small. The data definition language structure of the business-related tables can be directly input into the large model as the basis for table and field selection. Alternatively, if the user has sufficient technical capabilities in a big data department, the original table fields can be abstracted to create a new table, aiming to rebuild tables or views to meet the needs of the query data. The implementation method of step 401 in this disclosure is not limited to the above technical solutions. In fact, those skilled in the art can use any database lightweighting and field simplification techniques to implement this step, as long as it satisfies the requirements of reducing the database description size, retaining core business fields, and adapting to the context of large model processing.

[0055] The technical solution disclosed herein applies the principle of Locality Sensitive Hash (LSH) so that similar data are more likely to obtain the same hash value after hashing, while dissimilar data are more likely to have different hash values, thereby quickly finding samples similar to the query data in massive high-dimensional data.

[0056] Step 402 involves data chunking and keyword compression of the compressed query results to obtain the processed query results. This step addresses two technical issues in existing text summarization methods: firstly, the degradation of detail parsing ability in large language models when processing excessive data; and secondly, the negative correlation between model input length and comprehension ability.

[0057] Therefore, addressing the technical problems existing in the current field of text summarization, this disclosure proposes a data block segmentation strategy based on semantic dimensions in step 402. By addressing the data change description requirements in text summarization tasks, the structured data is reorganized using the GROUPBY operation. While retaining dynamic fields such as time and price, the remaining fields are fixedly grouped. This method ensures that all fields except time and price remain constant within the data block, forming structured semantic units.

[0058] Specifically, the keyword compression in step 402 includes a first compression and a second compression.

[0059] The first compression process simplifies the representation of repetitive or similar data values ​​by replacing duplicate content with the smallest bit placeholder, retaining only the original data that appears for the first time, thus reducing data character length and redundant information. This reduces data volume and character redundancy, making tables more concise and reducing the burden on the model for reading. It does not change the meaning of fields, add new fields, or alter the semantics of the data.

[0060] The second compression process involves performing calculations on the table, deriving new fields, replacing the original fields with more condensed statistical / result fields (such as price difference, trend, sum, and mean), and deleting the original fields. This extracts semantic information, transforming the original detailed data into result-oriented data more suitable for summarization, and helping the model complete inference in advance.

[0061] Specifically, the keyword compression in step 402 can be achieved through the following method.

[0062] In one specific embodiment, by retaining the metadata of the first row of the data block and the key fields of time and price, a placeholder replacement operation is performed on the remaining non-dynamic fixed fields: a preset specific character is used as a unified identifier to replace the original data content. Compared with the NaN data type conversion problem caused by emptying the data, this embodiment avoids model processing errors caused by abnormal data format, effectively reduces data redundancy, and improves data transmission efficiency while ensuring data integrity.

[0063] It is worth noting that the aforementioned specific characters are not limited to using a single character as a uniform identifier to replace the original data content. In fact, those skilled in the art will understand that the restrictions on specific characters can be set by those skilled in the art based on specific circumstances.

[0064] In one possible implementation, keyword compression includes retaining the first line of metadata and key fields of the data block after data slicing, and compressing the data block by replacing the remaining non-dynamically fixed fields of the data block with placeholders; wherein the replaced characters occupy no more than one token.

[0065] In another embodiment, keyword compression includes setting statistical indicators and embedding these indicators into the data blocks after data segmentation to compress the data blocks. Specifically, keyword compression is achieved by pre-embedding statistical indicators such as maximum, minimum, average, and weighted values ​​into the data blocks. Compared to real-time calculation of aggregated values ​​by the model, this significantly reduces the computational load during the model's inference phase. This strategy, by prioritizing complex computational tasks, allows the model to focus on the core text summarization task, greatly improving the model's accuracy in processing numerical data and its semantic understanding ability, ultimately achieving an effective improvement in the accuracy of the text summarization task.

[0066] And in step 403, the user's question and the processed query results are input into the text summary model to obtain a text summary.

[0067] In the above steps, through the progressive compression process, the database structure and original query data are first streamlined and simplified based on pattern linking and semantic retrieval. Then, data redundancy and processing complexity are further reduced through matrix transformation, duplicate value placeholders, pre-computation aggregation, and keyword compression. While fully preserving core business information, time series, and key numerical features, the model input size is significantly reduced, alleviating the limitations of large model context length and the problem of detail understanding decay. At the same time, complex computational logic is pre-positioned to reduce the burden of real-time model inference, ultimately effectively improving the response speed of text summarization, the accuracy of information extraction, and the overall task execution efficiency.

[0068] It should be noted that the method of this disclosure embodiment can be executed by a single device, such as a computer or server. The method of this embodiment can also be applied to a distributed scenario, where multiple devices cooperate to complete the task. In such a distributed scenario, one of these devices may execute only one or more steps of the method of this disclosure embodiment, and the multiple devices will interact with each other to complete the method described.

[0069] It should be noted that the above description describes some embodiments of this disclosure. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in a different order than that shown in the above embodiments and still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0070] Exemplary device Based on the same inventive concept, corresponding to the methods of any of the above embodiments, this disclosure 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 text summarization method described in any of the above embodiments.

[0071] Figure 5 This embodiment illustrates a more specific hardware structure of an electronic device, which may include a processor 1010, a memory 1020, an input / output interface 1030, a communication interface 1040, and a bus 1050. The processor 1010, memory 1020, input / output interface 1030, and communication interface 1040 are interconnected internally via the bus 1050.

[0072] The processor 1010 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this specification.

[0073] The memory 1020 can be implemented in the form of ROM (Read-Only Memory), RAM (Random Access Memory), static storage device, dynamic storage device, etc. The memory 1020 can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented by software or firmware, the relevant program code is stored in the memory 1020 and is called and executed by the processor 1010.

[0074] The input / output interface 1030 is used to connect input / output modules to realize information input and output. The input / output modules can be configured as components in the device (not shown in the figure) or externally connected to the device to provide corresponding functions. Input devices may include keyboards, mice, touch screens, microphones, various sensors, etc., and output devices may include displays, speakers, vibrators, indicator lights, etc.

[0075] The communication interface 1040 is used to connect a communication module (not shown in the figure) to enable communication between this device and other devices. The communication module can communicate via wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0076] Bus 1050 includes a pathway for transmitting information between various components of the device, such as processor 1010, memory 1020, input / output interface 1030, and communication interface 1040.

[0077] It should be noted that although the above-described device only shows the processor 1010, memory 1020, input / output interface 1030, communication interface 1040, and bus 1050, in specific implementations, the device may also include other components necessary for normal operation. Furthermore, those skilled in the art will understand that the above-described device may only include the components necessary for implementing the embodiments of this specification, and not necessarily all the components shown in the figures.

[0078] The electronic devices described above are used to implement the corresponding text summarization methods in any of the foregoing embodiments and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0079] The memory 1620 stores machine-readable instructions executable by the processor 1610. When the electronic device is running, the processor 1610 communicates with the memory 1620 via the bus 1630, causing the processor 1610 to execute the following instructions during operation: Receive user questions and data sources; Obtain the raw query results related to the user's question from the data source; Configure at least one target low-rank adaptive label for the user question; Based on the target low-rank adaptive label, obtain the customized low-rank adaptive weights corresponding to the user question; A text summary model is generated based on customized low-rank adaptive weights; The user's question and the original query results are input into the text summary model to obtain a text summary.

[0080] By using the above method, a large model that meets the needs of multiple domains can be generated through the dynamic combination of multiple low-rank adaptive modules. This can satisfy the needs of various types of text summarization and provide customized optimal model support for text summarization in various task scenarios. On the other hand, by performing multi-dimensional segmentation and intelligent compression on the data generated by Natural Language to Structured Query Language (NL2SQL), and inputting the processed data into the updated large language model that uses dynamically combined low-rank adaptive modules, the limitations of input length and memory capacity of existing large language models are overcome, enabling efficient and accurate processing of text summarization tasks.

[0081] Exemplary program product Based on the same inventive concept, corresponding to the methods of any of the above embodiments, this disclosure also provides a non-transitory computer-readable storage medium storing computer instructions for causing the computer to execute the text summarization method as described in any of the above embodiments.

[0082] The computer-readable medium of this embodiment includes permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device.

[0083] The aforementioned non-transitory computer-readable storage media can be any available medium or data storage device that a computer can access, including but not limited to magnetic storage (e.g., floppy disks, hard disks, magnetic tapes, magneto-optical disks (MOs), etc.), optical storage (e.g., CDs, DVDs, BDs, HVDs, etc.), and semiconductor storage (e.g., ROMs, EPROMs, EEPROMs, non-volatile memory (NAND flash), solid-state drives (SSDs)).

[0084] The computer instructions stored in the storage medium of the above embodiments are used to cause the computer to execute the text summarization method as described in any of the embodiments in the exemplary method section above, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0085] Based on the same inventive concept, corresponding to the text summarization method described in any of the above embodiments, this disclosure also provides a computer program product, which includes computer program instructions. In some embodiments, the computer program instructions can be executed by one or more processors of a computer to cause the computer and / or the processor to perform the text summarization method. Corresponding to the execution entity for each step in each embodiment of the text summarization method, the processor executing the corresponding step can belong to the corresponding execution entity.

[0086] The computer program products of the above embodiments are used to cause the computer and / or the processor to execute the text summarization method as described in any of the above embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0087] Those skilled in the art will recognize that embodiments of this disclosure can be implemented as a system, method, or computer program product. Therefore, this disclosure can be implemented as entirely hardware, entirely software (including firmware, resident software, microcode, etc.), or a combination of hardware and software, generally referred to herein as a "circuit," "module," or "system." Furthermore, in some embodiments, this disclosure can also be implemented as a computer program product contained in one or more computer-readable media, which includes computer-readable program code.

[0088] Any combination of one or more computer-readable media may be used. A computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. A computer-readable storage medium can be, for example,, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples (not exhaustive) of a computer-readable storage medium may include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this document, a computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in connection with an instruction execution system, apparatus, or device.

[0089] Computer-readable signal media may include data signals propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media may also be any computer-readable medium other than computer-readable storage media, capable of sending, propagating, or transmitting programs for use by or in connection with an instruction execution system, apparatus, or device.

[0090] Program code contained on a computer-readable medium may be transmitted using any suitable medium, including but not limited to wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0091] Computer program code for performing the operations of this disclosure can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0092] It should be understood that each block of a flowchart and / or block diagram, as well as combinations of blocks in a flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device to produce a machine that, when executed by a computer or other programmable data processing device, creates means for implementing the functions / operations specified in the blocks of the flowchart and / or block diagram.

[0093] These computer program instructions may also be stored in a computer-readable medium that enables a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable medium produce a product comprising an instruction apparatus that implements the functions / operations specified in the boxes of a flowchart and / or block diagram.

[0094] Computer program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operational steps to be performed on the computer, other programmable data processing apparatus, or other device to produce a computer-implemented process, such that the instructions that execute on the computer or other programmable apparatus can provide a process for implementing the functions / operations specified in the boxes of a flowchart and / or block diagram.

[0095] Furthermore, although the operations of the methods of this disclosure are described in a specific order in the accompanying drawings, this does not require or imply that these operations must be performed in that specific order, or that all of the operations shown must be performed to achieve the desired result. Rather, the steps depicted in the flowcharts may be executed in a different order. Additionally or alternatively, certain steps may be omitted, multiple steps may be combined into one step, and / or one step may be broken down into multiple steps.

[0096] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. Each block in a flowchart or block diagram may represent a module, segment, or portion of code, which contains one or more executable instructions for implementing a specified logical function. 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. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0097] It should be noted that although several modules or units for the device used to perform actions have been mentioned in the detailed description above, this division is not mandatory. In fact, according to the embodiments of this application, the features and functions of two or more modules or units described above can be embodied in one module or unit. Conversely, the features and functions of one module or unit described above can be further divided and embodied by multiple modules or units.

[0098] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of this application (including the claims) is limited to these examples; within the framework of this application, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of the embodiments of this application as described above, which are not provided in the details for the sake of brevity.

[0099] Additionally, to simplify the description and discussion, and to avoid obscuring the embodiments of this application, the well-known power / ground connections to integrated circuit (IC) chips and other components may or may not be shown in the provided drawings. Furthermore, the apparatus may be shown in block diagram form to avoid obscuring the embodiments of this application, and this also takes into account the fact that the details of the implementation of these block diagram apparatuses are highly dependent on the platform on which the embodiments of this application will be implemented (i.e., these details should be fully understood by those skilled in the art). While specific details (e.g., circuits) have been set forth to describe exemplary embodiments of this application, it will be apparent to those skilled in the art that the embodiments of this application can be implemented without these specific details or with variations thereof. Therefore, these descriptions should be considered illustrative rather than restrictive.

[0100] Although this application has been described in conjunction with specific embodiments thereof, many substitutions, modifications, and variations of these embodiments will be apparent to those skilled in the art from the foregoing description. For example, other memory architectures (e.g., dynamic RAM (DRAM)) may be used with the embodiments discussed.

[0101] The embodiments of this application are intended to cover all such substitutions, modifications, and variations that fall within the broad scope of the appended claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the embodiments of this application should be included within the protection scope of this application.

[0102] While the spirit and principles of this disclosure have been described with reference to several specific embodiments, it should be understood that this disclosure is not limited to the disclosed specific embodiments, and the division of aspects does not imply that features in these aspects cannot be combined for benefit; such division is merely for convenience of expression. This disclosure is intended to cover various modifications and equivalent arrangements included within the spirit and scope of the appended claims. The scope of the appended claims is to be interpreted in the broadest sense, thereby encompassing all such modifications and equivalent structures and functions.

Claims

1. A text summarization method, comprising: Receive user questions and data sources; Obtain the original query results related to the user's question from the data source; Configure at least one target low-rank adaptive label for the user question; Based on the target low-rank adaptive label, the customized low-rank adaptive weight corresponding to the user question is obtained; A text summary model is generated based on the customized low-rank adaptive weights. The user question and the original query results are input into the text summary model to obtain the text summary.

2. The method according to claim 1, wherein, The low-rank adaptive weights are stored in the low-rank adaptive weight storage; The low-rank adaptive weight storage includes several low-rank adaptive labels and low-rank adaptive modules adapted to each low-rank adaptive label; the low-rank adaptive module includes low-rank adaptive weights; wherein, the low-rank adaptive weights are used to indicate the specialized processing capability of the large model in the domain specified by the low-rank adaptive label adapted to the low-rank adaptive module.

3. The method according to claim 1, wherein, Configuring at least one target low-rank adaptive label for the user question includes: The user questions are semantically deconstructed and categorized to resolve them into at least one minimum business unit. Configure a target low-rank adaptive label for each of the smallest business units.

4. The method according to claim 2, wherein, Generating customized low-rank adaptive weights for the user problem based on the multiple adapted low-rank adaptive modules includes: In response to the fact that there is more than one target low-rank adaptive label configured for the user question, the multiple target low-rank adaptive labels are weighted and summed to obtain the customized low-rank adaptive weight corresponding to the user question.

5. The method according to any one of claims 1-4, wherein, The method also includes: In response to the original query result data volume exceeding the predetermined threshold, the original query result is subjected to pattern link compression to obtain compressed query result; The compressed query results are then divided into data chunks and compressed using keywords to obtain processed query results; and The user question and the processed query results are input into the text summary model to obtain the text summary.

6. The method according to claim 5, wherein, The keyword compression mentioned above includes: The first row of metadata and key fields of each data block after data segmentation are retained; these key fields are defined based on each text summarization task; and The data block is modified by removing the first row of metadata and replacing non-dynamic fixed fields of key fields with placeholders; wherein the replaced characters occupy no more than one word.

7. The method according to claim 5, wherein, The keyword compression includes: Set statistical indicators; and The statistical indicators are embedded in the data blocks after data segmentation, and the data blocks are compressed.

8. The method according to claim 1, wherein, The customized low-rank adaptive weight text summarization model includes: The customized low-rank adaptive weights are integrated into the base large model; the base large model keeps the original pre-trained weights frozen; and Based on the customized low-rank adaptive weights obtained from the access, the basic large model completes feature enhancement and forward propagation calculations to generate the text summary model.

9. A text summarization device, comprising: The information receiving module is configured to receive user questions and data sources; The data source query module is configured to obtain raw query results related to the user's question from the data source. The tag acquisition module obtains multiple low-rank adaptive tags carried by the user question; The large model acquisition module is configured to obtain customized low-rank adaptive weights corresponding to the user question based on the target low-rank adaptive label. A text summary model is generated based on the customized low-rank adaptive weights. The text summary module is configured to input the user question and the original query results into the text summary model to obtain the text summary.

10. An electronic device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the method as described in any one of claims 1 to 8.

11. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium stores computer instructions for causing the computer to perform the method of any one of claims 1 to 8.

12. A computer program product, characterized in that, It includes computer program instructions that, when executed on a computer, cause the computer to perform the method as described in any one of claims 1 to 8.

Images