Text generation method, document question answering method, electronic device and storage medium

Abstract

The present disclosure relates to the fields of large model technologies, text generation and intelligent question answering. Disclosed are a text generation method, a document question answering method, an electronic device and a storage medium. The method comprises: in response to an input instruction acting on an operation interface, acquiring original text and a target data set that correspond to the input instruction; and in response to a generation instruction acting on the operation interface, displaying target text on the operation interface, wherein the target text is generated on the basis of the original text and initial structured data, and the initial structured data is constructed on the basis of the original text and the target data set and conforms to a target structured data type corresponding to the original text. The present disclosure solves the technical problem in the related art of the accuracy of text generation being relatively low.

Description

Text generation method, document question answering method, electronic device, and storage medium TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of large model technology and text generation, in particular to a text generation method, a document question answering method, an electronic device, and a storage medium. BACKGROUND

[0002] In a document question answering task, as the model complexity increases, the traditional retrieval enhancement generation method faces challenges, especially when dealing with problems with scattered information and complex reasoning, its performance is limited. In the prior art, although attempts have been made to enhance retrieval through structured knowledge (such as graph structure), it is limited to a single knowledge type and the application scenario is limited. In addition, the process of directly extracting structured knowledge from unstructured documents will introduce a large amount of irrelevant noise, affecting the accuracy of the generated answer.

[0003] At present, no effective solution has been proposed to solve the above problems. SUMMARY

[0004] The embodiments of the present disclosure provide a text generation method, a document question answering method, an electronic device, and a storage medium to at least solve the technical problem of low accuracy of text generation in related technologies.

[0005] According to an aspect of an embodiment of the present disclosure, a text generation method is provided, comprising: in response to an input instruction acting on an operation interface, obtaining original text corresponding to the input instruction and identification information of a target data set; in response to a generation instruction acting on the operation interface, displaying target text on the operation interface, wherein the target text is generated based on the original text and initial structured data, the initial structured data is constructed based on the original text and the target data set, and conforms to a target structured data type corresponding to the original text.

[0006] According to another aspect of an embodiment of the present disclosure, a document question answering method is also provided, comprising: in response to an input instruction acting on an operation interface, obtaining inquiry information corresponding to the input instruction and identification information of a target document; in response to a generation instruction acting on the operation interface, displaying reply information on the operation interface, wherein the reply information is generated based on the inquiry information and initial structured data, the initial structured data is constructed based on the inquiry information and the target document, and conforms to a target structured data type corresponding to the inquiry information.

[0007] According to another aspect of an embodiment of the present disclosure, a text generation method is also provided, comprising: obtaining original text and a target data set; constructing initial structured data based on the original text and the target data set, wherein the initial structured data conforms to a target structured data type corresponding to the original text; generating target text based on the original text and the initial structured data.

[0008] According to another aspect of the embodiments of the present disclosure, a text generation method is also provided, including: obtaining an original text and a target data set by calling a first interface, wherein the first interface includes a first parameter, and a parameter value of the first parameter includes the original text and the target data set; constructing initial structured data based on the original text and the target data set, wherein the initial structured data conforms to a target structured data type corresponding to the original text; generating a target text based on the original text and the initial structured data; and outputting the target text by calling a second interface, wherein the second interface includes a second parameter, and a parameter value of the second parameter includes the target text.

[0009] According to another aspect of the embodiments of the present disclosure, a document question answering system is also provided, including: a client configured to obtain inquiry information corresponding to an input instruction and identification information of a target document; a server connected with the client and configured to construct initial structured data based on an original text and a target data set, and generate a target text based on the original text and the initial structured data, wherein the initial structured data conforms to a target structured data type corresponding to the original text; and the client is further configured to display the target text.

[0010] According to another aspect of the embodiments of the present disclosure, a computer terminal is also provided, including: a memory storing an executable program; and a processor configured to run the program, wherein the program performs the method in the various embodiments of the present disclosure when running.

[0011] According to another aspect of the embodiments of the present disclosure, a computer readable storage medium is also provided, including a stored executable program, wherein the computer readable storage medium controls a device where the computer readable storage medium is located to perform the method in the various embodiments of the present disclosure when the executable program runs.

[0012] According to another aspect of the embodiments of the present disclosure, a computer program product is also provided, including a computer program, which, when executed by a processor, implements the method in the various embodiments of the present disclosure.

[0013] According to another aspect of the embodiments of the present disclosure, a computer program product is also provided, including a non-volatile computer readable storage medium storing a computer program, which, when executed by a processor, implements the method in the various embodiments of the present disclosure.

[0014] According to another aspect of the embodiments of the present disclosure, a computer program is also provided, which, when executed by a processor, implements the method in the various embodiments of the present disclosure.

[0015] In the embodiments of the present disclosure, in response to an input instruction acting on the operation interface, the original text corresponding to the input instruction and the identification information of the target data set are obtained; in response to a generation instruction acting on the operation interface, the target text is displayed on the operation interface, wherein the target text is generated based on the original text and the initial structured data, the initial structured data is constructed based on the original text and the target data set, and conforms to the target structured data type corresponding to the original text, thereby achieving the purpose of improving the accuracy of text generation. It is easy to note that by using the target structured data type to construct the initial structured data, the system can more accurately understand and extract the key information in the original text, so as to provide more relevant and specific information as a reference when generating the target text, thereby ensuring that the model can reason and create based on accurate structured data when generating the target text, reducing the interference of irrelevant information, improving the accuracy and pertinence of the generated text, and thereby solving the technical problem of low accuracy of text generation in related technologies.

[0016] It should be noted that the above general description and the following detailed description are only for the purpose of illustrating and explaining the present disclosure, and do not constitute a limitation on the present disclosure. BRIEF DESCRIPTION OF DRAWINGS

[0017] The drawings described herein are used to provide further understanding of the present disclosure, and form a part of the present disclosure. The illustrative embodiments of the present disclosure and their descriptions serve to explain the present disclosure, and do not constitute an improper limitation on the present disclosure. In the drawings:

[0018] FIG. 1 is a schematic diagram of an application scenario of a text generation method according to an embodiment of the present disclosure;

[0019] FIG. 2 is a flowchart of a text generation method according to an embodiment of the present disclosure;

[0020] FIG. 3 is a schematic diagram of a retrieval enhancement framework according to an embodiment of the present disclosure;

[0021] FIG. 4 is a flowchart of a document question and answer method according to an embodiment of the present disclosure;

[0022] FIG. 5 is a flowchart of a text generation method according to an embodiment of the present disclosure;

[0023] FIG. 6 is a flowchart of a text generation method according to an embodiment of the present disclosure;

[0024] FIG. 7 is a schematic diagram of a text generation apparatus according to an embodiment of the present disclosure;

[0025] FIG. 8 is a schematic diagram of a document question and answer apparatus according to an embodiment of the present disclosure;

[0026] FIG. 9 is a schematic diagram of a text generation apparatus according to an embodiment of the present disclosure;

[0027] FIG. 10 is a schematic diagram of a text generation apparatus according to an embodiment of the present disclosure;

[0028] FIG. 11 is a structural block diagram of an electronic device according to an embodiment of the present disclosure;

[0029] FIG. 12 is a schematic diagram of a document question-answering system according to an embodiment of the present disclosure. DETAILED DESCRIPTION

[0030] In order to enable persons skilled in the art to better understand the present disclosure scheme, the technical solutions in the embodiments of the present disclosure will be clearly and completely described below in conjunction with the drawings in the embodiments of the present disclosure. Obviously, the described embodiments are only a part of the embodiments of the present disclosure, not all the embodiments. Based on the embodiments in the present disclosure, all other embodiments obtained by persons skilled in the art without creative labor should be within the scope of protection of the present disclosure.

[0031] It should be noted that the terms "first", "second", and the like in the specification and claims of the present disclosure and the above-described drawings are used to distinguish similar objects, and do not necessarily have to describe a specific order or sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments of the present disclosure described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product, or device that includes a series of steps or units does not have to be limited to only those steps or units clearly listed, but can include other steps or units that are not clearly listed or inherent to these processes, methods, products, or devices.

[0032] The technical solutions provided by the present disclosure are mainly implemented by using large model technology. Here, the large model refers to a deep learning model with a large number of model parameters, which can typically include hundreds of millions, tens of billions, hundreds of billions, thousands of billions, or even tens of trillions of model parameters. The large model can also be referred to as a foundation model. Through large-scale unlabeled corpus pre-training, a pre-trained model with hundreds of millions of parameters is output. Such a model can adapt to a wide range of downstream tasks and has good generalization ability. For example, large language models (LLM) and multi-modal pre-training models.

[0033] It should be noted that in actual application, the large model can be fine-tuned by a small amount of samples on the pre-trained model, so that the large model can be applied to different tasks. For example, the large model can be widely applied to natural language processing (NLP), computer vision, speech processing and the like, and can be applied to computer vision field tasks such as visual question answering (VQA), image caption (IC), image generation, and the like, and can be widely applied to natural language processing field tasks such as text-based sentiment classification, text summary generation, machine translation and the like. Therefore, the main application scenarios of the large model include but are not limited to digital assistants, intelligent robots, search, online education, office software, e-commerce, intelligent design and the like. In the embodiments of the present disclosure, the data processing by the large model in the text processing scenario is taken as an example for explanation and description.

[0034] First, some of the nouns or terms that appear in the description of the embodiments of the present disclosure are applicable to the following explanations:

[0035] Large language model (LLM) is a large-scale language model based on massive data and autoregressive method after a large amount of training.

[0036] Retrieval-augmentation Generation (RAG) is a commonly used technique that uses external knowledge bases to provide reliable information for large models and alleviate the hallucination problem of large models.

[0037] For document question answering tasks, with the continuous enhancement of the model capability in the field of natural language processing, the complexity of the task gradually increases. Early document question answering tasks are mainly single-hop question answering, which only needs to locate a certain position in the given document to directly obtain the answer. Then there appeared a slightly more complex multi-hop question answering task, which required the model to locate 2-4 relevant information in the given document according to the question, and to obtain the answer by reasoning based on the information. In recent period, the natural language community proposed a more complex document question answering task, which is more difficult than the previous multi-hop question answering task, mainly in two aspects, first, the information required by the complex document question answering task is more scattered, second, the model needs to make complex reasoning based on this information, for example, analyzing the development trend of a company based on financial data.

[0038] To solve the above complex document question answering task, the disclosure proposes to construct and use multiple types of structured knowledge as retrieval enhancement knowledge, to select a more matched structured data type for the current question, to use the structured knowledge of this type to improve the effective information content in the retrieval result and assist the generation module in reasoning, and thus to improve the performance on complex document tasks.

[0039] According to an embodiment of the disclosure, a text generation method is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described herein can be executed in an order different from that shown.

[0040] Considering that the model parameter amount of the large model is huge, and the operation resources of the mobile terminal are limited, the above-mentioned text generation method provided by the embodiment of the disclosure can be applied to the application scenario as shown in FIG. 1, but is not limited thereto. FIG. 1 is a schematic diagram of an application scenario of a text generation method according to an embodiment of the disclosure. In the application scenario as shown in FIG. 1, the large model is deployed in a server 10. The server 10 can be connected to one or more client devices 20 through a local area network connection, a wide area network connection, an Internet connection, or other types of data networks. The client device 20 herein can include, but is not limited to, a smartphone, a tablet computer, a notebook computer, a palm computer, a personal computer, a smart home device, a vehicle-mounted device, etc. The client device 20 can interact with the user through a graphical user interface to realize the calling of the large model, and thus realize the method provided by the embodiment of the disclosure.

[0041] In the embodiment of the disclosure, the system composed of the client device and the server can execute the following steps: the client device executes the input of the original text and the target data set on the operation interface. The server executes the generation of the target text according to the original text and the initialization structure data in the target data set, and constructs the initialization structure data conforming to the target structured data type corresponding to the original text based on the original text and the target data set. It should be noted that in the case that the running resources of the client device can meet the deployment and running conditions of the large model, the embodiment of the disclosure can be performed in the client device.

[0042] Under the above running environment, the disclosure provides a text generation method as shown in FIG. 2. FIG. 2 is a flowchart of a text generation method according to an embodiment of the disclosure. As shown in FIG. 2, the method can include the following steps:

[0043] Step S202, in response to the input instruction acting on the operation interface, obtaining the original text and the target data set corresponding to the input instruction;

[0044] The operation interface can be an interface for user interaction with the computer system, and can be a graphical user interface, a command line interface, or a web page interface, etc. In the present disclosure, the operation interface can refer to an interface for user input of a question and viewing of an answer.

[0045] The input instruction can be an instruction provided by the user on the operation interface, and can be text, voice, touch instruction, etc. The input instruction can be a question raised by the user.

[0046] The original text is unprocessed text input by the user, and can be a query, an instruction, a natural language sentence, etc. The original text can be a question input by the user on the operation interface. The original text refers to a document or text material to be processed, and the original text can be the basis for constructing structured knowledge and generating target text.

[0047] The target data set can refer to a database or information collection in the computer system for processing the original text or assisting in generating a response. The target data set can be a database containing original document information, which can be financial reports, abstracts, papers, etc. The target data set is a collection of multiple types of structured data, such as tables, graphs, and short text chunks, etc., which are defined in advance to guide how to extract and construct information from the original text.

[0048] In an optional embodiment, the system responds to the input instruction (i.e., the question) of the user on the operation interface, and displays the original question of the user and the original document information to be extracted from the knowledge on the operation interface. This step is the starting point of the entire question and answer process, and the user raises a question to the system through the operation interface, and the system takes the question and the target data set (the original document) as initial inputs, and prepares for subsequent knowledge selection, information extraction, and answer generation processes.

[0049] For example, when the user inputs a question, such as “How is the profit growth of Company A?” on the operation interface, the system will immediately respond to the input, and display the original question and the related original document (assuming a document containing financial information of Company A) on the interface. By displaying the original question and the related original document, it can be ensured that the user can confirm that the question and the data set being processed by the system are as expected. By displaying the original text and the target data set, a clear processing direction can be provided for subsequent complex processing steps.

[0050] It should be noted that after responding to the input instruction acting on the operation interface, the original text corresponding to the input instruction and the target data set can be obtained, and the title information of the target data set can be displayed on the operation interface.

[0051] Step S204, in response to the generation instruction acting on the operation interface, displaying the target text on the operation interface.

[0052] The target text is generated based on the original text and initial structured data, the initial structured data is constructed based on the original text and the target data set, and conforms to the target structured data type corresponding to the original text.

[0053] The above-mentioned generation instruction can be issued by a user or internally by a system, which is an instruction for triggering content creation, modification or display. The user can input a question in the operation interface and click the "submit" button. The system receives the generation instruction.

[0054] The above-mentioned target text can be a text output generated by the system based on specific input and processing logic. The target text is the final generated text prepared to answer a given question or meet a specific requirement. The target text can combine the content of the original text and the structured knowledge type selected from the target data set that is suitable for the current question.

[0055] The above-mentioned target data set can be a data collection for generating or retrieving information, which can be a structured database or an unstructured document collection.

[0056] The above-mentioned initial structured data can be information extracted from the original text and organized according to a specific structure.

[0057] The above-mentioned construction can be represented as a process of generating a new, structured representation from input data.

[0058] The above-mentioned target structured data type can be a structured form of data representation and processing that is more suitable for a specific task or question. The target structured data type includes but is not limited to tables, charts, paragraphs, time series and knowledge bases, etc., which are not limited here.

[0059] In an optional embodiment, when the user issues a generation instruction, such as submitting a question, the system will respond to the instruction and display the target text, i.e. the answer to the question, on the operation interface. This process involves the original text and the initial structured data in the target data set. The initial structured data is obtained based on the original text and the target data set through the construction process, and conforms to the target structured data type corresponding to the original text, that is, the system will select the appropriate structured form, such as tables, charts or paragraphs, according to the question and the document content, in order to organize and present the information.

[0060] By determining the initial structured data that matches the original text to the corresponding target structured data type, text noise can be more effectively reduced, aiding in reasoning, and thus providing more accurate and easier-to-understand answers to users on the operation interface. This method has shown significant advantages in complex document question answering tasks, as it can utilize information associations in structured knowledge to help the model conduct more in-depth analysis and reasoning.

[0061] During the interaction between the user and the system, the user can directly input the question (i.e., the original text) on the operation interface and specify or the system automatically selects the target dataset for retrieval and generation of answers. The target dataset here can refer to a dataset containing different structured knowledge (such as tables, graphs, paragraphs, etc.). By intuitively displaying the original text and the target dataset on the operation interface, the user can clearly see the content of the question and the source of the data processed by the system, thereby increasing the transparency and reliability of the question and answer process. By integrating the original question (original text) and the structured knowledge (initial structured data) selected from the target dataset, the system will generate an answer (i.e., the target text). The initial structured data here is the most suitable structured knowledge type, such as table data, relationship graph, or key information paragraph, determined and constructed by the knowledge selector and knowledge construction agent according to the characteristics of the question. The generated target text is not only based on the original text, but also utilizes highly relevant and structured knowledge to the question, which can reduce irrelevant information interference, improve the accuracy of the answer, and increase the efficiency of the reasoning process.

[0062] Since structured data helps the system understand the information associations within the document, it can conduct more accurate reasoning. By utilizing structured knowledge, the answer can be more accurate and comprehensive. The improvement of the user interface makes the display of the original text and the dataset more intuitive, enhancing the user's understanding of the answering process and improving the transparency of the system. By pre-constructing and screening structured knowledge, the system can quickly locate key information when generating answers, reducing the time spent on processing irrelevant text, and thus improving the overall question and answer efficiency.

[0063] By the above steps, in response to the input instruction acting on the operation interface, the original text and the target data set corresponding to the input instruction are obtained; in response to the generation instruction acting on the operation interface, the target text is displayed on the operation interface, wherein the target text is generated based on the original text and the initial structured data, the initial structured data is constructed based on the original text and the target data set, and conforms to the target structured data type corresponding to the original text, thereby achieving the purpose of improving the accuracy of text generation. It is easy to note that by using the target structured data type to construct the initial structured data, the system can more accurately understand and extract the key information in the original text, so as to provide more relevant and specific information as a reference when generating the target text, thereby ensuring that the model can reason and create based on accurate structured data when generating the target text, reducing the interference of irrelevant information, improving the accuracy and pertinence of the generated text, and thereby solving the technical problem of low accuracy of text generation in related technologies.

[0064] In the above embodiments of the present disclosure, the method further comprises: inputting the original text, the target data set and the target structured data type into the construction agent, and obtaining the initial structured data output by the construction agent; based on the initial structured data and the original text, the retrieval agent is used to obtain the target structured data, wherein the target structured data is used to represent the structured data filtered from the initial structured data, and the matching degree with the original text is greater than a preset threshold; based on the target structured data and the original text, the generation agent is used to obtain the target text.

[0065] The above-mentioned construction agent can be configured to extract relevant information from the original text and convert it into specific structured knowledge. This is achieved by problem-based instructions, ensuring that information extraction and structuring are specific to a particular task. For example, the construction agent can be instructed to "extract all data about income and expenses from the text and organize them into a table format", which is only an example and does not limit in any way.

[0066] The above-mentioned retrieval agent is configured to filter the initial structured data and retain information related to the problem. This is done by identifying keywords in the sub-problems and searching for associated information in the structured knowledge. For example, the retrieval agent can be required to "find all information related to historical income in the table". The above-mentioned target structured data is the output of the retrieval agent, which is a subset of the initial structured data highly related to the problem, and has a high matching degree with the original text, meaning that the information is more targeted and relevant.

[0067] The aforementioned generation agent is configured to conduct comprehensive reasoning based on the target structured data and the raw text to generate the final answer or analysis result. This step utilizes structured knowledge to assist reasoning, reducing noise and improving the accuracy of the answer. For example, the generation agent can be required to "analyze the company's current financial trends based on the filtered historical revenue data".

[0068] In an alternative embodiment, the raw text, target dataset, and target structured data type can be input together to a construction agent, which generates initial structured data based on these information. This step mainly realizes the preliminary structuring of information and knowledge extraction. The initial structured data and the raw text are input again to the retrieval agent, which further filters out structured data that is more relevant to the question, as target structured data, to ensure high matching degree with the raw text. The target structured data and the raw text are submitted to the generation agent for comprehensive analysis and reasoning, and finally the target text, i.e. the answer or analysis report of the question, is generated.

[0069] Through the above steps, irrelevant information interference in the retrieval and reasoning process can be effectively reduced, and the accuracy and efficiency of question answering and analysis based on structured knowledge can be improved. For questions that require integration of information across text segments, compared with using only unstructured text or a single type of structured knowledge, the present disclosure can handle more complex and scattered information to assist the model in making more accurate reasoning. By using diversified structured knowledge, the performance of complex document question answering tasks is significantly improved, especially in cases where complex reasoning is required, text noise can be effectively reduced to assist the model in making accurate judgments.

[0070] For example, suppose there is a task of analyzing a company's annual financial report to determine its growth prediction for the next three years. The raw text is this detailed financial report. First, the construction agent inputs the report and the target dataset (containing knowledge and examples related to financial analysis) together. The target structured data type is a table, because financial data is usually most intuitive in the form of a table. The initial structured data output by the construction agent can be a table containing annual income, expenditure, profit, etc. Second, the retrieval agent receives this table and the raw report, as well as the specific question ("analyze the company's growth prediction for the next three years"). It filters out the most relevant data to "growth prediction for the next three years", such as the income growth rate of the past five years, as target structured data. Finally, the generation agent conducts comprehensive analysis based on these selected structured data and the raw text, considering market trends, company strategies, etc., to generate the target text, i.e. an analysis report on the company's growth prediction for the next three years.

[0071] In this way, the entire framework ensures the accuracy, relevance, and depth of reasoning of the information, providing strong support for complex document question answering.

[0072] In the above embodiments of the present disclosure, after inputting the original text, the target data set and the target structured data type into the construction agent, the method further comprises: obtaining first description information of initial structured data output by the construction agent; inputting the original text and the first description information of the initial structured data into the disassembly agent, and disassembling the original text by using the disassembly agent to obtain a plurality of subtexts; inputting the initial structured data and the original text into the retrieval agent, and obtaining target structured data output by the retrieval agent, comprising: inputting the initial structured data and the plurality of subtexts into the retrieval agent, and obtaining sub-structured data corresponding to the plurality of subtexts output by the retrieval agent; inputting the target structured data and the original text into the generation agent, and obtaining target text output by the generation agent, comprising: inputting the plurality of subtexts and the sub-structured data into the generation agent, and generating target text by using the generation agent to process the plurality of subtexts and the sub-structured data.

[0073] The first description information described above refers to the summary or description of the initial structured data, which helps the subsequent disassembly agent to understand the content and structure of the structured data.

[0074] The disassembly agent described above is an intelligent system for decomposing complex problems or texts into simpler and more easily handled sub-problems or subtexts.

[0075] The retrieval agent described above refers to an intelligent system that can find and extract relevant information in a knowledge base according to a specific query or requirement.

[0076] The subtext described above is a smaller and more specific part of the original text after being processed by the disassembly agent.

[0077] The sub-structured data described above is a specific set of information extracted from the initial structured data for the subtext or sub-problem.

[0078] The generation agent described above refers to an intelligent system that can generate new text or answers based on the given information, usually based on a deep learning model.

[0079] In an alternative embodiment, the original text and the first description information of the initial structured data (summary of structured knowledge) can be input to the disassembling agent. The disassembling agent splits the original text (here specifically the question) into multiple more specific sub-questions. The initial structured data and the split sub-questions are input to the retrieval agent, which performs targeted retrieval on each sub-question and outputs corresponding sub-structured data. The sub-questions and the corresponding sub-structured data are input to the generation agent, which generates the final answer based on these information. Through the above process, not only the accuracy and information relevance of structured knowledge are utilized, but also through the splitting of questions and the fine matching of structured knowledge, it is ensured that the generated answer is based on sufficient and relevant information. This is particularly effective in handling complex document question answering tasks, such as financial report analysis, action plan development, etc., which require extraction and correlation of key information from a large amount of unstructured data. The text generation process can significantly reduce irrelevant noise in the text, reduce reasoning difficulty, and improve the accuracy of question answering tasks.

[0080] For example, assume that the original text is a 100-page company financial report, and the question is "analyze the company's financial trends over the past 5 years". First, the agent constructs the key financial data (such as revenue, expenditure, and profit) in the report into a table format and outputs a summary of structured knowledge. Then, the disassembling agent splits the question into sub-questions such as "changes in revenue over the past 5 years" and "changes in expenditure over the past 5 years". Next, the retrieval agent retrieves the most relevant sub-structured data for each sub-question in the initial structured data. Finally, the generation agent generates a detailed financial trend analysis report based on the sub-questions and sub-structured data. The entire process makes full use of the information relevance of structured knowledge and significantly improves the efficiency and accuracy of problem solving through the collaborative work of multiple agents, especially in complex document question answering scenarios.

[0081] In the above embodiments of the present disclosure, the disassembling agent is used to disassemble the original text to obtain multiple sub-texts, including: determining a disassembling strategy corresponding to the original text based on the first description information; and using the disassembling agent to disassemble the original text according to the disassembling strategy and a preset disassembling granularity to obtain multiple sub-texts.

[0082] The disassembling strategy described above can specify how to divide the original text into smaller, meaningful sub-portions according to the first description information and task requirements. The disassembling strategy can be a sub-question used to generate an answer based on the current structured knowledge.

[0083] The disassembling agent described above can be configured to split the original text according to the disassembling strategy to generate sub-texts.

[0084] The preset disassembly granularity can be a basic unit size or information amount set when the original text is disassembled, to ensure the processability and information integrity of the subtext. The preset disassembly granularity can be a sub-problem that cannot be further disassembled.

[0085] The plurality of subtexts can be smaller parts of the original text formed after disassembly, focusing on specific information or topics, for subsequent processing and analysis.

[0086] In an optional embodiment, the disassembly process of the original text is guided based on the first description information such as the question and the document title. First, by analyzing the first description information, the disassembly strategy most suitable for processing the current question is determined, such as splitting by time, theme or keyword. Then, the disassembly agent is used to disassemble the original text according to the established strategy and the preset disassembly granularity, to generate a series of subtexts that focus more on the core elements of the question, providing a more accurate information basis for subsequent analysis and reasoning.

[0087] Through the above steps, the efficiency and accuracy of the system can be improved. The computational cost of processing irrelevant information is reduced, and the reasoning errors caused by text noise are also reduced. The determination of the disassembly strategy and the application of the disassembly agent ensure the relevance of the subtext and the coherence of the information, thereby improving the performance on complex document question answering tasks.

[0088] In the above embodiments of the disclosure, the original text, the target data set and the target structured data type are input into the construction agent, and the initial structured data output by the construction agent is obtained, including: inputting the original text, the target data set and the target structured data type into the construction agent; using the construction agent to extract keywords from the original text to obtain target keywords; using the construction agent to extract target data matching the target keywords from the target data set; using the construction agent to construct the target data into initial structured data conforming to the target structured data type.

[0089] The target keyword refers to a keyword related to the query or question, used to locate or extract relevant information in the original text. For example, the target keyword can be "financial data", "profit trend", etc., used to locate information related to financial analysis from the annual report. This is only an example and is not limited.

[0090] The target data can be data elements extracted from the target data set and matching the target keywords. For example, the target data can be specific numerical values or text passages containing keywords such as "financial data" and "profit trend". This is only an example and is not limited.

[0091] In an alternative embodiment, the original text, the target data set and the target structured data type can be input into the construction agent, so that information highly relevant to the question can be extracted from the original text and constructed into a specific type of structured data. This process includes three sub-steps of keyword extraction, data extraction and data configuration, and the generated initial structured data will be used as input for further retrieval enhancement and question answering. Through the intelligent driving of the large model, deep understanding and information extraction of the original text can be achieved, effectively reducing irrelevant noise, and at the same time, unstructured information is converted into structured data, so that subsequent retrieval and reasoning can be carried out more efficiently. Through the pre-selection of the target structured data type, it is ensured that the generated structured data is most suitable for the solution of the current question, thereby assisting the model to carry out more accurate reasoning and answering.

[0092] For example, assuming that the original text is an xx annual report of company A, the target data set includes tables, graphs and short text segments extracted from the report, and the target structured data type is a table. The construction agent receives the question: "What is the net profit of company A in xx year?" and the original text and the target data set. The construction agent first identifies the target keywords "net profit" and "xx year", and then extracts the data containing these keywords from the target data set, such as the corresponding cells in the financial data table. Finally, these data are constructed into a table containing net profit information as initial structured data, providing accurate structured knowledge support for subsequent retrieval enhancement and question answering. This process not only reduces irrelevant information in the original report, but also ensures the accuracy of the answer, as the constructed table directly focuses on the specific financial data required by the question, thus avoiding the uncertainty of the model in searching for answers in a large amount of unstructured text.

[0093] In the above embodiments of the present disclosure, the method further comprises: obtaining second description information of the target data set; inputting the second description information and the original text into a type determination model, and obtaining the target structured data type output by the type determination model.

[0094] The above-mentioned second description information refers to information that can provide additional context or features about the target data set, which is used to assist the model to understand the structure, theme or context of the document.

[0095] The above-mentioned type determination model is a machine learning model that can predict or select the most suitable structured knowledge type according to the input description information and text requirements, which is used for subsequent information retrieval and answer generation. Optionally, the type determination model can be a knowledge selector that can determine whether to use tables, charts or text paragraphs to construct structured knowledge based on the question and the document title to assist the answer.

[0096] The target structured data type mentioned above refers to the structured knowledge type that is most beneficial for information retrieval and answer generation in the context of a specific problem and data set.

[0097] In an optional embodiment, the system first collects additional information about the target data set, such as document titles, to help more accurately understand the nature of the document. The question text is input into the knowledge selector along with the contextual information of the target data set, with the goal of enabling the model to judge the more appropriate structured knowledge type based on more comprehensive information. After the type determination model analyzes the input information, it outputs the data type that is more suitable as retrieval enhancement, thereby determining the target structured data type. By combining the question text and the contextual information of the target data set, the knowledge selector can make more accurate and targeted selections, thereby improving the accuracy of text generation. The present disclosure can significantly reduce irrelevant noise in retrieval, enhance the accuracy and reasonableness of answers, especially when dealing with complex document question answering tasks. With explicit document title information, the model can more accurately locate the relevant text part of the question, avoiding a large amount of irrelevant information that may occur in full document retrieval. Structured knowledge such as tables and charts provides clear data relationships and patterns, helping the model to perform more complex data analysis and trend prediction, thereby improving the depth and quality of answers.

[0098] For example, assume a user asks the question: "Please analyze the performance growth of Company A in the fourth quarter based on its 2022 annual financial report." The system first obtains the second description information of the target data set (the financial report), which is the document title "Annual Financial Report of Company A in 2022." Then, the question text and the document title are input into the knowledge selector (type determination model), and after analysis, the knowledge selector outputs the target structured data type as "table" because financial data is most intuitively presented in a table. Next, the system will extract the relevant table data from the document based on this determined type, perform further information retrieval and question decomposition, and finally generate a detailed analysis of the performance growth of Company A in the fourth quarter. This process fully utilizes the advantages of structured knowledge, significantly improving the accuracy and efficiency of answers.

[0099] In the above embodiments of the present disclosure, the type determination model is trained by preference data constructed based on multiple structured data types, and the preference data contains training texts corresponding to multiple structured data types and preference types corresponding to the training texts.

[0100] The type determination model mentioned above is a machine learning model that primarily functions to predict or select the most appropriate structured knowledge type given a question and a document. In this solution, the type determination model is trained using preference data generated by a large model to improve its understanding of the question scenario and the accuracy of knowledge type selection.

[0101] Structured data types refer to the conversion of unstructured text data into organized and queryable forms. Common types include tables, graphs, snippets, etc. Each type is suitable for handling different types of problems or information, such as tables for numerical and categorical information, graphs for relational data, and snippets for text content.

[0102] The above-mentioned preference data is a kind of training data, which contains the preferences or tendencies of the model for various structured knowledge types in different scenarios. These data are usually generated by experts or through specific algorithms, and are used to guide the model to learn when to use which type of structured knowledge.

[0103] The above-mentioned large model (Large Language Model, LLM for short) is a natural language processing model trained on a large amount of data, with strong language understanding and generation capabilities. In this disclosure, LLM is used to construct preference data and perform complex tasks such as information extraction, question decomposition, and answer generation.

[0104] In an optional embodiment, a decision process update algorithm (Direct Policy Optimization, DPO for short) can be used. DPO is a decision process improvement algorithm used to train the model to make better choices given input. In this disclosure, DPO can be used to train the type determination model to select the appropriate structured knowledge type based on the problem and document information.

[0105] The type determination model is trained by preference data constructed based on multiple structured data types to identify the optimal structured knowledge type for a given problem scenario. The generation of preference data is completed by LLM. Specifically, LLM first generates document information and questions covering different structured data types based on given documents and questions, then generates solutions based on each knowledge type, and compares and scores these solutions to determine the appropriate structured knowledge type for the current problem. By using the DPO algorithm, the type determination model can learn the decision logic in the preference data, improving its efficiency and accuracy in practical applications.

[0106] In the embodiments of the present disclosure, the type determination model can accurately determine and select the appropriate structured knowledge type based on the problem requirements and document content through the training of preference data. This improvement in capability enables effective reduction of irrelevant noise in search results when performing complex document question answering tasks, while utilizing information associations in structured knowledge to assist reasoning, thereby achieving higher performance in complex document question answering tasks.

[0107] In the above embodiments of the present disclosure, the preferred type is determined from the plurality of structured data types by comparing the generated text corresponding to the training text using the pre-trained model, the generated text is the text output by the pre-trained model by inputting the training text and the training data set corresponding to the training text into the pre-trained model, the training data set and the training text are the text output by the pre-trained model by inputting the structured data type corresponding to the training text into the pre-trained model.

[0108] The above-mentioned training text can be the input text used for training the knowledge selector, and can be one or more pairs of questions and related documents. The above-mentioned generated text refers to the output generated by the pre-trained model according to the training text, which is used to evaluate the performance of the model under different structured data types. The superordinate concept is the model output. The above-mentioned training data set contains a series of training texts and their corresponding structured knowledge types, which are used to teach the pre-trained model how to select the correct structured knowledge type under different questions. The above-mentioned structured data type refers to the process of converting unstructured text into structured formats such as tables, graphs, or short text segments.

[0109] When training the knowledge selector, the training text is the question and the document, and the pre-trained model can generate output based on these training texts and training data sets (containing questions, documents, and corresponding structured data type labels). The generated text here can include answers generated by the LLM based on the question and the document when using different structured data types. By comparing these generated texts, it can be evaluated which answer under which structured data type is the most accurate and most relevant, thereby determining the preferred type.

[0110] In an optional embodiment, the LLM can be used to generate document information and questions covering different structured data types. For each generated text, the LLM generates solutions based on each structured data type again. The LLM compares and scores these solutions to determine which structured data type is most suitable for the current question. Using the DPO algorithm, the knowledge selector is trained based on the LLM's score to automatically prefer the most suitable structured data type based on the question and document information. Through the comparison and evaluation of the LLM, the most suitable type is selected from the plurality of structured data types. This method of training the knowledge selector based on the simulated preference data generated by the LLM can improve the efficiency of text generation, ensure that the retrieved enhanced knowledge is highly relevant to the question, reduce irrelevant noise, assist reasoning, and thus improve performance on complex document question answering tasks.

[0111] In the above embodiments of the present disclosure, before responding to the generation instruction acting on the operation interface, the method further comprises: displaying a target structured data type corresponding to the original text on the operation interface; and in response to an adjustment instruction acting on the target structured data type, displaying an adjusted structured data type on the operation interface, wherein the adjusted structured data type is obtained by adjusting the target structured data type based on the adjustment instruction; and wherein the initial structured data is constructed based on the original text and the target data set, and conforms to the adjusted structured data type.

[0112] In an optional embodiment, when a user inputs or uploads an original text on the operation interface, the system first determines which type of structured knowledge is most suitable for the text based on the knowledge selector. This process may involve analysis of the text content and relevance assessment with the problem, with the goal of determining which structured form can most effectively assist subsequent question answering or reasoning tasks. After viewing the target structured data type automatically selected by the system, the user may believe that other types of structured knowledge are more suitable based on their own understanding or needs. Therefore, the operation interface should provide a mechanism for the user to issue adjustment instructions to indicate the system to convert the structured data type from one to another. For example, if the system initially converts the text into a table structure, the user may require a graph structure through interface operation to better represent the relationships between entities, which is only used as an example and is not limited. In response to these adjustment instructions, the system can re-analyze the original text and convert it according to the requirements of the new structured data type, and then display the adjusted structured data type on the operation interface.

[0113] The above adjustment instruction is a key to the interaction between the user and the system, which can be used to indicate the change of the structured data type, and may also contain preferences for specific data processing methods, such as requiring more detailed structured data type or more concise data overview. The system's response needs to be based on the adjustment instruction to re-analyze and structure the original text using the corresponding algorithm or model, ensuring that the adjusted structured data type better meets the user's needs and the characteristics of the task.

[0114] The above initial structured data is constructed based on the original text and the target data set specified by the user or selected by the system, which means that the system will consider the format and requirements of the target data set when constructing structured knowledge. When the structured data type is adjusted, the system needs to reconstruct the structured data according to the adjusted type, which may include information extraction, data reorganization and format conversion, etc. to ensure that the adjusted structured knowledge can still effectively support subsequent question answering or reasoning tasks.

[0115] Through the above steps, the system not only adaptively selects a more appropriate structured knowledge type for a given question, but also allows users to adjust the type according to specific scenarios and needs, thereby improving the flexibility and practicality of the framework and better serving complex document question answering and reasoning tasks.

[0116] In the above embodiments of the present disclosure, after displaying the adjusted structured data type on the operation interface, the method further includes adjusting the model parameters of the type determination model based on the adjusted structured data type.

[0117] Users can be allowed to adjust the initially determined structured data type according to more in-depth analysis. This can be manual (user interface adjustment) or automatic (system adjustment based on deeper understanding of the question and document). After the structured data type is adjusted, the system collects the adjustment information and uses it as training data to update the parameters of the type determination model. This step, through reinforcement learning or other machine learning methods, enables the model to more accurately identify and select the structured data type that best matches the question in the future. Through continuous parameter adjustment, the type determination model can continuously learn and improve, better adapting to different types of questions and documents. This not only improves the generalization ability of the model, but also enhances its performance in complex document question answering tasks. This process embodies the adaptability and learning ability of the system. By allowing adjustment and feedback, the system can adjust according to specific scenarios and user needs, improving the efficiency of structured knowledge utilization and the accuracy of question answering.

[0118] For the aforementioned agents, especially the knowledge selector, its parameters can be adjusted and updated through continuous training data. This continuous training and adjustment can be based on user feedback, which can provide information about the task, whether the selected knowledge type is appropriate, and which knowledge types are rejected or inappropriate. By collecting these feedbacks, more training data can be constructed, which is usually in the form of data(task, chosen knowledge type, reject knowledge type)>, where task is the specific problem or task, chosen knowledge type is the knowledge type selected by the knowledge selector to solve the problem, and reject knowledge type is the knowledge type that is considered inappropriate for the current task after trying. Using these partial order training data for the training of the knowledge selector can continuously improve its ability to make accurate and appropriate choices in diverse task scenarios. This method utilizes the feedback in the actual use process of the user, converts real-world data into resources for the learning of the knowledge selector, so that it can more accurately select the most effective knowledge type when facing new problems.

[0119] For complex document question answering tasks, the present disclosure proposes a structured knowledge retrieval-augmentation generation framework (StructRAG) to construct multiple types of structured knowledge based on the original document and use it as retrieval augmentation knowledge, thereby reducing text noise and reasoning difficulty, which can be applied to multiple practical scene tasks. The framework first uses a knowledge selector to determine which type of knowledge to use based on the question and the original document information, then uses a multi-intelligent driven system to disassemble the complex question and construct and utilize the corresponding structured knowledge to generate an answer.

[0120] FIG. 3 is a schematic diagram of retrieving an enhanced framework according to an embodiment of the present disclosure. As shown in FIG. 3, the question can be the original text, the document can be the target data set, the target text can be the final answer, the knowledge building agent can be the building agent described above, the question disassembling agent can be the disassembling agent, the sub-question can be the subtext obtained by disassembling the original text, the knowledge retrieval agent can be the retrieval agent, the accurate knowledge can be the target structured data, and the comprehensive reasoning agent can be the generation agent. The title information of the question and the document can be used as input, the knowledge selector can be used to determine which structured knowledge is more suitable for the current scene, thereby obtaining the knowledge type, the knowledge type and the document can be input into the knowledge building agent, thereby obtaining the structured knowledge concept, the structured knowledge concept and the question can be input into the question disassembling agent, the question disassembling agent can disassemble the input question, thereby obtaining the sub-question, the structured knowledge concept and the sub-question can be input into the knowledge retrieval agent, the accurate knowledge can be obtained by the knowledge retrieval agent, thereby inputting the accurate knowledge and the sub-question into the comprehensive reasoning agent to obtain the final answer.

[0121] The knowledge selector described above uses LLM as the base to directly generate the preferred knowledge type name (i.e., table, graph, chunk, etc.) after obtaining the input. In order to train the knowledge selector described above, the present disclosure can use LLM to construct preference data and then use DPO algorithm for training. Specifically, first, possible knowledge types are provided to LLM, and through a small number of examples (few-shot examples), the large model generates document information and questions covering these knowledge types. Then, the document information and the question are input into LLM to generate a solution to the question based on each type of knowledge. Finally, LLM is used to compare and score each solution, thereby determining which type of structured knowledge is most suitable for the current question, thereby obtaining the preference data required for DPO training.

[0122] The present disclosure can input the original document and the question into the agent responsible for building structured knowledge, and through artificially formulated instructions, the agent first extracts relevant information from the original document based on the question, then builds the information into structured knowledge, and generates a summary of the structured knowledge. Specifically, in order to ensure the performance of the knowledge building process, the artificially formulated instructions can include the following contents: explicitly requiring the building process, that is, “first extract the information related to the question in the original document, and then build it into structured knowledge”; explicitly indicating possible extreme cases, that is, “if it is believed that there is no any relevant information in the document, a small amount of text paragraph closest to the question requirement still needs to be retained”; and requiring the output content to “meet the preset format (json) and contain complete document title information”.

[0123] The question and the structured knowledge summary obtained in the previous step are input to an agent responsible for problem decomposition, which is instructed to split the original question based on the existing structured knowledge, obtaining a plurality of relatively simple sub-questions. Specifically, the instructions set for the agent include the following: a requirement for the granularity of the sub-questions, i.e., "the original question needs to be split into a plurality of sub-questions that cannot be further divided"; and ensuring that the sub-questions obtained by splitting can fully consider the existing structured knowledge, i.e., "according to the structured knowledge summary, split a combination of sub-questions that can easily generate answers using the current structured knowledge".

[0124] The structured knowledge and the plurality of sub-questions obtained in the previous step are input to an agent responsible for knowledge retrieval, which filters the structured knowledge constructed based on the needs of the sub-questions, thereby retaining the most relevant effective information for the question. Specifically, the instructions set for the agent include the following: a clear requirement for the retrieval process, i.e., "first identify the keywords in the sub-questions, then retain the information related to them in the structured knowledge", and the requirement that the agent "still organizes the retrieval results as structured knowledge".

[0125] All sub-questions and the retrieval results corresponding to each sub-question are input to an agent responsible for comprehensive reasoning, which reasons to obtain the final answer based on the input information.

[0126] The artificial instructions for all agents can be prompt words in the form of text, so that the LLM-based agent can execute the related requirements according to the prompts.

[0127] In the present disclosure, the simulated preference data generated based on the LLM is used to train the knowledge selector, which can accurately determine the most suitable type of knowledge according to the question. The multi-agent driven retrieval enhancement system includes the decomposition of complex questions, the construction and use of multiple types of structured knowledge.

[0128] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in the present disclosure are all information and data authorized by the user or authorized by all parties, and the collection, use and processing of related data need to comply with relevant laws, regulations and standards of relevant countries and regions, and provide corresponding operation portal for user to choose authorization or refusal.

[0129] It should be noted that for the foregoing method embodiments, for the sake of simple description, they are all expressed as a combination of a series of actions, but those skilled in the art should know that the disclosure is not limited by the order of the described actions, because according to the disclosure, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should know that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to the disclosure.

[0130] Through the description of the above embodiments, those skilled in the art can clearly understand that the method according to the above embodiments can be realized by means of software and the necessary general hardware platform, and of course it can also be realized by hardware. Based on such understanding, the technical solutions of the disclosure can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes a plurality of instructions for causing an end device (which can be a mobile phone, computer, server, or network device) to execute the method of each embodiment of the disclosure.

[0131] According to the embodiments of the disclosure, a document question answering method is also provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that here.

[0132] FIG. 4 is a flowchart of a document question answering method according to an embodiment of the disclosure. As shown in FIG. 4, the method includes:

[0133] Step S402, in response to an input instruction acting on the operation interface, obtaining inquiry information and a target document corresponding to the input instruction;

[0134] Step S404, in response to a generation instruction acting on the operation interface, displaying reply information on the operation interface.

[0135] Among them, the reply information is generated based on the inquiry information and the initial structured data, the initial structured data is constructed based on the inquiry information and the target document, and conforms to the target structured data type corresponding to the inquiry information.

[0136] For the question answering task of a given knowledge base, the disclosure proposes StructRAG that can adaptively construct and utilize multi-structured knowledge based on the analysis, decision and information extraction capabilities of large models. This framework can effectively reduce irrelevant noise in enhanced knowledge and utilize information association in structured knowledge to assist reasoning, thereby achieving higher performance on complex document question answering tasks.

[0137] By the above steps, the inquiry information and the target document corresponding to the input instruction are acquired in response to the input instruction acting on the operation interface; and the reply information is displayed on the operation interface in response to the generation instruction acting on the operation interface, wherein the reply information is generated based on the inquiry information and the initial structured data, the initial structured data is constructed based on the inquiry information and the target document, and conforms to the target structured data type corresponding to the inquiry information, thereby achieving the purpose of improving the accuracy of text generation. It is easy to note that by using the target structured data type to construct the initial structured data, the system can more accurately understand and extract the key information in the original text, thereby providing more relevant and specific information as a reference when generating the target text, so as to ensure that the model can reason and create based on accurate structured data when generating the target text, reduce the interference of irrelevant information, improve the accuracy and pertinence of the generated text, and thus solve the technical problem of low accuracy of text generation in the related art.

[0138] FIG. 5 is a flowchart of a text generation method according to an embodiment of the present disclosure. As shown in FIG. 5, the method includes:

[0139] Step S502, acquiring an original text and a target data set;

[0140] Step S504, constructing initial structured data based on the original text and the target data set;

[0141] The initial structured data conforms to a target structured data type corresponding to the original text.

[0142] Step S506, generating a target text based on the original text and the initial structured data.

[0143] By the above steps, the original text and the target data set are acquired; the initial structured data is constructed based on the original text and the target data set, wherein the initial structured data conforms to a target structured data type corresponding to the original text; and the target text is generated based on the original text and the initial structured data, thereby achieving the purpose of improving the accuracy of text generation. It is easy to note that by using the target structured data type to construct the initial structured data, the system can more accurately understand and extract the key information in the original text, thereby providing more relevant and specific information as a reference when generating the target text, so as to ensure that the model can reason and create based on accurate structured data when generating the target text, reduce the interference of irrelevant information, improve the accuracy and pertinence of the generated text, and thus solve the technical problem of low accuracy of text generation in the related art.

[0144] FIG. 6 is a flowchart of a text generation method according to an embodiment of the present disclosure. As shown in FIG. 6, the method includes:

[0145] In step S602, the original text and the target data set are obtained by calling the first interface.

[0146] The first interface includes a first parameter, and a parameter value of the first parameter includes the original text and the target data set.

[0147] In step S604, the initial structured data is constructed based on the original text and the target data set.

[0148] The initial structured data conforms to a target structured data type corresponding to the original text.

[0149] In step S606, the target text is generated based on the original text and the initial structured data.

[0150] In step S608, the target text is output by calling the second interface.

[0151] The second interface includes a second parameter, and a parameter value of the second parameter includes the target text.

[0152] Through the above steps, the original text and the target data set are obtained by calling the first interface, the first interface includes the first parameter, and the parameter value of the first parameter includes the original text and the target data set. The initial structured data is constructed based on the original text and the target data set, the initial structured data conforms to the target structured data type corresponding to the original text, the target text is generated based on the original text and the initial structured data, and the target text is output by calling the second interface, the second interface includes the second parameter, and the parameter value of the second parameter includes the target text. The purpose of improving the accuracy of text generation is achieved. It is easy to note that by using the target structured data type to construct the initial structured data, the system can more accurately understand and extract the key information in the original text, so as to provide more relevant and specific information as a reference when generating the target text, thereby ensuring that the model can reason and create according to accurate structured data when generating the target text, reducing the interference of irrelevant information, improving the accuracy and pertinence of the generated text, and thereby solving the technical problem of low accuracy of text generation in related technologies.

[0153] According to the embodiments of the present disclosure, a text generation apparatus for implementing the above text generation method is also provided. FIG. 7 is a schematic diagram of a text generation apparatus according to an embodiment of the present disclosure. As shown in FIG. 7, the apparatus 700 includes an obtaining module 702 and a display module 704.

[0154] The obtaining module is configured to obtain original text and a target data set corresponding to an input instruction acting on the operation interface; and the display module is configured to display a target text on the operation interface in response to a generation instruction acting on the operation interface, wherein the target text is generated based on the original text and initial structured data in the target data set, the initial structured data is constructed based on the original text and the target data set, and conforms to a target structured data type corresponding to the original text.

[0155] It should be noted that the obtaining module 702 and the display module 704 correspond to steps S202 to S204 in the above embodiment, and the two modules have the same instances and application scenarios as the corresponding steps, but are not limited to the content disclosed in the above embodiment. It should be noted that the above modules or units can be hardware components or software components stored in the memory and processed by one or more processors, and the above modules can also be run in the server 10 provided in the above embodiment as part of the device.

[0156] In the above embodiment of the present disclosure, the obtaining module is further configured to input the original text, the target data set, and the target structured data type into the construction agent, and obtain initial structured data output by the construction agent; based on the initial structured data and the original text, the retrieval agent is used to obtain target structured data, wherein the target structured data is used to represent structured data filtered from the initial structured data, and the matching degree with the original text is greater than a preset threshold; based on the target structured data and the original text, the generation agent is used to obtain the target text.

[0157] In the above embodiment of the present disclosure, the obtaining module is further configured to obtain first description information of the initial structured data output by the construction agent; the original text and the first description information of the initial structured data are input into the disassembly agent, and the original text is disassembled by the disassembly agent to obtain a plurality of subtexts, wherein the first description information is output by the construction agent; the initial structured data and the original text are input into the retrieval agent, and target structured data output by the retrieval agent is obtained, including: the initial structured data and the plurality of subtexts are input into the retrieval agent, and sub-structured data corresponding to the plurality of subtexts output by the retrieval agent is obtained; the target structured data and the original text are input into the generation agent, and target text output by the generation agent is obtained, including: the plurality of subtexts and the sub-structured data are input into the generation agent, and the plurality of subtexts and the sub-structured data are text generated by the generation agent to obtain the target text.

[0158] In the above embodiment of the present disclosure, the obtaining module is further configured to determine a disassembly strategy corresponding to the original text based on the first description information; and the disassembly agent is used to disassemble the original text according to the disassembly strategy and a preset disassembly granularity to obtain a plurality of subtexts.

[0159] In the above embodiment of the present disclosure, the obtaining module is further configured to input the original text, the target data set and the target structured data type into the constructed agent; perform keyword extraction on the original text by using the constructed agent to obtain target keywords; extract target data matching the target keywords from the target data set by using the constructed agent; and construct the target data into initial structured data conforming to the target structured data type by using the constructed agent.

[0160] In the above embodiment of the present disclosure, the obtaining module is further configured to obtain second description information of the target data set; input the second description information and the original text into the type determination model, and obtain the target structured data type output by the type determination model.

[0161] In the above embodiment of the present disclosure, the type determination model is trained by preference data constructed based on a plurality of structured data types, and the preference data includes training texts corresponding to the plurality of structured data types and preference types corresponding to the training texts.

[0162] In the above embodiment of the present disclosure, the preference type is a structured data type determined from the plurality of structured data types by comparing a generated text corresponding to the training text by using a pre-training model, and the generated text is a text output by the pre-training model by inputting the training text and a training data set corresponding to the training text into the pre-training model, wherein the training data set and the training text are texts output by the pre-training model by inputting a structured data type corresponding to the training text into the pre-training model.

[0163] In the above embodiment of the present disclosure, the device further comprises a third display module.

[0164] The third display module is configured to display the target structured data type corresponding to the original text on the operation interface; and display an adjusted structured data type on the operation interface in response to an adjustment instruction acting on the target structured data type, wherein the adjusted structured data type is obtained by adjusting the target structured data type based on the adjustment instruction; and the initial structured data is constructed based on the original text and the target data set, and conforms to the adjusted structured data type.

[0165] In the above embodiment of the present disclosure, the device further comprises an adjustment module.

[0166] The adjustment module is configured to adjust model parameters of the type determination model based on the adjusted structured data type.

[0167] According to the embodiments of the present disclosure, a document question answering device for implementing the above document question answering method is further provided. FIG. 8 is a schematic diagram of a document question answering device according to an embodiment of the present disclosure. As shown in FIG. 8, the device 800 includes an acquisition module 802 and a display module 804.

[0168] The acquisition module is configured to acquire the query information and the target document corresponding to the input instruction acting on the operation interface in response to the input instruction acting on the operation interface. The display module is configured to display the reply information on the operation interface in response to the generation instruction acting on the operation interface. The reply information is generated based on the query information and the initial structured data. The initial structured data is constructed based on the query information and the target document, and conforms to the target structured data type corresponding to the query information.

[0169] It should be noted that the acquisition module 802 and the display module 804 correspond to steps S402-S404 in the above embodiments, and the two modules have the same instances and application scenarios as the corresponding steps, but are not limited to the content disclosed in the above embodiments. It should be noted that the above modules or units can be hardware components or software components stored in a memory and processed by one or more processors. The above modules can also be a part of the device and can run in the server 10 provided in the above embodiments.

[0170] According to the embodiments of the present disclosure, a text generation device for implementing the above text generation method is further provided. FIG. 9 is a schematic diagram of a text generation device according to an embodiment of the present disclosure. As shown in FIG. 9, the device 900 includes an acquisition module 902, a construction module 904, and a generation module 906.

[0171] The acquisition module is configured to acquire the original text and the target data set. The construction module is configured to construct the initial structured data based on the original text and the target data set. The initial structured data conforms to the target structured data type corresponding to the original text. The generation module is configured to generate the target text based on the original text and the initial structured data.

[0172] It should be noted that the acquisition module 902, the construction module 904, and the generation module 906 correspond to steps S502-S506 in the above embodiments, and the three modules have the same instances and application scenarios as the corresponding steps, but are not limited to the content disclosed in the above embodiments. It should be noted that the above modules or units can be hardware components or software components stored in a memory and processed by one or more processors. The above modules can also be a part of the device and can run in the server 10 provided in the above embodiments.

[0173] According to the embodiments of the present disclosure, a text generation apparatus for implementing the above-mentioned text generation method is also provided. FIG. 10 is a schematic diagram of a text generation apparatus according to an embodiment of the present disclosure. As shown in FIG. 10, the apparatus 1000 includes an obtaining module 1002, a constructing module 1004, a generating module 1006, and an outputting module 1008.

[0174] The obtaining module is configured to obtain the original text and the target data set by calling a first interface, the first interface including a first parameter, and a parameter value of the first parameter including the original text and the target data set. The constructing module is configured to construct initial structured data based on the original text and the target data set, the initial structured data conforming to a target structured data type corresponding to the original text. The generating module is configured to generate the target text based on the original text and the initial structured data. The outputting module is configured to output the target text by calling a second interface, the second interface including a second parameter, and a parameter value of the second parameter including the target text.

[0175] It should be noted that the obtaining module 1002, the constructing module 1004, the generating module 1006, and the outputting module 1008 correspond to steps S602-S608 in the above-mentioned embodiments, and the four modules have the same instances and application scenarios as the corresponding steps, but are not limited to the content disclosed in the above-mentioned embodiments. It should be noted that the above-mentioned modules or units can be hardware components or software components stored in a memory and processed by one or more processors. The above-mentioned modules can also be a part of the apparatus and can run in the server 10 provided in the above-mentioned embodiments.

[0176] It should be noted that the preferred embodiments involved in the above-mentioned embodiments of the present disclosure have the same scheme, application scenario, and implementation process as the above-mentioned embodiments, but are not limited to the schemes provided in the above-mentioned embodiments.

[0177] The embodiments of the present disclosure can provide an electronic device, which can be any one of the electronic devices in the electronic device group. Alternatively, in the present embodiment, the electronic device can also be replaced by a terminal device such as a mobile terminal.

[0178] Alternatively, in the present embodiment, the electronic device can be located in at least one network device of a plurality of network devices of a computer network.

[0179] In the present embodiment, the computer terminal can execute the program code in the method.

[0180] Optionally, FIG. 11 is a structural block diagram of an electronic device according to an embodiment of the present disclosure. As shown in FIG. 11, the electronic device A can include one or more (only one is shown in FIG. 11) processors 102, a memory 104, a storage controller, and a peripheral interface, wherein the peripheral interface is connected with a radio frequency module, an audio module, and a display.

[0181] The memory can be used to store software programs and modules, such as program instructions / modules corresponding to the method and device in the embodiments of the present disclosure. The processor executes various functions and data processing by running the software programs and modules stored in the memory, that is, implements the method in the above embodiments. The memory can include a high-speed random access memory, and can further include a non-volatile memory, such as one or more magnetic storage devices, flash memories, or other non-volatile solid-state memories. In some examples, the memory can further include a memory remotely arranged with respect to the processor, which can be connected to the terminal A through a network. Examples of the above network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and a combination thereof.

[0182] The processor can call information and application programs stored in the memory through the transmission device to perform the following steps: in response to an input instruction acting on the operation interface, displaying the original text and the target data set on the operation interface; in response to a generation instruction acting on the operation interface, displaying the target text on the operation interface, wherein the target text is generated based on the original text and the initial structured data in the target data set, the initial structured data is constructed based on the original text and the target data set, and conforms to the target structured data type corresponding to the original text.

[0183] By using the embodiments of the present disclosure, in response to an input instruction acting on the operation interface, the identification information of the original text and the target data set corresponding to the input instruction is obtained; in response to a generation instruction acting on the operation interface, the target text is displayed on the operation interface, wherein the target text is generated based on the original text and the initial structured data, the initial structured data is constructed based on the original text and the target data set, and conforms to the target structured data type corresponding to the original text, which achieves the purpose of improving the accuracy of text generation. It is easy to note that by using the target structured data type to construct the initial structured data, the system can more accurately understand and extract the key information in the original text, so as to provide more relevant and specific information as a reference when generating the target text, thereby ensuring that the model can reason and create according to accurate structured data when generating the target text, reducing the interference of irrelevant information, improving the accuracy and pertinence of the generated text, and thereby solving the technical problem of low accuracy of text generation in related technologies.

[0184] Those skilled in the art can understand that the structure shown in FIG. 11 is only schematic, and the electronic device can also be a terminal device such as a smart phone, a tablet computer, a palm computer, a Mobile Internet Device (MID), a PAD, and the like. The FIG. 11 does not limit the structure of the electronic device. For example, the electronic device A can further include more or less components (such as a network interface, a display device, and the like) than those shown in the FIG. 11, or have a different configuration from that shown in the FIG. 11.

[0185] Those skilled in the art can understand that all or part of the steps in the various methods of the above embodiments can be completed by instructing the terminal device related hardware through a program, and the program can be stored in a computer readable storage medium, which can include a flash disk, a Read-Only Memory (ROM), a Random Access Memory (RAM), a magnetic disk or an optical disk, and the like.

[0186] The embodiments of the present disclosure further provide a computer readable storage medium. Optionally, in the present embodiment, the above computer readable storage medium can be used to save the program code executed by the method provided by the above embodiments.

[0187] Optionally, in the present embodiment, the above storage medium can be located in any one of the electronic devices in the group of electronic devices in the computer network, or in any one of the mobile terminals in the group of mobile terminals.

[0188] Optionally, in the present embodiment, the computer readable storage medium is configured to store program code for performing the following steps: in response to an input instruction acting on the operation interface, displaying the original text and the target data set on the operation interface; and in response to a generation instruction acting on the operation interface, displaying the target text on the operation interface, wherein the target text is generated based on the original text and initial structured data in the target data set, the initial structured data is constructed based on the original text and the target data set, and conforms to a target structured data type corresponding to the original text.

[0189] According to the embodiment of the present disclosure, in response to an input instruction acting on the operation interface, original text corresponding to the input instruction and identification information of a target data set are obtained; in response to a generation instruction acting on the operation interface, a target text is displayed on the operation interface, wherein the target text is generated based on the original text and initial structured data, the initial structured data is constructed based on the original text and the target data set, and conforms to a target structured data type corresponding to the original text, thereby achieving the purpose of improving the accuracy of text generation. It is easy to note that by using the target structured data type to construct the initial structured data, the system can more accurately understand and extract the key information in the original text, thereby providing more relevant and specific information as a reference when generating the target text, so as to ensure that the model can reason and create based on accurate structured data when generating the target text, reduce the interference of irrelevant information, improve the accuracy and pertinence of the generated text, and thereby solve the technical problem of low accuracy of text generation in related technologies.

[0190] The embodiment of the present disclosure also provides a document question answering system, and FIG. 12 is a schematic diagram of a document question answering system according to the embodiment of the present disclosure. As shown in FIG. 12, the system 1200 includes a client 1202 and a server 1204.

[0191] The client 1202 is configured to obtain query information corresponding to an input instruction, and the server 1204 is connected with the client and is configured to construct initial structured data based on original text and a target data set, and generate a target text based on the original text and the initial structured data, wherein the initial structured data conforms to a target structured data type corresponding to the original text, and the client is further configured to display the target text.

[0192] The embodiment of the present disclosure also provides a computer program product. Optionally, in the embodiment, the computer program product can include a computer program, and the computer program is executed by a processor to implement the method provided by the above embodiment.

[0193] The embodiment of the present disclosure also provides a computer program product. Optionally, the computer program product can include a non-volatile computer readable storage medium, and the non-volatile computer readable storage medium can be used to store a computer program, and the computer program is executed by a processor to implement the method provided by the above embodiment.

[0194] The embodiment of the present disclosure also provides a computer program. Optionally, in the embodiment, the computer program is executed by a processor to implement the method provided by the above embodiment.

[0195] In the above-described embodiments of the present disclosure, the description of each embodiment focuses on different aspects, and the parts not described in detail in a certain embodiment can be referred to the relevant description of other embodiments.

[0196] In several embodiments provided by the present disclosure, it should be understood that the disclosed technology can be implemented in other manners. The described embodiments of the device are merely schematic, and the division of units is merely logical function division, and there can be other division manners in actual implementation. For example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the displayed or discussed mutual couplings or direct couplings or communication connections can be indirect couplings or communication connections through some interfaces, and electrical or other forms.

[0197] The units described as separate components can or can not be physically separate, and the components shown as units can or can not be physical units, that is, can be located in one place, or can be distributed on a plurality of network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the embodiments.

[0198] In addition, each functional unit in the various embodiments of the present disclosure can be integrated in one processing unit, or each unit can exist physically, or two or more units can be integrated in one unit. The integrated unit can be implemented in the form of hardware or in the form of a software functional unit.

[0199] When the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the present disclosure essentially or substantially, or all or part of the technical solutions, can be embodied in the form of a software product. The computer software product is stored in a storage medium, and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present disclosure. The foregoing storage medium includes: U disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), mobile hard disk, magnetic disk or optical disk, and various media that can store program codes.

[0200] The above only describes the preferred embodiments of the present disclosure, and it should be pointed out that, for those skilled in the art, without departing from the principles of the present disclosure, a number of improvements and refinements can be made, and these improvements and refinements should be considered as the protection scope of the present disclosure.

Claims

1. A text generation method, comprising: in response to an input instruction acting on an operation interface, obtaining original text and a target data set corresponding to the input instruction; in response to a generation instruction acting on the operation interface, displaying a target text on the operation interface, wherein the target text is generated based on the original text and initial structured data, the initial structured data is constructed based on the original text and the target data set, and conforms to a target structured data type corresponding to the original text.

2. The method of claim 1, wherein, The method further comprises: inputting the original text, the target data set and the target structured data type into a construction agent, and obtaining the initial structured data output by the construction agent; based on the initial structured data and the original text, using a retrieval agent to obtain target structured data, wherein the target structured data is used to represent structured data filtered from the initial structured data, and the matching degree of the structured data with the original text is greater than a preset threshold; based on the target structured data and the original text, using a generation agent to obtain the target text.

3. The method of claim 2, wherein, After inputting the original text, the target data set and the target structured data type into the construction agent, the method further comprises: obtaining first description information of the initial structured data output by the construction agent; inputting the original text and the first description information of the initial structured data into a disassembly agent, and using the disassembly agent to disassemble the original text to obtain a plurality of subtexts; the inputting the initial structured data and the original text into the retrieval agent, and obtaining the target structured data output by the retrieval agent, comprises: inputting the initial structured data and the plurality of subtexts into the retrieval agent, and obtaining sub-structured data corresponding to the plurality of subtexts output by the retrieval agent; the inputting the target structured data and the original text into the generation agent, and obtaining the target text output by the generation agent, comprises: inputting the plurality of subtexts and the sub-structured data into the generation agent, using the generation agent to generate text for the plurality of subtexts and the sub-structured data, and obtaining the target text.

4. The method of claim 3, wherein, the disassembling the original text using the disassembly agent to obtain a plurality of subtexts, comprises: based on the first description information, determining a disassembly strategy corresponding to the original text; using the disassembly agent to disassemble the original text according to the disassembly strategy and a preset disassembly granularity to obtain the plurality of subtexts.

5. The method of claim 2, wherein, the inputting the original text, the target data set and the target structured data type into the construction agent, and obtaining the initial structured data output by the construction agent, comprises: inputting the original text, the target data set and the target structured data type into the construction agent; using the construction agent to extract target keywords from the original text; extracting, by the constructed agent, target data matching the target keyword from the target data set; constructing, by the constructed agent, the target data into the initial structured data conforming to the target structured data type.

6. The method of claim 1, wherein, The method further comprises: obtaining second description information of the target data set; inputting the second description information and the original text into a type determination model, and obtaining the target structured data type output by the type determination model.

7. The method of claim 6, wherein, The type determination model is trained by preference data constructed based on a plurality of structured data types, and the preference data includes training texts of a plurality of structured data types and preference types corresponding to the training texts.

8. The method of claim 7, wherein, The preference type is a structured data type determined from the plurality of structured data types by comparing a generated text corresponding to the training text with a pre-training model, wherein the generated text is a text output by the pre-training model by inputting the training text and a training data set corresponding to the training text into the pre-training model, and the training data set and the training text are texts output by the pre-training model by inputting a structured data type corresponding to the training text into the pre-training model.

9. The method of any one of claims 6 to 8, wherein, Before responding to the generation instruction acting on the operation interface, the method further comprises: displaying the target structured data type corresponding to the original text on the operation interface; displaying an adjusted structured data type on the operation interface in response to an adjustment instruction acting on the target structured data type, wherein the adjusted structured data type is obtained by adjusting the target structured data type based on the adjustment instruction; wherein the initial structured data is constructed based on the original text and the target data set, and conforms to the adjusted structured data type.

10. The method of claim 9, wherein, After displaying the adjusted structured data type on the operation interface, the method further comprises: adjusting model parameters of the type determination model based on the adjusted structured data type.

11. A document question answering method, comprising: in response to an input instruction acting on an operation interface, obtaining inquiry information and a target document corresponding to the input instruction; in response to a generation instruction acting on the operation interface, displaying reply information on the operation interface, wherein the reply information is generated based on the inquiry information and initial structured data, the initial structured data is constructed based on the inquiry information and the target document, and conforms to a target structured data type corresponding to the inquiry information.

12. A text generation method, comprising: obtaining an original text and a target data set; constructing initial structured data based on the original text and the target data set, wherein the initial structured data conforms to a target structured data type corresponding to the original text; generating a target text based on the original text and the initial structured data.

13. A text generation method, comprising: obtaining an original text and a target data set by calling a first interface, wherein the first interface comprises a first parameter, and a parameter value of the first parameter comprises the original text and the target data set; constructing initial structured data based on the original text and the target data set, wherein the initial structured data conforms to a target structured data type corresponding to the original text; generating a target text based on the original text and the initial structured data; outputting the target text by calling a second interface, wherein the second interface comprises a second parameter, and a parameter value of the second parameter comprises the target text.

14. A document question-answering system, comprising: a client configured to obtain inquiry information corresponding to an input instruction; a server connected to the client and configured to construct initial structured data based on an original text and a target data set, and generate a target text based on the original text and the initial structured data, wherein the initial structured data conforms to a target structured data type corresponding to the original text; the client is further configured to display the target text.

15. An electronic device, comprising: a memory storing an executable program; a processor configured to run the program, wherein the program performs the following method when executed: in response to an input instruction acting on an operation interface, obtaining an original text and a target data set corresponding to the input instruction; and in response to a generation instruction acting on the operation interface, displaying a target text on the operation interface, wherein the target text is generated based on the original text and initial structured data, the initial structured data is constructed based on the original text and the target data set, and conforms to a target structured data type corresponding to the original text.

16. The electronic device of claim 15, wherein, The program further performs the following method when executed: inputting the original text, the target data set, and the target structured data type into a construction agent, and obtaining the initial structured data output by the construction agent; based on the initial structured data and the original text, obtaining target structured data by using a retrieval agent, wherein the target structured data is used to represent structured data filtered from the initial structured data, and a matching degree of the structured data with the original text is greater than a preset threshold; based on the target structured data and the original text, obtaining the target text by using a generation agent.

17. The electronic device of claim 16, wherein, The program further performs the following method when executed: obtaining first description information of the initial structured data output by the construction agent; inputting the original text and the first description information of the initial structured data into a disassembly agent, and disassembling the original text by using the disassembly agent to obtain a plurality of subtexts; the inputting the initial structured data and the original text into the retrieval agent, and obtaining the target structured data output by the retrieval agent, comprises: inputting the initial structured data and the plurality of subtexts into the retrieval agent, and obtaining sub-structured data corresponding to the plurality of subtexts output by the retrieval agent; The inputting the target structured data and the original text into a generation agent and obtaining the target text output by the generation agent comprises: inputting the plurality of subtexts and the substructured data into the generation agent, performing text generation on the plurality of subtexts and the substructured data by using the generation agent, and obtaining the target text.

18. The electronic device of claim 17, wherein, The program also executes the following method when running: Based on the first description information, determine the disassembly strategy corresponding to the original text; Using the disassembly agent to disassemble the original text according to the disassembly strategy and the preset disassembly granularity to obtain the plurality of subtexts.

19. A computer readable storage medium comprising a stored executable program, wherein, The computer readable storage medium is executed by the device when the executable program is running to control the computer readable storage medium to execute the method of any one of claims 1-13.

20. A computer program product comprising a computer program which, when executed by a processor, implements the method of any one of claims 1-13.

Images