Literature summary text generation method, sample generation method and related devices

Abstract

The invention provides a document summary text generation method, a sample generation method and related devices. The method comprises the steps of obtaining a plurality of target text blocks arranged according to a text sequence for a received literature; wherein each target text block comprises at least one text statement; generating a summary outline according to the target text block summary; wherein the summary outline comprises a plurality of titles; the title corresponds to a text block identifier; the text block identifier is used for identifying a text block; respectively generating summary sub-texts corresponding to a plurality of titles in the summary outline to form a document summary text of the document; wherein the summary sub-text is generated according to the text block corresponding to the corresponding title. The content accuracy of the literature summary text can be improved to a certain degree.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a method for generating literature summary text, a sample generation method, and related apparatus. Background Technology

[0002] When using a literature reading tool, users will first read the summary text of the literature to decide whether they need to read the detailed content.

[0003] In related technologies, the summary content generated directly from the entire document by an artificial intelligence model has the problem of inaccuracy. Summary of the Invention

[0004] In view of this, one or more embodiments of this application provide a method for generating literature summary text, a sample generation method, and related apparatus, which can improve the accuracy of the content of literature summary text to a certain extent.

[0005] In a first aspect, one or more embodiments of this application propose a method for generating a literature summary text, comprising: for a received document, obtaining a plurality of target text blocks arranged in text order; wherein each target text block includes at least one text statement; generating a summary outline based on the target text blocks; wherein the summary outline includes a plurality of headings; each heading corresponds to a text block identifier; the text block identifier is used to identify the text block; generating summary sub-texts corresponding to the plurality of headings in the summary outline to form the literature summary text of the document; wherein the summary sub-texts are generated based on the text blocks corresponding to the respective headings.

[0006] Secondly, one or more embodiments of this application propose a training sample generation method, the method comprising: for a document, obtaining a document summary text generated according to the aforementioned document summary text generation method; receiving a corrected summary text after a first correction for the document summary text; calling a specified large model on the corrected summary text to determine the reasonableness of deriving the corrected summary text from multiple target text blocks of the document by the specified large model; wherein the reasonableness suggestion is used to express the reasonableness of deriving the corrected summary text from sentences included in the multiple target text blocks; receiving a target summary text determined by a second correction based on the reasonableness suggestion, wherein the summary outline and corresponding text blocks of the target summary text constitute a first training sample, and the target summary text, the summary outline, and the multiple text blocks constitute a second training sample; wherein the first training sample is used to train an artificial intelligence model for generating a summary outline, and the second training sample is used to train an artificial intelligence model for generating document summary text.

[0007] Thirdly, one or more embodiments of this application propose an apparatus for generating a literature summary text, comprising: an acquisition module, configured to acquire, for a received document, a plurality of target text blocks arranged in text order; wherein each target text block includes at least one text statement; an outline generation module, configured to generate a summary outline based on the target text blocks; wherein the summary outline includes a plurality of headings; each heading corresponds to a text block identifier; the text block identifier is used to identify the text block; and a summary generation module, configured to generate summary sub-texts corresponding to the plurality of headings in the summary outline, to form the literature summary text of the document; wherein the summary sub-texts are generated based on the text blocks corresponding to the respective headings.

[0008] Fourthly, one or more embodiments of this application propose a training sample generation apparatus, comprising: a summary text acquisition module, configured to acquire a literature summary text generated according to the aforementioned literature summary text generation method for a given literature; a first receiving module, configured to receive a corrected summary text after a first correction for the literature summary text; a calling module, configured to call a specified large model to the corrected summary text, so that the specified large model can determine the reasonableness suggestions for deriving the corrected summary text from multiple target text blocks of the literature; wherein the reasonableness suggestions are used to express the reasonableness of deriving the corrected summary text from sentences included in the multiple target text blocks; a second receiving module, configured to receive a target summary text determined by a second correction based on the reasonableness suggestions, wherein the summary outline and corresponding text blocks of the target summary text constitute a first training sample, and the target summary text, the summary outline, and the multiple text blocks constitute a second training sample; wherein the first training sample is used to train an artificial intelligence model for generating a summary outline, and the second training sample is used to train an artificial intelligence model for generating literature summary text.

[0009] Fifthly, one or more embodiments of this application provide a computer device including a memory and a processor, wherein the memory stores at least one computer program, which is loaded and executed by the processor to implement the method as described above.

[0010] In a sixth aspect, one or more embodiments of this application provide a computer program product including computer instructions that, when executed by a processor, implement the method as described above.

[0011] In a seventh aspect, one or more embodiments of this application provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, causes the processor to implement the method as described above.

[0012] As can be seen from the above embodiments, the multiple embodiments in this application obtain multiple target text blocks arranged in text order for the received documents, generate a summary outline containing text block identifiers based on the target text blocks, and then generate summary sub-texts according to the text blocks corresponding to each title in the summary outline to form the document summary text. This realizes the document summary process of generating a document summary by constraining the document content according to a structured outline, and achieves the effect of improving the accuracy of the summary content when generating the document summary text. Attached Figure Description

[0013] Figure 1 This is a schematic diagram illustrating the process of generating a document summary text according to one embodiment of this application.

[0014] Figure 2 This is a schematic diagram illustrating the process of a training sample generation method provided in one embodiment of this application.

[0015] Figure 3 This is a flowchart illustrating a method for generating a document summary text according to one embodiment of this application.

[0016] Figure 4 This is a flowchart illustrating a training sample generation method provided in one embodiment of this application.

[0017] Figure 5 This is a schematic diagram of a module for generating a document summary text according to an embodiment of this application.

[0018] Figure 6 This is a schematic diagram of a training sample generation device provided in one embodiment of this application.

[0019] Figure 7 This is a schematic diagram of a computer device provided in one embodiment of this application. Detailed Implementation

[0020] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments.

[0021] In the description of the embodiments of this application, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, features defined with "first" and "second" may explicitly or implicitly include one or more of the stated features. In the description of the embodiments of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0022] The user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use and processing of the relevant data must comply with the relevant laws, regulations and standards of the relevant countries and regions, and corresponding operation entry points are provided for users to choose to authorize or refuse.

[0023] In related technologies, when users obtain literature information through literature reading tools, they often prioritize reading the summary text of the literature in order to quickly determine whether it is necessary to read the full text within a limited time. In the past, it was possible to rely on manual reading of the literature and writing of summaries. Although this could ensure the accuracy of the content to a certain extent, the whole process was time-consuming and labor-intensive. Moreover, different annotators had different understandings of the key points, which easily led to inconsistent quality of the summary texts, making it difficult to promote its use in large-scale, multi-literacy reading scenarios.

[0024] With the development of artificial intelligence technology, especially the widespread application of Large Language Models (LLM), solutions have gradually emerged that allow AI models to automatically generate summary texts of documents by directly taking the entire document as input. This type of solution has advantages in terms of generation efficiency and ease of use, quickly producing summary content without requiring manual writing of each document. However, because these methods often treat the entire document as a single input, AI models are prone to omitting key conclusions, confusing chapter information, or even generating "illusionary" content that contradicts the original text when processing lengthy and complex documents. This can lead to discrepancies between the generated summary text and the actual content of the document, potentially misleading users' judgment of its value.

[0025] In summary, existing technologies still suffer from low accuracy in summarizing literature texts, which requires further improvement.

[0026] In several embodiments provided in this application, the methods for generating literature summary text and training samples can be applied to electronic devices with certain computing and network access capabilities. These electronic devices can be desktop computers, laptops, tablets, smartphones, or servers. Specifically, the electronic device includes a processor, memory, and a network access module for network communication. The server can be an electronic device with strong data processing capabilities; alternatively, it can refer to a server cluster formed by multiple electronic devices, or a quantum server built using a quantum computer.

[0027] Please see Figure 1 and Figure 2This application provides an example application scenario for a literature summary text generation device and a training sample generation device. The literature summary text generation device and the training sample generation device can be deployed in an online literature reading platform system.

[0028] For example, staff at an online literature reading platform might want to generate a Chinese summary text for an English medical journal article about a randomized controlled clinical trial of a novel antihypertensive drug. This would allow doctors using the online literature reading tool to first read the summary text and gain a preliminary understanding of the research conclusions and key information. The online literature reading platform can provide this PDF-formatted medical journal article as input to the summary text generation device. Upon receiving the article, the summary text generation device first parses it, dividing it into multiple text blocks according to textual order. These blocks can correspond sequentially to different sections such as title, abstract, introduction, research methods, main results, secondary results, safety analysis, discussion, conclusion, author information, funding support, acknowledgments, copyright statement, and a list of references. After obtaining these text blocks, the device uses an artificial intelligence model to identify and remove text blocks irrelevant to the full summary. For example, text blocks containing only author affiliations, funding numbers, acknowledgments, copyright statements, or pure reference entries do not contain content suitable for a full summary. After removing text blocks that are irrelevant to the overall summary, the remaining text blocks are used as multiple target text blocks for subsequent summary processing.

[0029] In identifying text blocks unrelated to the full-text summary, the document summary text generation device does not statically judge each text block individually. Instead, it uses overlapping sliding recognition windows to identify text blocks according to their order. Specifically, the device can select a specified number of consecutive text blocks for recognition in each round, following the text order of the documents. For example, it might select 10 consecutive text blocks as a recognition window per round, with the specified number being less than the total number of text blocks. After recognizing the 10 text blocks in the current recognition window, the device moves the recognition window along the text order in steps less than a specified number, such as moving forward 5 text blocks each time. This ensures that there are some overlapping text blocks between recognition windows in adjacent rounds, but the number of overlapping text blocks is less than the specified number. During each round of recognition, the device provides a result on whether each of the 10 text blocks in the recognition window is related to the full-text summary, but it discards the recognition results of the first and last text blocks in the recognition window, only using the recognition results of the text blocks in the middle of the window. As the recognition window slides continuously, the same text block will be in the middle of the recognition window in one or more rounds. At this time, the recognition result of the text block is obtained under the premise of sufficient context. Based on these recognition results, the document summary text generation device can finally determine which text blocks belong to text blocks that are not related to the full text summary and remove them, and which text blocks belong to multiple target text blocks and retain them.

[0030] After obtaining multiple target text blocks, the literature summary text generation device can generate a summary outline based on these blocks. Taking the clinical trial literature of this antihypertensive drug as an example, the device can perform thematic analysis and content aggregation on multiple target text blocks, automatically extracting multiple headings, such as "Research Background and Objectives," "Study Design and Subject Baseline Information," "Primary Efficacy Endpoint Results," "Safety and Adverse Events," and "Research Conclusions and Clinical Significance." The device assigns one or more text block identifiers to each heading in the summary outline. Each identifier identifies a corresponding target text block, clearly recording in the summary outline which target text blocks each heading is associated with, making the source of the summary content clear and traceable.

[0031] After generating the summary outline, the literature summary text generation device generates summary sub-texts for each of the multiple headings in the summary outline to form the literature summary text for that literature. Specifically, the literature summary text generation device follows the order of the headings in the summary outline. For each heading, it locates the corresponding set of target text blocks in multiple target text blocks based on the text block identifier corresponding to that heading, and calls the artificial intelligence model for generating literature summary texts to perform content understanding and compression on these target text blocks, generating summary sub-texts corresponding to the topic of that heading. For example, under the heading "Study Design and Subject Baseline Information," the literature summary text generation device can generate summary sub-texts summarizing information such as trial type, randomization method, sample size, inclusion criteria, control group setup, and follow-up time based on the target text blocks associated with that heading; under the heading "Primary Efficacy Endpoint Results," the literature summary text generation device can generate summary sub-texts summarizing changes in major blood pressure indicators, efficacy rate, and statistical significance; under the heading "Safety and Adverse Events," the literature summary text generation device can generate summary sub-texts summarizing the incidence of adverse events, serious adverse events, and typical adverse reaction types. During the generation of each summary subtext, the document summary text generation device internally maintains a one-to-one correspondence between the summary subtext and the text block identifiers of the content source text blocks. When outputting the document summary text, the generation device assigns a corresponding text block identifier to each summary subtext within the document summary text, allowing doctors to easily trace back to the original target text block for detailed verification when needed. Finally, the document summary text generation device can combine multiple summary subtexts into a structured document summary text according to the heading order of the summary outline, and display this document summary text in the document display interface of the online document reading platform system.

[0032] In the actual operation of the online literature reading platform system, to further improve the performance of the AI ​​models for generating summary outlines, generating literature summary texts, and identifying text blocks unrelated to the full-text summary from multiple text blocks, the online literature reading platform system can utilize a training sample generation device to organize the data from the above processes and construct multiple training samples. The training sample generation device can obtain the literature summary text generated by the literature summary text generation method for the same document and assign it to reviewers with medical backgrounds for manual correction. Reviewers can combine the original content of the document with multiple target text blocks to modify the literature summary text, forming a corrected summary text. For example, reviewers can add certain clinically crucial but briefly summarized results indicators, correct insufficiently rigorous statements, or delete inferential statements lacking supporting evidence. The training sample generation device can receive the corrected summary text and use it as the basis for subsequent rationality judgments and training sample construction. Of course, in some implementations, during a cold start, reviewers can also directly divide the document into multiple text blocks, filter out target text blocks, and create summary outlines and literature summary texts based on the target text blocks.

[0033] The training sample generation device can invoke a designated large model to perform a reasonableness assessment of the corrected summary text based on multiple target text blocks. The device can construct prompts, using multiple target text blocks corresponding to the literature as reference knowledge, sentences in the corrected summary text as viewpoints to be judged, and inputting both the reference knowledge and viewpoints to be judged into the designated large model, requesting the model to output reasonableness suggestions. These reasonableness suggestions express the reasonableness of each summary statement in the corrected summary text derived from the sentences included in the multiple target text blocks. For example, if the corrected summary text contains a statement such as "This antihypertensive drug is significantly superior to the control drug in patients of all ages," but the multiple target text blocks only provide efficacy results for the overall population without age-group comparison data, the designated large model might output an "unreasonable" reasonableness suggestion, indicating that the summary statement does not completely match the evidence in the original text. The training sample generation device can then provide feedback on these reasonableness suggestions to reviewers, who can then perform secondary corrections based on the corrected summary text and the reasonableness suggestions, deleting, modifying, or supplementing unreasonable summary statements to ultimately form the target summary text.

[0034] After obtaining the target summary text, the training sample generation device can generate a first training sample based on the target summary text, the corresponding summary outline, and the target text blocks associated with each heading in the summary outline. The first training sample can include multiple target text blocks and their corresponding summary outlines, used to train an AI model for generating summary outlines. This enables the AI ​​model to learn and output a summary outline corresponding to the target summary text when multiple target text blocks are input. Furthermore, the training sample generation device can construct a second training sample. The second training sample can include the target summary text, the corresponding summary outline, and multiple text blocks obtained by dividing the document according to text order. This second training sample is used to train an AI model for generating document summary text, enabling the AI ​​model to output a document summary text that is structurally and content-wise essentially identical to the target summary text when multiple text blocks and their corresponding summary outlines are input.

[0035] Furthermore, the training sample generation device can also construct a third training sample to train an artificial intelligence model that identifies and extracts text blocks unrelated to the full-text summary from multiple text blocks. The training sample generation device can utilize the multiple text blocks and multiple target text blocks obtained in the aforementioned steps by the document summary text generation device. It associates the multiple text blocks with their text order and whether they are retained as target text blocks, marking those retained as "related to the full-text summary" and those to be removed as "unrelated to the full-text summary." The training sample generation device can then use these multiple text blocks and their labels to form a third training sample, training the artificial intelligence model to identify and extract text blocks unrelated to the full-text summary from multiple text blocks. After training, the artificial intelligence model that identifies and extracts text blocks unrelated to the full-text summary from multiple text blocks can automatically receive multiple text blocks obtained by dividing documents according to text order in new document processing scenarios, and output the text blocks that need to be removed and are unrelated to the full-text summary, thereby improving the ability of the online document reading platform system to automatically filter target text blocks in large-scale document processing scenarios.

[0036] As can be seen from the above application scenario examples, the literature summary text generation device utilizes steps such as text block segmentation, sliding window recognition, target text block filtering, summary outline generation, and summary sub-text generation to generate well-structured and traceable literature summary texts for the online literature reading platform system. The training sample generation device uses the corrected summary text, reasonable suggestions, and target summary text to construct the first training sample, the second training sample, and the third training sample. These are used to continuously optimize the artificial intelligence model for generating the summary outline, the artificial intelligence model for generating the literature summary text, and the artificial intelligence model for identifying text blocks unrelated to the full text summary from multiple text blocks. This continuously improves the accuracy of literature summary text generation in the online literature reading platform system.

[0037] Please see Figure 1 and Figure 3 One embodiment of this application provides a method for generating literature summary text. The method for generating literature summary text can be applied to a generating device, which can be applied to the aforementioned electronic device possessing certain computing power and network access capabilities. Of course, in some embodiments, the generating device can also be software running on the electronic device. The method for generating literature summary text may include the following steps.

[0038] Step S110: For the received document, obtain multiple target text blocks arranged in text order; wherein each target text block includes at least one text statement.

[0039] Step S120: Generate a summary outline based on the target text block; wherein the summary outline includes multiple titles; each title corresponds to a text block identifier; the text block identifier is used to identify the text block.

[0040] Step S130: Generate summary sub-texts for each of the multiple headings in the summary outline to form the literature summary text of the document; wherein, the summary sub-texts are generated based on the text blocks corresponding to the respective headings.

[0041] In this embodiment, the generation device can be used to obtain multiple target text blocks arranged in text order from the received document. A target text block can be understood as a text segment obtained after parsing and segmenting the document content, used to participate in generating the summary text. Each target text block includes at least one text statement and maintains consistency with the original text order of the document. Specifically, the generation device can first parse the document's format, extracting the original text content from different parts such as the title, abstract, introduction, methods, results, discussion, and conclusions, and then divide the original text content into multiple text blocks arranged in text order according to paragraph boundaries, sentence structure, or preset segmentation rules. A text block can be understood as a basic content unit obtained after preliminary segmentation of the document, used to accommodate a small segment of text that is relatively continuous in the document's structure.

[0042] In some implementations, the generation device can further identify and remove text blocks unrelated to the full-text summary from the aforementioned multiple text blocks, thereby obtaining target text blocks for subsequent processing. Text blocks unrelated to the full-text summary can be understood as text fragments that do not involve content that can serve as the full-text summary, such as only containing author lists, funding information, institution names, acknowledgments, copyright statements, footnotes, reference information, and figure / table numbering descriptions. Since such text content is generally not used by users to judge the reliability of research conclusions, methods, or findings when reading literature summary texts, it does not need to participate in the summary generation process. In this implementation, the generation device can analyze multiple text blocks based on preset filtering rules, text feature recognition models, or other artificial intelligence models, mark text blocks unrelated to the full-text summary, and remove these marked text blocks from the subsequent processing flow, preventing them from being used as target text blocks. After the above identification and removal operations, the remaining text blocks constitute the target text blocks, which provide content sources for subsequent summary outline generation and literature summary text generation.

[0043] In this embodiment, the generating device can be used to generate a summary outline based on target text blocks. The summary outline can be understood as a hierarchical outline used to indicate the overall structure of the literature summary text. Specifically, the generating device can perform thematic analysis and content aggregation based on the text content of multiple target text blocks, extracting multiple headings to summarize different content themes, and constructing a summary outline accordingly. Each heading in the summary outline can be used to represent a theme or section in the literature summary text, such as "Research Background and Objectives," "Research Methods," "Experimental Results," "Safety Analysis," "Conclusions and Outlook," etc. In this embodiment, the generating device can configure a text block identifier for each heading in the summary outline. The text block identifier is used to identify the target text block associated with that heading, indicating the scope of content sources that should be referenced when generating the summary. The text block identifier can be marking information used to uniquely identify the target text block, such as the sequential number, offset position, or other distinguishable markers of the target text block in the literature, thereby explicitly recording the "heading-text block" correspondence in the summary outline.

[0044] In this embodiment, the generation device can also be used to generate summary sub-texts for each of the multiple headings in the summary outline, thus forming the literature summary text. A summary sub-text can be understood as a partial summary of a specific heading and its associated target text blocks, used to describe key information under the corresponding topic. Specifically, the generation device can read each heading and its corresponding text block identifier one by one according to the order of the headings in the summary outline, and locate the corresponding target text blocks in the target text blocks based on the text block identifiers, using these target text blocks as the content source corresponding to the heading. Based on this, the generation device can call an artificial intelligence model to perform semantic analysis and information compression on the corresponding target text blocks, generating summary sub-texts that match the topic of the heading. For example, when the heading is "Research Methods," the generation device can generate a summary sub-text summarizing the research design, sample inclusion criteria, and main interventions based on the target text blocks associated with the heading; when the heading is "Research Results," the generation device can generate a summary sub-text summarizing the primary endpoint indicators, secondary endpoint indicators, and significance results based on the corresponding target text blocks. In some implementations, to ensure a clear correspondence between the summary subtexts and the original text content, the generation device can continuously use text block identifiers to constrain the generation scope of the summary subtexts within its internal processing flow, ensuring that each summary subtext is primarily generated based on the target text block of the information source. The generation device can then concatenate multiple summary subtexts or combine them in a structured manner according to the order of the headings in the summary outline, thereby constructing a document summary text for external display.

[0045] In this embodiment, the literature summary text always maintains a correspondence with the target text block and summary outline during the generation process, which helps to improve the accuracy of the literature summary text while removing content that is irrelevant to the full summary.

[0046] In some implementations, the generating device may select a specified number of text blocks for recognition in each round according to the text order; wherein the specified number is less than the total number of the multiple text blocks; between adjacent rounds, there are some repeated text blocks among the specified number of text blocks, and the number of repeated text blocks is less than the specified number.

[0047] In this embodiment, to identify and remove text blocks irrelevant to the full text summary from multiple text blocks to obtain target text blocks for subsequent processing, the generation device can perform batch recognition processing on multiple text blocks using a sliding recognition window-like approach. Specifically, the generation device can sort the multiple text blocks obtained from the document division according to the text order, and select a specified number of text blocks arranged consecutively in text order in each round. These specified number of text blocks are then treated as text blocks in a single recognition window for relevance recognition. The specified number is less than the total number of text blocks, limiting the number of text blocks included in each round of recognition, allowing the generation device to perform comprehensive recognition of the entire document through multiple rounds of recognition windows without processing all text blocks at once.

[0048] In this embodiment, after completing the recognition of the current round of the recognition window, the generation device can move the recognition window along the text sequence to construct the recognition window for the next round. Specifically, the step size of the window movement can be set to be less than the aforementioned specified number, so that there are some overlapping text blocks in the recognition windows of adjacent rounds. That is, several text blocks located in the middle position of the current round of the recognition window are still included in the window range of the next round of the recognition window. Since the number of overlapping text blocks is less than the specified number, overlapping areas are only generated between adjacent recognition windows on some text blocks, thereby achieving the coverage recognition of multiple text blocks in the process of gradually sliding the recognition window along the text sequence. Through the above-mentioned sliding recognition window method with overlapping areas, the generation device can combine the context information of continuous text blocks in the window in each round to identify whether each text block in the window is related to the full text summary. Based on the recognition results of multiple rounds, text blocks that are not related to the full text summary are eliminated, so that the remaining text blocks constitute the target text blocks, providing a more accurate content basis for the subsequent generation of summary outlines and literature summary texts.

[0049] In some implementations, the generating device discards the recognition results of the first and last text blocks in a specified number of text blocks according to the text order during each round of recognition.

[0050] In this embodiment, during the process of identifying and removing text blocks unrelated to the full text summary from multiple text blocks using a sliding recognition window to obtain the target text block, the generation device can further filter the recognition results of each text block within each round of the recognition window to improve the reliability of the recognition results. Specifically, when selecting a specified number of consecutively arranged text blocks as a round of recognition windows according to text order, the generation device can discard the recognition results of the first and last text blocks within the specified number of text blocks after completing the recognition of whether each text block within the recognition window is related to the full text summary. That is, in each round of recognition, only the recognition results of one or more text blocks located in the middle of the current recognition window are used for subsequent elimination decisions, instead of using the recognition results corresponding to the first and last text blocks in the current recognition window.

[0051] In this embodiment, since the aforementioned recognition window slides sequentially along the text with a step size less than a specified number, there are partially overlapping text blocks between adjacent recognition windows, allowing the same text block to appear in different positions in different recognition windows. By discarding the recognition results of the first and last text blocks in each round, the generation device can ensure that each text block participates in recognition at least in one round by being located in the middle of the recognition window. This allows the device to combine the contextual information provided by its preceding and following text blocks when determining whether a text block is related to the full text summary, avoiding the instability problem caused by relying solely on truncated context. In some embodiments, the generation device can determine whether a text block is to be discarded as a text block unrelated to the full text summary based solely on the recognition result obtained when the text block is in the middle of the recognition window; for text blocks located at the beginning or end of the recognition window in certain rounds, their corresponding recognition results are not directly adopted. By abandoning the recognition results of the first and last text blocks of the recognition window in each round of recognition, the generation device can further improve the accuracy of identifying and removing text blocks that are irrelevant to the full text summary within the framework of the sliding recognition window. This makes the target text blocks that are ultimately retained provide a more reliable content foundation for the subsequent generation of summary outlines and literature summary texts.

[0052] In some implementations, each summary subtext in the literature summary text is assigned a text block identifier corresponding to the source text block of the content.

[0053] In this embodiment, the generating device can also explicitly retain the correspondence between summary sub-texts and their content sources in the literature summary text. Specifically, while generating summary sub-texts based on multiple headings in the summary outline, the generating device can associate the information of the target text block used to generate the summary sub-text with the summary sub-text based on the text block identifier corresponding to the heading. This results in a text block identifier for each summary sub-text in the literature summary text, corresponding to a text block with a corresponding content source. The text block identifier can then be used to identify one or more target text blocks associated with the content of the summary sub-text, thereby recording the range of information sources upon which the summary sub-text is based at the summary sub-text level.

[0054] In some implementations, the generation device can output the text content representing the summary subtext along with the corresponding text block identifiers when constructing the literature summary text. For example, the generation device can maintain a field containing a list of text block identifiers for each summary subtext in its internal data structure, which is used to characterize which target text blocks the summary subtext was generated from. When displaying externally, the text block identifiers can be attached to the summary subtext in the form of tags, annotations, or metadata, or presented in the front-end display layer as reference annotations, floating prompts, etc. By setting text block identifiers for the corresponding content source text blocks for each summary subtext in the literature summary text, on the one hand, a traceable relationship can be maintained between the summary subtext and its source target text blocks, facilitating subsequent verification, review, or correction of the summary content; on the other hand, it also provides basic data support for constructing training samples and optimizing artificial intelligence models based on the correspondence between the literature summary text and target text blocks.

[0055] Please see Figure 2 and Figure 4 One or more embodiments of this application also provide a training sample generation method. The training sample generation method can be applied to a training sample generation device, which can be applied to the aforementioned electronic device possessing certain computing power and network access capabilities. Of course, in some embodiments, the training sample generation device can also be software running on the electronic device. The training sample generation method may include the following steps.

[0056] Step S210: For the literature, obtain the literature summary text generated according to the literature summary text generation method described above.

[0057] Step S220: Receive the literature summary text, which is a corrected summary text after one correction.

[0058] Step S230: Call the specified large model to call the correction summary text, so that the specified large model can determine the reasonableness of the correction summary text derived from multiple target text blocks of the document; wherein, the reasonableness suggestion is used to express the reasonableness of the correction summary text derived from the sentences included in the multiple target text blocks.

[0059] Step S240: Receive the target summary text determined by secondary correction based on the rationality suggestion. The summary outline and corresponding text blocks of the target summary text constitute the first training sample. The target summary text, the summary outline, and the multiple text blocks constitute the second training sample. The first training sample is used to train the artificial intelligence model that generates the summary outline, and the second training sample is used to train the artificial intelligence model that generates the literature summary text.

[0060] In this embodiment, the training sample generation device can be used to obtain a literature summary text generated according to the aforementioned method for generating literature summary texts for a target document. Specifically, the training sample generation device can obtain the literature summary text and its internal structure information from the generation device through a preset interface. The literature summary text is generated based on multiple target text blocks of the document, and a correspondence has been established between the summary outline and the target text blocks during the generation process. The training sample generation device can use the obtained literature summary text as one of the basic samples for constructing training data based on subsequent training requirements.

[0061] In this embodiment, to improve the quality of the summary text in the training samples, the training sample generation device can receive a corrected summary text after the literature summary text has undergone one round of manual correction. The corrected summary text can be understood as the text result obtained by a human reviewer after the first round of manual correction based on the literature summary text. It is used to correct omissions, unclear expressions, or inconsistencies with the original text in the literature summary text. Specifically, the human reviewer can refer to the original literature and target text blocks to supplement or delete from the literature summary text, such as supplementing missing key conclusions, adjusting inaccurate expressions, or merging redundant summary paragraphs. After completing the correction, the corrected summary text is provided to the training sample generation device. By introducing one round of manual correction, subsequent training samples can be based on higher-quality summary texts.

[0062] In this embodiment, the training sample generation device can also be used to call a designated large model to retrieve the corrected summary text, allowing the designated large model to determine the reasonableness of deriving the corrected summary text from multiple target text blocks of the document. The designated large model can be understood as an artificial intelligence model used to perform content consistency and evidence support judgments, such as a large language model. The training sample generation device can organize the corrected summary text and the multiple target text blocks corresponding to the document into prompts that meet the input requirements of the designated large model. Sentences included in the multiple target text blocks can serve as reference knowledge, and sentences in the corrected summary text can serve as viewpoints or conclusions to be tested. After receiving the prompts, the designated large model can determine whether one or more sentences in the corrected summary text can be reasonably derived from the sentences included in the multiple target text blocks, and output reasonableness suggestions. Reasonableness suggestions can be used to express the reasonableness of deriving the corrected summary text from multiple target text blocks, such as indicating whether a certain summary statement lacks corresponding original text evidence, whether there are inferences inconsistent with the content of the target text blocks, or marking which parts need further verification and modification. The training sample generation device can associate and store reasonableness suggestions with the corrected summary text for subsequent manual revision.

[0063] In this embodiment, the training sample generation device can be used to receive the target summary text determined through secondary correction based on reasonableness suggestions. The target summary text can be understood as the final summary result obtained after a second round of correction, either manually or semi-automatically, based on the corrected summary text and reasonableness suggestions output by a specified large model. Specifically, human reviewers can refer to the reasonableness suggestions to modify, supplement, or delete sentences deemed lacking evidence or containing expression biases, ensuring that every key conclusion in the target summary text can find reasonable support from multiple target text blocks in the literature. By introducing a secondary correction process based on reasonableness suggestions, the proportion of unreasonable inferences or "illusionary" content in the summary text can be significantly reduced, making the target summary text more closely aligned with the facts of the original text.

[0064] In this embodiment, the training sample generation device can construct training samples for different artificial intelligence models based on the target summary text. Specifically, the training sample generation device can first extract the corresponding summary outline from the target summary text, and determine the set of text blocks corresponding to each title in the summary outline based on the association between each title and the corresponding text block identifier in the summary outline, so that the summary outline and corresponding text blocks of the target summary text constitute the first training sample. The first training sample can be used to train the artificial intelligence model that generates the summary outline, that is, by taking multiple target text blocks of the document as input and the summary outline corresponding to the target summary text as training labels, the artificial intelligence model learns how to automatically generate a well-structured and clearly divided summary outline based on the target text blocks.

[0065] In this embodiment, the training sample generation device can also construct a second training sample for training the generation of literature summary text. Specifically, the training sample generation device can combine the target summary text, the corresponding summary outline, and multiple text blocks of the document as the second training sample. The multiple text blocks can include all text blocks divided according to the document's text order, providing a more complete original content context for the artificial intelligence model; the summary outline provides structural constraints for the literature summary text; and the target summary text serves as the output result that the model needs to fit during training. The second training sample can be used to train the artificial intelligence model for generating literature summary text, enabling the model to output literature summary text consistent with the content and structure of the target summary text when inputting multiple text blocks and corresponding summary outlines.

[0066] By constructing the first and second training samples described above, the training sample generation device can provide high-quality supervision data for the AI ​​models that generate summary outlines and literature summary texts, while ensuring the quality of the summary text and the rationality of the evidence. This is beneficial to improving the overall effect of automatically generating summary outlines and literature summary texts. In some embodiments, the first and second training samples can be used to train the same large language model, so that the large language model can generate summary outlines and literature summary texts.

[0067] In some implementations, the training sample generation device can acquire multiple text blocks of the document divided according to textual order; wherein, the multiple text blocks and the multiple target text blocks constitute a third training sample; the third training sample is used to train an artificial intelligence model to identify and extract text blocks that are unrelated to the full text summary from the multiple text blocks.

[0068] In this embodiment, based on the construction of a first training sample for training the generation of a summary outline and a second training sample for training the generation of a literature summary text, the training sample generation device can also be used to construct a third training sample for training an artificial intelligence model for filtering text blocks, so that the artificial intelligence model can automatically identify and output text blocks that are not related to the full text summary from multiple text blocks.

[0069] Specifically, the training sample generation device can acquire multiple text blocks of the document, divided according to textual order. These multiple text blocks can be consistent with the source of the text blocks obtained by the generation device in the aforementioned method for generating the document summary text. That is, they are the basic content units obtained by segmenting the original text content of the title, abstract, introduction, methods, results, discussion, and conclusion sections according to the document's textual order. After completing the document summary text generation and target text block determination processes, the training sample generation device can have multiple text blocks divided according to textual order, and multiple target text blocks obtained by identifying and removing text blocks irrelevant to the full-text summary from these multiple text blocks. Based on the inclusion relationship between the multiple text blocks and the multiple target text blocks, the training sample generation device can automatically generate labeling information for each text block indicating its relevance to the full-text summary. For example, text blocks belonging to the set of multiple target text blocks are labeled as "relevant to the full-text summary," and text blocks not belonging to the set of target text blocks are labeled as "irrelevant to the full-text summary."

[0070] In this embodiment, the training sample generation device can organize the aforementioned multiple text blocks and their correspondence with multiple target text blocks into a third training sample. The third training sample can be understood as a supervised data set constructed for a text block-level filtering task. It uses multiple text blocks as input features of the model and whether each text block belongs to a text block unrelated to the full-text summary as the output label. This is used to train an artificial intelligence model that identifies and extracts text blocks unrelated to the full-text summary from multiple text blocks. This artificial intelligence model can learn the semantic features, positional features, and statistical patterns of the correlation between different types of text blocks in multiple documents and the full-text summary during the training phase. Therefore, when receiving multiple text blocks obtained by dividing any document according to text order during the inference phase, it can automatically identify and output which text blocks belong to text blocks unrelated to the full-text summary.

[0071] Please see Figure 5 One or more embodiments of this application also provide an apparatus for generating literature summary text. The apparatus for generating literature summary text includes: an acquisition module, an outline generation module, and a summary generation module.

[0072] The acquisition module is used to acquire multiple target text blocks arranged in text order for the received documents; wherein each target text block includes at least one text statement.

[0073] The outline generation module is used to generate a summary outline based on the target text block; wherein, the summary outline includes multiple titles; each title corresponds to a text block identifier; the text block identifier is used to identify the text block.

[0074] The summary generation module is used to generate summary sub-texts for each of the multiple headings in the summary outline, so as to form the literature summary text of the document; wherein, the summary sub-texts are generated based on the text blocks corresponding to the respective headings.

[0075] In this embodiment, the functions and effects of the document summary text generation device can be explained in comparison with the aforementioned embodiments, and will not be repeated here.

[0076] Please see Figure 6 One or more embodiments of this application also provide a training sample generation apparatus. The training sample generation apparatus may include: a summary text acquisition module, a first receiving module, a calling module, and a second receiving module.

[0077] The summary text acquisition module is used to acquire literature summary texts generated according to the literature summary text generation method described above.

[0078] The first receiving module is used to receive the corrected summary text of the literature summary text after one correction.

[0079] The calling module is used to call a specified large model to the correction summary text, so that the specified large model can determine the reasonableness of the correction summary text derived from multiple target text blocks of the document; wherein, the reasonableness suggestion is used to express the reasonableness of deriving the correction summary text from the sentences included in the multiple target text blocks.

[0080] The second receiving module is used to receive the target summary text determined by secondary correction based on the rationality suggestions. The summary outline and corresponding text blocks of the target summary text constitute the first training sample, and the target summary text, the summary outline, and the multiple text blocks constitute the second training sample. The first training sample is used to train an artificial intelligence model to generate the summary outline, and the second training sample is used to train an artificial intelligence model to generate the literature summary text.

[0081] In this embodiment, the functions and effects of the training sample generation device can be explained in comparison with the aforementioned embodiments, and will not be repeated here.

[0082] Please see Figure 7This application also provides a computer device comprising: a memory and a processor, wherein the memory stores at least one computer program, and the at least one computer program is loaded and executed by the processor to implement the method described above.

[0083] The memory, processor, and communication interface in the computer device can communicate with each other via the system bus and network communication.

[0084] In this embodiment, the functions and effects implemented by the computer device can be explained by referring to the foregoing embodiments, and will not be repeated here.

[0085] This application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, causes the processor to implement the method as described above.

[0086] The functions and effects achieved in this embodiment can be explained by referring to other embodiments, and will not be repeated here.

[0087] This application also provides a computer program product containing instructions, including a computer program / instructions that, when executed by a processor, implement the method as described above.

[0088] The functions and effects achieved in this embodiment can be explained by referring to other embodiments, and will not be repeated here.

[0089] It is understood that the specific examples in this document are only intended to help those skilled in the art better understand the embodiments of this application, and are not intended to limit the scope of the invention.

[0090] It is understood that in the various embodiments of this application, the sequence number of each process does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0091] It is understood that the various implementation methods described in this application can be implemented individually or in combination, and the implementation methods in this application are not limited in this respect.

[0092] Unless otherwise stated, all technical and scientific terms used in the embodiments of this application have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to limit the scope of this application. The term "and / or" as used in this application includes any and all combinations of one or more of the associated listed items. The singular forms "a," "the," and "the" as used in the embodiments of this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise.

[0093] It is understood that the processor in the embodiments of this application can be an integrated circuit chip with signal processing capabilities. During implementation, each step of the above method embodiments can be completed by the integrated logic circuits in the processor's hardware or by instructions in software form. The processor can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this application can be directly embodied in the execution of a hardware decoding processor, or executed by a combination of hardware and software modules in the decoding processor. The software modules can be located in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. This storage medium is located in memory; the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method.

[0094] It is understood that the memory in the embodiments of this application may be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. Specifically, non-volatile memory may be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. Volatile memory may be random access memory (RAM). It should be noted that the memory in the systems and methods described herein is intended to include, but is not limited to, these and any other suitable types of memory.

[0095] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0096] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the aforementioned method implementations, and will not be repeated here.

[0097] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0098] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0099] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0100] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0101] The above description is merely a specific embodiment of this application, but the scope of protection of this invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this invention should be determined by the scope of the claims.

Claims

1. A method for generating a literature summary text, characterized in that, include: For the received document, obtain multiple target text blocks arranged in text order; wherein each target text block includes at least one text statement; A summary outline is generated based on the target text block; wherein, the summary outline includes multiple headings; each heading corresponds to a text block identifier; the text block identifier is used to identify the text block; Each of the multiple headings in the summary outline is used to generate a summary sub-text to form the literature summary text of the document; wherein, the summary sub-text is generated based on the text block corresponding to the respective heading.

2. The method according to claim 1, characterized in that, For the received documents, obtain multiple target text blocks arranged in text order, including: The document is divided into multiple text blocks according to the text order; The target text block is obtained by identifying and removing text blocks that are irrelevant to the full text summary from the plurality of text blocks; wherein the removed text blocks do not contain content that can be used as a full text summary.

3. The method according to claim 1, characterized in that, The target text block is obtained by identifying and removing text blocks that are irrelevant to the full text summary from the plurality of text blocks, including: According to the text order, a specified number of text blocks are selected for recognition in each round; wherein, the specified number is less than the total number of the multiple text blocks; between adjacent rounds, there are some repeated text blocks among the specified number of text blocks, and the number of repeated text blocks is less than the specified number.

4. The method according to claim 3, characterized in that, The process of identifying and removing text blocks that are irrelevant to the full text summary from the plurality of text blocks to obtain the target text block also includes: in each round of identification, discarding the identification results of the first and last text blocks that are in the specified number of text blocks according to the text order.

5. The method according to claim 1, characterized in that, Each summary subtext in the literature summary text is assigned a text block identifier corresponding to the source text block.

6. A method for generating training samples, characterized in that, The method includes: For the literature, obtain the literature summary text generated according to any one of claims 1 to 5; Receive the summary text of the literature, which is a corrected summary text after one correction. The correction summary text is invoked by a specified large model, which then determines the reasonableness of deriving the correction summary text from multiple target text blocks in the document; wherein, the reasonableness suggestion is used to express the reasonableness of deriving the correction summary text from the sentences included in the multiple target text blocks; The system receives a target summary text determined through secondary correction based on the aforementioned rationale suggestions. The summary outline and corresponding text blocks of the target summary text constitute a first training sample, and the target summary text, the summary outline, and the multiple text blocks constitute a second training sample. The first training sample is used to train an artificial intelligence model for generating a summary outline, and the second training sample is used to train an artificial intelligence model for generating a literature summary text.

7. The method according to claim 6, characterized in that, The method further includes: The document is divided into multiple text blocks according to text order; wherein, the multiple text blocks and the multiple target text blocks constitute a third training sample; the third training sample is used to train an artificial intelligence model to identify and extract text blocks that are unrelated to the full text summary from the multiple text blocks.

8. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by a processor, causes the processor to implement the method as described in any one of claims 1 to 7.

9. A computer device, characterized in that, The computer device includes a memory and a processor, the memory storing at least one computer program, the at least one computer program being loaded and executed by the processor to implement the method as described in any one of claims 1 to 7.

10. A computer program product, characterized in that, Includes computer instructions that, when executed by a processor, implement the method as described in any one of claims 1 to 7.

Images