Automatic related work generation system based on multi-agent framework

Abstract

The invention discloses an automatic related work generation system based on a multi-agent framework. The automatic related work generation system comprises a selector, a reader and a writing device, wherein the selector is a first agent in a multi-agent framework and is responsible for determining a next literature part needing to be read; the reader is a second agent in the multi-agent framework and is responsible for reading the literature part indicated by the selector and updating the read content into the working memory; and the writer is the last agent in the multi-agent framework and is used for generating a complete related working part based on the final working memory after the selector and the reader finish document content processing. According to the multi-agent framework, related working parts with reasonable structures and logic coherence can be effectively generated, and the defects of literature processing and relation modeling in the prior art are overcome. According to the technical scheme, a systematic and efficient solution is provided, and high-quality generation of related work is ensured.

Description

Technical Field

[0001] The present invention relates to the technical fields of natural language processing (NLP) and artificial intelligence (AI), and particularly to an automated related work generation system based on a multi-agent framework. Background Art

[0002] With the continuous increase in the number of academic publications, researchers are faced with a large number of documents that need to be read and sorted for writing the related work section of a paper. Automatically generating related work can greatly save the time and effort of researchers and improve the writing efficiency of papers.

[0003] In academic writing, the related work section is used to summarize and analyze the research results in the field, and to demonstrate the innovation and frontier of the current research. However, writing a high-quality related work section requires in-depth understanding and comparison of a large amount of literature content and reasonable organization of their relationships. The traditional manual writing process is time-consuming and laborious, especially when dealing with multiple documents, it is very difficult to manually organize and compare their relationships. Therefore, how to generate the related work section efficiently and accurately has become an important challenge in the current research field.

[0004] The existing automatic related work generation methods mainly include two types: extraction-based and generation-based.

[0005] The extraction-based method extracts key sentences or paragraphs from the cited documents and stitches them together to form the related work section. This method is usually relatively straightforward and relies on selecting the most representative information from the documents to construct the related work. However, the sentences obtained by the extraction-based method are all pre-written in the cited documents, and it is impossible to deeply understand the internal relationships of the literature content and lacks a systematic analysis of the relationships between the documents.

[0006] On the other hand, the generation-based method uses generative language models (such as recurrent neural networks, Transformer models) to input the content of all reference documents and output the generated related work section. These models can understand the key information of the documents and summarize and organize them to better organize and display the logical relationships of the reference documents, and have improved in the fluency and structure of the generated text compared to the extraction-based method. In the generation-based method, some studies have also tried to better capture the relationships between reference documents by introducing graph structures, such as adding graph neural networks to the language model to model the relationships between reference documents.

[0007] The extraction-based method usually relies on extracting sentences from specific parts of the reference documents, and these sentences are all pre-written in the reference documents. Stitching these sentences together will result in the final text lacking fluency, and it is impossible to deeply understand the internal relationships of the literature content and lacks a systematic analysis of the relationships between the documents.

[0008] Although the generative method has improved in terms of the fluency and structure of the generated text compared to the extractive method, there are still some drawbacks: (1) Incomplete information capture. Existing generative language models have limitations in the input window size, which results in the model usually being able to input only a small part of the reference document, rather than being able to obtain all the content of the document. For example, some methods only input the abstract part of the reference document, or only input the introduction and conclusion parts of the document, which makes the model unable to fully obtain the information in each part of the document. Especially when it comes to detailed research methods and experimental results, it is unable to provide sufficient depth and details for the generated related work part.

[0009] (2) Lack of effective modeling of document relationships and insufficient structure in the generated content. A high-quality related work part requires a deep understanding and comparison of different reference documents. Although some studies have tried to capture the relationships between documents by introducing graph structures, these methods usually implicitly integrate the graph structure in the model architecture, lacking explicit and effective graph structure guidance and optimization strategies, and failing to effectively capture and integrate the complex relationships between documents. These methods cannot deeply reveal the logical associations between reference documents, resulting in a loose logical relationship between documents in the generated related work part, and ultimately leading to poor systematicness and logic in the generated results.

[0010] These drawbacks limit the practical application effect of the existing technologies. Especially when a large number of reference documents need to be integrated and their internal relationships need to be revealed, it is difficult for the existing methods to automatically generate a high-quality related work part. To address these problems, the present invention proposes an automatic related work generation method based on a multi-agent framework, aiming to overcome the defects of the existing technologies by optimizing the reading order of reference documents and making full use of graph structure information, so as to achieve the generation of a more accurate and coherent related work part.

[0011] In summary, the existing automatic related work generation methods have several significant drawbacks, especially in terms of document content capture and document relationship modeling. First, due to the input window size limitation of the language model, the existing methods cannot process the complete content of the reference document, resulting in a lack of depth and incomplete information organization in the generated related work part. Second, the existing methods have insufficient modeling of the complex relationships between documents and are unable to effectively capture the complex and close internal connections between documents. This leads to a lack of logical coherence in the related work part generated by the existing methods, a loose relationship between documents, and unreasonable information organization.

[0012] The information disclosed in this background art section is only intended to deepen the understanding of the overall background art of the present invention and should not be regarded as an admission or any form of implication that this information constitutes the prior art known to those skilled in the art. Summary of the Invention

[0013] The object of the present invention is to provide an automated related work generation system based on a multi-agent framework to solve the technical problems existing in the prior art.

[0014] To achieve the above object, the present invention adopts the following technical solutions: The present invention provides an automated related work generation system based on a multi-agent framework, including: a selector, a reader, and a writer; wherein, The selector is the first agent in the multi-agent framework and is responsible for determining the next part of the literature to be read; The reader is the second agent in the multi-agent framework and is responsible for reading the part of the literature indicated by the selector and updating the read content into the working memory; The writer is the last agent in the multi-agent framework, and its function is to generate a complete related work part based on the final working memory after the selector and the reader complete the processing of the literature content.

[0015] Furthermore, the selector makes a selection based on the following information: 1) Literature abstracts: the abstracts of all cited literatures and the abstract of the literature currently being written; 2) Read literature content: i.e., the information in the working memory; 3) Graph structure information: including citation graphs and co-occurrence graphs, which are used to capture the citation relationships and co-occurrence relationships between literatures.

[0016] Furthermore, the workflow of the selector is as follows: Step 1: Receive the abstracts of all literatures, the current working memory, and the graph structure between literatures; Step 2: Based on the graph structure information, select a certain chapter of a certain literature for reading; Step 3: Feed back the selected literature and chapter information to the reader; When the selector decides not to continue reading more literatures, it outputs a "termination" signal <end>, indicating the end of the iterative process and entering the writing stage.

[0017] Furthermore, the citation graph contains the citation relationships between all cited documents. Nodes represent documents, and edges represent the citation relationships between documents. The construction of the citation graph is used to help the selector jump between documents and identify which documents are most closely related. The co-occurrence graph is used to represent the co-occurrence relationships between documents, especially the contexts co-cited in the documents. Edges represent that two documents co-occur in the same sentence of other documents, and nodes represent documents. By capturing the co-occurrence relationships between documents, the co-occurrence graph helps the selector better select the reading order of documents and ensure that the relevant work section covers the close connections between documents.

[0018] Furthermore, the working process of the reader is as follows: Step 1: Receive the document part from the selector, including specific chapters or paragraphs of the document. Step 2: Extract key information from the document and update the working memory. The updated working memory needs to eliminate irrelevant information to ensure that only the information closely related to generating the relevant work section is retained. Step 3: Feed the updated working memory back to the selector and the writer for reference in subsequent operations.

[0019] Furthermore, the working process of the writer is as follows: Step 1: Receive the final working memory and obtain the complete document content information. Step 2: Organize the information according to the generation rules to avoid describing each document in isolation. The writer should ensure that the relationships between documents are clearly presented in the generated relevant work section, highlighting the innovation points. Step 3: Generate the relevant work section and ensure logical coherence and smooth language.

[0020] Furthermore, the selector, reader, and writer share a working memory to ensure that each agent can obtain the current information during the processing and adjust subsequent operations according to the previous content. The working memory is used to store the content of the read documents and is updated throughout the iterative process. To process long documents, the capacity of the working memory is limited to a fixed size to avoid overflow. During each iteration, the reader and the selector update and remove irrelevant information in the working memory to ensure the accuracy and conciseness of the information in the working memory.

[0021] Furthermore, when generating the relevant work section, the writer follows a series of rules to ensure that the generated text complies with academic norms and meets the requirements of the target task, specifically as follows: Avoid isolated descriptions: The writer needs to avoid describing each document in isolation one by one, but to organically organize the relationships between the documents to form a well-structured related work; Emphasize relationships between the literature: Make sure that the connections between the literature are highlighted in the text, including similarities, differences, and their impact on the current study.

[0022] By adopting the above technical solution, the present invention has the following beneficial effects: 1. Comprehensively process the content of the literature. By utilizing the complete content of the literature instead of just relying on a small part of the literature (such as the abstract, introduction, etc.), the information richness and accuracy of the generated related work section are improved, and errors caused by insufficient information are avoided.

[0023] 2. Explicitly modeling the relationship between documents. By introducing citation graphs and co-occurrence graphs, the present invention explicitly models the complex relationships between documents, overcoming the problem of insufficient relationship capture caused by implicit modeling graph structures in existing methods, thereby improving the coherence and logic of the generated results.

[0024] 3. Optimize the generation process. The multi-agent framework optimizes the reading order of the literature, avoiding the problems of loose logic and improper information organization in the traditional method, and ensuring that the generated related work parts are well structured and well organized. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0026] Figure 1 A system flow chart of an automated related work generation system based on a multi-agent framework provided by an embodiment of the present invention; Figure 2 A schematic diagram of a graph-guided selector and two graph structures provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0027] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0028] The following will describe in detail the specific embodiments of the present invention with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only for the purpose of illustrating and explaining the present invention, and are not intended to limit the present invention.

[0029] For the convenience of understanding this application, the following explains the technical terms involved in this application: Related Work: In academic papers, it refers to the part where the author summarizes and discusses the previously published work related to the current research.

[0030] Multi-agent framework: A system composed of multiple cooperating agents, where each agent undertakes different tasks, and is usually used for task allocation and execution in complex systems. In the field of natural language processing, agents are generally served by advanced large language models.

[0031] Citation graph: A graph structure used to represent the citation relationship between documents.

[0032] Co-occurrence graph: A graph structure used to represent the co-occurrence relationship between documents, usually based on the context of co-citation of documents.

[0033] Combined Figure 1 As shown, this embodiment provides an automated related work generation system based on a multi-agent framework, which includes three core agents: a Selector, a Reader, and a Writer. They cooperate to generate the related work part of an academic paper. Each agent undertakes different functions and gradually completes the processing of documents and the final content generation in an iterative manner, as shown in the flow chart. All three agents share a Working Memory to ensure that each agent can obtain the current information during the processing and adjust subsequent operations based on the previous content.

[0034] Selector: The Selector is the first agent in the multi-agent framework and is responsible for deciding the next part of the document to be read. The Selector makes a selection based on the following information: Document abstracts: The abstracts of all cited documents and the abstract of the document currently being written.

[0035] Read document content: That is, the information in the Working Memory.

[0036] Graph structure information: Includes a Citation Graph and a Co-occurrence Graph, which are used to capture the citation relationship and co-occurrence relationship between documents.

[0037] The main function of the selector is to optimize the reading order of the literature by analyzing this information. The selector can select the next part of the literature to be read based on the information of the graph structure and the relevance of the literature. The information of the graph structure restricts the selector to be able to intelligently jump to other literatures closely related to the current literature, avoiding the reading of irrelevant parts and improving the overall efficiency.

[0038] Workflow of the selector: Step 1: Receive the abstracts of all literatures, the current working memory, and the graph structure between the literatures.

[0039] Step 2: Based on the graph structure information (citation graph, co-occurrence graph), select a certain chapter of a certain literature for reading.

[0040] Step 3: Feed back the selected literature and chapter information to the reader.

[0041] When the selector decides not to continue reading more literatures, it will output a "termination" signal <end>, indicating the end of the iterative process and entering the writing stage.

[0042] Graph Structure: Citation Graph and Co-occurrence Graph. An important innovation in the present invention is to optimize the reading order of documents by explicitly modeling the citation relationship and co-occurrence relationship between documents. Specifically, the present invention designs two types of graphs, as Figure 2 shown: Citation Graph: This graph contains the citation relationships between all cited documents. Nodes represent documents, and edges represent the citation relationships between documents. The construction of the citation graph can help the selector jump between documents and identify which documents are most closely related.

[0043] Co-occurrence Graph: This graph is used to represent the co-occurrence relationship between documents, especially the context co-cited in documents. Edges represent that two documents co-occur in the same sentence of other documents, and nodes represent documents. The co-occurrence graph helps the selector better select the document reading order by capturing the co-occurrence relationship between documents, ensuring that the relevant work part covers the close connection between documents.

[0044] Under the limitation of the graph structure, the selector selects the next part of the document to be read. As Figure 2 (a) shows, the selector first selects the initial paper. At each step, the selector is located at a certain paper node on the graph. At this time, the selector can only choose to continue reading the current paper or jump to an adjacent paper on the graph. This limits the selection range of the selector by the graph structure, thus optimizing the reading order of documents.

[0045] The introduction of these two graph structures enables the present invention to effectively capture and organize the complex relationships between documents, avoiding the insufficient handling of document relationships in existing methods.

[0046] Reader: The reader is the second agent in the multi-agent framework, responsible for reading the part of the document indicated by the selector and updating the read content into the working memory. After each reading, the reader extracts key information from the document and updates the working memory. To process long documents, the reader also performs content screening and updating according to the set maximum working memory size to ensure that the memory limit is not exceeded during the processing.

[0047] Workflow of the Reader: Step 1: Receive the part of the document (including the specific chapter or paragraph of the document) from the selector.

[0048] Step 2: Extract key information from the document and update the working memory. The updated working memory needs to eliminate irrelevant information to ensure that only the information closely related to generating the relevant work part is retained.

[0049] Step 3: Feed the updated working memory back to the selector and the writer for reference in subsequent operations.

[0050] The working memory is designed with an important feature: its content is optimized and updated in each iteration to ensure rich information without excessive redundancy.

[0051] Writer: The writer is the last agent in the multi-agent framework. Its role is to generate the complete relevant work part based on the final working memory after the selector and the reader have completed the processing of the literature content. The writer integrates the information of each literature part to generate a well-organized and logically coherent relevant work. During the generation process, the writer follows certain rules to ensure that the generated content conforms to academic writing norms and can accurately reflect the relationships between different literatures.

[0052] Workflow of the writer: Step 1: Receive the final working memory and obtain the complete literature content information.

[0053] Step 2: Organize the information according to the generation rules to avoid describing each literature in isolation. The writer should ensure that the relationships between literatures are clearly presented in the generated relevant work part, highlighting the innovation points.

[0054] Step 3: Generate the relevant work part and ensure logical coherence and smooth language.

[0055] The output of the writer is the final relevant work part, which not only contains a summary of the literature but also clearly shows the connections between the literatures and is reasonably organized according to the similarities and differences of the literatures.

[0056] Working memory management: Due to the complexity of this generation task and the large volume of literature, the present invention introduces the concept of working memory (Working Memory). The working memory is used to store the read literature content and is updated throughout the iteration process. To handle long literatures, the capacity of the working memory is limited to a fixed size (such as 4096 lengths) to avoid overflow. In each iteration, the reader and the selector update and remove irrelevant information in the working memory to ensure the accuracy and conciseness of the information in the working memory.

[0057] Generation rules and optimization strategies: When generating the relevant work part, the writer follows a series of rules to ensure that the generated text complies with academic norms and meets the requirements of the target task: Avoid isolated description: The writer needs to avoid describing each literature in isolation one by one, but rather to organically organize the relationships between the literatures to form a reasonably structured relevant work.

[0058] Emphasize the relationships between the literatures: Ensure that the connections between the literatures are highlighted in the text, such as similarities, differences, and their impacts on the current research.

[0059] Through the above steps, the multi-agent framework of the present invention can effectively generate a relevant work part with reasonable structure and logical coherence, making up for the deficiencies of the prior art in literature processing and relationship modeling. This technical solution provides a systematic and efficient solution, ensuring the high-quality generation of the relevant work.

[0060] Compared with the prior art, the innovation points of the present invention mainly lie in: Design of the multi-agent framework: A multi-agent collaborative working mechanism based on a selector, a reader, and a writer is proposed. Through iterative information extraction and update, the efficiency and quality of generating the relevant work part are greatly improved.

[0061] Optimized reading order of the graph structure: By referring to the citation graph and the co-occurrence graph, the relationships between the literatures are explicitly modeled, optimizing the reading order of the literatures and ensuring that the relationships between the literatures can be effectively captured and presented.

[0062] Efficient memory management and content generation: A limited working memory (with a length of 4096) is adopted, and the most relevant information is maintained through dynamic update and deduplication, ensuring that the generated relevant work part is rich in information and clear in structure.

[0063] Currently, the multi-agent framework of the present application is the optimal solution to achieve the purpose of the present invention. Although other methods such as the single-selector-based solution or other forms of graph structures (such as knowledge graphs) can also be used for literature relationship modeling, these solutions often cannot achieve such high efficiency in information extraction, literature relationship capture, and result generation as the present invention. Therefore, the multi-agent framework and the graph structure optimization method of the present invention constitute the core innovation points of the present invention.

[0064] Finally, it should be noted that: The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: They can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.< / end> < / end>

Claims

1. An automated related work generation system based on a multi-agent framework, characterized in that, including: a selector, a reader, and a writer; among which, the selector is the first agent in the multi-agent framework, responsible for deciding the next part of the literature to be read; the reader is the second agent in the multi-agent framework, responsible for reading the part of the literature indicated by the selector and updating the content read into the working memory; the writer is the last agent in the multi-agent framework, and its role is to generate a complete relevant work part based on the final working memory after the selector and the reader have completed the processing of the literature content.

2. The automated related work generation system based on a multi-agent framework according to claim 1, wherein The selector makes a selection based on the following information: 1) Literature abstracts: the abstracts of all cited literatures and the abstract of the literature currently being written; 2) Read literature content: that is, the information in the working memory; 3) Graph structure information: including citation graphs and co-occurrence graphs, used to capture the citation relationships and co-occurrence relationships between literatures.

3. The automated related work generation system based on a multi-agent framework according to claim 2, characterized in that, The workflow of the selector is as follows: Step 1: Receive the abstracts of all literatures, the current working memory, and the graph structure between literatures; Step 2: Based on the graph structure information, select a certain chapter of a certain literature for reading; Step 3: Feed back the selected literature and chapter information to the reader; When the selector decides not to continue reading more documents, it outputs a "termination" signal <end>, indicating the end of the iterative process and entering the writing stage. < / end> 4. The automated related work generation system based on a multi-agent framework according to claim 2, wherein The citation graph contains the citation relationships between all cited literatures, where nodes represent literatures and edges represent the citation relationships between literatures; the construction of the citation graph is used to help the selector jump between literatures and identify which literatures have the closest relationships; The co-occurrence graph is used to represent the co-occurrence relationships between literatures, especially the contexts co-cited in the literatures; edges represent that two literatures co-occur in the same sentence of other literatures, and nodes represent literatures; The co-occurrence graph helps the selector better select the literature reading order by capturing the co-occurrence relationships between literatures, ensuring that the relevant work part covers the close connections between literatures.

5. The automated related work generation system based on a multi-agent framework according to claim 1, characterized in that, The workflow of the reader is as follows: Step 1: Receive the part of the literature from the selector, including the specific chapter or paragraph of the literature; Step 2: Extract key information from the literature and update the working memory; the updated working memory needs to eliminate irrelevant information to ensure that only the information closely related to generating the relevant work part is retained; Step 3: Feed back the updated working memory to the selector and the writer for reference in subsequent operations.

6. The automated related work generation system based on a multi-agent framework according to claim 1, wherein The workflow of the writer is as follows: Step 1: Receive the final working memory and obtain the complete literature content information; Step 2: Organize the information according to the generation rules to avoid describing each literature in isolation;; the writer should ensure that the relationships between literatures are clearly presented in the generated relevant work part, highlighting the innovation points; Step 3: Generate the relevant work part and ensure logical coherence and smooth language.

7. The automated related work generation system based on a multi-agent framework according to claim 1, wherein The selector, reader, and writer share a working memory to ensure that each agent can access the current information during processing and adjust subsequent operations based on previous content. The working memory is used to store the content of the read literature and is updated throughout the iteration process. To handle long literature, the capacity of the working memory is limited to a fixed size to avoid overflow. During each iteration, the reader and selector update and remove irrelevant information from the working memory to ensure the information in the working memory is accurate and concise.

8. The automated related work generation system based on a multi-agent framework according to claim 1, characterized in that When generating the relevant work section, the writer follows a series of rules to ensure that the generated text complies with academic norms and meets the requirements of the target task, as follows: Avoid isolated descriptions: The writer needs to avoid describing each literature piece by piece in isolation, but rather organically organize the relationships between the literatures to form a well-structured relevant work. Highlight the relationships between literatures: Ensure that the connections between literatures, including similarities, differences, and their impacts on the current research, are emphasized in the text.

Images