Intelligent research and report writing method based on multi-agent cooperative work

Abstract

The invention provides an intelligent research report writing method based on multi-agent cooperative work, and the method is characterized in that the method comprises the following steps: S1, carrying out the preprocessing of a plurality of existing research reports, and obtaining a plurality of preprocessed research report samples; s2, training a text classification model according to the preprocessed research and report sample in combination with manual annotation to obtain a trained text classification model; s3, extracting a paragraph representing a causal relationship from each preprocessed research report sample through a trained text classification model; s4, summarizing all paragraphs through a large model to generate causal knowledge; s5, constructing a training data set according to the causal knowledge, and training the large language model through the training data set to obtain a plurality of different agents; and S6, generating a target research report according to all the agents in combination with the industry field and the research report type. In short, the method can efficiently generate research reports.

Description

Technical Field

[0001] The present invention relates to an intelligent research report writing method, and in particular to an intelligent research report writing method based on the collaborative work of multiple intelligent agents. Background Art

[0002] Research reports, or research reports, are of great value in all fields. For example, in the financial field, research reports are key communication media and are of great value to investors and financial institutions. They not only provide the latest news on specific markets, industries and companies, but also conduct in-depth analysis of this information, thus becoming an important basis for financial decision-making. Especially in today's era of rapid information development, the role of research reports is even more important. Through in-depth analysis of market trends, investment opportunities and potential risks, research reports help readers make more informed and informed investment decisions. However, the traditional research report writing process faces many challenges: due to the rapid changes in the market, manually written research reports often find it difficult to keep up with the latest market dynamics; choosing the right analytical method to interpret the data and draw conclusions is a very challenging task, which requires the writer to have deep professional knowledge and rich experience; in addition, the writing of research reports also involves a lot of data collection and fact verification work, which is not only time-consuming but also prone to errors.

[0003] Therefore, how to efficiently generate timely, multi-angle and accurate research reports is an urgent problem that needs to be solved. Summary of the invention

[0004] The present invention is made to solve the above-mentioned problems, and its purpose is to provide an intelligent research report writing method based on the collaborative work of multiple agents.

[0005] The present invention provides an intelligent research report writing method based on the collaborative work of multiple agents, which is used to generate a target research report according to industry fields and research report types, and has the following characteristics, including the following steps: Step S1, pre-processing multiple existing research reports respectively to obtain multiple pre-processed research report samples; Step S2, training a text classification model based on the pre-processed research report samples in combination with manual annotation to obtain a trained text classification model; Step S3, extracting paragraphs representing causal relationships from each pre-processed research report sample through the trained text classification model; Step S4, summarizing all paragraphs through a large model to generate causal knowledge; Step S5, constructing a training data set based on causal knowledge, and training a large language model through the training data set to obtain multiple different agents; Step S6, generating a target research report based on all agents in combination with industry fields and research report types.

[0006] In the intelligent research report writing method based on multi-agent collaborative work provided by the present invention, it can also have the following characteristics: wherein, multiple different agents include a chief analyst agent, a researcher agent, an assistant researcher agent, an intern agent, a reviewer agent and a summarizer agent, and step S6 includes the following sub-steps: step S6-1, according to the industry field and the research report type, the chief analyst agent and the intern agent perform research report design to generate a research report outline; step S6-2, according to the research report outline, the researcher agent, the assistant researcher agent and the intern agent perform research report writing to generate an initial research report; step S6-3, according to the initial research report, the chief analyst agent and the reviewer agent perform research report review to generate a verification research report; step S6-4, according to the verification research report, the summarizer agent performs research report induction to obtain a target research report.

[0007] The intelligent research report writing method based on multi-agent collaborative work provided by the present invention may also have the following features: wherein, in step S6-1, the intern agent is used to collect industry news and the latest announcements of related companies as industry information data according to the industry field and the research report type,

[0008] The chief analyst agent is used to generate a research report outline based on the industry field, research report type and industry information data, combined with a preset standardized template. In step S6-2, the researcher agent is used to generate the core content data of the research report based on the research report outline. The assistant researcher agent is used to collect existing data and generate corresponding supporting data based on the core content data of the research report. The intern agent is used to graph the complex data of the core content data of the research report and generate corresponding chart data. The initial research report includes the research report outline, the core content data of the research report, supporting data and chart data. In step S6-3, the chief analyst agent is used to review the logic and accuracy of the initial research report. The reviewer agent is used to review the grammar and calculation errors of the initial research report. In step S6-4, the summarizer agent is used to extract key information and main points from the verified research report, and generate the target research report by combining the key information, main points and verified research report.

[0009] The intelligent research report writing method based on multi-agent collaborative work provided by the present invention may also have the following features: wherein, step S1 includes the following sub-steps: step S1-1, language screening of all existing research reports, and screening out existing research reports in a specified language; step S1-2, converting the existing research reports in each specified language into a text in a specific format; step S1-3, performing paragraph-level deduplication processing on the text in each specific format, and generating a corresponding preprocessed research report sample.

[0010] The intelligent research report writing method based on multi-agent collaborative work provided by the present invention may also have the following feature: wherein, in step S1-1, language screening is performed through a rule-based method combined with the Fasttext tool.

[0011] The intelligent research report writing method based on multi-agent collaborative work provided by the present invention may also have the following features: wherein, the format of the existing research report is pdf, and in step S1-2, the existing research report in the specified language is parsed by an OCR parsing tool, and the existing research report in the specified language is converted into a plain text txt format.

[0012] The intelligent research report writing method based on multi-agent collaborative work provided by the present invention may also have the following feature: wherein, in step S1-3, deduplication processing is performed using the MiniHash algorithm.

[0013] In the intelligent research report writing method based on multi-agent collaborative work provided by the present invention, it can also have the following characteristics: wherein, step S2 includes the following sub-steps: step S2-1, selecting multiple sentences in the pre-processed research report sample and manually annotating them to obtain classification training samples; step S2-2, training the text classification model through all classification training samples to obtain an optimized text classification model; step S2-3, judging whether the training termination condition is met, if so, obtaining a trained text classification model, if not, executing step S2-4; step S2-4, inputting the unannotated pre-processed research report sample into the optimized text classification model to obtain annotated paragraph-level data; step S2-5, subjecting the annotated paragraph-level data to manual review as a new classification training sample, and executing step S2-2.

[0014] In the intelligent research report writing method based on multi-agent collaborative work provided by the present invention, it can also have the following characteristics: wherein the text classification model is the BGE-TextCNN model, and the large model is the Baichuan2-53B model.

[0015] The intelligent research report writing method based on multi-agent collaborative work provided by the present invention may also have the following characteristics: a plurality of different agents are obtained by role-playing the same large language model, or by fine-tuning different professional knowledge of each large language model.

[0016] Functions and Effects of the Invention

[0017] According to the intelligent research report writing method based on the collaborative work of multiple agents involved in the present invention, on the one hand, causal knowledge is extracted from a large number of existing research reports through text classification models and large models, and the causal knowledge is used to train the large language model, so that the trained large language model can better understand the dynamic changes of the industry and make more accurate predictions; on the other hand, through multiple agents based on the trained large language model, multiple agents are used to collaboratively complete the target research report, avoiding the problem of semantic incoherence caused by a single large language model being limited by the length of the context, and being able to write the entire research report at one time, and being able to write the research report from multiple angles, and screen effective data and draw charts. Therefore, the intelligent research report writing method based on the collaborative work of multiple agents of the present invention can efficiently generate research reports. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a flowchart of an intelligent research report writing method based on multi-agent collaborative work in an embodiment of the present invention;

[0019] Figure 2 is a schematic diagram of a flow chart of training a text classification model in an embodiment of the present invention;

[0020] Figure 3 It is a schematic diagram of multi-agent generation of target research reports in an embodiment of the present invention. DETAILED DESCRIPTION

[0021] In order to make the technical means, creative features, objectives and effects achieved by the present invention easy to understand, the following embodiments and the accompanying drawings specifically illustrate the intelligent research report writing method based on the collaborative work of multiple intelligent agents of the present invention.

[0022] This embodiment provides an intelligent research report writing method based on multi-agent collaborative work, which is used to generate a target research report according to the industry field and the research report type. In this embodiment, the target research report is a financial research report, and the industry fields involved are all walks of life.

[0023] Figure 1 It is a flow chart of an intelligent research report writing method based on multi-agent collaborative work in an embodiment of the present invention.

[0024] like Figure 1 As shown in the figure, the intelligent research report writing method based on multi-agent collaborative work includes the following steps:

[0025] Step S1, preprocessing multiple existing research reports respectively to obtain multiple preprocessed research report samples.

[0026] Step S1 includes the following sub-steps:

[0027] Step S1-1, language screening is performed on all existing research reports to obtain existing research reports in a specified language.

[0028] In step S1-1, language screening is performed by combining a rule-based method with the Fasttext tool. In this embodiment, the specified language is Chinese.

[0029] Step S1-2, converting existing research reports in each designated language into texts in a specific format.

[0030] The existing research report is in the format of pdf. In step S1-2, the existing research report in the specified language is parsed by an OCR parsing tool, and the existing research report in the specified language is converted into a plain text txt format.

[0031] Step S1-3, perform paragraph-level deduplication processing on the texts in each specific format to generate corresponding pre-processed research report samples.

[0032] Among them, in step S1-3, deduplication processing is performed using the MiniHash algorithm.

[0033] Step S2: training a text classification model based on the preprocessed research report samples and manual annotation to obtain a trained text classification model, wherein the text classification model is a BGE-TextCNN model.

[0034] Figure 2 It is a schematic diagram of the process of training a text classification model in an embodiment of the present invention.

[0035] like Figure 2 As shown, step S2 includes the following sub-steps:

[0036] Step S2-1, select multiple sentences in the pre-processed research report sample and manually annotate them to obtain classification training samples. In this embodiment, the sentences in the research report are divided into 11 categories by manual annotation, such as risk warning, company and personal information, causal relationship, etc.

[0037] Step S2-2, training the text classification model using all classification training samples to obtain an optimized text classification model.

[0038] Step S2-3, determine whether the training termination condition is met, if so, obtain the trained text classification model, if not, execute step S2-4.

[0039] Step S2-4, input the unlabeled preprocessed research report sample into the optimized text classification model to obtain labeled paragraph-level data.

[0040] Step S2-5: The annotated paragraph-level data is manually reviewed as new classification training samples, and step S2-2 is executed.

[0041] Step S3, extracting paragraphs representing causal relationships from each preprocessed research report sample using the trained text classification model.

[0042] Step S4, summarize all paragraphs through the large model to generate causal knowledge. In this embodiment, the model is selected as the large model based on the compression rate of the research report by the model's Tokenizer and the real test of this scenario. That is, considering the model's ability to follow instructions and its ability to summarize causal knowledge, the large model of this embodiment is the Baichuan2-53B model.

[0043] Step S5, constructing a training data set based on causal knowledge, and training the large language model through the training data set to obtain multiple different intelligent agents.

[0044] Step S6, generating a target research report based on all agents, combined with industry fields and research report types. Among them, multiple different agents include chief analyst agent, researcher agent, assistant researcher agent, intern agent, reviewer agent and summarizer agent. Multiple different agents are obtained by role-playing the same large language model, or by fine-tuning different professional knowledge of each large language model.

[0045] Figure 3 It is a schematic diagram of multi-agent generation of target research reports in an embodiment of the present invention.

[0046] like Figure 3 As shown, step S6 includes the following sub-steps:

[0047] Step S6-1, i.e. Design, designs the research report based on the industry field and the type of research report through the chief analyst agent and the intern agent to generate a research report outline.

[0048] Among them, the intern agent is used to collect industry news and the latest announcements of related companies as industry information data according to the industry field and research report type. The chief analyst agent is used to generate a research report outline based on the industry field, research report type and industry information data, combined with the preset standardized template.

[0049] Step S6-2, namely Compile, generates an initial research report by compiling the research report according to the research report outline through the researcher agent, assistant researcher agent and intern agent. In Compile, the researcher agent, assistant researcher agent and intern agent collaborate and interact with each other to generate an initial research report.

[0050] Among them, the researcher agent is used to generate the core content data of the research report according to the research report outline. The assistant researcher agent is used to collect existing data and generate corresponding supporting data based on the core content data of the research report. The intern agent is used to graph the complex data of the core content data of the research report and generate corresponding chart data. The initial research report includes the research report outline, the core content data of the research report, the supporting data and the chart data.

[0051] In this example, the intern agent uses various tools such as Wind volume and price data, stock exchange announcement query system, and graph database query capabilities to search for data. The assistant researcher agent uses various efficient tools such as calculators and MarkDown table generators to present complex data in a graphical way.

[0052] Step S6-3, namely Examination, is to generate a verification report by conducting a research report review through the chief analyst agent and the reviewer agent based on the initial research report. In the Examination, the chief analyst agent and the reviewer agent collaborate and exchange data with each other to generate a verification report.

[0053] Among them, the chief analyst agent is used to review the logic and accuracy of the initial research report, and the reviewer agent is used to review the grammar and calculation errors of the initial research report.

[0054] Step S6-4, namely Summary, summarizes the research report based on the verified research report and obtains the target research report.

[0055] Among them, the summarizer agent is used to extract key information and main ideas from the verification research report, and generate the target research report by combining the key information, main ideas and verification research report.

[0056] Functions and Effects of the Embodiments

[0057] According to the intelligent research report writing method based on multi-agent collaboration involved in this embodiment, on the one hand, causal knowledge is extracted from a large number of existing research reports through text classification models and large models, and the large language model is trained with causal knowledge, so that the trained large language model can better understand the dynamic changes of the industry and make more accurate predictions; on the other hand, through multiple agents based on the trained large language model, multiple agents are used to collaborate to complete the target research report, avoiding the problem of semantic incoherence caused by a single large language model being limited by the length of the context, and being able to write the research report from multiple angles, and screen effective data and draw charts. In short, this method can efficiently generate research reports.

[0058] Those skilled in the art should understand that the present invention is not limited to the above embodiments, and the above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, and these changes and improvements fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. An intelligent research report writing method based on multi-agent collaborative work, used to generate target research reports according to industry fields and research report types, characterized in that: The following steps are involved: Step S1, preprocessing a plurality of existing research reports respectively to obtain a plurality of preprocessed research report samples; Step S2, training a text classification model based on the pre-processed research report samples and combining manual annotation to obtain a trained text classification model; Step S3, extracting paragraphs representing causal relationships from each of the pre-processed research report samples using the trained text classification model; Step S4, summarizing all the paragraphs through the large model to generate causal knowledge; Step S5, constructing a training data set according to the causal knowledge, and training a large language model using the training data set to obtain a plurality of different intelligent agents; Step S6, generating the target research report based on all the intelligent agents, combining the industry field and the research report type.

2. According to claim 1, the intelligent research report writing method based on multi-agent collaborative work, Features: The multiple different agents include a chief analyst agent, a researcher agent, an assistant researcher agent, an intern agent, a reviewer agent, and a summarizer agent. The step S6 comprises the following sub-steps: Step S6-1, according to the industry field and the research report type, the chief analyst agent and the intern agent design the research report to generate a research report outline; Step S6-2, according to the research report outline, the researcher agent, the assistant researcher agent and the intern agent write the research report to generate an initial research report; Step S6-3, based on the initial research report, the chief analyst agent and the reviewer agent conduct a research report review to generate a verification research report; Step S6-4, based on the verification research report, the research report is summarized by the summarizer intelligent agent to obtain the target research report.

3. The intelligent research report writing method based on multi-agent collaborative work according to claim 2 is characterized by: in, In step S6-1, the intern agent is used to collect industry news and the latest announcements of related companies as industry information data according to the industry field and the research report type. The chief analyst agent is used to generate the research report outline according to the industry field, the research report type and the industry information data in combination with a preset standardized template. In step S6-2, the researcher agent is used to generate the core content data of the research report according to the research report outline. The assistant researcher agent is used to collect existing data and generate corresponding supporting data based on the core content data of the research report. The intern agent is used to graph the complex data of the core content data of the research report and generate corresponding chart data. The initial research report includes the research report outline, the core content data of the research report, the supporting data and the chart data. In step S6-3, the chief analyst agent is used to review the logic and accuracy of the initial research report. The reviewer agent is used to perform grammar review and calculation error review on the initial research report. In step S6-4, the summarizer agent is used to extract key information and main viewpoints from the verification research report, and generate the target research report by combining the key information, the main viewpoints and the verification research report.

4. The intelligent research report writing method based on multi-agent collaborative work according to claim 1 is characterized by: in, The step S1 comprises the following sub-steps: Step S1-1, performing language screening on all the existing research reports to obtain existing research reports in a specified language; Step S1-2, converting each existing research report in the specified language into a text in a specific format; Step S1-3, performing paragraph-level deduplication processing on each text in the specific format to generate the corresponding pre-processed research report sample.

5. The intelligent research report writing method based on multi-agent collaborative work according to claim 4 is characterized by: in, In the step S1 - 1 , the language screening is performed by using a rule-based method in combination with the Fasttext tool.

6. The intelligent research report writing method based on multi-agent collaborative work according to claim 4 is characterized by: in, The format of the existing research report is pdf. In the step S1-2, the existing research report in the specified language is parsed by an OCR parsing tool, and the existing research report in the specified language is converted into a plain text txt format.

7. The intelligent research report writing method based on multi-agent collaborative work according to claim 4 is characterized by: in, In the step S1-3, the deduplication process is performed using the MiniHash algorithm.

8. The intelligent research report writing method based on multi-agent collaborative work according to claim 1 is characterized by: in, The step S2 comprises the following sub-steps: Step S2-1, selecting multiple sentences from the pre-processed research report sample and manually annotating them to obtain classification training samples; Step S2-2, training the text classification model using all the classification training samples to obtain an optimized text classification model; Step S2-3, determining whether the training termination condition is met, if so, obtaining the trained text classification model, if not, executing step S2-4; Step S2-4, inputting the unlabeled preprocessed research report sample into the optimized text classification model to obtain labeled paragraph-level data; Step S2-5: manually review the annotated paragraph-level data as new classification training samples, and execute step S2-2.

9. The intelligent research report writing method based on multi-agent collaborative work according to claim 1 is characterized by: in, The text classification model is a BGE-TextCNN model. The large model is the Baichuan2-53B model.

10. The intelligent research report writing method based on multi-agent collaborative work according to claim 1 is characterized by: in, The multiple different agents are obtained by role-playing the same large language model, or by fine-tuning different professional knowledge on each of the large language models.

Images