A system and method for generating a duty officer document based on a large reasoning model

Abstract

The application discloses a kind of on-duty document generation system and method based on reasoning big model, the system intention acquisition module, knowledge base module, interactive generation module and style conversion module: intention acquisition module is used to obtain task requirement text, extension mark and retention mark;Knowledge base module is used to store the pre-set business knowledge text, and based on the external input text to be searched is searched, and retrieval result is obtained;Interactive generation module is used to store the pre-set expert template set, and based on task requirement text, extension mark, retention mark and retrieval result, prompt word generation and reasoning are carried out, and target on-duty document is obtained;Style conversion module is used to carry out style conversion processing on initial on-duty document based on pre-set style template text, and target on-duty document is obtained.The application can improve the richness and integrity of the generated on-duty document, and generate personalized and industry-specific on-duty document.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of text data processing, and specifically relates to a shift clerk generation system and method based on a reasoning large model. BACKGROUND

[0002] In recent years, with the rapid development of the technology field, deep learning and natural language processing (NLP) technology have also made remarkable progress. In this context, reasoning large models such as GPT-4o and Qwen 2.5 have emerged in the field of text generation, and their capabilities have improved significantly. These reasoning large models, supported by powerful algorithms and massive data, can output high-quality results in various types of text generation tasks such as news reporting, novel writing, and advertising copywriting, thereby greatly improving writing efficiency and quality. At the same time, they also have the feature of accurately understanding structured instructions and can automatically fill in the corresponding content according to pre-set templates, which undoubtedly lays a solid technical foundation for the generation of standardized documents, making reasoning large models a promising application direction in the field of document generation.

[0003] In today's society, whether it is an enterprise or various institutions, document work is an essential part of daily operations, but this part often takes up a lot of time and effort. From report writing to meeting minutes compilation, from timely release of news releases to careful writing of academic papers, each step requires high-quality text output as support. Traditional manual writing methods not only have relatively low efficiency, but are also easily affected by factors such as the personal emotions and professional knowledge levels of the writers, resulting in uneven text quality. In addition, the amount of information in today's era is growing explosively, which poses unprecedented challenges to the accuracy and timeliness of document content.

[0004] Shift daily report documents, as a type of daily document, refer to documents filled out by shift personnel during their shift to record the work done that day, event handling, important information, etc. It is usually required to be completed as soon as possible after the shift ends to ensure the timeliness and accuracy of the information. This provides the latest reference for subsequent work and is an important carrier for information transmission, work handover, and responsibility tracing.

[0005] However, for daily shift documents, a large number of frontline personnel write non-standardly, and a large number of personal daily reports are generated every day, which takes time and effort to count, and not counting can easily miss some risk factors. In this context, the document generation capability of reasoning large models is timely, like a timely rain, providing a new and promising solution to improve document work efficiency and quality, and is expected to completely transform the traditional document work mode and inject strong momentum into the efficient operation of enterprises and institutions.

[0006] Existing document generation methods based on large-scale reasoning models first extract key information from raw data, then process this information using the large-scale reasoning model to obtain semantically coherent text content. Finally, standard document templates are used to precisely fill in or logically combine the generated text content to obtain the final document. However, as large-scale reasoning models become increasingly integrated with daily business operations and the need for a deeper understanding of business needs grows, the number of daily report scenarios that can be adapted to by a limited number of template-based writing methods is gradually decreasing, failing to meet the increasingly diverse customization needs of users. Furthermore, for companies with special industry backgrounds or professional field requirements, such as healthcare, finance, and law, general template writing methods cannot fully consider the characteristics and needs of the industry, appearing rigid and failing to generate high-quality documents that meet industry standards and dynamic requirements. Summary of the Invention

[0007] The technical problem to be solved by this invention is to provide a duty roster document generation system and method based on a large reasoning model, which can generate personalized duty roster documents that conform to industry characteristics.

[0008] To address the aforementioned technical problems, the first aspect of this invention discloses a duty document generation system, the system comprising an intent acquisition module, a knowledge base module, an interactive generation module, and a style conversion module;

[0009] The intent acquisition module is data-connected to the knowledge base module and the interaction generation module, and is used to acquire task requirement text, extended flags and reserved flags;

[0010] The knowledge base module is data-connected to the interactive generation module, used to store preset business knowledge text, and to perform retrieval processing based on externally input text to obtain retrieval results; the retrieval results include a set of cited texts and a set of reconstructed texts;

[0011] The interactive generation module is data-connected to the style conversion module and is used to store a preset set of expert templates. Based on the task requirement text, the extended flag, the retained flag, and the search results, it generates and infers prompt words to obtain the target duty document. The set of expert templates includes N expert templates, where N is an integer greater than 1.

[0012] The style conversion module is used to perform style conversion processing on the initial duty document based on a preset style template text to obtain the target duty document.

[0013] As an optional implementation, in the first aspect of the present invention, the knowledge base module includes a text segmentation unit, a vector index construction unit, a citation counting unit, a similarity calculation unit, a retrieval unit, and a text recombination unit;

[0014] The text segmentation unit is data-connected to the vector index construction unit and the retrieval unit, and is used to segment the business knowledge text to obtain a text block set; the text block set includes P text blocks; P is an integer greater than 1;

[0015] The vector index construction unit is data-connected to the similarity calculation unit and the text recombination unit, and is used to process the task requirement text and the text block set to obtain a task vector and an index vector set, and to process the text to be retrieved in response to receiving external input to obtain a retrieval vector; the index vector set includes the index vector corresponding to each text block;

[0016] The citation counting unit is data-connected to the similarity calculation unit and the text recombination unit, and is used to store the citation count set and update the citation count set based on the recombined text set from the text recombination unit; the citation count set includes the citation count corresponding to each text block;

[0017] The similarity calculation unit is data-connected to the retrieval unit and is used to process the vector to be retrieved, the set of index vectors, and the set of reference counts using a similarity calculation model to obtain a similarity set; the similarity set includes P similarity values;

[0018] The retrieval unit is data-connected to the text recombination unit and is used to generate a set of cited texts and a set of non-cited texts based on the similarity set and the set of text blocks; the set of cited texts and the set of non-cited texts each include Q1 and Q2 text blocks respectively; Q1 + Q2 <= P, and Q1 and Q2 are both integers greater than 1;

[0019] The text recombination unit is used to process the cited text set, the non-cited text set, the task vector, and the retrieval vector to obtain the recombined text set.

[0020] As an optional implementation, in the first aspect of the present invention, the text recombination unit includes a relevance calculation subunit, a filtering subunit, and a merging subunit:

[0021] The relevance calculation subunit is data-connected to the filtering subunit and is used to process the non-referenced text set, the index vector set, the vector to be retrieved, and the task vector using the relevance calculation model to obtain a relevance set; the relevance set includes Q2 relevance values;

[0022] The filtering subunit is data-connected to the merging subunit and is used to sort the non-referenced text set based on the relevance set to obtain a sorted text set, and to delete the last Q3 text blocks of the sorted text set to obtain a supplementary text set; Q3 is an integer greater than 1 and less than Q2;

[0023] The merging subunit is used to calculate the union of the cited text set and the supplementary text set to obtain the recombined text set.

[0024] As an optional implementation, in the first aspect of the present invention, the expression of the correlation calculation model is:

[0025] CD j =cos β (VI j ,VC)(1-cos(VI j ,VA)) β

[0026] In the formula, CD j The j-th correlation value in the correlation set CD; VI j The index vector in the index vector set corresponds to the j-th text block in the non-referenced text set I; VA and VC are the vector to be retrieved and the task vector, respectively; β is a preset correlation coefficient; j is an integer from 1 to Q2.

[0027] As an optional implementation, in the first aspect of the present invention, the interaction generation module includes an expert template storage unit, a first prompt word generation unit, a second prompt word generation unit, a third prompt word generation unit, a fourth prompt word generation unit, a fifth prompt word generation unit, and a reasoning unit;

[0028] The expert template storage unit is used to store a preset set of expert templates;

[0029] The first prompt word generation unit is data-connected to the intent acquisition module, the knowledge base module and the reasoning unit, and is used to perform a first combination processing on the task requirement text and the reference text set to obtain the initial prompt word;

[0030] The second prompt word generation unit is data-connected to the reasoning unit and is used to perform a second combination processing on the externally input current question text, current expert template, and current text block set to obtain question expansion prompt words;

[0031] The third prompt word generation unit is data-connected to the intent acquisition module and the reasoning unit, and is used to perform a third combination processing on the task requirement text and the current expert template input externally to obtain associated question prompt words;

[0032] The fourth prompt word generation unit is data-connected to the reasoning unit and is used to perform a fourth combination processing on the externally input current question text, the current expert template, and the current text block set to obtain the current prompt word.

[0033] The fifth prompt word generation unit is data-connected to the reasoning unit and is used to perform a fifth combination processing on the externally input current answer text and the current duty document to obtain document update prompt words;

[0034] The reasoning unit is used to process the document construction prompts to obtain the corresponding answer text; the document construction prompts are the initial prompts, the question expansion prompts, the related question prompts, the current prompts, or the document update prompts.

[0035] A second aspect of this invention discloses a method for generating duty roster documents, the method comprising:

[0036] S1. Use the intent acquisition module to obtain the task requirement text;

[0037] S2. Utilize the knowledge base module to store preset business knowledge text;

[0038] S3. Using the interactive generation module and the knowledge base module, construct the current duty document based on the task requirement text;

[0039] S4. Using the intent acquisition module, the interaction generation module, and the knowledge base module, update the current duty document to obtain the updated current duty document;

[0040] S5. Repeat S4 until the user inputs an end command;

[0041] S6. Determine the initial duty document as the current duty document;

[0042] S7. Using the style conversion module, based on the preset style template text, perform style conversion processing on the initial duty document to obtain the target duty document.

[0043] As an optional implementation, in the second aspect of the present invention, the step of constructing the current duty document based on the task requirement text using the interactive generation module and the knowledge base module includes:

[0044] S31. Using the knowledge base module, the task requirement text is retrieved and processed to obtain the corresponding set of cited texts and the set of reconstructed texts;

[0045] S32. Using the first prompt word generation unit, perform a first combination process on the task requirement text and the set of referenced texts to obtain initial prompt words;

[0046] S33. Using the reasoning unit, process the initial prompt words to obtain the current duty document.

[0047] As an optional implementation, in a second aspect of the present invention, updating the current duty document using the intent acquisition module, the interaction generation module, and the knowledge base module to obtain the updated current duty document includes:

[0048] S41. Initialize the current text block set to empty; initialize the loop count l to 1; initialize the current question text to the task requirement text;

[0049] S42. Set the current expert template as the l-th expert template in the expert template set;

[0050] S43. Using the intent acquisition module and the reasoning unit, update the current question text based on the current expert template, the current text block set, and the task requirement text;

[0051] S44. Using the knowledge base module, perform retrieval processing on the current question text to obtain the corresponding set of referenced texts and the set of reconstructed texts; update the current text block set to the set of reconstructed texts;

[0052] S45. Using the reasoning unit, the fourth prompt word generation unit, and the fifth prompt word generation unit, update the current duty document based on the current expert template, the current text block set, and the current question text;

[0053] S46. Increment the value of l by 1;

[0054] S47. Repeat S42 to S46 until l is greater than N.

[0055] As an optional implementation, in a second aspect of the present invention, the step of using the intent acquisition module and the reasoning unit to update the current question text based on the current expert template, the current text block set, and the task requirement text includes:

[0056] S431. Obtain the extended flag using the intent acquisition module;

[0057] S432. Determine the extended flag;

[0058] When the extension flag is true, execute S433;

[0059] If the extension flag is not specified, execute S435;

[0060] S433. Using the second prompt word generation unit, perform a second combination process on the current question text, the current expert template, and the current text block set to obtain extended prompt words for the question.

[0061] S434. Using the reasoning unit, process the extended prompt words of the question to obtain the extended question text; update the current question text to the extended question text; execute S44;

[0062] S435. Using the third prompt word generation unit, the task requirement text and the current expert template are combined in a third way to obtain associated question prompt words;

[0063] S436. Using the reasoning unit, process the associated question prompt words to obtain associated question text; update the current question text to the associated question text.

[0064] As an optional implementation, in a second aspect of the present invention, updating the current duty document based on the current expert template, the current text block set, and the current question text using the reasoning unit, the fourth prompt word generation unit, and the fifth prompt word generation unit includes:

[0065] S451. Using the fourth prompt word generation unit, the current question text, the current expert template, and the current text block set are combined in a fourth way to obtain the current prompt word.

[0066] S452. Using the reasoning unit, the current prompt word is processed to obtain the current answer text;

[0067] S453. Using the intent acquisition module, obtain the retention flag;

[0068] S454. Determine the retention flag;

[0069] When the reservation flag is true, execute S455;

[0070] When the retention flag is not set, the current duty document remains unchanged; execute S46;

[0071] S455. Using the fifth prompt word generation unit, the current answer text and the current duty document are combined in a fifth combination process to obtain the document update prompt word;

[0072] S456. Using the reasoning unit, process the document update prompt words to obtain intermediate text; replace the content of the current duty document with the intermediate text.

[0073] Compared with the prior art, the embodiments of the present invention have the following beneficial effects:

[0074] By introducing a user feedback mechanism in multi-round dialogues, the duty roster documents can be dynamically generated and updated, which can improve the richness and completeness of the generated duty roster documents and make full use of users' deep understanding of the industry to generate personalized duty roster documents that conform to industry characteristics. Attached Figure Description

[0075] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0076] Figure 1 This is a schematic diagram of the structure of a duty document generation system disclosed in an embodiment of the present invention.

[0077] Figure 2 This is a schematic diagram of the structure of the knowledge base module of a duty document generation system disclosed in an embodiment of the present invention.

[0078] Figure 3 This is a schematic diagram of the structure of a text reorganization unit in a duty document generation system disclosed in an embodiment of the present invention.

[0079] Figure 4 This is a schematic diagram of the structure of the interactive generation module of a duty document generation system disclosed in an embodiment of the present invention.

[0080] Figure 5 This is a flowchart illustrating a method for generating duty documents according to an embodiment of the present invention. Detailed Implementation

[0081] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0082] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0083] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0084] Example 1

[0085] Please see Figures 1-4 . Figure 1 This is a schematic diagram of the structure of a duty document generation system disclosed in an embodiment of the present invention; Figure 2 This is a schematic diagram of the knowledge base module of a duty document generation system disclosed in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of a text reorganization unit in a duty document generation system disclosed in an embodiment of the present invention; Figure 4 This is a schematic diagram of the interactive generation module of a duty document generation system disclosed in an embodiment of the present invention. Figure 1 The described duty roster document generation system is applied in the field of text data processing, such as duty roster document generation; however, this embodiment of the invention is not limited to this application. Figure 1 As shown, the system includes an intent acquisition module, a knowledge base module, an interaction generation module, and a style transfer module.

[0086] The intent acquisition module is connected to the knowledge base module and the interaction generation module to acquire task requirement text, extended flags, and reserved flags.

[0087] It should be noted that the above intent acquisition module is used to acquire the task requirement text, extension flag, and retention flag input by the user, wherein: (1) The task requirement text is used to describe the user's requirements for the duty document to be generated in text form, such as "Generate a duty document, the content of which is a priority list of targets". (2) The value of the extension flag is yes or no. When the value is yes, it means that the question asked by the previous expert needs to be expanded to obtain an extended question text adapted to the current expert's identity, and the current expert is asked a question based on the extended question text; when the value is no, it means that a related question text needs to be generated based on the overall task requirement text for the current expert's identity, and the current expert is asked a question using the related question text, so as to avoid topic repetition or narrowing of content scope after multiple rounds of dialogue. (3) The value of the retention flag is yes or no. When the value is yes, it means that the user is satisfied with the content of the current expert's answer and needs to merge it into the current duty document; when the value is no, it means that the user is not satisfied with the content of the current expert's answer and does not need to change the current duty document.

[0088] It is evident that by using extension and retention flags, allowing users to decide whether to expand topics and retain content, a user feedback mechanism can be introduced into the generation process of duty documents, increasing user participation. This also avoids topic repetition or narrowing after multiple rounds of dialogue, improving the richness and completeness of the generated duty documents. Furthermore, it can fully leverage users' deep understanding of the industry to generate personalized duty documents that conform to industry characteristics.

[0089] The aforementioned knowledge base module is data-connected to the aforementioned interactive generation module. It is used to store preset business knowledge text and perform retrieval processing based on externally input text to obtain retrieval results. The aforementioned retrieval results include a set of cited texts and a set of reconstructed texts.

[0090] It should be noted that the aforementioned business knowledge text includes industry-specific business knowledge, such as descriptive texts of specific concepts, daily production data, and business logic, which are not limited in this embodiment of the invention.

[0091] The aforementioned interactive generation module is data-connected to the aforementioned style conversion module. It is used to store a preset set of expert templates and to generate and reason about prompt words based on the aforementioned task requirement text, the aforementioned extended flags, the aforementioned reserved flags, and the aforementioned search results to obtain the target duty document. The aforementioned set of expert templates includes N expert templates, where N is an integer greater than 1.

[0092] It should be noted that each expert template is used to describe the expert role in a specific subfield of the profession in text form. For example, the expert template for a weather expert could be in the following form:

[0093]

[0094] The aforementioned style conversion module is used to perform style conversion processing on the initial duty document based on a preset style template text, so as to obtain the target duty document.

[0095] It should be noted that the above-mentioned style template text can be historical duty documents or other reference documents, and this embodiment of the invention does not limit it.

[0096] In an optional embodiment, such as Figure 2 As shown, the knowledge base module includes a text segmentation unit, a vector index construction unit, a citation counting unit, a similarity calculation unit, a retrieval unit, and a text recombination unit.

[0097] The aforementioned text segmentation unit is data-connected to the aforementioned vector index construction unit and the aforementioned retrieval unit, and is used to segment the aforementioned business knowledge text to obtain a text block set; the aforementioned text block set includes P text blocks; P is an integer greater than 1.

[0098] It should be noted that the above-mentioned block processing involves dividing the business knowledge text into a group of paragraphs containing multiple statements, where each paragraph constitutes a text block.

[0099] The aforementioned vector index construction unit is data-connected to the aforementioned similarity calculation unit and the aforementioned text recombination unit. It is used to process the aforementioned task requirement text and the aforementioned text block set to obtain a task vector and an index vector set. In response to receiving external input text to be retrieved, it processes the aforementioned text to be retrieved to obtain a retrieval vector. The aforementioned index vector set includes the index vector corresponding to each of the aforementioned text blocks.

[0100] It should be noted that the above vector index building unit can be built based on embedding models such as BERT or OpenAI Embeddings and vector databases such as FAISS or Pinecone, wherein: (1) the embedding model is used to process the task requirement text and the text to be retrieved to obtain the task vector and the text to be retrieved, respectively, and to process each text block in the text block set to obtain the corresponding index vector; (2) the vector database is used to store the obtained task vector, text to be retrieved and index vector set.

[0101] The aforementioned citation counting unit is data-connected to the aforementioned similarity calculation unit and the aforementioned text recombination unit, and is used to store the citation count set and update the citation count set based on the recombined text set from the aforementioned text recombination unit; the aforementioned citation count set includes the citation count corresponding to each of the aforementioned text blocks.

[0102] It should be noted that the citation count for each text block is initialized to 0.

[0103] It should be noted that the above update of the citation count set based on the recombined text set means that after each text recombining unit calculates the recombined text set, the citation count of each text block in the recombined text set is incremented by 1 in the citation count set.

[0104] The similarity calculation unit is connected to the retrieval unit and is used to process the target vector, the index vector set, and the citation count set using a similarity calculation model to obtain a similarity set; the similarity set includes P similarity values.

[0105] The aforementioned retrieval unit is data-connected to the aforementioned text recombination unit and is used to generate a set of cited texts and a set of non-cited texts based on the aforementioned similarity set and the aforementioned set of text blocks; the aforementioned set of cited texts and the set of non-cited texts respectively include Q1 and Q2 of the aforementioned text blocks; Q1+Q2≤P, and Q1 and Q2 are both integers greater than 1.

[0106] It should be noted that the above process of generating the cited text set and the non-cited text set based on the similarity set and the text block set involves sorting the text block set in descending order of similarity value to obtain the sorted text block set. Then, the first Q1 and the first Q2 text blocks in the sorted text block set are taken in turn to obtain the cited text set and the non-cited text set.

[0107] The aforementioned text reorganization unit is used to process the aforementioned set of cited texts, the aforementioned set of non-cited texts, the aforementioned task vector, and the aforementioned vector to be retrieved to obtain the aforementioned reorganized text set.

[0108] It should be noted that the vector indexing unit, similarity calculation unit, retrieval unit, and text recombination unit will execute their respective functions in sequence only after receiving external input text to be retrieved, and finally drive the citation counting unit to update the citation count set.

[0109] In another optional embodiment, the expression for the above similarity calculation model is:

[0110]

[0111] In the formula, SM i λ is the i-th similarity value in the aforementioned similarity set; i is an integer from 1 to P; λ is a preset weight coefficient; VA is the aforementioned vector to be retrieved; VB i The i-th index vector in the aforementioned set of index vectors; α is a preset adjustment coefficient; u i This refers to the i-th reference in the aforementioned set of reference counts.

[0112] Optionally, the above weighting coefficients can be set to a value range of [0.7, 0.9].

[0113] Preferably, the weighting coefficient is 0.8.

[0114] Optionally, the range of the above adjustment coefficient is [0.1, 1].

[0115] Preferably, the adjustment coefficient is 0.6.

[0116] It should be noted that the expression of the above similarity calculation model includes two parts: cosine similarity and citation-based similarity. The closer the value of the cosine similarity component is to 1, the better the similarity (VA) and citation count (VB). i In semantic space, the closer the two parts are, the more similar the retrieved text and the i-th text block are. Similarity based on citation counts takes into account the number of times the corresponding text block is cited. The similarity value of the same text block gradually increases after multiple citations, placing it higher in the ranked text block set and thus increasing its importance. Furthermore, by using weighted summations of the two parts, the relative weights of the two parts in the similarity value can be adjusted.

[0117] It is evident that by calculating similarity values ​​through a similarity calculation model, semantic similarity and citation frequency can be comprehensively considered. Thus, during retrieval, it is possible to prioritize the recall of text blocks with high semantic similarity and further prioritize the use of high-quality text blocks that are frequently cited.

[0118] In yet another alternative embodiment, such as Figure 3 As shown, the above text recombination unit includes a correlation calculation subunit, a filtering subunit, and a merging subunit.

[0119] The aforementioned relevance calculation subunit is data-connected to the aforementioned filtering subunit and is used to process the aforementioned non-referenced text set, the aforementioned index vector set, the aforementioned search vector, and the aforementioned task vector using the relevance calculation model to obtain a relevance set; the aforementioned relevance set includes Q2 relevance values.

[0120] The above filtering subunit is data-connected with the above merging subunit. It is used to sort the above non-referenced text set based on the above relevance set to obtain a sorted text set, and delete the last Q3 text blocks of the sorted text set to obtain a supplementary text set; Q3 is an integer greater than 1 and less than Q2.

[0121] It should be noted that the above sorting is based on the corresponding relevance value in descending order, sorting the text blocks in the non-referenced text collection.

[0122] The aforementioned merging subunit is used to calculate the union of the aforementioned set of cited texts and the aforementioned set of supplementary texts to obtain the recombined text set.

[0123] In yet another optional embodiment, the expression for the above correlation calculation model is:

[0124] CD j =cos β (VI j ,VC)(1-cos(VI j ,VA)) β

[0125] In the formula, CD j VI is the j-th correlation value in the aforementioned correlation set CD; j The index vector in the index vector set corresponds to the j-th text block in the non-referenced text set I; VA and VC are the search vector and the task vector, respectively; β is the preset correlation coefficient; j is an integer from 1 to Q2.

[0126] Optionally, the correlation coefficient can be in the range of [0,1].

[0127] Preferably, the correlation coefficient is 0.5.

[0128] It should be noted that the above correlation calculation model includes two multiplied factors, where the first factor represents cosine similarity. The closer this value is to 1, the stronger the VI. j The closer VI is to VC in semantic space, the more similar the task requires the j-th text block in the set of text and non-referenced text; the second factor represents VI. j The degree of difference between VA and the target text is considered; the larger this value, the greater the difference between the j-th text block in the searched text and the non-referenced text set. Furthermore, calculating the β power for each of the two factors can amplify the influence of the two factors on the relevance value. The closer β is to 0, the more stable the calculated relevance value; conversely, the closer β is to 1, the more discriminative the calculated relevance value.

[0129] As can be seen, the above-mentioned correlation calculation model can comprehensively reflect the degree of difference between the text blocks in the non-referenced text set and the text to be retrieved, as well as the degree of similarity with the text required by the task. This results in text blocks in the sorted text set that are similar to the overall task requirements and have a high degree of difference from the text to be retrieved being placed higher, ensuring that the text blocks in the supplementary text set do not deviate from the overall task requirements and contain as much correlation information as possible.

[0130] In yet another alternative embodiment, such as Figure 4As shown, the interactive generation module includes an expert template storage unit, a first prompt word generation unit, a second prompt word generation unit, a third prompt word generation unit, a fourth prompt word generation unit, a fifth prompt word generation unit, and a reasoning unit.

[0131] The aforementioned expert template storage unit is used to store a preset set of expert templates.

[0132] The aforementioned first prompt word generation unit is data-connected to the aforementioned intent acquisition module, the aforementioned knowledge base module, and the aforementioned reasoning unit, and is used to perform a first combination processing on the aforementioned task requirement text and the aforementioned set of referenced texts to obtain the initial prompt word.

[0133] It should be noted that the first combined processing described above involves filling the referenced text set and the task requirement text into a preset first prompt template. The first prompt template can be in the following form, where positions ① and ② are used to fill the referenced text set and the task requirement text, respectively:

[0134]

[0135] The second prompt word generation unit is connected to the reasoning unit and is used to perform a second combination processing on the externally input current question text, current expert template and current text block set to obtain question expansion prompt words.

[0136] It should be noted that the second combination process described above involves filling the current question text, the current expert template, and the current text block set into a preset second prompt word template. The second prompt word template can be in the following form, where positions ③, ④, and ⑤ are used to fill the current text block set, the current expert template, and the current question text, respectively:

[0137]

[0138] The aforementioned third prompt word generation unit is data-connected to the aforementioned intent acquisition module and the aforementioned reasoning unit, and is used to perform a third combination processing on the aforementioned task requirement text and the aforementioned current expert template input from the outside to obtain associated question prompt words.

[0139] It should be noted that the third combination process described above involves filling the task requirement text and the current expert template into a preset third prompt template. This third prompt template can take the following form, where positions ⑥ and ⑦ are used to fill in the current expert template and the task requirement text, respectively:

[0140]

[0141] The aforementioned fourth prompt word generation unit is data-connected to the aforementioned reasoning unit and is used to perform a fourth combination processing on the aforementioned current question text, the aforementioned current expert template, and the aforementioned current text block set input from the outside to obtain the current prompt word.

[0142] It should be noted that the fourth combination process described above involves filling the current question text, the current expert template, and the current text block set into a preset fourth prompt word template. This fourth prompt word template can be in the following form, where positions ⑧, ⑨, and ⑩ are used to fill the current text block set, the current expert template, and the current question text, respectively:

[0143]

[0144] The aforementioned fifth prompt word generation unit is data-connected to the aforementioned reasoning unit and is used to perform a fifth combination processing on the externally input current answer text and the current duty document to obtain the document update prompt word.

[0145] It should be noted that the fifth combination process described above involves filling the current answer text and the current duty roster document into a preset fifth prompt word template. This fifth prompt word template can be in the following form, where the position... and Used to populate the current answer text and the current duty roster text, respectively:

[0146]

[0147] The aforementioned reasoning unit is used to process the document construction prompts to obtain the corresponding answer text; the aforementioned document construction prompts are the aforementioned initial prompts, the aforementioned question expansion prompts, the aforementioned related question prompts, the aforementioned current prompts, or the aforementioned document update prompts.

[0148] It should be noted that the above-mentioned inference units are large inference models such as GPT-4o and Qwen 2.5, and the embodiments of the present invention are not limited thereto.

[0149] In another optional embodiment, the style transfer module is a pre-trained style transfer model; the style transfer model includes a style embedding vector generation unit, a style encoder unit, a source embedding vector set generation unit, a source encoder unit, a feature vector construction unit, a decoder unit, and a concatenation unit.

[0150] The aforementioned style embedding vector generation unit is used to first split the preset style template text into several sentences, calculate the embedding vector of each sentence in the style template text, and then set the style embedding vector to the mean of the embedding vectors of all style template sentences.

[0151] It should be noted that the above-mentioned style template text can be historical duty text or other specified text, and this embodiment of the invention does not impose any limitations.

[0152] The style encoder unit described above is used to perform style encoding on the style embedding vector to obtain the style vector.

[0153] Optionally, the style encoder unit described above is built based on an MLP (Multilayer Perceptron).

[0154] The aforementioned source embedding vector set generation unit is used to split the initial duty document into several clauses and calculate the embedding vector of each clause of the initial duty document to obtain the source embedding vector set.

[0155] The aforementioned source encoder unit is used to process each clause of the initial duty document to obtain the corresponding encoding vector.

[0156] Optionally, the above source encoder unit is built based on a BiLSTM network.

[0157] The aforementioned feature vector construction unit is used to concatenate the style vector with each encoding vector in sequence to obtain the feature vector corresponding to each encoding vector.

[0158] The aforementioned decoder unit is used to decode each feature vector sequentially to obtain the corresponding target clause.

[0159] Optionally, the above decoder is built based on an LSTM network.

[0160] The above splicing unit is used to sequentially splice the target clauses to obtain the target duty document.

[0161] It is evident that by implementing the duty document generation system disclosed in this embodiment of the invention, the richness and completeness of the generated duty documents can be improved by introducing a user feedback mechanism in multi-round dialogues. Furthermore, it can fully utilize users' deep understanding of the industry to generate personalized duty documents that conform to industry characteristics.

[0162] Example 2

[0163] Please see Figure 5 , Figure 5 This is a flowchart illustrating a method for generating duty roster documents according to an embodiment of the present invention. Figure 5 The flowchart illustrating the described method for generating duty documents is applicable to text data processing, such as generating duty documents, and is not limited to this embodiment of the invention. Figure 5 As shown, the method for generating duty roster documents includes:

[0164] S1. Use the intent acquisition module to obtain the task requirement text.

[0165] S2. Use the knowledge base module to store preset business knowledge text.

[0166] S3. Using the interactive generation module and the aforementioned knowledge base module, construct the current duty document based on the task requirement text.

[0167] S4. Using the above-mentioned intent acquisition module, the above-mentioned interaction generation module and the above-mentioned knowledge base module, update the above-mentioned current duty document to obtain the updated above-mentioned current duty document.

[0168] S5. Repeat S4 until the user inputs an end command.

[0169] S6. Determine that the above-mentioned initial duty document is the above-mentioned current duty document.

[0170] S7. Using the style conversion module, based on the preset style template text, perform style conversion processing on the above initial duty document to obtain the target duty document.

[0171] In an optional embodiment, the current duty roster document is constructed based on the task requirement text using the interactive generation module and the knowledge base module, including:

[0172] S31. Using the knowledge base module described above, the text required by the task is retrieved and processed to obtain the corresponding set of cited texts and the set of reconstructed texts.

[0173] It should be noted that the above-mentioned use of the knowledge base module to retrieve the task requirement text involves inputting the task requirement text as the text to be retrieved into the knowledge base module, so that the vector index construction unit, similarity calculation unit, and retrieval unit are executed sequentially to obtain the corresponding set of cited text and set of reconstructed text.

[0174] S32. Using the first prompt word generation unit, perform a first combination process on the above task requirement text and the above set of referenced texts to obtain the initial prompt word.

[0175] S33. Using the reasoning unit, process the initial prompts to obtain the current duty document.

[0176] It should be noted that the above-mentioned current duty document is the response text obtained by the reasoning unit after processing the initial prompt words.

[0177] As can be seen, since the text blocks in the referenced text set retain the information most similar to the task requirement text, constructing the first prompt word based on this and generating the current duty document can make the initially generated current duty document contain the most accurate information possible.

[0178] In another optional embodiment, the above-mentioned intent acquisition module, interaction generation module, and knowledge base module are used to update the current duty document to obtain the updated current duty document, including:

[0179] S41. Initialize the current text block set to empty; initialize the loop count l to 1; initialize the current question text to the task requirement text mentioned above.

[0180] S42. Set the current expert template to the l-th expert template in the above expert template set.

[0181] S43. Using the intent acquisition module and the reasoning unit, update the current problem text based on the current expert template, the current text block set, and the task requirement text.

[0182] S44. Using the knowledge base module described above, perform retrieval processing on the current problem text to obtain the corresponding set of referenced texts and the set of reconstructed texts; update the current text block set to the set of reconstructed texts.

[0183] It should be noted that the above-mentioned use of the knowledge base module to retrieve the current question text involves inputting the current question text as the text to be retrieved into the knowledge base module, so that the vector index construction unit, similarity calculation unit, and retrieval unit are executed sequentially to obtain the corresponding set of cited texts and the set of reconstructed texts.

[0184] S45. Using the above-mentioned reasoning unit, fourth prompt word generation unit and fifth prompt word generation unit, based on the above-mentioned current expert template, the above-mentioned current text block set and the above-mentioned current question text, update the above-mentioned current duty document.

[0185] S46. Increment the value of l by 1.

[0186] S47. Repeat S42 to S46 until l is greater than N.

[0187] In another optional embodiment, the above-described intention acquisition module and reasoning unit update the current question text based on the current expert template, the current text block set, and the task requirement text, including:

[0188] S431. Using the above-mentioned intent acquisition module, obtain the above-mentioned extended flag.

[0189] S432. Determine the above extended flags;

[0190] When the above extension flag is true, execute S433;

[0191] If the above extended flag is not specified, execute S435.

[0192] S433. Using the second prompt word generation unit, perform a second combination process on the current question text, the current expert template, and the current text block set to obtain the question expansion prompt word.

[0193] S434. Using the above reasoning unit, process the above-mentioned question extension prompts to obtain the extended question text; update the above-mentioned current question text to the above-mentioned extended question text; execute S44.

[0194] It should be noted that the extended question text mentioned above is the answer text obtained after the reasoning unit processes the extended question prompts.

[0195] It should be noted that the text blocks in the referenced text set obtained in step S44 retain the information most similar to the current question text, while the reconstructed text set adds richer related information on top of the referenced text set. Therefore, by updating the current text block set to this reconstructed text set, further constructing question expansion prompts, generating expanded question text, and finally updating the current question text to the expanded question text, the topic scope of the current question text after multiple rounds of dialogue can be effectively expanded.

[0196] S435. Using the third prompt word generation unit, the above task requirement text and the above current expert template are combined in a third way to obtain related question prompt words.

[0197] S436. Using the above reasoning unit, process the above-mentioned related question prompts to obtain related question text; update the above-mentioned current question text to the above-mentioned related question text.

[0198] It should be noted that the above-mentioned related question text is the answer text obtained by the reasoning unit after processing the related question prompts.

[0199] In another optional embodiment, the above-mentioned use of the inference unit, the fourth prompt word generation unit, and the fifth prompt word generation unit to update the current duty document based on the current expert template, the current text block set, and the current question text includes:

[0200] S451. Using the fourth prompt word generation unit, the current question text, the current expert template, and the current text block set are combined in a fourth way to obtain the current prompt word.

[0201] S452. Using the above reasoning unit, process the current prompt word to obtain the current answer text.

[0202] S453. Using the intent acquisition module described above, obtain the retention flag.

[0203] S454. Determine the above-mentioned retention flags;

[0204] When the above reservation flag is true, execute S455;

[0205] When the above retention flag is not set, the above current duty document remains unchanged; execute S46.

[0206] S455. Using the fifth prompt word generation unit, perform the fifth combination processing on the above current answer text and the above current duty document to obtain the document update prompt word.

[0207] S456. Using the above reasoning unit, process the above document update prompt words to obtain intermediate text; replace the content of the above current duty document with the above intermediate text.

[0208] It should be noted that the above intermediate text is the response text obtained by the reasoning unit after processing the document update prompt words.

[0209] It is evident that by implementing the duty roster document generation method disclosed in this embodiment of the invention, and by introducing a user feedback mechanism in multi-round dialogues to dynamically generate and update duty roster documents, the richness and completeness of the generated duty roster documents can be improved. Furthermore, it can fully utilize users' deep understanding of the industry to generate personalized duty roster documents that conform to industry characteristics.

[0210] The device embodiments described above are merely illustrative. The modules described as separate components may or may not be physically separate, and the components shown as modules may or may not be physical modules; that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0211] Through the detailed description of the above embodiments, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, including read-only memory (ROM), random access memory (RAM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), one-time programmable read-only memory (OTPROM), electrically-erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, disk storage, magnetic tape storage, or any other computer-readable medium that can be used to carry or store data.

[0212] Finally, it should be noted that the duty document generation system and method based on a large reasoning model disclosed in the embodiments of the present invention are merely preferred embodiments of the present invention, and are only used to illustrate the technical solutions of the present invention, not to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A duty roster document generation system, characterized in that, It includes an intent acquisition module, a knowledge base module, an interaction generation module, and a style transfer module: The intent acquisition module is data-connected to the knowledge base module and the interaction generation module, and is used to acquire task requirement text, extended flags and reserved flags; The knowledge base module is data-connected to the interactive generation module, used to store preset business knowledge text, and to perform retrieval processing based on externally input text to obtain retrieval results; the retrieval results include a set of cited texts and a set of reconstructed texts; The knowledge base module includes a text segmentation unit, a vector index construction unit, a citation counting unit, a similarity calculation unit, a retrieval unit, and a text recombination unit; The text segmentation unit is data-connected to the vector index construction unit and the retrieval unit, and is used to segment the business knowledge text to obtain a text block set; the text block set includes P text blocks; P is an integer greater than 1; The vector index construction unit is data-connected to the similarity calculation unit and the text recombination unit, and is used to process the task requirement text and the text block set to obtain a task vector and an index vector set, and to process the text to be retrieved in response to receiving external input to obtain a retrieval vector; the index vector set includes the index vector corresponding to each text block; The citation counting unit is data-connected to the similarity calculation unit and the text recombination unit, and is used to store the citation count set and update the citation count set based on the recombined text set from the text recombination unit; the citation count set includes the citation count corresponding to each text block; The similarity calculation unit is data-connected to the retrieval unit and is used to process the vector to be retrieved, the set of index vectors, and the set of reference counts using a similarity calculation model to obtain a similarity set; the similarity set includes P similarity values; The retrieval unit is data-connected to the text recombination unit and is used to generate a set of cited texts and a set of non-cited texts based on the similarity set and the set of text blocks; the set of cited texts and the set of non-cited texts each include Q1 and Q2 text blocks; Q1 + Q2 ≤ P, and Q1 and Q2 are both integers greater than 1; The text recombination unit is used to process the cited text set, the non-cited text set, the task vector, and the retrieval vector to obtain the recombined text set; The interactive generation module is data-connected to the style conversion module and is used to store a preset set of expert templates, and to generate and reason about prompt words based on the task requirement text, the extended flag, the retained flag and the search results to obtain the target duty document; The expert template set includes N expert templates; N is an integer greater than 1. The style conversion module is used to perform style conversion processing on the initial duty document based on the preset style template text to obtain the target duty document.

2. The duty document generation system according to claim 1, characterized in that, The text recombination unit includes a relevance calculation subunit, a filtering subunit, and a merging subunit: The correlation calculation subunit is data-connected to the filtering subunit and is used to apply the correlation calculation model to the non-referenced text set and the index vector set. 、 The vector to be retrieved and the task vector are processed to obtain a set of relevance values; the set of relevance values ​​includes Q2 relevance values. The filtering subunit is data-connected to the merging subunit and is used to sort the non-referenced text set based on the relevance set to obtain a sorted text set, and to delete the last Q3 text blocks of the sorted text set to obtain a supplementary text set; Q3 is an integer greater than 1 and less than Q2; The merging subunit is used to calculate the union of the cited text set and the supplementary text set to obtain the recombined text set.

3. The duty document generation system according to claim 2, characterized in that, The expression for the correlation calculation model is: In the formula, CD j For the set of correlation degrees CD The first in j The aforementioned correlation value; VI j For the non-referenced text collection I The first in j The text blocks, and the corresponding index vectors in the set of index vectors; VA and VC These are the vector to be retrieved and the task vector, respectively. β The preset correlation coefficient; j It is an integer from 1 to Q2.

4. The duty document generation system according to claim 1, characterized in that, The interactive generation module includes an expert template storage unit, a first prompt word generation unit, a second prompt word generation unit, a third prompt word generation unit, a fourth prompt word generation unit, a fifth prompt word generation unit, and a reasoning unit; The expert template storage unit is used to store a preset set of expert templates; The first prompt word generation unit is data-connected to the intent acquisition module, the knowledge base module and the reasoning unit, and is used to perform a first combination processing on the task requirement text and the reference text set to obtain the initial prompt word; The second prompt word generation unit is data-connected to the reasoning unit and is used to perform a second combination processing on the externally input current question text, current expert template, and current text block set to obtain question expansion prompt words; The third prompt word generation unit is data-connected to the intent acquisition module and the reasoning unit, and is used to perform a third combination processing on the task requirement text and the current expert template input externally to obtain associated question prompt words; The fourth prompt word generation unit is data-connected to the reasoning unit and is used to perform a fourth combination processing on the externally input current question text, the current expert template, and the current text block set to obtain the current prompt word. The fifth prompt word generation unit is data-connected to the reasoning unit and is used to perform a fifth combination processing on the externally input current answer text and the current duty document to obtain document update prompt words; The reasoning unit is used to process the document construction prompts to obtain the corresponding answer text; the document construction prompts are the initial prompts, the question expansion prompts, the related question prompts, the current prompts, or the document update prompts.

5. A method for generating duty roster documents, characterized in that, The method, applied to the duty document generation system according to any one of claims 1-4, comprises: S1. Use the intent acquisition module to obtain the task requirement text; S2. Utilize the knowledge base module to store preset business knowledge text; S3. Using the interactive generation module and the knowledge base module, construct the current duty document based on the task requirement text; S4. Using the intent acquisition module, the interaction generation module, and the knowledge base module, update the current duty document to obtain the updated current duty document; S5. Repeat S4 until the user inputs an end command; S6. Determine the initial duty document as the current duty document; S7. Using the style conversion module, based on the preset style template text, perform style conversion processing on the initial duty document to obtain the target duty document.

6. The method for generating duty roster documents according to claim 5, characterized in that, The process of constructing the current duty roster document based on the task requirement text using the interactive generation module and the knowledge base module includes: S31. Using the knowledge base module, the task requirement text is retrieved and processed to obtain the corresponding set of cited texts and the set of reconstructed texts; S32. Using the first prompt word generation unit, perform a first combination process on the task requirement text and the set of referenced texts to obtain initial prompt words; S33. Using the reasoning unit, process the initial prompt words to obtain the current duty document.

7. The method for generating duty roster documents according to claim 5, characterized in that, The step of updating the current duty document using the intent acquisition module, the interaction generation module, and the knowledge base module to obtain the updated current duty document includes: S41. Initialize the current text block set to empty; set the loop count... l Initialize to 1; initialize the current question text to the task requirement text; S42. Set the current expert template as the first one in the expert template set. l The expert templates described above; S43. Using the intent acquisition module and the reasoning unit, update the current question text based on the current expert template, the current text block set, and the task requirement text; S44. Using the knowledge base module, perform retrieval processing on the current question text to obtain the corresponding set of referenced texts and the set of reconstructed texts; update the current text block set to the set of reconstructed texts; S45. Using the reasoning unit, the fourth prompt word generation unit, and the fifth prompt word generation unit, update the current duty document based on the current expert template, the current text block set, and the current question text; S46, will l Increment the value by 1; S47. Repeat S42~S46 until... l Greater than N.

8. The method for generating duty roster documents according to claim 7, characterized in that, The step of using the intent acquisition module and the reasoning unit to update the current question text based on the current expert template, the current text block set, and the task requirement text includes: S431. Obtain the extended flag using the intent acquisition module; S432. Determine the extended flag; When the extension flag is true, execute S433; If the extension flag is not specified, execute S435; S433. Using the second prompt word generation unit, perform a second combination process on the current question text, the current expert template, and the current text block set to obtain extended prompt words for the question. S434. Using the reasoning unit, process the extended prompt words of the question to obtain the extended question text; update the current question text to the extended question text; execute S44; S435. Using the third prompt word generation unit, the task requirement text and the current expert template are combined in a third way to obtain associated question prompt words; S436. Using the reasoning unit, process the associated question prompt words to obtain associated question text; update the current question text to the associated question text.

9. The method for generating duty roster documents according to claim 7, characterized in that, The step of updating the current duty roster document using the reasoning unit, the fourth prompt word generation unit, and the fifth prompt word generation unit, based on the current expert template, the current text block set, and the current question text, includes: S451. Using the fourth prompt word generation unit, the current question text, the current expert template, and the current text block set are combined in a fourth way to obtain the current prompt word. S452. Using the reasoning unit, the current prompt word is processed to obtain the current answer text; S453. Using the intent acquisition module, obtain the retention flag; S454. Determine the retention flag; When the reservation flag is true, execute S455; When the retention flag is not set, the current duty document remains unchanged; execute S46; S455. Using the fifth prompt word generation unit, the current answer text and the current duty document are combined in a fifth combination process to obtain the document update prompt word; S456. Using the reasoning unit, process the document update prompt words to obtain intermediate text; replace the content of the current duty document with the intermediate text.

Images