End side summary expansion method, device and equipment based on retrieval enhancement and storage medium

Abstract

The invention relates to the technical field of text expansion, in particular to an end-side summary expansion method and device based on retrieval enhancement, equipment and a storage medium. Comprising the steps of obtaining an original voice text and an initial summary text; constructing a local vector library based on the original text blocks; transferring the initial summary text to a local vector library, and performing association retrieval on the initial summary text and each original text block through the local vector library to obtain a candidate text block sequence; obtaining each candidate text block, calculating a comprehensive weight corresponding to each candidate text unit, and selecting the candidate text unit with the highest correlation as a target text unit; inputting the target text unit and the corresponding comprehensive weight into a first language model, and extracting to obtain a retrieval abstract text; and determining a to-be-expanded text block, and inputting the to-be-expanded text block and the retrieval abstract text into a second language model to obtain an expanded summary text. According to the method, data security and lightweight deployment are taken into consideration, and the retrieval precision and the generated content quality are improved.

Description

Technical Field

[0001] This application relates to the technical field of text expansion, and in particular to a method, apparatus, device and storage medium for end-side summary expansion based on retrieval enhancement. Background Technology

[0002] With the rapid evolution of LLM (Large Language Model) technology, artificial intelligence has been widely penetrated into various industries such as finance, healthcare, and enterprise management, becoming a key driver for improving work efficiency. In enterprise meeting scenarios, using AI to automatically generate meeting minutes has become a mainstream application model. This process typically includes two preliminary steps: speech-to-text and speaker separation. The original audio is converted into structured text, and then the model performs summarization and refinement to form meeting results that can be archived and executed.

[0003] However, speech transcription often introduces noise and slips of the tongue, resulting in transcriptions that frequently contain interfering information. Furthermore, for data security reasons, confidential meeting content cannot often be uploaded to the cloud to access large language models, while locally deployed small-parameter models struggle to balance generation quality with factual accuracy. Therefore, to suppress model illusions and avoid noise interference, the industry tends to have models output only highly summarized, concise summaries.

[0004] However, while this simplification can filter out irrelevant information and mitigate the impact of errors to some extent, it inevitably leads to the loss of crucial discussion details, decision-making basis, and potential disagreements. Overly concise minutes fail to fully recreate the entire meeting and cannot meet the in-depth needs of enterprises in areas such as task traceability, accountability clarification, and subsequent execution. In other words, existing technological approaches struggle to achieve an effective balance between security, lightweight design, accuracy, and information integrity. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, this application provides a method, apparatus, device, and storage medium for edge-side minutes expansion based on retrieval enhancement, which achieves a balance between data security and lightweight deployment, improves retrieval accuracy and the quality of generated content, and provides an efficient, secure, and intelligent solution for meeting minutes processing.

[0006] The technical solution adopted by this application to solve its technical problem is: Firstly, this application provides a method for expanding client-side minutes based on retrieval enhancement, the method comprising: Obtain the original speech text and initial summary text corresponding to the original speech data; the original speech text includes multiple original text blocks; Each of the original text blocks is encoded to obtain multiple text block vectors, and a local vector library is constructed based on all the text block vectors. The initial summary text is used as the query input to the local vector library. The local vector library is used to perform association retrieval between the initial summary text and each original text block to obtain a candidate text block sequence including multiple candidate text blocks. Obtain the candidate text unit corresponding to each candidate text block, calculate the context association strength of each candidate text unit to obtain the comprehensive weight corresponding to each candidate text unit, and select several candidate text units with the highest semantic relevance to the initial summary text as target text units. All target text units and their corresponding comprehensive weights are input into a preset first language model, and the first language model extracts the retrieval summary text based on the target text units and the comprehensive weights. The text block to be expanded in the initial summary text is determined, and the text block to be expanded and the search summary text are input into a preset second language model. The second language model expands the text block to be expanded based on the search summary text to update the initial summary text to an expanded summary text.

[0007] Optionally, the associated retrieval includes semantic retrieval paths and lexical retrieval paths; The step of associating the initial summary text with each of the original text blocks through the local vector library to obtain a sequence of candidate text blocks including multiple candidate text blocks includes: Encode the initial minutes text into a query vector; The semantic approximation score between the query vector and each text block vector is calculated using the semantic retrieval path, and all original text blocks are sorted according to the semantic approximation score to obtain a semantic path sequence. Furthermore, the vocabulary matching score between the initial summary text and each of the original text blocks is calculated through the vocabulary retrieval path, and all original text blocks are sorted according to the vocabulary matching score to obtain a vocabulary path sequence; Based on a preset smoothing factor, a first sorting score is calculated according to the sorting position of each original text block in the semantic path sequence, and a second sorting score is calculated according to the sorting position of each original text block in the lexical path sequence. Based on the first sorting score and the second sorting score, a comprehensive sorting score is determined for each of the original text blocks. All the original text blocks are sorted according to all the comprehensive sorting scores to obtain the candidate text block sequence.

[0008] Optionally, in the step of obtaining the candidate text unit corresponding to each candidate text block and calculating the contextual association strength of each candidate text unit to obtain the comprehensive weight corresponding to each candidate text unit, the steps performed for each candidate text block include: Obtain the preceding and following text blocks adjacent to the candidate text block from the original speech text, and combine the preceding text block, the candidate text block, and the following text block into the candidate text unit; Based on a preset multidimensional similarity index, the multidimensional association strength between the initial summary text and the preceding adjacent text block, the candidate text block, and the following adjacent text block is evaluated respectively. Based on the multidimensional association strength, the weight scores of the preceding adjacent text block, the candidate text block, and the following adjacent text block are determined, and the comprehensive weight corresponding to the candidate text unit is generated based on the weight scores.

[0009] Optionally, the step of selecting several candidate text units with the highest semantic relevance to the initial summary text as target text units includes: The initial summary text is concatenated with each of the candidate text blocks to form multiple query candidate pairs; Each of the query candidate pairs is input into a preset reordering model, and the intent score corresponding to each of the query candidate pairs is calculated through the reordering model. The candidate text block sequence is reordered in descending order based on all the intent scores to obtain an intent matching sequence, and the candidate text units corresponding to the first preset number of candidate text blocks in the intent matching sequence are selected as the target text units.

[0010] Optionally, the first language model is configured with a first preset parameter; The step of obtaining the retrieval summary text by extracting it from the target text unit and the comprehensive weight using the first language model includes: The first preset prompt word is input into the first language model, and the first language model determines the first target task based on the first preset prompt word. Constrained by the first preset parameters, the first target task is performed based on the target text unit and the comprehensive weight to obtain the retrieval summary text.

[0011] Optionally, the initial minutes text includes multiple minutes text blocks; The step of determining the text block to be expanded in the initial minutes text includes: Extract the information completeness corresponding to each of the aforementioned minutes text blocks, and determine whether the information completeness reaches a preset completeness threshold; If the information completeness does not reach the completeness threshold, the corresponding minutes text block will be determined as the text block to be expanded.

[0012] Optionally, the second language model is configured with a second preset parameter; The step of expanding the text block to be expanded based on the retrieved summary text to update the initial summary text to the expanded summary text includes: The second preset prompt word is input into the second language model, and the second language model determines the second target task based on the second preset prompt word. Constrained by the second preset parameter, the second target task is performed based on the text block to be expanded and the search summary text to obtain the expanded text block corresponding to the text block to be expanded; The expanded summary text is obtained by updating the initial summary text based on all the expanded text blocks.

[0013] Secondly, this application provides a retrieval-enhanced end-side summary expansion device, comprising: The text acquisition module is used to acquire the original speech text and initial summary text corresponding to the original speech data; the original speech text includes multiple original text blocks. The vector encoding module is used to encode each of the original text blocks to obtain multiple text block vectors, and to build a local vector library based on all the text block vectors; The associated retrieval module is used to input the initial summary text as a query into the local vector library, and to perform associated retrieval between the initial summary text and each of the original text blocks through the local vector library to obtain a candidate text block sequence including multiple candidate text blocks; The unit filtering module is used to obtain the candidate text units corresponding to each candidate text block, calculate the context association strength of each candidate text unit to obtain the comprehensive weight corresponding to each candidate text unit, and select several candidate text units with the highest semantic relevance to the initial summary text as target text units. The summary and extraction module is used to input all the target text units and their corresponding comprehensive weights into a preset first language model, and to extract the retrieval summary text based on the target text units and the comprehensive weights through the first language model. The minutes expansion module is used to determine the text block to be expanded in the initial minutes text, input the text block to be expanded and the search summary text into a preset second language model, and expand the text block to be expanded based on the search summary text through the second language model to update the initial minutes text into expanded minutes text.

[0014] Thirdly, this application provides an electronic device, comprising: One or more processors; One or more memory units; And one or more computer programs, wherein the one or more computer programs are stored in the one or more memories, and the one or more computer programs include instructions that, when executed by the one or more processors, cause the electronic device to perform the methods described above.

[0015] Fourthly, this application provides a computer-readable storage medium storing a program or instructions that, when executed, implement the above-described method.

[0016] This application effectively bridges the gap between security, accuracy, and information integrity in existing meeting minutes generation technologies by constructing a user-friendly, enhanced retrieval generation process. Its core working principle is as follows: First, the original audio text is stored in a structured manner using a local vector library, and the initial minutes text is used as a query trigger to retrieve information from the original content. Then, by calculating the contextual relevance strength of candidate text units and assigning comprehensive weights, several target text units most semantically relevant to the initial minutes text are selected. Next, a first language model is used to refine these weighted highly relevant contents, generating a denoised and focused retrieval summary text. Finally, a second language model, combined with this summary, provides evidence-based expansion of the text blocks to be expanded in the initial minutes text, thereby outputting a more detailed and logically complete expanded minutes text.

[0017] The beneficial effects of this application are as follows: Firstly, all data processing is completed locally, eliminating the need to upload the original speech text to the cloud. Combined with a local vector library built based on the original speech text, this fundamentally ensures the data security of confidential meeting content and meets the rigid requirements of edge deployment. Secondly, through targeted retrieval based on the initial minutes text, the system can accurately recall supporting context from the noisy original transcribed text and guide the first language model to prioritize highly relevant segments with comprehensive weights, effectively filtering out interference information such as slips of the tongue and redundancy, and significantly reducing the risk of model illusion caused by poor input quality. More importantly, the retrieved summary text provides reliable factual anchors for the second language model, enabling its expansion behavior to no longer rely on fuzzy memory or generalized guesswork, but rather to reasonably extend the original discussion details while faithfully reproducing them. This avoids the problem of the initial minutes text losing key decision-making basis, differing viewpoints, and other deep information due to over-generalization.

[0018] Thus, this method achieves in-depth restoration and structural enhancement of meeting content without sacrificing data security and lightweight deployment. It avoids the shortcomings of insufficient generation capabilities of pure terminal small models and overcomes the drawbacks of the lack of information in traditional concise initial minutes text, making the automatically generated meeting results truly have traceable, executable, and archiveable business value. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating a typical end-side amplification method provided in this application; Figure 2 This is a schematic flowchart of the end-side summary expansion method based on retrieval enhancement provided in the embodiments of this application; Figure 3 This is a schematic diagram of the virtual structure of the end-side summary expansion device based on retrieval enhancement provided in this application; Figure 4 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0020] The present application will be further described below with reference to the accompanying drawings and embodiments.

[0021] The following will clearly and completely describe the concept, specific structure, and resulting technical effects of this application in conjunction with embodiments and accompanying drawings, so as to fully understand the purpose, features, and effects of this application. Obviously, the described embodiments are only a part of the embodiments of this application, not all of them. Other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are all within the scope of protection of this application. Furthermore, all connections / linkages involved in the patent do not simply refer to direct contact between components, but rather to the ability to form a better connection structure by adding or reducing connecting accessories according to specific implementation conditions. The various technical features in this application can be combined interactively without contradicting each other.

[0022] With the rapid iteration and popularization of large language model technology, artificial intelligence has deeply empowered enterprise office scenarios such as meeting management. Among them, automated meeting minutes technology based on speech-to-text and text generation has become a key link in improving efficiency.

[0023] Current mainstream automated meeting minutes technologies typically follow a two-stage processing flow: first, speech-to-text technology converts meeting recordings into text, and speaker separation technology distinguishes different speakers; then, based on the transcribed text, a large language model is used to summarize the content and generate meeting minutes. However, this process faces multiple challenges in practical applications: the speech-to-text stage often introduces noise, slips of the tongue, and other errors, directly affecting the quality of the generated minutes; simultaneously, meeting content often involves sensitive corporate information, making it impossible to rely on large cloud models for processing due to data security concerns; furthermore, existing high-performance models typically have large parameters, making efficient deployment on lightweight edge devices difficult, thus limiting the technology's applicability.

[0024] To avoid transcription noise and the potential "illusion" problems caused by large language models, current mainstream solutions tend to control the length of generated content, outputting highly concise meeting minutes. (See reference...) Figure 1 , Figure 1 This is a flowchart illustrating a typical edge-side amplification method provided in this application. Such simplified minutes generation methods typically begin by inputting the original speech into an edge-side language model, which then performs simple extraction to obtain the simplified minutes text. While this improves the stability and security of the output to some extent, it introduces new drawbacks: overly simplified minutes often lose crucial discussion details, decision-making logic, task allocation, and differing opinions, making it difficult to comprehensively and accurately reflect the entire meeting. This not only affects the completeness of the meeting records but also weakens the practical value of the minutes in subsequent task tracing, execution follow-up, and knowledge accumulation.

[0025] Therefore, existing technologies still have significant limitations in the field of meeting minutes generation: on the one hand, how to effectively suppress speech-to-text noise and model illusions to ensure content accuracy while ensuring data security and lightweight deployment on the device side; on the other hand, how to generate detailed, complete, and logically clear minutes to meet enterprises' deep-seated needs for structured retention and utilization of meeting information without oversimplifying the content. These issues collectively restrict the reliable application of automated meeting minutes technology in scenarios with high security requirements and high information density.

[0026] To address the aforementioned technical deficiencies, this application proposes a method for intelligently expanding minutes based on simplified minutes text, referring to... Figure 2 , Figure 2 This is a schematic flowchart of the client-side summary expansion method based on retrieval enhancement provided in the embodiments of this application. Figure 2 The document illustrates several key steps involved in the intelligent expansion method provided in this application, which are described in detail below: In step S1, the original speech text and initial summary text corresponding to the original speech data are obtained; Raw audio data refers to unprocessed audio signals collected during the meeting, which typically includes the voice content of multiple speakers.

[0027] The original speech text is a written, structured text obtained by transcribing the original speech data into speech-to-text and speaker diarization, preserving the original expressions of each speaker during the meeting. The original speech text includes several original text blocks, which are continuous text units segmented according to semantic or temporal order. Each text block represents a relatively complete speech or discussion segment, facilitating subsequent fine-grained information retrieval and contextual analysis.

[0028] The initial minutes text is a highly summarized and concise meeting summary generated by LLM (Large Language Model) based on the original speech text. It usually contains only core issues, key conclusions or action items, with high information density but lacking details.

[0029] Specifically, the system first converts raw speech data into raw speech text. This process includes two sub-stages: speech recognition and speaker identification, ensuring that the output text retains both the content of the speech and the corresponding speaker's identity. Subsequently, the raw speech text is further divided into multiple raw text blocks. The division can be based on speaking turn, semantic boundaries, or fixed-length windows, aiming to ensure each block has independent contextual integrity and facilitates efficient subsequent retrieval. Furthermore, the system uses a lightweight large language model (on the device side) to summarize the entire raw speech text, generating an initial summary text. This initial summary text serves as a guiding anchor for subsequent expansion; although brief, it reflects the core information of the meeting to a certain extent.

[0030] In one specific embodiment, suppose a company holds a project progress meeting, and the original audio data is a full recording of the meeting. Using speech-to-text technology, the recording is converted into original audio text containing the statements of all participants, which may include statements such as "Project A's R&D progress is lagging, mainly due to delayed parts delivery from supplier B, which is expected to be resolved next week." The initial minutes text generated based on this original audio text might read, "Project A's R&D progress is lagging, and supplier B's parts delivery is delayed."

[0031] Subsequently, the original audio text was divided into multiple original text blocks, such as "Project A's R&D progress is lagging behind," "The main reason is the delayed delivery of parts by supplier B," and "It is expected to be resolved next week." At this point, the original audio text retained key details such as "It is expected to be resolved next week," and the initial summary text clearly defined the core issue of the expanded text as "the reasons for the lagging R&D progress of Project A." The division of the original text blocks provided structured units for subsequent retrieval.

[0032] In step S2, each of the original text blocks is encoded to obtain multiple text block vectors, and a local vector library is constructed based on all the text block vectors.

[0033] Among them, the text block vector is mapped to a dense vector of fixed dimension by an embedding model. The dense vector can represent the semantic information of the text block and is used to support the semantic retrieval of subsequent query inputs. The local vector library refers to an efficient similarity retrieval structure composed of all text block vectors that is built and stored on the terminal device. The embodiments of this application propose that it can be implemented based on lightweight vector indexing tools such as FAISS (Facebook AI Similarity Search), which can support fast semantic retrieval without relying on cloud services.

[0034] Specifically, the system first invokes a lightweight embedding model (such as bge-small-zh or embedding-gemma) to semantically encode each original text block, transforming it into a corresponding text block vector. These vectors retain the core semantic features of the original text blocks and express their positions in the high-dimensional semantic space in numerical form. Subsequently, all generated text block vectors are loaded into the FAISS framework to build a local vector library. This vector library employs index structures such as inverted index (IVF) or hierarchical approximate nearest neighbor (HNSW), combined with vector quantization technology, to significantly reduce memory usage and computational overhead while ensuring retrieval accuracy. The entire process is completed on the client side without any data upload, ensuring content confidentiality.

[0035] More specifically, this step utilizes a small-parameter embedding model to achieve efficient and low-resource-consumption semantic vectorization, and relies on FAISS to build a localized high-dimensional vector index system, enabling the system to quickly respond to subsequent semantic retrieval requests on terminal devices. This "local vectorization + local index" design not only meets the enterprise's security requirement of data not leaving the domain, but also provides information support for subsequent information retrieval based on semantic relevance.

[0036] In step S3, the initial summary text is used as a query input to the local vector library. The local vector library is used to perform association retrieval between the initial summary text and each of the original text blocks to obtain a candidate text block sequence including multiple candidate text blocks.

[0037] Among them, the candidate text block sequence refers to an ordered set of multiple original text blocks that are semantically or keyword-wise related to the initial summary text and are retrieved from the local vector library.

[0038] Specifically, the initial minutes text is used as a semantic query to perform an association retrieval in the local vector library to recall original text blocks related to its content. Specifically, the initial minutes text is first fed into the same embedding model used when building the local vector library, converting it into a query vector. Then, this query vector is compared with the text block vectors of all original text blocks in the local vector library for similarity calculation, using metrics such as cosine similarity. The system ranks the original text blocks based on the similarity scores and selects the top-ranked results to form a candidate text block sequence. This process essentially maps the highly compressed initial minutes text back to its context in the original conference speeches, thus providing factual basis for subsequent expansion.

[0039] More specifically, in the embodiments of this application, the association retrieval adopts a multi-strategy fusion retrieval architecture to combine the dual advantages of semantic understanding and lexical matching. Under this concept, the association retrieval includes two retrieval paths: a semantic retrieval path and a lexical retrieval path. The semantic retrieval path refers to a retrieval method based on semantic similarity in vector space. By encoding the text into dense vectors (such as using embedding models like Sentence-BERT), the distance or similarity between the query and the original text block in the high-dimensional semantic space is calculated to measure their deep semantic association. The lexical retrieval path refers to a retrieval method based on keyword matching. Traditional information retrieval methods such as BM25 (Best Matching 25, a classic information retrieval algorithm) are used to evaluate the relevance at the surface lexical level by statistically analyzing the word frequency, inverse document frequency, and word overlap between the initial summary text and the original text block.

[0040] Based on this, in this embodiment of the application, the step of performing association retrieval between the initial summary text and each of the original text blocks through the local vector library to obtain a candidate text block sequence including multiple candidate text blocks includes: The initial minutes text is encoded into a query vector.

[0041] Specifically, the initial summary text is input into the embedding model and encoded into a query vector, which serves as the input for semantic retrieval and is used to perform semantic similarity calculations in the vector space.

[0042] Furthermore, the semantic approximation score between the query vector and each text block vector is calculated through the semantic retrieval path, and all original text blocks are sorted according to the semantic approximation score to obtain a semantic path sequence.

[0043] The semantic approximation score is a numerical value calculated from the semantic retrieval path, reflecting the degree of semantic similarity between the query vector and a certain text block vector.

[0044] Specifically, in the semantic retrieval path, the system uses the text block vectors stored in the local vector library to calculate the semantic approximation score between the query vector and each text block vector. This calculation can use cosine similarity, and based on this, all original text blocks are sorted from high to low according to the score to form a semantic path sequence.

[0045] It's worth noting that in semantic retrieval, the system focuses on "similarity of meaning," rather than literal identicalness. Therefore, even if the original statement doesn't use words like "postponed" or "user survey," as long as it expresses a similar intent, it may be retrieved with a high ranking. For example, the first few items in a semantic path sequence might include: "Currently, we are short-staffed, so the user interview originally scheduled to start this week will probably have to be postponed." "The marketing department suggests focusing on competitor analysis first, and that user feedback collection can wait for a week." "Because the questionnaire has not yet passed legal review, user research cannot proceed for the time being." These original text blocks did not completely reproduce the keywords in the initial summary text, but they all expressed the core semantics of "a research activity related to the user was postponed for some reason". Therefore, they were close to the query vector in the vector space, had a high semantic approximation score, and were ranked higher.

[0046] In addition, the vocabulary matching score between the initial summary text and each of the original text blocks is calculated through the vocabulary retrieval path, and all the original text blocks are sorted according to the vocabulary matching score to obtain the vocabulary path sequence.

[0047] Among them, the word matching score is a value calculated from the word retrieval path, reflecting the strength of keyword co-occurrence and term matching between the initial summary text and a certain original text block.

[0048] Specifically, in the vocabulary retrieval path, the system performs Chinese word segmentation and stop word filtering on the initial summary text and each original text block, then uses the BM25 algorithm to calculate the vocabulary matching score between each pair of texts, and sorts all the original text blocks accordingly to form a vocabulary path sequence.

[0049] It's worth noting that in the vocabulary retrieval path, the system pays more attention to word overlap, especially the frequency and matching strength of key terms. Therefore, original text blocks containing words such as "user survey," "postponed," and "next month" will be prioritized for retrieval. For example, the first few items in the vocabulary path sequence might include: "The user survey was originally scheduled to begin this week, but it has now been postponed to the beginning of next month." "Two rounds of user surveys are scheduled for next month, and we need to prepare the equipment in advance." "Should we postpone the user survey? — Yes, postpone it to June." These sentences highly overlap with the initial minutes text at the lexical level. The BM25 algorithm will assign high lexical matching scores due to their keyword co-occurrence. Even if some sentences are not actually related to "decision postponement" (such as sentence 2 which only states that there will be a survey next month), they may still be ranked at the top due to word frequency advantage.

[0050] The comparison shows that semantic path sequences are better at capturing meaning. Figure 1 While related content with similar wording but different expressions is highly resistant to lexical variation, lexical path sequences are sensitive to explicit information such as terms, time, and action words, providing precise location but are easily limited by literal meaning. Combining the two can retrieve contexts with similar semantics but different wording, as well as fragments with consistent wording but requiring contextual discernment, thereby jointly improving the overall relevance and coverage of candidate text blocks.

[0051] Subsequently, based on a preset smoothing factor, a first sorting score is calculated according to the sorting position of each original text block in the semantic path sequence, and a second sorting score is calculated according to the sorting position of each original text block in the lexical path sequence.

[0052] Among them, the smoothing factor is a constant parameter (e.g., the default value is 60) used to adjust the decay rate of position weights when fusing multi-way ranking results, so as to avoid the scores of lower-ranked items approaching zero and improve the robustness of fusion; the first ranking score is the score calculated based on the ranking position of the original text block in the semantic path sequence, combined with the smoothing factor; the second ranking score is the score calculated based on the ranking position of the original text block in the lexical path sequence, combined with the smoothing factor.

[0053] Specifically, the system introduces a preset smoothing factor (e.g., k=60) and calculates a score for the ranking position of each original text block in the two sequences. In the embodiments of this application, a formula is used. Calculate the ranking score of each original text block in each sequence, where As a smoothing factor, This represents the sequence position of the currently computed original text block within the current processing sequence.

[0054] More specifically, if a text block ranks in the semantic path sequence... If the position is 1, then its first ranking score is 1. If it ranks first in the word path sequence If the second sort score is 1, then the second sort score is 1. This scoring design based on reverse order ensures that text blocks ranked higher receive higher weights, while a smoothing factor prevents tail-ranked text blocks from being completely ignored due to low scores.

[0055] Finally, based on the first sorting score and the second sorting score, a comprehensive sorting score for each of the original text blocks is determined, and all the original text blocks are sorted according to all the comprehensive sorting scores to obtain the candidate text block sequence.

[0056] The comprehensive ranking score is the total score obtained by adding the first ranking score and the second ranking score, which is used for the final ranking to form a sequence of candidate text blocks.

[0057] Specifically, the first sorting score and the second sorting score of each original text block are added together to obtain its comprehensive sorting score. Based on this score, all original text blocks are re-sorted globally, and the sorted result is finally output as a candidate text block sequence.

[0058] More specifically, by executing two complementary retrieval strategies—semantic and lexical—in parallel and employing a location-based weighted fusion mechanism, this approach retains the sensitivity of semantic retrieval to synonyms and contextual reasoning while inheriting the high-precision capture capability of lexical retrieval for explicit information such as key terms, numbers, and names. Especially in situations where conference speech-to-text contains noise such as slips of the tongue, omissions, and unclear references, a single retrieval path can easily miss important contextual information. Dual-path fusion significantly improves the recall completeness and ranking accuracy of relevant segments, providing a more reliable and comprehensive factual basis for subsequent expansion.

[0059] In one specific embodiment, suppose the initial summary text is: "Project A is delayed due to resource conflict." This sentence is brief and does not specify the specific resource type or conflicting party. In the semantic retrieval path, the system may recall an original text block stating "We cannot simultaneously support the GPU requirements of Projects A and B," because it semantically expresses the causal relationship between "resource conflict" and "delay," even though the word "delay" does not appear. In the lexical retrieval path, the system may recall an original text block stating "Project A needs to call 3 A100s, but they have already been reserved by Project C," because it contains keywords such as "Project A" and "resources." The former ranks higher in the semantic path, and the latter ranks higher in the lexical path. After calculating the first and second ranking scores of both in the two sequences using a smoothing factor (e.g., k=60), both have high combined ranking scores and are therefore included in the candidate text block sequence. It is precisely because this hybrid retrieval mechanism considers both lexical and semantic aspects that these two key segments, which are different in perspective but complementary in content, are effectively recalled, thus providing sufficient basis for subsequent expansion and avoiding information loss due to retrieval bias.

[0060] In step S4, candidate text units corresponding to each candidate text block are obtained, the contextual association strength of each candidate text unit is calculated to obtain the comprehensive weight of each candidate text unit, and several candidate text units with the highest semantic relevance to the initial summary text are selected as target text units.

[0061] In this embodiment, a candidate text unit consists of a candidate text block and its adjacent context in the original speech text. In this embodiment, it includes a triplet structure consisting of the preceding original text block as the preamble, the current candidate text block as the core content, and the following original text block as the adjournment, which is used to preserve local contextual information. Among them, the contextual relevance strength measures the degree of relevance between a candidate text unit as a whole and the initial summary text in multiple dimensions, taking into account a variety of indicators such as semantic similarity, word overlap, keyword matching, edit distance, TF-IDF distribution and Jaccard coefficient. The comprehensive weight is a value calculated based on the contextual relevance strength, used to characterize the support of the candidate text unit for the expansion task. The higher the weight, the more critical and relevant the information is.

[0062] Among them, the target text units are selected from all candidate text units, and are the units with the highest comprehensive weight and the strongest semantic relevance to the initial summary text. They are used as input for subsequent summary extraction. The number of target text units selected here is determined based on a preset value.

[0063] Specifically, after obtaining the candidate text block sequence, the system does not directly use these isolated text blocks, but instead constructs a candidate text unit containing its context for each candidate text block. Specifically, for each candidate text block, the system locates its position in the original speech text and automatically extracts the preceding original text block as the "preceding context", itself as the "core content", and the following original text block as the "following context", and the three are combined to form a complete candidate text unit.

[0064] Subsequently, the system calculates the contextual relevance strength between the unit as a whole and the initial summary text using multidimensional similarity indicators. These indicators are weighted or normalized and then fused into a unified contextual relevance strength value, which is used to assign a corresponding comprehensive weight to each candidate text unit. Finally, the system sorts the candidate text units from high to low according to their comprehensive weights and selects the top-ranked candidate text units as the target text units.

[0065] In one specific embodiment, suppose the initial minutes text is: "Approved an increase of 200,000 yuan in the budget for server procurement." A candidate text block reads "An additional 200,000 yuan is needed to buy GPU servers," which is semantically clear on its own. However, if the preceding text is "The marketing department proposed increasing advertising spending," and the following text is "However, finance believes the cost is too high," then the entire candidate text unit may reflect a suggestion that was not adopted, rather than a final decision. Conversely, another candidate text block, "An additional 200,000 yuan for server procurement has been confirmed," is more likely to correspond to a real decision if the preceding text is "The CTO reported insufficient computing power," and the following text is "No objections from all staff." By calculating the comprehensive performance of these two candidate text units under six similarity indicators, the system finds that the latter scores higher in keyword matching, semantic consistency, and contextual coherence, thus assigning it a higher overall weight and selecting it as the target text unit.

[0066] It can be seen that by introducing contextual unit structure and multidimensional association strength assessment, the system can effectively distinguish pragmatic differences such as "proposal" and "resolution", "question" and "confirmation", thereby ensuring that the selected target text unit truly reflects the meeting conclusions, avoiding the introduction of erroneous information due to deviating from the context, and ultimately improving the accuracy and credibility of the expanded results.

[0067] More specifically, in the embodiments of this application, the steps of obtaining the candidate text unit corresponding to each candidate text block, calculating the contextual association strength of each candidate text unit, and obtaining the comprehensive weight corresponding to each candidate text unit, include the following steps for each candidate text block: Obtain the preceding and following text blocks adjacent to the candidate text block from the original speech text, and combine the preceding text block, the candidate text block, and the following text block into the candidate text unit.

[0068] Among them, the preceding adjacent text block refers to the original text block that immediately precedes a candidate text block in the original speech text, and is used to provide the preceding context of the candidate content; the following adjacent text block refers to the original text block that immediately follows a candidate text block in the original speech text, and is used to provide the following context of the candidate content.

[0069] Specifically, the system first locates the position of each candidate text block in the original speech text, and extracts its immediate preceding and following text blocks. These three are then combined sequentially to form a structured candidate text unit. For example, suppose the initial minutes text is: "It is determined that Engineer Zhang will be responsible for the development of the new module." A candidate text block is "I will take over the new module," and the speaker is Engineer Zhang. Its immediate preceding text block is "Manager Li asked: Who is willing to take responsibility?", and its following adjacent text block is "Everyone applauded to show their support."

[0070] Furthermore, based on a preset multidimensional similarity index, the multidimensional association strength between the initial summary text and the preceding adjacent text block, the candidate text block, and the following adjacent text block is evaluated respectively.

[0071] Among them, the multidimensional similarity index refers to a set of calculation methods used to quantify the relevance between texts from different perspectives, including six indicators: cosine similarity (measures the degree of proximity in the semantic vector space), Jaccard coefficient (assesses the overlap ratio of the vocabulary set), BM25, edit distance (calculates the minimum number of character operations between two strings), TF-IDF (Term Frequency-Inverse Document Frequency, used to measure the importance of words in the text), and keyword overlap (detects whether core terms co-occur).

[0072] Furthermore, based on the multidimensional association strength, the weight scores of the preceding adjacent text block, the candidate text block, and the following adjacent text block are determined, and a comprehensive weight corresponding to the candidate text unit is generated based on the weight scores.

[0073] The weighted score is a numerical value assigned to the multidimensional association strength between the preceding adjacent text block, the candidate text block, and the following adjacent text block and the initial summary text, reflecting their information contribution in the current context unit; while the comprehensive weight is a single value obtained by weighted fusion of the weighted scores of the three components, used to characterize the support strength of the entire candidate text unit for the initial summary text.

[0074] Specifically, for each of the three components of this unit—the preceding adjacent text block, the candidate text block, and the following adjacent text block—the association strength with the initial summary text is calculated independently under a preset multi-dimensional similarity index. That is, for each component, the cosine similarity, Jaccard coefficient, BM25 score, edit distance, TF-IDF similarity, and keyword overlap with the initial summary text are calculated to form a multi-dimensional association strength vector. Based on these multi-dimensional association strengths, the system assigns a weighted score to each component, typically through a weighted average or a learning-based fusion strategy, ensuring that components with strong semantic relevance, high keyword matching, and good contextual coherence receive higher scores. Finally, the weighted scores of the three components are integrated (e.g., through weighted summation or taking the maximum value) to generate the overall comprehensive weight of the candidate text unit.

[0075] More specifically, some candidate text blocks may be highly relevant to the initial summary text, but their context contains irrelevant or even contradictory information; conversely, some candidate text blocks may appear to have only minor relevance, but their surrounding text provides crucial background information, making them more interpretable overall. By evaluating and then merging components, the system can more finely distinguish these situations and dynamically adjust the contribution weights of each part, thereby generating a more accurate and robust overall weight. This not only improves the accuracy of target text unit selection but also provides a more context-aware input foundation for subsequent summary generation.

[0076] For example, regarding the candidate text units in the aforementioned specific embodiments, the system evaluates the multidimensional correlation strength between these three parts and the initial minutes text. One possible scenario is that the candidate text block, containing "new module" and semantically pointing to "responsible," scores highly in cosine similarity and keyword overlap. The preceding adjacent text block, although lacking keywords, can identify the initiation point of the action "responsible" through TF-IDF and contextual analysis, thus possessing certain contextual value. The following adjacent text block, while lacking direct keywords, implies that a resolution has been reached through "support," enhancing the certainty of the statement. Therefore, all three parts receive reasonable weight scores, resulting in a high overall weight.

[0077] In contrast, another candidate text block, "I can try the new module," while semantically similar, is followed by the text block "However, there might not be enough time," indicating uncertainty. The adjacent text block scores lower in the multi-dimensional indicators, dragging down the overall weight. It is precisely because the system performs multi-dimensional evaluations of the three parts of the context and generates differentiated weighted scores that it can effectively distinguish between "explicit commitments" and "tentative statements." This ensures that the final selected target text unit truly reflects the meeting's decisions, avoids misjudging the speaker's intent due to ignoring context, and significantly improves the factual accuracy and pragmatic rationality of the expanded results.

[0078] After completing the initial mixed retrieval and constructing candidate text units, in order to further improve the accuracy of relevance judgment, the system introduces a refined reordering stage to achieve the step of selecting several candidate text units with the highest semantic relevance to the initial summary text as target text units, specifically including: The initial summary text is concatenated with each of the candidate text blocks to form multiple query candidate pairs.

[0079] The query candidate pair is a combined text composed of the initial summary text and a candidate text block, which is used as input to the reordering model to evaluate the consistency of the two in terms of overall semantic intent. The reordering model is a pre-trained and fine-tuned sequence pair classification model (e.g., based on the Transformer architecture) specifically designed to determine whether two text fragments are highly matched in semantic intent and output a numerical value that reflects the strength of their relevance, namely the intent score.

[0080] Furthermore, each of the query candidate pairs is input into a preset reordering model, and the intent score corresponding to each of the query candidate pairs is calculated through the reordering model.

[0081] The intent score is a scalar score calculated by the reordering model for a given query candidate pair. It is used to quantify the degree of matching between the initial summary text and the corresponding candidate text block at the deep semantic intent level. The higher the score, the more likely the two are to express the same decision, event, or conclusion.

[0082] Further, the candidate text block sequence is reordered in descending order according to all the intent scores to obtain an intent matching sequence, and the candidate text units corresponding to the first preset number of candidate text blocks in the intent matching sequence are selected as the target text units.

[0083] The intent matching sequence refers to the new sequence obtained by reordering the original candidate text block sequence according to the intent score corresponding to each candidate text block, and is arranged in descending order of intent score from high to low; the target text unit is the finally selected candidate text unit corresponding to the first preset number of candidate text blocks in the intent matching sequence, which serves as a high-confidence context source for subsequent processing.

[0084] Specifically, after completing the initial mixed retrieval and constructing candidate text units, the system introduces a refined reordering stage to further improve the accuracy of relevance judgment. In this stage, the initial summary text is first concatenated with each candidate text block to form multiple independent query candidate pairs. Each pair is usually in the form of "[initial summary text][SEP][candidate text block]", where [SEP] is the separator to help the model distinguish between the two parts.

[0085] Subsequently, these query candidate pairs are input in batches into a pre-defined reordering model. This reordering model, based on its internal deep semantic understanding capabilities (e.g., attention mechanisms focusing on key semantic components), jointly encodes each text pair and outputs an intent score. This score reflects whether the candidate text blocks truly support or correspond to the core intent expressed in the initial summary text, rather than merely superficially similar. Next, the system globally reorders the original candidate text block sequence based on all intent scores, generating an intent matching sequence. Finally, a pre-defined number (e.g., 3) of the top-ranked candidate text blocks are selected from this sequence, and their corresponding complete candidate text units (including the preceding, core, and following adjacent text blocks) are taken as the target text units.

[0086] In one specific embodiment, suppose the initial minutes text is: "The meeting rejected the proposal to outsource testing." In the candidate text block sequence, there are two text blocks with high recall scores: A is "It is recommended that testing be outsourced to a third-party company," and B is "Everyone agreed that testing must be done by the internal team." From a lexical or semantic vector perspective only, A, containing keywords such as "outsource" and "testing," may rank higher in the initial search; B, although not mentioning "outsource," expresses the opposite stance. During the re-ranking stage, the system concatenates the initial minutes text with A and B respectively to form query candidate pairs.

[0087] The reordering model identified that the combined intent of A was "the proposal was made," while the initial minutes' intent was "the proposal was rejected," indicating a conflict between the two intents and resulting in a lower intentional intent score. However, in the combination of B, "internal completion" implicitly rejects outsourcing, aligning with the decision-making intent of "rejecting the outsourcing proposal," thus resulting in a significantly higher intent score. Ultimately, B ranked higher in the intent matching sequence, and its corresponding candidate text unit was selected as the target text unit. By introducing an intent alignment mechanism based on the reordering model, the system can penetrate literal wording and accurately capture the true conclusions of the meeting, thereby avoiding misjudging "discussed proposals" as "adopted resolutions." This ensures that the expanded content faithfully reflects the meeting results, greatly improving the factual accuracy and logical consistency of the generated minutes.

[0088] In step S5, all the target text units and their corresponding comprehensive weights are input into a preset first language model, and the first language model extracts the retrieval summary text based on the target text units and the comprehensive weights.

[0089] The first language model refers to a pre-trained large language model specifically designed for information condensation and noise filtering. In this step, it undertakes the task of summarizing and extracting information. Its input includes target text units and their corresponding comprehensive weights, and its output is a redundancy-free, high-fidelity retrieval summary text. The retrieval summary text is a concise text generated by the first language model that integrates and extracts the content of all target text units. It focuses on the core facts related to the initial summary text, removes duplicate, irrelevant, or low-confidence information, and serves as an enhanced context for subsequent expansion.

[0090] Specifically, firstly, all target text units and their corresponding comprehensive weights are input into the first language model. The model automatically identifies text units with higher comprehensive weights and prioritizes their core content. Then, through optimized prompting engineering, the model is guided to remove redundant information, background noise, slips of the tongue, and other noise from the text units, while retaining key discussion details, decision-making basis, and other effective information. Finally, based on the priority ranking of comprehensive weights, the model organizes the filtered content into a coherent and accurate search summary text.

[0091] More specifically, in this embodiment, units with high overall weight are considered more credible and relevant sources of fact, and their content is preferentially retained and emphasized in the summary; while parts with lower weight, even if included, are weakened or omitted. Simultaneously, the model is explicitly instructed to perform the role of a "knowledge condensation assistant," tasked with removing repetitive statements, filtering out verbal errors and redundancy, eliminating irrelevant details, and outputting only concise, coherent summaries that faithfully reflect the original statement.

[0092] Based on this concept, this application further proposes that the first language model is configured with first preset parameters; the first preset parameters refer to a set of generation control parameters set for the first language model to constrain its output behavior and ensure that the summary extraction process is stable, concise, and focused. In this embodiment, the first preset parameters include a temperature set to 0.3 to reduce randomness, a top-p sampling set to 0.85 to balance diversity and relevance, a repetition penalty coefficient set to 1.05 to suppress redundancy, and a maximum number of generated tokens limited to within 400 to ensure concise output.

[0093] Based on the aforementioned first preset parameters, the step of obtaining the retrieval summary text through the first language model based on the target text unit and the comprehensive weight includes: The first preset prompt word is input into the first language model, and the first language model determines the first target task based on the first preset prompt word.

[0094] The first preset prompt is a pre-designed structured instruction text used to guide the first language model to clarify its role and task objectives. For example, "You are a knowledge condensation assistant. Based on the following weighted context information, please remove duplicate and irrelevant content and generate a concise, accurate, and coherent summary without explaining the source or using a list format." The first target task is the core generation task defined by the first preset prompt, which is to intelligently filter and fuse the input target text units and their comprehensive weights to output a denoised, deredundant, and high-fidelity search summary text.

[0095] Subsequently, constrained by the first preset parameters, the first target task is executed based on the target text unit and the comprehensive weight to obtain the retrieval summary text.

[0096] Specifically, by coordinating the first preset prompt words with the first preset parameters, the system precisely guides the first language model to perform the summarization task. Specifically, the system first inputs the first preset prompt words into the first language model, enabling it to understand its role as a "knowledge condensation assistant" before generation and to comprehend the task requirements: The model focuses on high-weighted content, removes repetitive or off-topic information, and maintains sentence coherence. Within this semantic framework, the model receives all target text units and their corresponding weights as contextual input, where the structure (preceding context, core content, following context) and weight score of each unit are explicitly presented. Subsequently, under strict constraints of the first preset parameters, the model initiates a controlled text generation process: low temperature (0.3) ensures stable output and avoids conjecture; top-p=0.85 allows for moderate lexical diversity but prevents deviation from the topic; a repetition penalty mechanism suppresses multiple expressions of the same viewpoint; and a maximum token limit forces the model to complete information compression within a finite length.

[0097] In one specific embodiment, assume the target text unit contains two statements: a high-weight unit (overall weight 0.9) states "An additional 200,000 yuan has been confirmed for server procurement, to be used to purchase A100 GPUs," and a low-weight unit (overall weight 0.5) states "Someone suggested using H80, but this was rejected due to its long delivery cycle." A first preset prompt instructs the model to "only retain the decision result and key evidence, removing the discussion process and rejected options." Under the constraints of the first preset parameters (temperature 0.3, maximum 400 tokens), the model ignores the "H80" suggestion details in the low-weight section, focuses on the confirmation information in the high-weight section, and generates the search summary text: "The meeting confirmed an additional 200,000 yuan for server procurement, to be used to purchase A100 GPUs." Without pre-defined prompts, the model might improvise, adding unnecessary background information such as "despite other suggestions"; without temperature restrictions, it might generate uncertain statements like "perhaps we could consider the A100 or other models." Because the first pre-defined prompts clearly define the task objective of "condensing facts and removing extraneous details," and the first pre-defined parameters enforce a concise and definite output style, the search summary text accurately focuses on the resolved items, avoiding information overload and factual distortion, thus providing credible enhanced input for subsequent expansion.

[0098] In step S6, the text block to be expanded in the initial summary text is determined, and the text block to be expanded and the search summary text are input into a preset second language model. The second language model expands the text block to be expanded based on the search summary text to update the initial summary text to expanded summary text.

[0099] Among them, the text blocks to be expanded refer to the segments in the initial minutes text that are judged to have insufficient information density, lack of details, or need to be supplemented with background. They are usually characterized by high summarization, lack of decision-making basis, action subject, or execution conditions, etc.

[0100] The second language model refers to a pre-trained large language model specifically designed for content expansion and language refinement. At this stage, it is responsible for transforming concise statements into detailed and standardized meeting minutes, and its generation behavior is strictly constrained by the retrieved abstract text. The expanded minutes text is an updated version of the meeting minutes obtained after processing by the second language model. It retains the original structure and core conclusions while incorporating specific details, logical chains, and contextual background into the parts to be expanded, resulting in a complete and fluent formal output.

[0101] Specifically, the core of this step lies in using the retrieved summary text as fact anchors to drive the second language model to perform evidence-based augmentation on the text blocks to be expanded in the initial summary text. The system first analyzes the initial summary text to identify which text blocks belong to the text to be expanded. Then, these text blocks to be expanded, together with the retrieved summary text, are organized into input prompts and fed into the second language model. In the prompts, the text blocks to be expanded are placed at the core, with the retrieved summary text serving as background supplement, forming a clearly hierarchical contextual structure. Based on this, the second language model, using the facts in the retrieved summary text as the sole reliable source, semantically expands the text blocks to be expanded. Finally, the expanded content replaces the corresponding parts in the original initial summary text, generating the complete expanded summary text.

[0102] More specifically, in this application embodiment, the criteria for identifying the text to be expanded include whether it contains substantive content, whether the information is too brief, and whether it lacks an execution subject or background description. Based on this concept, this application embodiment further proposes that the step of determining the text block to be expanded in the initial summary text includes: Extract the information completeness corresponding to each of the aforementioned minutes text blocks, and determine whether the information completeness reaches a preset completeness threshold.

[0103] Among them, the minutes text block refers to the independent sentences or paragraphs that are logically or semantically segmented from the initial minutes text. Each block expresses a relatively complete meeting matter, decision, or action item. Information completeness is a numerical indicator used to quantify the sufficiency of key elements (such as subject, action, object, time, reason, condition, etc.) contained in a minutes text block. It can be evaluated through rule matching, keyword coverage, syntactic structural completeness, or lightweight classification models.

[0104] Furthermore, if the information completeness does not reach the completeness threshold, the corresponding minutes text block is determined as the text block to be expanded.

[0105] The completeness threshold is a preset judgment standard value used to distinguish whether the minutes text block has sufficient information density; if the information completeness is lower than the threshold, it is considered that the content is too brief and needs to be expanded.

[0106] Specifically, the system first divides the initial minutes text into multiple minutes text blocks, and then independently extracts the information completeness of each block. This extraction process can be based on multi-dimensional features: for example, whether the implementing entity is clearly indicated (e.g., "Zhang Gong is in charge"), whether it contains specific action verbs (e.g., "approved", "postponed", "allocated"), whether it specifies the object or scope (e.g., "200,000 yuan budget", "user survey"), and whether it includes contextual elements such as time, conditions, or basis. These features are comprehensively calculated into a normalized information completeness score.

[0107] Next, the system compares the score with a preset completeness threshold: if the information completeness of a certain minutes text block does not reach the threshold, it indicates that its content is missing key details and is insufficient to support subsequent task execution or result traceability, and is therefore marked as a text block to be expanded; conversely, if the information completeness meets the threshold, it is retained as is and does not enter the expansion process. This on-demand filtering mechanism ensures that expansion resources are only applied to fragments that truly need supplementation, avoiding redundant or disruptive modifications to already complete content.

[0108] In one specific embodiment, assume that the initial summary text contains two summary text blocks: Block A is "The Marketing Department will submit a promotion plan before June 10th," and Block B is "User experience needs to be optimized." The system analyzes Block A and identifies that it contains a clear subject ("Marketing Department"), a specific action ("submission"), an object ("Promotion Plan"), and a deadline ("before June 10th"), with an information completeness score of 0.88, which is higher than the preset completeness threshold of 0.6, so it is not considered as a text block to be expanded; while Block B only contains a vague action ("optimization") and a generalized object ("User Experience"), lacking optimization direction, responsible person, measurement criteria, or time nodes, with an information completeness score of 0.32, which does not meet the threshold, so it is identified as a text block to be expanded.

[0109] In subsequent processes, the system will only expand block B in conjunction with the search summary text (e.g., "The product manager proposed that the homepage loading speed needs to be improved to within 2 seconds, to be completed by the front-end team within two weeks"), generating "The front-end team needs to optimize the product homepage loading speed to within 2 seconds within two weeks to improve user experience; this task was proposed and followed up by the product manager." It is precisely because of this precise filtering through comparison of information completeness and thresholds that the system avoids unnecessary rewriting of block A (which could potentially disrupt its clear structure), while ensuring that block B receives sufficient and actionable supplementary details, thereby improving the overall professionalism, operability, and factual accuracy of the expanded summary text.

[0110] After selecting the text blocks that need to be expanded, the second language model is guided to complete a high-quality, controlled expansion task through the synergistic effect of prompting engineering and parameter constraints. In this embodiment, the second language model is configured with second preset parameters; these second preset parameters refer to a set of generation control parameters set for the second language model to balance creativity, fluency, and factual fidelity during the expansion process. These parameters can be configured to include a temperature of 0.4 to allow moderate linguistic diversity but avoid excessive divergence, a top-p sampling of 0.85 to focus on high-probability word sequences, a repetition penalty coefficient of 1.1 to suppress redundant expressions, and a maximum generated token limit of 1024 to ensure sufficient content expansion without excessive verbosity.

[0111] In this configuration, the step of expanding the text block to be expanded based on the retrieved summary text to update the initial summary text to the expanded summary text includes: The second preset prompt word is input into the second language model, and the second language model determines the second target task based on the second preset prompt word.

[0112] The second preset prompt is a structured instruction text used to clarify the role and output specifications of the second language model. For example, "You are a professional meeting secretary. Based solely on the provided background information, please expand the following abbreviated minutes items into a formal, complete, and coherent meeting minutes text, without adding explanations, using lists, or including thought processes, and only output the expanded content." Specifically, the system first inputs the second preset prompt word into the second language model, so that it understands its role as "conference secretary" before generation and understands the task boundaries: it can only expand based on the facts in the search summary text and must not introduce external knowledge or subjective speculation; the output must be plain text and is prohibited from containing meta-descriptions (such as "according to the above" or "it can be inferred"), list symbols or model reflexive statements.

[0113] Furthermore, constrained by the second preset parameter, the second target task is performed based on the text block to be expanded and the search summary text to obtain the expanded text block corresponding to the text block to be expanded.

[0114] The second objective task is the core task defined by the second preset prompt, which is to enrich the semantics and standardize the language of the text block to be expanded, under the premise of strictly following the facts provided by the search summary text, and generate an expanded text block that is detailed, logically clear and formatted in accordance with regulations.

[0115] Specifically, under the strict constraints of the second preset parameters, the model receives the text block to be expanded and the retrieved summary text as contextual input, and initiates the expansion generation process. The text block to be expanded serves as the starting point for expansion, and the retrieved summary text serves as the only reliable source of facts. The model needs to extract supporting details from these, including but not limited to specific figures, responsible persons, technical terms, and decision-making basis, and organically integrate them into the expanded result. The generation process is controlled by parameters such as temperature, top-p, repetition penalty, and length limit to ensure that the language is natural and fluent while maintaining factual accuracy and structural compactness.

[0116] Further, the initial summary text is updated based on all the expanded text blocks to obtain the expanded summary text.

[0117] The expanded text block is an enhanced version of a text block to be expanded, obtained after processing by a second language model. It includes key elements such as the subject, object, time, reason, and conditions, and the language style conforms to the requirements of formal meeting minutes.

[0118] Specifically, each text block to be expanded is replaced with its corresponding expanded text block, while all non-expanded parts remain unchanged, together forming a complete expanded minutes text.

[0119] More specifically, through role-based definition, fact-anchoring, and parameter constraints, the second language model can still stably output professional and credible expanded content even under lightweight deployment on the device. Prompt words eliminate the illusion of unrestrained expression, the retrieved summary text provides high-quality factual evidence, and preset parameters ensure consistency and efficiency in output style.

[0120] In one specific embodiment, assuming the text block to be expanded is "approved procurement", the retrieved summary text is "The CTO applied to purchase two A100 servers with a budget of 600,000 yuan to accelerate large model training. After financial review, it was unanimously approved by all members." The second preset prompt requires "only outputting the expanded text, without explaining the source". Under the constraints of the second preset parameters (temperature 0.4, maximum 1024 tokens), the second language model generates the expanded text block: "The meeting approved the procurement application submitted by the CTO, agreeing to purchase two A100 servers with a budget of 600,000 yuan, specifically for large model training tasks. This plan has passed financial review and has been unanimously approved by all members." If the second preset prompt is not used, the model may output informal statements containing meta-descriptions, such as "Based on the discussion, we believe it is acceptable to purchase..."; if the temperature is not restricted, it may speculate on unmentioned information such as "expected delivery in Q3".

[0121] Because the second preset prompt clearly defines the output specifications, the second preset parameters suppress randomness and redundancy, and the expansion is strictly limited to the factual scope of the search summary text, the final expanded text block is both detailed and standardized, and completely faithful to the original meeting content, thus ensuring that the expanded minutes text can be directly used for task assignment, budget execution, and audit traceability.

[0122] Reference Figure 3 , Figure 3 This is a virtual structural diagram of the retrieval-enhanced end-side note-taking device provided in this application. A second aspect of this application provides a retrieval-enhanced end-side note-taking device, comprising: The text acquisition module 100 is used to acquire the original speech text and initial summary text corresponding to the original speech data; the original speech text includes multiple original text blocks. The vector encoding module 200 is used to encode each of the original text blocks to obtain multiple text block vectors, and to build a local vector library based on all the text block vectors; The association retrieval module 300 is used to input the initial summary text as a query into the local vector library, and to perform association retrieval between the initial summary text and each original text block through the local vector library to obtain a candidate text block sequence including multiple candidate text blocks; The unit filtering module 400 is used to obtain the candidate text units corresponding to each candidate text block, calculate the context association strength of each candidate text unit to obtain the comprehensive weight corresponding to each candidate text unit, and select several candidate text units with the highest semantic relevance to the initial summary text as target text units. The summary and extraction module 500 is used to input all the target text units and their corresponding comprehensive weights into a preset first language model, and extract the retrieval summary text based on the target text units and the comprehensive weights through the first language model. The minutes expansion module 600 is used to determine the text block to be expanded in the initial minutes text, input the text block to be expanded and the search summary text into a preset second language model, and expand the text block to be expanded based on the search summary text through the second language model to update the initial minutes text into an expanded minutes text.

[0123] The retrieval-enhanced end-side timekeeping device described in this application embodiment can execute the retrieval-enhanced end-side timekeeping method provided in the above embodiments. The retrieval-enhanced end-side timekeeping device has the corresponding functional steps and beneficial effects of the retrieval-enhanced end-side timekeeping method described in the above embodiments. For details, please refer to the embodiments of the retrieval-enhanced end-side timekeeping method described above. The embodiments of this application will not be repeated here.

[0124] This application also provides an electronic device, please refer to... Figure 4 , Figure 4This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. The electronic device may include a processor and a memory, which can be connected via a bus or other means. The processor may be a Central Processing Unit (CPU). The processor may also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, or combinations of the above types of chips. The memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs, non-transitory computer-executable programs, and modules, such as the program instructions / modules corresponding to the retrieval-enhanced end-side timesheet extension method in the embodiments of this application. The processor executes various functional applications and data processing by running the non-transitory software programs, instructions, and modules stored in the memory, thereby implementing the retrieval-enhanced end-side timesheet extension method in the above method embodiments.

[0125] The memory may include a program storage area and a data storage area. The program storage area may store the operating system and at least one application program required for a function; the data storage area may store data created by the processor, etc. Furthermore, the memory may include high-speed random access memory and non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. The one or more modules are stored in the memory and, when executed by the processor, perform the retrieval-enhanced edge-side timesheet extension method as described in the above method embodiments. Specific details of the above electronic device can be understood by referring to the corresponding descriptions and effects in the above method embodiments, and will not be repeated here. Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it may include the processes of the embodiments of the above methods. The storage medium may be a read-only memory (ROM), a random access memory (RAM), a flash memory, a hard disk drive (HDD), or a solid-state drive (SSD), etc.; the storage medium may also include a combination of the above types of memory.

[0126] Numerous specific details are set forth in the specification provided herein. However, it will be understood that embodiments of this application may be practiced without these specific details. In some instances, well-known methods, structures, and techniques have not been shown in detail so as not to obscure the understanding of this specification.

[0127] Similarly, it should be understood that, in order to streamline this disclosure and aid in understanding one or more of the various inventive aspects, in the above description of exemplary embodiments of this application, various features of this application are sometimes grouped together in a single embodiment, figure, or description thereof. However, this approach to disclosure should not be construed as reflecting an intention that the claimed application requires more features than expressly recited in each claim. Rather, as reflected in the claims, inventive aspects lie in fewer than all features of a single foregoing disclosed embodiment. Therefore, the claims following the detailed description are hereby expressly incorporated into that detailed description, wherein each claim itself is a separate embodiment of this application.

[0128] It should be noted that the above embodiments are illustrative of this application and not restrictive of this application, and that those skilled in the art can devise alternative embodiments without departing from the scope of the appended claims.

Claims

1. A method for expanding end-side minutes based on retrieval enhancement, characterized in that, The method includes: Obtain the original speech text and initial summary text corresponding to the original speech data; the original speech text includes multiple original text blocks; Each of the original text blocks is encoded to obtain multiple text block vectors, and a local vector library is constructed based on all the text block vectors. The initial summary text is used as the query input to the local vector library. The local vector library is used to perform association retrieval between the initial summary text and each original text block to obtain a candidate text block sequence including multiple candidate text blocks. Obtain the candidate text unit corresponding to each candidate text block, calculate the context association strength of each candidate text unit to obtain the comprehensive weight corresponding to each candidate text unit, and select several candidate text units with the highest semantic relevance to the initial summary text as target text units. All target text units and their corresponding comprehensive weights are input into a preset first language model, and the first language model extracts the retrieval summary text based on the target text units and the comprehensive weights. The text block to be expanded in the initial summary text is determined, and the text block to be expanded and the search summary text are input into a preset second language model. The second language model expands the text block to be expanded based on the search summary text to update the initial summary text to an expanded summary text.

2. The endpoint summary expansion method based on retrieval enhancement according to claim 1, characterized in that, The associated retrieval includes semantic retrieval paths and lexical retrieval paths; The step of associating the initial summary text with each of the original text blocks through the local vector library to obtain a sequence of candidate text blocks including multiple candidate text blocks includes: Encode the initial minutes text into a query vector; The semantic approximation score between the query vector and each text block vector is calculated using the semantic retrieval path, and all original text blocks are sorted according to the semantic approximation score to obtain a semantic path sequence. Furthermore, the vocabulary matching score between the initial summary text and each of the original text blocks is calculated through the vocabulary retrieval path, and all original text blocks are sorted according to the vocabulary matching score to obtain a vocabulary path sequence; Based on a preset smoothing factor, a first sorting score is calculated according to the sorting position of each original text block in the semantic path sequence, and a second sorting score is calculated according to the sorting position of each original text block in the lexical path sequence. Based on the first sorting score and the second sorting score, a comprehensive sorting score is determined for each of the original text blocks. All the original text blocks are sorted according to all the comprehensive sorting scores to obtain the candidate text block sequence.

3. The endpoint summary expansion method based on retrieval enhancement according to claim 1, characterized in that, In the step of obtaining the candidate text unit corresponding to each candidate text block and calculating the contextual association strength of each candidate text unit to obtain the comprehensive weight corresponding to each candidate text unit, the steps performed for each candidate text block include: Obtain the preceding and following text blocks adjacent to the candidate text block from the original speech text, and combine the preceding text block, the candidate text block, and the following text block into the candidate text unit; Based on a preset multidimensional similarity index, the multidimensional association strength between the initial summary text and the preceding adjacent text block, the candidate text block, and the following adjacent text block is evaluated respectively. Based on the multidimensional association strength, the weight scores of the preceding adjacent text block, the candidate text block, and the following adjacent text block are determined, and the comprehensive weight corresponding to the candidate text unit is generated based on the weight scores.

4. The endpoint summary expansion method based on retrieval enhancement according to claim 1, characterized in that, The step of selecting the candidate text units that have the highest semantic relevance to the initial summary text as the target text units includes: The initial summary text is concatenated with each of the candidate text blocks to form multiple query candidate pairs; Each of the query candidate pairs is input into a preset reordering model, and the intent score corresponding to each of the query candidate pairs is calculated through the reordering model. The candidate text block sequence is reordered in descending order based on all the intent scores to obtain an intent matching sequence, and the candidate text units corresponding to the first preset number of candidate text blocks in the intent matching sequence are selected as the target text units.

5. The end-side summary expansion method based on retrieval enhancement according to claim 1, characterized in that, The first language model is configured with a first preset parameter; The step of obtaining the retrieval summary text by extracting it from the target text unit and the comprehensive weight using the first language model includes: The first preset prompt word is input into the first language model, and the first language model determines the first target task based on the first preset prompt word. Constrained by the first preset parameters, the first target task is performed based on the target text unit and the comprehensive weight to obtain the retrieval summary text.

6. The end-side summary expansion method based on retrieval enhancement according to claim 1, characterized in that, The initial minutes text includes multiple minutes text blocks; The step of determining the text block to be expanded in the initial minutes text includes: Extract the information completeness corresponding to each of the aforementioned minutes text blocks, and determine whether the information completeness reaches a preset completeness threshold; If the information completeness does not reach the completeness threshold, the corresponding minutes text block will be determined as the text block to be expanded.

7. The end-side summary expansion method based on retrieval enhancement according to claim 1, characterized in that, The second language model is configured with a second preset parameter; The step of expanding the text block to be expanded based on the retrieved summary text to update the initial summary text to the expanded summary text includes: The second preset prompt word is input into the second language model, and the second language model determines the second target task based on the second preset prompt word. Constrained by the second preset parameter, the second target task is performed based on the text block to be expanded and the search summary text to obtain the expanded text block corresponding to the text block to be expanded; The expanded summary text is obtained by updating the initial summary text based on all the expanded text blocks.

8. A terminal-side summary expansion device based on retrieval enhancement, characterized in that, include: The text acquisition module is used to acquire the original speech text and initial summary text corresponding to the original speech data; The original speech text includes multiple original text blocks; The vector encoding module is used to encode each of the original text blocks to obtain multiple text block vectors, and to build a local vector library based on all the text block vectors; The associated retrieval module is used to input the initial summary text as a query into the local vector library, and to perform associated retrieval between the initial summary text and each of the original text blocks through the local vector library to obtain a candidate text block sequence including multiple candidate text blocks; The unit filtering module is used to obtain the candidate text units corresponding to each candidate text block, calculate the context association strength of each candidate text unit to obtain the comprehensive weight corresponding to each candidate text unit, and select several candidate text units with the highest semantic relevance to the initial summary text as target text units. The summary and extraction module is used to input all the target text units and their corresponding comprehensive weights into a preset first language model, and to extract the retrieval summary text based on the target text units and the comprehensive weights through the first language model. The minutes expansion module is used to determine the text block to be expanded in the initial minutes text, input the text block to be expanded and the search summary text into a preset second language model, and expand the text block to be expanded based on the search summary text through the second language model to update the initial minutes text into expanded minutes text.

9. An electronic device, characterized in that, include: One or more processors; One or more memory units; And one or more computer programs, wherein the one or more computer programs are stored in the one or more memories, the one or more computer programs including instructions that, when executed by the one or more processors, cause the electronic device to perform the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The storage medium stores a program or instructions that, when executed, implement the method as described in any one of claims 1 to 7.

Images