A robot-based intelligent meeting memorandum generation method

Abstract

The present application provides a robot-based intelligent conference memorandum generation method, which includes: collecting multi-channel voice signals through a microphone array, and using beamforming technology to enhance the target speaker's voice; using an adaptive filtering algorithm to separate background noise and improve the quality of the voice signal; extracting voice features and adjusting language model parameters to generate a transcribed text; parsing the text structure, extracting the meeting topic, participants, and resolution content to form structured information; calculating the semantic similarity between knowledge graph nodes, and if it is higher than a preset threshold, associating historical records to generate extended information; constructing a structured memorandum based on the extended information, organizing the meeting topic, participants, resolution content, and associated historical records to generate an initial memorandum; monitoring memorandum editing operations, updating knowledge graph node relationships, and generating a final memorandum document. The present application effectively improves the accuracy, completeness, and usability of meeting records, and significantly improves meeting efficiency.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to a robot-based intelligent conference memorandum generation method. Background Art

[0002] The real-time transcription of voice, extraction of key information, and generation of intelligent memos by portable robots in multi-scenario calls are important research areas in the field of artificial intelligence and voice interaction. The core of this research is to provide users with convenient meeting record-keeping and information management support in complex acoustic environments through efficient voice processing and intelligent analysis technologies. This technology has key application value in scenarios such as business meetings, remote collaboration, and daily communication, and can significantly improve information processing efficiency and user experience. However, existing solutions have obvious limitations in terms of multi-scenario adaptability, real-time performance, and intelligence. Many systems experience reduced speech recognition accuracy in noisy environments, making it difficult to effectively separate the target speaker's voice from background noise. At the same time, the accuracy of key information extraction is insufficient, and the generated memos lack structure and intelligence, which cannot meet users' needs for fast query and editing.

[0003] Therefore, the key issue of this study is how to achieve efficient processing of voice signals, accurate extraction of key information, and automatic generation of structured intelligent memos in multi-scenario calls by optimizing microphone array noise reduction, real-time speech transcription, natural language processing, and semantic understanding technologies under the limited computing resources of portable robots. Summary of the Invention

[0004] The present invention provides a robot-based intelligent conference memorandum generation method, which mainly includes S101. Collecting original voice data through a microphone array to obtain a multi-channel voice signal, adjusting the directional parameters based on the voice signal using beamforming technology, calculating the target speaker's direction angle, and obtaining an enhanced voice signal; S102. Using the enhanced voice signal, using an adaptive filtering algorithm to separate background noise to obtain a high signal-to-noise ratio voice signal; S103. Extracting voice features from the high signal-to-noise ratio voice signal, using a pre-trained model combined with conference domain corpus for fine-tuning, and generating a transcribed text; S104. Parsing the sentence structure of the transcribed text to extract the meeting topic, participants, and resolution content to form a structured information set;

[0005] S105. Based on the structured information set, calculate the semantic similarity between the knowledge graph nodes. If the similarity is higher than a second preset threshold, generate an extended information set by associating the historical records; S106. Construct a structured memorandum based on the extended information set, organize the meeting topic, participants, resolution content and associated historical records, and generate an initial memorandum document; S107. Monitor the editing operations of the initial memorandum document, update the knowledge graph node relationship, and generate a memorandum document.

[0006] Preferably, the step S101 includes:

[0007] Collecting original speech data through the microphone array to obtain a multi-channel signal, and if noise exists in the multi-channel signal, preprocessing it using a spatial filtering technique to remove background noise and obtain a first speech signal;

[0008] According to the first speech signal, the directivity parameters are adjusted using beamforming technology, and the spatial focusing is optimized using a minimum variance distortionless response algorithm to obtain a second speech signal;

[0009] If the direction of the target speaker in the second speech signal is unclear, using a direction angle estimation technique to calculate the direction angle based on phase difference analysis to determine the position of the target speaker;

[0010] adjusting the beamforming parameters according to the direction angle, and enhancing the target speaker signal through spatial filtering to obtain a third speech signal;

[0011] For the third speech signal, noise suppression is performed on the spectral subtraction technology combined with a deep learning noise reduction model to remove residual noise and obtain an enhanced speech signal.

[0012] Preferably, the step S102 includes:

[0013] Based on the voice signal, decomposing the voice signal using a time domain analysis technique to extract a main voice signal;

[0014] For the main voice signal, using spectrum analysis technology to calculate the interference frequency distribution and determine the interference frequency band;

[0015] According to the interference frequency band, a notch filter technology is used to suppress the interference of a specific frequency to obtain a filtered voice signal;

[0016] For the filtered speech signal, an adaptive filtering technology is used to adjust the filter coefficient. If the signal-to-noise ratio is lower than a second preset threshold, the filter coefficient is iteratively updated to obtain a speech signal with a high signal-to-noise ratio.

[0017] Preferably, the step S103 includes:

[0018] Extracting time domain features from the high signal-to-noise ratio speech signal, and obtaining a time-frequency feature distribution by using a feature extraction method combining Mel-frequency cepstral coefficients and a convolutional neural network;

[0019] According to the time-frequency feature distribution, principal component analysis technology is used to perform dimensionality reduction processing to obtain reduced-dimensional speech features;

[0020] For the reduced-dimensional speech feature, if the feature vector variance is lower than a third preset threshold, optimizing the feature by using a regularization technique to obtain an optimized speech feature;

[0021] Extracting semantic information from the optimized speech features, performing feature mapping using a pre-trained language model, and obtaining a semantic feature vector;

[0022] Based on the semantic feature vector, a pre-trained model is used in combination with conference domain corpus for fine-tuning to generate a preliminary transcription text;

[0023] For the preliminary transcribed text, the text semantic similarity is calculated using word embedding technology. If the text semantic similarity is lower than a fourth preset threshold, the language model parameters are iteratively updated until the similarity of the generated transcribed text is greater than or equal to the preset threshold.

[0024] Preferably, the step S104 includes:

[0025] Parsing the sentence structure in the transcribed text using dependency syntax to obtain a syntactic dependency tree;

[0026] Extracting a subject-verb-object structure according to the syntactic dependency tree, identifying the subject, predicate, and object through syntactic role labeling technology, and obtaining a subject-verb-object structure;

[0027] If the subject or object in the subject-verb-object structure contains a noun phrase, decomposing the noun phrase using a keyword extraction technique to obtain a conference keyword set;

[0028] Obtaining named entities from the syntactic dependency tree, annotating names of people and organizations using named entity recognition technology, and obtaining a list of participants;

[0029] According to the subject-verb-object structure and the conference keyword set, a semantic role labeling technology is used to identify core verbs and their associated arguments to obtain a resolution content set;

[0030] With respect to the resolution content set and the participant list, a structured information set is generated through information integration technology.

[0031] Preferably, the step S105 includes:

[0032] Obtaining knowledge graph nodes from the structured information set, and calculating semantic similarities between nodes using semantic analysis technology to obtain a similarity set;

[0033] If the similarity exceeds the preset threshold, the relevant node association data is extracted from the historical records to obtain a set of associated records;

[0034] According to the associated record set, the knowledge graph expansion technology is used to generate the expansion nodes to obtain the expansion node set;

[0035] Node attributes are obtained from the extended node set, and extended information is generated using information integration technology to obtain an extended information set.

[0036] Preferably, the step S106 includes:

[0037] Acquire meeting data from the extended information set, the meeting data including meeting topics, participants, resolution content, and historical records, and generate an extracted data set using data extraction technology;

[0038] generating a classified data set based on the extracted data set by using an information classification technology to obtain a classified data set;

[0039] Obtaining classified data from a classified data set, using integrity verification technology to verify whether the classified data meets a preset integrity condition, and if the verification result indicates that the classified data meets the integrity condition, generating an initial memorandum document using a structured template;

[0040] Obtaining a memo document from an initial memo document, generating a formatted memo using a formatting technique, and determining the relevance between the formatted memo and the historical record using a correlation analysis technique to obtain a correlation analysis result;

[0041] According to the correlation analysis result, the correlation data is obtained from the historical records, and the correlation data is integrated into the formatted memo using information integration technology to obtain an initial memo document.

[0042] Preferably, the step S107 includes:

[0043] Obtaining editing operations from the initial memo document, extracting operation data using an operation capture technology, and obtaining an operation data set;

[0044] Extracting operation data from the operation data set and generating an operation feature set using data analysis technology;

[0045] Acquire operation features from the operation feature set, and use graph mapping technology to update the node relationship of the knowledge graph to obtain an updated node relationship set;

[0046] Extracting node relationships from the updated node relationship set, and using relationship verification technology to determine whether the node relationships meet a preset consistency condition, and if so, determining an updated knowledge graph;

[0047] Node relationships are obtained from the updated knowledge graph, and memorandum documents are generated using document generation technology to obtain a memorandum document set.

[0048] The technical solution provided by the embodiment of the present invention may include the following beneficial effects: The present invention discloses a robot-based intelligent conference memorandum generation method. In response to the challenges of speech recognition and content extraction in a multi-person conference environment, beamforming technology and adaptive filtering algorithms are used to improve the quality of voice signals, and domain-adaptive language models are used to achieve accurate transcription. Through semantic analysis, key information of the meeting is extracted and a knowledge graph is constructed to generate a structured memorandum. The present invention also integrates historical record extension information, optimizes the memorandum content and establishes a semantic index, and finally forms an interactive dynamic memorandum document. This method can effectively improve the accuracy, completeness and usability of meeting records, provide enterprises with intelligent meeting management solutions, and significantly improve meeting efficiency and knowledge accumulation. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 This is a flowchart of a robot-based intelligent meeting memorandum generation method of the present invention.

[0050] Figure 2 1 is a flow chart of step S101 of the present invention.

[0051] Figure 3 1 is a flow chart of step S106 of the present invention. DETAILED DESCRIPTION

[0052] The following will describe the technical solutions in the embodiments of the present invention in detail with reference to the accompanying drawings. The described embodiments are only a part of the embodiments of the present invention.

[0053] As attached Figure 1 In this embodiment, a robot-based intelligent conference memorandum generation method may specifically include:

[0054] Step S101 , collecting original speech data through a microphone array to obtain a multi-channel speech signal, using beamforming technology to adjust the directivity parameters, calculating the direction angle of the target speaker, and obtaining an enhanced speech signal.

[0055] In this embodiment, see the attached Figure 2, collect original voice data through a microphone array to obtain a multi-channel voice signal. If there is noise in the multi-channel signal, spatial filtering technology is used for preprocessing to remove background noise and obtain a first voice signal. According to the first voice signal, beamforming technology is used to adjust the directivity parameters, and the minimum variance distortion-free response algorithm is used to optimize spatial focusing to obtain a second voice signal. If the direction of the target speaker in the second voice signal is unclear, the direction angle is calculated based on phase difference analysis through directional angle estimation technology to determine the position of the target speaker. According to the directional angle, the beamforming parameters are adjusted, and the target speaker signal is enhanced through spatial filtering to obtain a third voice signal. For the third voice signal, the spectral subtraction technology is combined with the deep learning noise reduction model to perform noise suppression, remove residual noise, and obtain an enhanced voice signal.

[0056] For example, a microphone array can be used to collect multi-channel speech signals. A linear microphone array consisting of four microphones spaced 5 cm apart and placed inside a portable robot can be used to collect raw speech data. The signal sampling rate is 16 kHz, and time series data is generated for each channel. The collected signal contains the target speaker's voice, air conditioning noise, and distant noise, requiring further processing to improve speech quality.

[0057] In one possible implementation, the noise in the multi-channel signal can be pre-processed by spatial filtering, for example, using a delay and sum filter to suppress background noise from non-target directions based on the time difference between microphones.

[0058] Specifically, assuming the noise comes primarily from the side, the filter parameters are adjusted to attenuate the signal at a 90-degree angle, retaining the front voice and outputting the first voice signal. This process effectively reduces ambient noise and improves the signal-to-noise ratio.

[0059] Specifically, based on the first voice signal, beamforming technology can be used to optimize directivity.

[0060] Preferably, the minimum variance distortion-free response algorithm is used to focus on the target direction. For example, if the beam is set to point toward the center of a conference table, the minimum variance distortion-free response algorithm adaptively adjusts weights to suppress lateral interference and output a second speech signal. This signal has higher speech clarity in the target direction and is suitable for multi-person conference scenarios.

[0061] It should be noted that if the target speaker's direction is unclear in the second speech signal, directional angle estimation can be used to locate the speaker. For example, based on phase difference analysis, the arrival time difference of the signals received by each microphone is calculated to estimate the directional angle. Assuming the calculated angle is 30 degrees, this indicates that the speaker is located slightly to the right of the front of the array. The beamforming parameters are adjusted to focus on the 30-degree direction, and the target signal is enhanced through spatial filtering to output the third speech signal. This method improves the separation of the target speech and is suitable for dynamic scenes.

[0062] For example, spectral subtraction may be used to suppress the residual noise in the third speech signal.

[0063] In one embodiment, the signal spectrum is analyzed to identify noise-dominated frequency bands, such as high-frequency air conditioning noise. The estimated noise spectrum is then subtracted, preserving the dominant speech frequency components to produce an enhanced speech signal. This significantly reduces background noise, resulting in clearer speech and suitable for input into a speech recognition system.

[0064] As you can see, the combination of these technologies forms a complete voice enhancement chain, progressively improving voice quality from acquisition to final output. Spatial filtering reduces background interference, beamforming focuses on the target, directional angle estimation dynamically adjusts, and spectral subtraction refines noise removal. The overall effect is an approximately 10dB improvement in signal-to-noise ratio and improved speech intelligibility, making it suitable for intelligent conferencing systems and ensuring efficient and robust far-field voice interaction.

[0065] Step S102, using an adaptive filtering algorithm to separate background noise based on the enhanced speech signal to obtain a speech signal with a high signal-to-noise ratio;

[0066] In an embodiment of the present application, based on the enhanced speech signal, the time domain analysis technology is used to decompose the speech signal and extract the main speech signal; for the main speech signal, the spectrum analysis technology is used to calculate the interference frequency distribution and determine the interference frequency band; based on the interference frequency band, the notch filtering technology is used to suppress specific frequency interference to obtain a filtered speech signal; for the filtered speech signal, the adaptive filtering technology is used to adjust the filter coefficient. If the signal-to-noise ratio is lower than a second preset threshold, the filter coefficient is iteratively updated to obtain a high signal-to-noise ratio speech signal.

[0067] For example, for decomposing signal components using a time domain analysis technique, the enhanced speech signal may be decomposed into a plurality of time domain sub-bands by short-time Fourier transform to extract the main speech signal.

[0068] In one possible implementation, consider a conference room scenario where a microphone samples a signal at a 16kHz rate, containing the target speaker's voice and background noise, such as the low-frequency hum of an air conditioner. Time-domain analysis first identifies periods of high signal energy and identifies them as the primary speech region. For example, a 20-millisecond time window is set. The energy peaks in each window are analyzed, and the energy-concentrated regions are extracted as the primary speech signal, generating the primary speech signal. This process effectively separates speech from persistent, low-energy noise. For example, if the spectrum shows that the air conditioner noise dominates the 200Hz to 300Hz frequency band, it can be marked as an interference band. This analysis provides precise targets for subsequent filtering, avoiding accidental interference with speech bands.

[0069] In one embodiment, for interference frequency bands, notch filtering technology can effectively suppress specific frequency interference.

[0070] Preferably, a notch filter with a center frequency of 250Hz and a bandwidth of 50Hz is designed to attenuate signals in this frequency band. For example, in a conference scenario, the filtered voice signal output after filtering significantly reduces air conditioning noise, while retaining the main voice frequency components, such as those from 500Hz to 2kHz. This method accurately eliminates interference and maintains the naturalness of speech.

[0071] For filtered speech signals, adaptive filtering technology can further optimize signal quality. In one possible implementation, a recursive least squares (RLS) algorithm is used to dynamically adjust the filter coefficients. Assuming an initial signal-to-noise ratio (SNR) of 5dB, which is 10dB below a preset threshold, the algorithm iteratively updates the coefficients, gradually enhancing the target speech. For example, each iteration analyzes 100 milliseconds of the signal and adjusts the filter to suppress residual high-frequency noise, outputting a high SNR speech signal. This process improves speech clarity and adapts to dynamic noise environments.

[0072] Step S103 , extracting speech features from the high signal-to-noise ratio speech signal, using a pre-trained model combined with conference domain corpus for fine-tuning, and generating a transcribed text.

[0073] In this embodiment, time-domain features are extracted from the high signal-to-noise ratio speech signal, and the time-domain features are decomposed using a feature extraction method combining Mel-frequency cepstral coefficients (MFCCs) and a convolutional neural network (CNN) to obtain a time-frequency feature distribution. Based on the time-frequency feature distribution, principal component analysis is used to perform dimensionality reduction processing to obtain reduced-dimensional speech features. For the reduced-dimensional speech features, if the feature vector variance is lower than a third preset threshold, regularization is used to optimize the features to obtain optimized speech features. Semantic information is extracted from the optimized speech features, and feature mapping is performed using a pre-trained language model to obtain a semantic feature vector. Based on the semantic feature vector, a preliminary transcription is generated using the pre-trained model and fine-tuned with conference-related corpus. For the preliminary transcription, text semantic similarity is calculated using word embedding technology. If the text semantic similarity is lower than a fourth preset threshold, the language model parameters are iteratively updated until the similarity of the generated transcription is greater than or equal to the preset threshold. For example, when extracting time-domain features from a high signal-to-noise ratio speech signal, a feature extraction method combining Mel-frequency cepstral coefficients and a convolutional neural network can be used to decompose the signal. Specifically, the original speech signal (sampling rate 16kHz) is framed with a 25ms frame length (400 samples) and a 10ms frame shift (160 samples). A Hamming window is used to reduce spectral leakage. A power spectrum is obtained by performing a FFT on each frame. Twenty triangular filters (frequency range 0-8kHz) are applied to simulate the human auditory characteristics and compress the frequency domain information. The filter bank energy is logarithmically computed and then subjected to a DCT to obtain 13-dimensional MFCC coefficients. First- and second-order differences are then added to form a 39-dimensional feature vector, forming a time-frequency feature matrix. This feature matrix is ​​then converted into a time-frequency image (time axis × frequency axis) as input to a convolutional neural network. The output is a 64-dimensional time-domain feature vector for subsequent classification or recognition.

[0074] In one possible implementation, principal component analysis (PCA) is used to reduce dimensionality based on the time-frequency feature distribution. PCA calculates the feature covariance matrix and extracts the principal components that contribute most to the variance, preserving the key speech features. For example, if the time-frequency features contain 1000 dimensions, PCA retains the first 50 dimensions, covering 90% of the variance, to generate reduced-dimensional speech features. This dimensionality reduction reduces computational complexity while preserving core speech information.

[0075] It should be noted that if the variance of the feature vector of the reduced-dimensional speech feature falls below a third preset threshold, such as 0.8, regularization techniques can be used to adjust the feature weights. Regularization enhances the discriminability of features by introducing a penalty term. For example, L2 regularization can be used to adjust the weights to enhance the influence of low-variance features, thereby generating optimized speech features. This approach improves the robustness of the feature and adapts it to different speaker styles.

[0076] Specifically, semantic information can be extracted from optimized speech features using pre-trained language models for feature mapping. Pre-trained models such as BERT use deep networks to convert speech features into semantic feature vectors. For example, by inputting optimized speech features into the model, semantic feature vectors are generated that capture the meaning of key words in a meeting, such as "plan" and "budget." This mapping ensures that subsequent text generation is more semantically accurate.

[0077] In one embodiment, based on semantic feature vectors, a general speech or language pre-trained model is loaded and model parameters are adjusted in conjunction with conference corpus. Speech features are then input to generate accurate text for conference scenarios. Specifically, the pre-trained model type is adapted as follows: Speech recognition model: Open source pre-trained models (such as Wav2Vec2.0 and DeepSpeech) are selected to directly process speech features to generate text. Language model: If combined with speech recognition post-processing, general language models (such as the GPT series and BERT) can be used to optimize the syntax and semantics of the transcribed text. Pre-trained parameters are loaded from public weight libraries (such as HuggingFace) as the basis for feature extraction and text generation. Adaptive strategy: Parameter Efficient Fine-tuning (PEFT): Most parameters of the pre-trained model are frozen, and only specific layers (such as the last few layers of the language model) are fine-tuned to reduce computational effort. Adversarial training: Domain Adversarial Network (DANN) is introduced, and a discriminator is used to distinguish between general and conference domain data, enabling the model to learn domain-independent features. Transfer learning: Tasks are customized for conference scenarios (such as correcting incorrect professional terminology or identifying conference-specific sentence patterns), and parameters are updated through backpropagation. Fine-tune parameter settings: Learning rate: Use a dynamic adjustment strategy (such as the AdamW optimizer, with an initial learning rate of 1e-5); training rounds: Set to 3-8 rounds based on the data size, combined with early stopping on the validation set to prevent overfitting, and quickly generate structured conference transcripts. This approach effectively adapts the pre-trained model to the speech transcription needs of conferences, significantly improving the accuracy and professionalism of text generation in high-signal-to-noise ratio speech scenarios.

[0078] Preferably, for the preliminary transcribed text, the text semantic similarity is calculated using word embedding technology. Word embedding converts the text into vectors and calculates the cosine similarity between the vectors. For example, if the similarity between the preliminary transcribed text and the reference text is lower than a fourth preset threshold, such as 0.9, the language model parameters are iteratively updated to continuously optimize the transcribed text, wherein the fourth preset threshold is set based on the extremely high similarity standard of the medical scenario and the extremely strict requirements for semantic consistency. This method ensures the semantic consistency of the text.

[0079] Step S104: parsing the sentence structure of the transcribed text, extracting the meeting theme, participants, and resolution content, and forming a structured information set;

[0080] In an embodiment of the present application, dependency syntax is used to parse the sentence structure in the transcribed text to obtain a syntactic dependency tree; based on the syntactic dependency tree, a subject-predicate-object structure is extracted, and the subject, predicate and object are identified by syntactic role labeling technology to obtain a subject-predicate-object structure; if the subject or object in the subject-predicate-object structure contains a noun phrase, the noun phrase is decomposed by keyword extraction technology to obtain a conference theme word set; named entities are obtained from the syntactic dependency tree, and named entity recognition technology is used to annotate the names of people and organizations to obtain a list of participants; based on the subject-predicate-object structure and the conference theme word set, semantic role labeling technology is used to identify core verbs and their associated arguments to obtain a resolution content set; for the resolution content set and the list of participants, a structured information set is generated by information integration technology.

[0081] Exemplarily, a syntactic dependency tree is obtained from the transcribed text. Dependency parsing technology constructs a tree structure by analyzing the grammatical relationships between words in a sentence. For example, in the meeting record text "Manager Zhang proposed to speed up the project progress," dependency parsing identifies "propose" as the predicate center, "Manager Zhang" as the subject of the proposal, and "speed up the project progress" as the object, generating a syntactic dependency tree that includes the dependencies between the words. This tree structure clearly displays the grammatical hierarchy of the sentence, providing a foundation for subsequent analysis.

[0082] In one possible implementation, the subject-verb-object structure is extracted from the syntactic dependency tree. Syntactic role labeling technology identifies the subject, predicate, and object by annotating words according to their grammatical roles. For example, in the sentence above, "Manager Zhang" is labeled as the subject, "proposal" as the predicate, and "accelerate the project progress" as the object, resulting in a subject-verb-object structure. This structural extraction preserves the core semantics of the sentence, facilitating further processing.

[0083] It's important to note that if the subject or object contains a noun phrase, keyword extraction technology decomposes the noun phrase to generate a set of conference theme words. For example, "accelerate project progress" contains the noun phrase "project progress." Keyword extraction decomposes this into "project" and "progress," forming a set of conference theme words. This decomposition focuses on the core content of the meeting and improves the accuracy of theme extraction.

[0084] Specifically, named entities are obtained from the syntactic dependency tree, and named entity recognition technology annotates the names of people and organizations to generate a list of participants. For example, in the text "Engineer Li and the R&D department discussed technical solutions," named entity recognition annotates "Engineer Li" as a person and "R&D department" as an organization, generating a list of participants. This method accurately identifies the participants involved in the meeting and supports information integration.

[0085] In one embodiment, semantic role labeling technology uses the subject-verb-object structure and a set of conference keywords to identify core verbs and their associated arguments, generating a set of resolution content. For example, in the sentence "Director Wang decided to optimize budget allocation," semantic role labeling identifies "decided" as the core verb and "optimize budget allocation" as the argument, generating the resolution content "Optimize budget allocation." This analysis captures the core intent of the meeting resolution.

[0086] Preferably, information integration technology generates a structured information set for the resolution content set and the list of participants. For example, "Manager Zhang," "project progress," and "accelerate" can be integrated into the structured information: {Personnel: Manager Zhang, Topic: Project Progress, Resolution: Accelerate}. This structured information clearly organizes the meeting content and facilitates subsequent processing.

[0087] Step S105: Calculate the semantic similarity between knowledge graph nodes based on the structured information set. If the similarity is higher than a first preset threshold, associate the historical records to generate an extended information set.

[0088] In this embodiment, knowledge graph nodes are obtained from the structured information set, and semantic analysis technology is used to calculate the semantic similarity between nodes to obtain a similarity set; if the similarity exceeds a first preset threshold, relevant node association data is extracted from the historical records to obtain an association record set; based on the association record set, knowledge graph extension technology is used to generate extended nodes to obtain an extended node set; node attributes are obtained from the extended node set, and information integration technology is used to generate extended information to obtain an extended information set.

[0089] Exemplarily, knowledge graph nodes are obtained from a structured information set, and semantic analysis technology calculates similarity by analyzing the semantic features of the node content. Assume that the meeting minutes contain structured information: {personnel: Manager Zhang, subject: project progress, resolution: speed up}, which is mapped to a knowledge graph node, and the node attributes include "Manager Zhang", "project progress", and "speed up". The semantic analysis technology is based on the word vector model, compares "project progress" with the historical node "development plan", and calculates the semantic similarity to be 0.85. The first preset threshold is 0.8, and because 0.85 exceeds the threshold, historical record extraction is triggered, where the first preset threshold is set based on the medical scenario and the need for nearly confirmed associations.

[0090] In one possible implementation, the associated data for relevant nodes is extracted from the historical records to form a set of associated records. The historical records contain associated data for the nodes "Development Plan" and "Engineer Li" and "Adjustment Plan." This extraction creates the set: {Development Plan, Engineer Li, Adjustment}. This set provides the foundation for subsequent expansion, focusing on the meeting management domain and avoiding interference from irrelevant business.

[0091] Specifically, knowledge graph expansion technology generates expansion nodes based on a set of related records. Based on the similarity between "Development Plan" and "Project Progress," the expansion node "Technology Development Progress" is generated, with the attributes "Technology" and "Progress." The expansion node set also includes the association between "Engineer Li" and "Technology Development Progress," ensuring logical consistency between nodes.

[0092] It's important to note that information integration technology extracts attributes from the extended node set to generate extended information. For example, for "Technical Development Progress," the attributes "Technology" and "Progress" are integrated with the associated person "Engineer Li," resulting in the following extended information: {Subject: Technical Development Progress, Person: Engineer Li, Resolution: Accelerate}. This information preserves the core semantics of the meeting, facilitating further processing.

[0093] Step S106, constructing a structured memorandum based on the extended information set, organizing the meeting topic, participants, resolution content and related historical records, and generating an initial memorandum document;

[0094] In the embodiments of this application, see the attached Figure 3 , obtain meeting data from the extended information set, the meeting data including the meeting theme, participants, resolution content and historical records, and use data extraction technology to generate an extracted data set; based on the extracted data set, generate a classified data set through information classification technology to obtain a classified data set; obtain classified data from the classified data set, and use integrity verification technology to verify whether the classified data meets the preset integrity conditions. If the verification result shows that the classified data meets the integrity conditions, an initial memorandum document is generated through a structured template, wherein the structured template is usually based on field requirements, standardization specifications and historical data precipitation; obtain a memorandum document from the initial memorandum document, use formatting technology to generate a formatted memo, use correlation analysis technology to judge the correlation between the formatted memo and the historical records, and obtain a correlation analysis result; according to the correlation analysis result, obtain related data from the historical records, use information integration technology to integrate the related data into the formatted memo, and obtain an initial memorandum document.

[0095] Exemplarily, the data extraction technology generates an initial data set by parsing structured data to obtain the meeting topic, participants, resolution, and historical records from the extended information set. For example, the extended information set contains a meeting record with the topic "Product Optimization," participants "Manager Wang, Engineer Chen," resolution "Adjustment Plan," and historical records containing nodes related to "Product Development." The data extraction technology parses this information into an initial data set: the topic "Product Optimization," participants "Manager Wang, Engineer Chen," resolution "Adjustment Plan," and historical records "Product Development."

[0096] In a possible implementation, information classification technology classifies the initial data set to generate a classified data set.

[0097] Specifically, classification technology categorizes information into four categories based on data attributes: topics, personnel, resolutions, and historical records. For example, the topic might be classified as "product optimization," the personnel as "Manager Wang, Engineer Chen," the resolutions as "adjustment plan," and the historical records as "product development." This clearly separates data types in a classified data set, facilitating subsequent processing.

[0098] It's important to note that integrity verification technology determines whether categorized data meets pre-set integrity criteria. These include a valid topic, at least one participant, and a clear resolution. For example, if a categorized data set contains the complete topic "Product Optimization," two participants, and a clear resolution "Adjustment Plan," the verification result indicates that the criteria are met. If the topic is missing, the verification fails, and additional data is required.

[0099] Preferably, the structured template technology generates a memorandum document based on the verified classification data.

[0100] For example, using the template, enter the topic "Product Optimization," the participants "Manager Wang, Engineer Chen," and the resolution "Adjustment Plan" to generate an initial memorandum document: Product Optimization Meeting, Participants: Manager Wang and Engineer Chen, Resolution: Adjustment Plan. This document concisely captures the core content of the meeting.

[0101] Step S107: monitoring the editing operation of the initial memorandum document, updating the knowledge graph node relationship, and generating a memorandum document;

[0102] In an embodiment of the present application, an editing operation is obtained from the initial memo document, and operation data is extracted using an operation capture technology to obtain an operation data set; operation data is extracted from the operation data set, and an operation feature set is generated using a data parsing technology; operation features are obtained from the operation feature set, and a graph mapping technology is used to update the node relationship of the knowledge graph to obtain an updated node relationship set;

[0103] Node relationships are extracted from the updated node relationship set, and relationship verification technology is used to determine whether the node relationships meet the preset consistency conditions. If so, the updated knowledge graph is determined; node relationships are obtained from the updated knowledge graph, and memorandum documents are generated using document generation technology to obtain a memorandum document set.

[0104] For example, when capturing edit actions from interactive memo documents, action capture technology extracts action data by monitoring user interactions in real time. For example, in a memo from a product launch meeting, a user might add a comment to the "User Experience Optimization" section or modify the responsible person information. Action capture technology records these actions and generates action data sets, such as "Add comment: Optimize interface design" or "Change responsible person: Change from Manager Wang to Manager Li." This approach ensures that action data fully reflects user intent, providing a foundation for subsequent analysis.

[0105] In a possible implementation, the data parsing technology extracts operation data from the operation data set and generates an operation feature set.

[0106] Specifically, the technology analyzes the type, goal, and context of an operation, generating features such as "Operation Type: Add Annotation," "Goal: User Experience Optimization," and "Context: Interface Design." For example, when parsing "Add Annotation: Optimize Interface Design," the technology identifies its relevance to the theme of "User Experience" and generates a feature set. This analysis ensures accurate features and facilitates subsequent graph updates.

[0107] It's important to note that graph mapping technology updates the node relationships in the knowledge graph based on operational features. For example, the original node "User Experience Optimization" in the knowledge graph is associated with "Manager Wang." Based on the operational feature "Change Responsible Person: Manager Li," the technology updates the node relationship to "User Experience Optimization - Manager Li." The updated node relationship set reflects the latest responsibility assignments. This mapping approach maintains the dynamic nature of the graph.

[0108] Specifically, relationship verification technology checks whether node relationships meet consistency conditions. For example, it verifies whether "User Experience Optimization - Manager Li" is consistent with the document context. If the document mentions that Manager Li is responsible for interface design, the condition is met, and the updated knowledge graph is determined. This verification ensures the logical rigor of the graph.

[0109] In one embodiment, document generation technology utilizes the updated knowledge graph to generate memo documents. For example, based on the "User Experience Optimization - Manager Li" node, a document is generated containing the content "User Experience Optimization is the responsibility of Manager Li," generating a collection of memo documents. This generation method ensures that the document content is consistent with the graph.

[0110] Obviously, those skilled in the art may make various changes and modifications to the embodiments of the present application without departing from the spirit and scope of the embodiments of the present application. Thus, if these modifications and variations of the embodiments of the present application fall within the scope of the claims of the present application and their equivalents, the present application also intends to include such modifications and variations.

Claims

1. A robot-based intelligent conference memorandum generation method, characterized in that: The method comprises the following steps: S101 collects original voice data through the microphone array to obtain a multi-channel voice signal, adjusts the directional parameters based on the voice signal using beamforming technology, calculates the direction angle of the target speaker, and obtains an enhanced voice signal; S102. Based on the enhanced speech signal, an adaptive filtering algorithm is used to separate background noise to obtain a high signal-to-noise ratio speech signal; S103 extracts speech features from the high signal-to-noise ratio speech signal, uses a pre-trained model combined with conference domain corpus fine-tuning, and generates a transcribed text; S104. Parsing the sentence structure of the transcribed text, extracting the meeting topic, participants and resolution content, and forming a structured information set; S105. Calculate the semantic similarity between knowledge graph nodes based on the structured information set, and if the similarity is higher than a first preset threshold, generate an extended information set by associating historical records; S106. Construct a structured memorandum based on the extended information set, organize the meeting topic, participants, resolution content and related historical records, and generate an initial memorandum document; S107. Monitor the editing operation of the initial memorandum document, update the knowledge graph node relationship, and generate a memorandum document; The step S105 includes: Obtaining a knowledge graph node from the structured information set, and calculating the semantic similarity between the node and the historical node using semantic analysis technology to obtain a similarity set; If the similarity exceeds a first preset threshold, extracting the relevant node association data from the historical records to obtain an associated record set; According to the associated record set, the knowledge graph expansion technology is used to generate the expansion nodes to obtain the expansion node set; Node attributes are obtained from the extended node set, and extended information is generated using information integration technology to obtain an extended information set.

2. The method according to claim 1, characterized in that The step S101 includes: Collecting original speech data through the microphone array to obtain a multi-channel signal, and if noise exists in the multi-channel signal, preprocessing it using a spatial filtering technique to remove background noise and obtain a first speech signal; According to the first speech signal, the directivity parameters are adjusted using beamforming technology, and the spatial focusing is optimized using a minimum variance distortionless response algorithm to obtain a second speech signal; If the direction of the target speaker in the second speech signal is unclear, using a direction angle estimation technique to calculate the direction angle based on phase difference analysis to determine the position of the target speaker; adjusting the beamforming parameters according to the direction angle, and enhancing the target speaker signal through spatial filtering to obtain a third speech signal; For the third speech signal, noise suppression is performed on the spectral subtraction technology combined with a deep learning noise reduction model to remove residual noise and obtain an enhanced speech signal.

3. The method according to claim 1, characterized in that The step S102 includes: Based on the voice signal, decomposing the voice signal using a time domain analysis technique to extract a main voice signal; For the main voice signal, using spectrum analysis technology to calculate the interference frequency distribution and determine the interference frequency band; According to the interference frequency band, a notch filter technology is used to suppress the interference of a specific frequency to obtain a filtered voice signal; For the filtered speech signal, an adaptive filtering technology is used to adjust the filter coefficient. If the signal-to-noise ratio is lower than a second preset threshold, the filter coefficient is iteratively updated to obtain a speech signal with a high signal-to-noise ratio.

4. The method according to claim 1, wherein The step S103 includes: Extracting time domain features from the high signal-to-noise ratio speech signal, and obtaining a time-frequency feature distribution by using a feature extraction method combining Mel-frequency cepstral coefficients and a convolutional neural network; According to the time-frequency feature distribution, principal component analysis technology is used to perform dimensionality reduction processing to obtain reduced-dimensional speech features; For the reduced-dimensional speech feature, if the feature vector variance is lower than a third preset threshold, optimizing the feature by using a regularization technique to obtain an optimized speech feature; Extracting semantic information from the optimized speech features, performing feature mapping using a pre-trained language model, and obtaining a semantic feature vector; Based on the semantic feature vector, a pre-trained model is used in combination with conference domain corpus for fine-tuning to generate a preliminary transcription text; For the preliminary transcribed text, the text semantic similarity is calculated using word embedding technology. If the text semantic similarity is lower than a fourth preset threshold, the language model parameters are iteratively updated until the similarity of the generated transcribed text is greater than or equal to the preset threshold.

5. The method according to claim 1, wherein The step S104 includes: Parsing the sentence structure in the transcribed text using dependency syntax to obtain a syntactic dependency tree; Extracting a subject-verb-object structure according to the syntactic dependency tree, identifying the subject, predicate, and object through syntactic role labeling technology, and obtaining a subject-verb-object structure; If the subject or object in the subject-verb-object structure contains a noun phrase, decomposing the noun phrase using a keyword extraction technique to obtain a conference keyword set; Obtaining named entities from the syntactic dependency tree, annotating names of people and organizations using named entity recognition technology, and obtaining a list of participants; According to the subject-verb-object structure and the conference keyword set, a semantic role labeling technology is used to identify core verbs and their associated arguments to obtain a resolution content set; With respect to the resolution content set and the participant list, a structured information set is generated through information integration technology.

6. The method according to claim 1, wherein The step S106 includes: Acquire meeting data from the extended information set, the meeting data including meeting topics, participants, resolution content, and historical records, and generate an extracted data set using data extraction technology; generating a classified data set based on the extracted data set by using an information classification technology to obtain a classified data set; Obtaining classified data from a classified data set, using integrity verification technology to verify whether the classified data meets a preset integrity condition, and if the verification result indicates that the classified data meets the integrity condition, generating an initial memorandum document using a structured template; Obtaining a memo document from an initial memo document, generating a formatted memo using a formatting technique, and determining the relevance between the formatted memo and the historical record using a correlation analysis technique to obtain a correlation analysis result; According to the correlation analysis result, the correlation data is obtained from the historical records, and the correlation data is integrated into the formatted memo using information integration technology to obtain an initial memo document.

7. The method according to claim 1, characterized in that The step S107 includes: Obtaining editing operations from the initial memo document, extracting operation data using an operation capture technology, and obtaining an operation data set; Extracting operation data from the operation data set and generating an operation feature set using data analysis technology; Acquire operation features from the operation feature set, and use graph mapping technology to update the node relationship of the knowledge graph to obtain an updated node relationship set; Extracting node relationships from the updated node relationship set, and using relationship verification technology to determine whether the node relationships meet a preset consistency condition, and if so, determining an updated knowledge graph; Node relationships are obtained from the updated knowledge graph, and memorandum documents are generated using document generation technology to obtain a memorandum document set.

Images