Large model-based conference content real-time processing method and system, and conference device

Abstract

The present application relates to the field of conference intelligent processing, and more particularly to a conference content real-time processing method and system based on a large model and a conference device, the method comprising: constructing a conference field article library, extracting a field word library and calculating the degree of subordination of word pairs; converting the speech of a main speaker into text, extracting word sequences, fusing multiple features to construct a model to obtain the degree of main line of each word, and clustering to extract main line word sequences; real-time processing of dialogue speech, calculating deviation index and continuous slope, judging whether it deviates from the main line, and realizing real-time processing of conference content. The present application quantifies the degree of subordination of word pairs by constructing a field article library, fuses multiple dimensions of features to cluster and extract main line word sequences, calculates the deviation index and continuous slope of the dialogue to realize real-time correction, solves the problems of insufficient field adaptation, low main line word recall rate and inability to correct immediately in the prior art, and improves conference efficiency and the accuracy of core information extraction.

Description

Technical Field

[0001] This invention relates to the field of intelligent conference processing. In particular, it relates to methods, systems, and conference equipment for real-time processing of conference content based on large-scale models. Background Technology

[0002] Online meetings in cloud-based collaborative models have become an important form of work communication, and their efficient operation relies on the support of real-time transcription and topic extraction technologies. These technologies can quickly convert the audio content of a meeting into text and extract the core discussion points, providing a foundation for subsequent meeting summaries, task organization, and knowledge accumulation. They are key to improving meeting collaboration efficiency and ensuring communication effectiveness, and are applied in various office scenarios.

[0003] Real-time transcription and topic extraction of meeting minutes are directly related to the realization of meeting value. Accurate transcription can completely preserve key information from the meeting, avoiding information distortion caused by omissions or errors in manual recording; while effective topic extraction can extract core issues and discussion lines from complex speech content, helping participants quickly focus on key points and clarify logical relationships. The combination of these two functions not only reduces the workload of post-meeting organization, but also provides a clear basis for decision-making and task allocation, while helping to form reusable meeting knowledge assets, playing an irreplaceable role in improving overall work efficiency.

[0004] Current mainstream solutions still face several technical bottlenecks in meeting transcription and topic extraction. First, they lack prior knowledge of domain affiliation, making it impossible to quantify the hierarchical weight of any two words within the main topic and hindering adaptation to the vocabulary association logic of specialized fields. Second, they rely solely on keyword frequency statistics for speaker presentations, ignoring the coupling characteristics of word spacing, coherence, and vocabulary complexity, resulting in low recall rates for key words and inaccurate extraction of core information. Third, the dialogue phase lacks a continuous deviation slope monitoring mechanism. When conflicts, digressions, or disputes arise, the system fails to provide timely alerts and prompts for a return to the main theme, ultimately leading to decreased meeting efficiency, fragmented conclusions, and difficulty in achieving meeting objectives. Summary of the Invention

[0005] To address the problems of low recall rate of main words and difficulty in timely correction due to lack of domain-level quantification and neglect of coupling features such as word spacing, resulting in low meeting efficiency and fragmented conclusions, this invention provides solutions in the following aspects.

[0006] In the first aspect, the real-time processing method for conference content based on a large model includes: constructing a domain article library corresponding to the conference domain; extracting a domain lexicon from the domain article library; calculating the degree of subordination between any two words based on their co-occurrence features and correlation coefficients in the domain article library; acquiring the speaker's voice information in the conference; converting the voice information into text information and extracting word sequences; analyzing the word frequency ratio, distribution penetration, text content complexity index, and average degree of subordination of each word with other words in the word sequences; using these as feature inputs to construct a main theme degree calculation model for each word; obtaining the main theme degree of each word; performing cluster analysis on the main theme degree of all words; extracting the main theme word sequence; and accurately extracting the main theme vocabulary of the conference content; acquiring the voice information of the conference dialogue stage in real time and converting it into dialogue text information; analyzing the dialogue deviation index based on the main theme word sequence; calculating the slope of the deviation index of continuous dialogue using the least squares method; and determining whether it deviates from the main theme, thus completing the real-time processing of the conference content.

[0007] Preferably, the method for calculating the degree of subordination includes:

[0008] Taking any conference as the target conference, extract a domain thesaurus from the domain article library corresponding to the target conference, take any word in the domain thesaurus as the target word, and the other words as words to be analyzed, sort all articles in the domain article library in ascending order of the number of words to be analyzed, and obtain an ordered article sequence; extract the occurrence count of words to be analyzed from article by article in the ordered article sequence to form a sequence of word counts to be analyzed, and similarly extract the occurrence count of target words and the target word count sequence;

[0009] The word frequency percentage of articles containing the target term in the domain article library corresponding to the target conference is calculated. The target term number sequence and the target term number sequence are extracted from all articles in the domain article library after being sorted by the number of times the target term is contained. The Pearson correlation coefficient between the target term number sequence and the target term number sequence is calculated. The word frequency percentage and the Pearson correlation coefficient are multiplied to obtain the degree of dependence of the target term on the target term.

[0010] Iterate through all words to obtain the degree of dependence of the target word on all words to be analyzed, and obtain the degree of dependence between any two words in the domain thesaurus.

[0011] Preferably, the calculation method for the degree of the main line includes:

[0012] Using any word in the domain thesaurus as the target word, calculate the ratio of the number of occurrences of the target word to the total number of words in the text information to obtain the word frequency percentage of the target word;

[0013] The standard deviation of the spacing sequence of target words in the text information is calculated, and the result is input into the hyperbolic tangent function to obtain the distribution penetration of the target words;

[0014] Calculate the information entropy value of the word frequency ratio of all words in the text information, sum them and take the negative value, and input it into the negative exponential function to obtain the complexity index of the text content;

[0015] Calculate the average degree of subordination between the target word and all other words in the domain thesaurus to obtain the domain relevance of the target word;

[0016] The degree of the main thread of the target word is obtained by multiplying the word frequency ratio, distribution penetration, text content complexity index and domain relevance of the target word.

[0017] Preferably, the step of performing cluster analysis on the degree of main theme of all words and extracting the main theme word sequence includes:

[0018] Sort all words in the text information corresponding to the speaker's speech according to their degree of importance from smallest to largest to form a sequence of degree of importance.

[0019] The main line degree sequence is clustered using an ordered sample clustering algorithm to obtain multiple subsequences. The mean of the main line degree of each subsequence is calculated to obtain the overall main line association strength of the words in each subsequence. The subsequence with the highest overall main line association strength is selected to form the main line word sequence, where each subsequence contains one or more words.

[0020] Preferably, the calculation method for the dialogue deviation index includes:

[0021] Determine the total number of words in the dialogue text information corresponding to a single dialogue phase and the total number of subsequences corresponding to the speaker's main word sequence;

[0022] For each word in a single dialogue text, extract the number of its subsequence, calculate the ratio of the subsequence number to the total number of subsequences, and obtain the hierarchical association strength of each word;

[0023] The product of the hierarchical association strength and the main line degree of the corresponding words is used to obtain the main line fit contribution value of each word; the main line fit contribution values ​​of all words contained in the dialogue text information are summed and averaged, and the result of 1 minus the average value is used as the dialogue deviation index of a single dialogue relative to the speaker's text content.

[0024] Preferably, the determination of whether it deviates from the main theme includes:

[0025] When the deviation index slope is less than the preset deviation value, it indicates that the dialogue discussion has deviated from the speaker's main topic and is going further and further away, and the topic has become divergent. An alarm should be issued to remind the participants that they have deviated from the main topic. Conversely, when the deviation index slope is greater than or equal to the preset deviation value, it indicates that the dialogue discussion is still on the speaker's main topic.

[0026] Preferably, the domain article library supports dynamic updates. When new domain-related articles are added, the degree of subordination between any two words in the domain thesaurus is recalculated, and the domain semantic association data is updated.

[0027] Secondly, a real-time conference content processing system based on a large model includes a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the aforementioned real-time conference content processing method based on a large model is implemented.

[0028] Thirdly, the conference equipment integrates the aforementioned real-time conference content processing system and includes: a domain corpus acquisition and processing module: used to determine the conference discussion domain and acquire relevant articles within the corresponding domain to construct a domain article library; a speech processing and main word sequence extraction module: used to collect the speaker's audio information, transcribe the audio into text information through a semantic recognition model; calculate the main theme degree of each word; after sorting the main theme degree of all words, use an ordered sample clustering algorithm to obtain multiple subsequences, and select the subsequence with the largest mean of main theme degree to form the main word sequence; a real-time dialogue monitoring and deviation index calculation module: used to collect the dialogue audio of the participants in real time, transcribe it into text information, use the main theme degree calculation method to obtain the main theme degree of each word in the dialogue, and calculate the dialogue deviation index. For each single dialogue deviation index, the least squares method is used to fit the deviation index of multiple consecutive dialogues to calculate the slope of the deviation index; and a deviation judgment and alarm prompt module: used to determine whether the dialogue deviates from the main theme to ensure the normal progress of the conference.

[0029] The present invention has the following effects:

[0030] 1. This invention makes up for the lack of prior knowledge in existing solutions by accurately quantifying the hierarchical relationship between domain words, constructing a domain article library and calculating the degree of subordination of word pairs, thus making the lexical association analysis adapt to the logic of the professional domain and providing a reliable semantic foundation for subsequent main line extraction.

[0031] 2. This invention improves the accuracy of main word extraction by fusing features from multiple dimensions. It constructs a main word degree calculation model by integrating word frequency ratio, distribution penetration, complexity index and domain relevance. Combined with ordered sample clustering, it achieves accurate extraction of main word sequences, effectively solving the problems of low main word recall and inaccurate extraction of core information caused by single frequency statistics.

[0032] 3. This invention monitors and corrects dialogue deviations in real time. By calculating the dialogue deviation index and the slope of continuous deviation, a dynamic deviation judgment mechanism is established. When the dialogue shows a divergent trend, an alarm can be triggered in time to prevent the meeting from deviating from the main line, significantly improve the efficiency of meeting progress, and reduce the fragmentation of conclusions. Attached Figure Description

[0033] Figure 1 This is a flowchart of steps S1-S3 in the real-time processing method of meeting content based on a large model according to an embodiment of the present invention. Detailed Implementation

[0034] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0035] Reference Figure 1 The real-time processing method for meeting content based on a large model includes steps S1-S3, as follows:

[0036] The conference corresponds to a specific discussion area, and the domain articles, as the carriers of professional knowledge and expression logic in that domain, contain words with distinct domain-specific semantic relationships. If two words appear frequently in a domain article within the same domain, it means that they are often used together to express relevant core information in professional scenarios, that is, their semantic relationship is closer. Conversely, if two words appear infrequently in a domain article, it indicates that they lack synergy in the expression logic of that domain, and their semantic relationship is weak or non-existent.

[0037] The domain-specific relationships between words directly determine the accuracy of conference theme extraction. Only by clearly defining the hierarchy and strength of association between words within a domain can we effectively distinguish between core topic words and auxiliary words, avoiding misjudging high-frequency colloquialisms and irrelevant words as core topics. Therefore, the key issue to be addressed is: based on articles in the corresponding domain, to obtain the degree of subordination between any two words in that domain, providing a core semantic basis for subsequent extraction of main conference terms and judgment of dialogue deviations. The specific steps are as follows:

[0038] S1: Construct a domain article library corresponding to the conference domain, extract a domain thesaurus from the domain article library, and calculate the degree of subordination between any two words in the domain article library based on the co-occurrence characteristics and correlation coefficients of any two words in the domain thesaurus.

[0039] First, a domain thesaurus is extracted from the domain article library corresponding to the target conference. This thesaurus contains the core professional vocabulary of the target conference and forms the basis for subsequent lexical association analysis. A word is arbitrarily selected from the thesaurus as the target word, and all other words in the thesaurus are treated as words to be analyzed. The degree of dependence of the target word on each word to be analyzed is calculated.

[0040] For example, the sources of the domain article library include: public literature, internal corporate documents, and online resources), and the selection criteria include publication time, authority, and relevance.

[0041] For each word to be analyzed, first sort all articles in the domain article library in ascending order of the number of words to be analyzed, forming an ordered article sequence; then extract the occurrence count of the word to be analyzed from each article in the ordered article sequence to form a sequence of word counts, and at the same time extract the occurrence count of the target word from each article to form a sequence of target word counts.

[0042] Next, the proportion of articles containing the target term in the domain article library is calculated, reflecting the core coverage of the target term in the target domain. The larger the value, the wider the coverage and the higher the frequency of the word to be analyzed in articles within the domain; the more significant the attribute of the word to be analyzed as a core vocabulary of the domain; the stronger the main feature corresponding to the word to be analyzed; and the higher the probability that the target word forms a subordinate relationship with this core vocabulary. Conversely, The smaller the value, the weaker the popularity and core nature of the term being analyzed within the field, the less obvious the main characteristics, and the lower the likelihood that the target term will become a subordinate term. Calculate the Pearson correlation coefficient. This is the linear correlation coefficient between the sequence of target word counts and the sequence of word counts to be analyzed, reflecting the synergistic consistency in the frequency of occurrence of the target word and the word to be analyzed. The larger the value, the more synchronously the number of occurrences of the target word increases as the number of occurrences of the words to be analyzed in the domain articles increases, and the growth rate of the two is consistent. This indicates that there is a strong synergistic association between the target word and the words to be analyzed in the professional expression of the domain, and the semantic relationship is closer. Conversely, if the correlation coefficient value is smaller (approaching 0 or negative), it indicates that there is no obvious correlation between the occurrence frequency of the target word and the words to be analyzed, or that they are inversely correlated. This means that the two lack synergy in the semantic logic of the target domain, or even have no relationship.

[0043] Next, the word frequency percentage is multiplied by the Pearson correlation coefficient to obtain the degree of dependence of the target word on the word to be analyzed. The degree of dependence quantifies the semantic dependence of the target word on the word to be analyzed. The larger the product result, the more semantically the target word tends to depend on the word to be analyzed, and the clearer the relationship between the two in the domain representation.

[0044] Finally, all words to be analyzed in the domain lexicon are traversed in the same way, and the degree of dependence of the target word on each word to be analyzed is calculated one by one, thereby obtaining the degree of dependence between any two words in the domain lexicon. Through this process, the semantic relationships of words in the domain can be transformed into quantifiable values, providing accurate domain semantic relationship basis for subsequent extraction of main theme words and judgment of topic deviation, effectively solving the technical problem that traditional solutions cannot quantify the hierarchical weight of words.

[0045] Specifically, the degree of subordination satisfies the following relationship:

[0046] ;

[0047] In the formula, This indicates the degree of subordination of the target word to the analyzed word. This indicates the percentage of articles containing the term to be analyzed within the domain's article library in terms of word frequency. Represents the sequence of target word counts. This represents the sequence of the number of words to be analyzed. This represents the Pearson correlation coefficient.

[0048] During a target meeting, the speaker typically begins by introducing the core topic and developing the main argument. The entire discussion process essentially revolves around this main argument, which serves as the core carrier of the meeting's core information. Furthermore, from a linguistic perspective, when elaborating on the main content, the speaker often uses multiple content words to collaboratively convey the core meaning. The more content words there are and the closer their semantic connections, the more likely these words are to be key elements supporting the main argument, and the higher their probability of being the main argument's core term.

[0049] However, existing solutions only count the frequency of keywords in the speaker's speech, ignoring coupling features such as word spacing, coherence, and lexical complexity. This results in a low recall rate for key words, making it difficult to accurately identify the core vocabulary that truly supports the main theme of the meeting. Therefore, it is necessary to combine the linguistic features of the speaker's speech with domain semantic relationships to quantitatively evaluate the strength of the main theme association (i.e., the degree of main theme) of each word in the speaker's speech. Based on this degree of main theme association, the key words of the meeting can be accurately identified and determined, providing a reliable basis for the dynamic maintenance of the meeting theme and the judgment of dialogue deviations. The specific steps are as follows:

[0050] S2: Obtain the speech information of the speaker in the meeting, convert the speech information into text information, extract word sequences, analyze the word frequency ratio, distribution penetration, text content complexity index and average subordination degree of each word in the word sequence, and use them as feature input to construct a main theme degree calculation model for each word, obtain the main theme degree of each word, and perform cluster analysis on the main theme degree of all words to extract the main theme word sequence, thus completing the accurate extraction of the main theme vocabulary of the meeting content.

[0051] Calculating word frequency percentage: Using any word selected from the domain thesaurus as the target word, the number of times the target word appears in the text information corresponding to the speaker's speech is counted, and the ratio is calculated to the total number of words in the text information to obtain the word frequency percentage of the target word. This reflects the frequency of the target word in the speech content. The higher the word frequency percentage, the more often the target word is mentioned in the speech, and the higher the basic probability that it supports the core theme.

[0052] Calculating distribution penetration: Extract the spacing sequence of target words in the text information, which is the sequence consisting of the number of other words contained between any two adjacent target words. Calculate the standard deviation of this spacing sequence to reflect the dispersion of the target word's distribution. Input the standard deviation result into a hyperbolic tangent function for nonlinear transformation to obtain the distribution penetration of the target word. This reflects the dispersion of the target word's distribution in the speech content and its ability to cover the entire text. The greater the distribution penetration, the more evenly the target word is distributed in the speech, the stronger its penetration, and the more comprehensively it supports the expression of the core theme.

[0053] Calculating the text content complexity index involves statistically analyzing the proportion of each word in the text, specifically the ratio of the number of occurrences of each word to the total number of words in the text. The information entropy value of this proportion is then negatively calculated and input into a negative exponential function to obtain the text content complexity index. This index reflects the degree of thematic focus in the speech content. A higher complexity index indicates a greater variety of vocabulary, more dispersed content, and a more divergent theme, indirectly reflecting the core focus of the target word's context.

[0054] Calculating Domain Relevance: Based on the domain lexicon and its degree of subordination, the average degree of subordination between the target word and all other words in the domain lexicon is calculated to obtain the domain relevance of the target word. This reflects the core relevance strength of the target word in the domain's semantic system. The higher the domain relevance, the closer the semantic connection between the target word and other professional terms in the domain, and the higher its importance in the core theme of the domain.

[0055] The degree of focus of a target word is calculated by multiplying its frequency proportion, distribution coverage, text content complexity index, and domain relevance. This degree quantifies the strength of the target word's support for the core theme of the conference. A larger product indicates better overall performance of the target word in terms of frequency of occurrence, distribution coverage, and domain semantic relevance, and a higher probability that the target word will serve as the main theme word of the conference. This provides a precise quantitative basis for subsequent extraction of the main theme word sequence.

[0056] Specifically, the degree of the main line satisfies the following relationship:

[0057] ;

[0058] In the formula, Indicates the degree of importance of the target word. Indicates the percentage of the target word's frequency. This represents the spacing sequence of target words in textual information. Indicates the first in the text information The proportion of each word to all words, Indicates standard deviation, Represents the hyperbolic tangent function. Represented by natural constant An exponential function with base 0. Represents the logarithmic function with base 2. This represents the total number of words in the domain thesaurus excluding the target word. This indicates that the target word is the first word in the domain thesaurus. The degree of subordination of each word to be analyzed.

[0059] The degree of main theme is sorted in ascending order to construct a main theme degree sequence. An ordered sample clustering algorithm is then used to perform segmented clustering on this sequence. The core characteristic of the ordered sample clustering algorithm is to maintain the continuity of the positions of words within each subsequence after clustering within the original main theme degree sequence, ensuring that the correspondence between words and the semantic logic of the speaker's speech is not disrupted. Multiple subsequences are obtained after clustering. The mean of the main theme degree in each subsequence is calculated; the mean value represents the overall main theme association strength of the words within the corresponding subsequence. The larger the mean value, the stronger the overall main theme association of the words within the subsequence. The stronger the degree of ...

[0060] During the dialogue and discussion sessions of a meeting, participants' remarks often revolve around the meeting's theme. However, in situations where opinions differ or conflict, there's a tendency to focus on refuting flaws in others' statements, deviating from the core discussion direction. This causes the meeting to stray from the main theme set by the speaker, resulting in reduced meeting efficiency and difficulty in reaching core conclusions. Therefore, to promptly correct deviations in dialogue and ensure the meeting progresses smoothly around the main theme, a theme correction and prompting mechanism needs to be established. Through real-time monitoring and precise prompts, this mechanism guides participants back to the core discussion topic. The specific steps are as follows:

[0061] S3: Real-time acquisition of voice information during the conference dialogue phase, conversion into dialogue text information, analysis of dialogue deviation index based on the main word sequence, calculation of the deviation index slope of continuous dialogue using the least squares method, and determination of whether it deviates from the main theme, thus completing the real-time processing of conference content.

[0062] The total number of words contained in the text information of a single dialogue phase is determined as the normalization basis for subsequent mean calculation. The total number of subsequences corresponding to the main word sequence of the speaker is determined. The subsequences are obtained by clustering the main word sequence of the speaker's speech using an ordered sample clustering algorithm, which provides a fixed reference standard for calculating the hierarchical association strength.

[0063] Hierarchical association strength calculation: For each word in the dialogue text information of a single dialogue phase, its number in the subsequence corresponding to the speaker's main word sequence is extracted. The hierarchical association strength of each word is obtained by calculating the ratio of the subsequence number to the total number of subsequences. This reflects the semantic proximity level between the dialogue vocabulary and the speaker's main theme. The larger the ratio, the higher the average degree of main theme connection of the corresponding word's subsequence, the more core its association level in the speaker's main semantic system, and the stronger its proximity to the main theme.

[0064] Calculation of Mainline Alignment Contribution Value: The hierarchical association strength of each word is multiplied by the word's own degree of alignment with the mainline to obtain the mainline alignment contribution value of each word. This reflects the dual characteristics of a word's proximity to the mainline hierarchy and its own mainline attributes. The larger the product result, the higher the word's contribution to the alignment with the mainline of the dialogue.

[0065] Dialogue Deviation Index Synthesis: The contribution values ​​of all words in the dialogue text to the main theme are summed, and then the sum is divided by the total number of words to obtain the average value. The result of 1 minus the average value is used as the dialogue deviation index of a single dialogue relative to the speaker's text content. This reflects a standardized comprehensive quantitative indicator. The larger the value, the higher the overall degree of alignment between the single dialogue and the speaker's main theme; the smaller the value, the more serious the deviation from the main theme. This provides accurate quantitative basis for subsequent continuous deviation monitoring of single-round dialogues.

[0066] Specifically, the dialogue deviation index satisfies the following relationship:

[0067] ;

[0068] in, This represents the deviation index of a single dialogue session from the speaker's written content. This indicates the total number of words contained in the dialogue text. Indicates the first in the dialogue text information The subsequence number to which each word belongs. This represents the total number of subsequences. Indicates the first in the dialogue text information The degree of importance of each word as the main theme.

[0069] During real-time monitoring of the meeting dialogue, the deviation index of multiple consecutive dialogues is fitted and calculated using the least squares method to obtain the deviation index slope. The deviation index slope reflects the dynamic trend of the dialogue content relative to the speaker's main topic: when the deviation index slope is less than a preset deviation threshold (the preset deviation threshold is...), the deviation index slope is calculated. When the deviation index slope is greater than or equal to the preset deviation threshold, it indicates that the current dialogue content is becoming less and less aligned with the speaker's main topic, and there has been a significant deviation that is gradually intensifying. The topic is showing a divergent trend, and if not guided in time, it will cause the meeting to deviate from the core discussion direction. When the deviation index slope is greater than or equal to the preset deviation threshold, it indicates that the dialogue deviation index remains stable or is on an upward trend, meaning that the current dialogue content is always focused on the main topic set by the speaker, and there has been no substantial deviation. The direction of the meeting discussion is in line with expectations.

[0070] Based on the above judgment results, when the following conditions are met When this happens, the system immediately triggers an alarm, alerting participants that the discussion has strayed from the main topic and guiding them back to the core discussion; when the conditions are met... In this scenario, the system does not trigger alarms, and the meeting proceeds according to the normal process, ensuring that the discussion always focuses on the main topic and improving meeting efficiency. The preset deviation threshold in this embodiment is obtained through statistical verification and model training based on 120 sets of real meeting data covering multiple fields and types. Users can customize and adjust the deviation threshold according to field characteristics, meeting type, and organizational management needs. To help users set the threshold accurately, the system has built-in scenario-based recommendation templates that users can directly select or fine-tune based on. After the user modifies the threshold, the system will display the expected recognition accuracy and false alarm rate (calculated based on historical data of similar meetings) corresponding to that threshold in real time, helping users judge the rationality of the adjustment.

[0071] The present invention also provides a real-time conference content processing system based on a large model. The system includes a processor and a memory, the memory storing computer program instructions. When the computer program instructions are executed by the processor, the real-time conference content processing method based on a large model according to the first aspect of the present invention is implemented. The system also includes other components well known to those skilled in the art, such as a communication bus and a communication interface, the settings and functions of which are known in the art and therefore will not be described further here.

[0072] This invention also provides a real-time conference content processing device based on a large model. The device includes: a domain corpus acquisition and processing module: used to determine the conference discussion domain and acquire relevant articles within that domain to construct a domain article library; a speech processing and main word sequence extraction module: used to collect the speaker's audio information, transcribe the audio into text information using a semantic recognition model; calculate the main theme degree of each word; after sorting the main theme degree of all words, use an ordered sample clustering algorithm to obtain multiple subsequences, and select the subsequence with the largest mean main theme degree to form the main word sequence; a real-time dialogue monitoring and deviation index calculation module: used to collect the dialogue audio of participants in real time, transcribe it into text information, use the main theme degree calculation method to obtain the main theme degree of each word in the dialogue, and calculate the dialogue deviation index. For each single dialogue deviation index, the least squares method is used to fit the deviation index of multiple consecutive dialogues to calculate the deviation index slope; and a deviation judgment and alarm prompt module: used to determine whether the dialogue deviates from the main theme to ensure the normal progress of the conference.

[0073] Workflow Example: The domain corpus acquisition and processing module first determines the conference domain and acquires relevant articles to build a domain article library; the speech processing and main word sequence extraction module collects the speaker's audio, transcribes it into text through semantic recognition, calculates the main theme degree of each word, sorts it, and obtains subsequences through ordered sample clustering, selecting the subsequence with the largest mean to form the main word sequence; the real-time dialogue monitoring and deviation index calculation module collects and transcribes the dialogue audio of the participants in real time, uses the same main theme degree calculation method to obtain the main theme degree of the dialogue words, then calculates the deviation index of a single dialogue, and then uses the least squares method to fit the deviation index of continuous dialogues to obtain the deviation index slope; the deviation judgment and alarm prompt module determines whether the dialogue deviates from the main theme based on the slope to ensure the normal progress of the meeting.

[0074] It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept, and these all fall within the scope of protection of this invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A method for real-time processing of meeting content based on a large model, characterized in that, include: Construct a domain article library corresponding to the conference field, extract a domain thesaurus from the domain article library, and calculate the degree of subordination between any two words in the domain article library based on the co-occurrence features and correlation coefficients of any two words in the domain thesaurus. The system acquires the speaker's voice information during the meeting, converts the voice information into text information, extracts word sequences, analyzes the word frequency ratio, distribution penetration, text content complexity index, and average subordination degree of each word with other words in the word sequence, and uses these as feature inputs to construct a main theme degree calculation model for each word, obtains the main theme degree of each word, performs cluster analysis on the main theme degree of all words, extracts the main theme word sequence, and completes the accurate extraction of the main theme vocabulary of the meeting content. The system acquires voice information from the conference dialogue phase in real time and converts it into dialogue text information. It analyzes the dialogue deviation index based on the main word sequence, calculates the slope of the deviation index of continuous dialogue using the least squares method, and determines whether it deviates from the main theme, thus completing the real-time processing of conference content. The calculation method for the degree of subordination includes: Taking any conference as the target conference, extract a domain thesaurus from the domain article library corresponding to the target conference, take any word in the domain thesaurus as the target word, and the other words as words to be analyzed, sort all articles in the domain article library in ascending order of the number of words to be analyzed, and obtain an ordered article sequence; extract the occurrence count of words to be analyzed from article by article in the ordered article sequence to form a sequence of word counts to be analyzed, and similarly extract the occurrence count of target words and the target word count sequence; The word frequency percentage of articles containing the target term in the domain article library corresponding to the target conference is calculated. The target term number sequence and the target term number sequence are extracted from all articles in the domain article library after being sorted by the number of times the target term is contained. The Pearson correlation coefficient between the target term number sequence and the target term number sequence is calculated. The word frequency percentage and the Pearson correlation coefficient are multiplied to obtain the degree of dependence of the target term on the target term. Iterate through all words to be analyzed to obtain the degree of dependence of the target word on all words to be analyzed, and obtain the degree of dependence of any two words in the domain thesaurus; The calculation method for the degree of the main line includes: Using any word in the domain thesaurus as the target word, calculate the ratio of the number of occurrences of the target word to the total number of words in the text information to obtain the word frequency percentage of the target word; The standard deviation of the spacing sequence of target words in the text information is calculated, and the result is input into the hyperbolic tangent function to obtain the distribution penetration of the target words; Calculate the information entropy value of the word frequency ratio of all words in the text information, sum them and take the negative value, and input it into the negative exponential function to obtain the complexity index of the text content; Calculate the average degree of subordination between the target word and all other words in the domain thesaurus to obtain the domain relevance of the target word; The degree of the main thread of the target word is obtained by multiplying the word frequency ratio, distribution penetration, text content complexity index and domain relevance of the target word.

2. The real-time meeting content processing method based on a large model according to claim 1, characterized in that, The clustering analysis of the degree of main theme of all words, and the extraction of the main theme word sequence, includes: Sort all words in the text information corresponding to the speaker's speech according to their degree of importance from smallest to largest to form a sequence of degree of importance. The main line degree sequence is clustered using an ordered sample clustering algorithm to obtain multiple subsequences. The mean of the main line degree of each subsequence is calculated to obtain the overall main line association strength of the words in each subsequence. The subsequence with the highest overall main line association strength is selected to form the main line word sequence, where each subsequence contains one or more words.

3. The real-time meeting content processing method based on a large model according to claim 1, characterized in that, The calculation method for the dialogue deviation index includes: Determine the total number of words in the dialogue text information corresponding to a single dialogue phase and the total number of subsequences corresponding to the speaker's main word sequence; For each word in a single dialogue text, extract the number of its subsequence, calculate the ratio of the subsequence number to the total number of subsequences, and obtain the hierarchical association strength of each word; The product of the hierarchical association strength and the main line degree of the corresponding words is used to obtain the main line fit contribution value of each word; the main line fit contribution values ​​of all words contained in the dialogue text information are summed and averaged, and the result of 1 minus the average value is used as the dialogue deviation index of a single dialogue relative to the speaker's text content.

4. The real-time processing method for meeting content based on a large model according to claim 1, characterized in that, The determination of whether it deviates from the main theme includes: When the deviation index slope is less than the preset deviation value, it indicates that the dialogue discussion has deviated from the speaker's main topic and is going further and further away, and the topic has become divergent. An alarm should be issued to remind the participants that they have deviated from the main topic. Conversely, when the deviation index slope is greater than or equal to the preset deviation value, it indicates that the dialogue discussion is still on the speaker's main topic.

5. The real-time processing method for meeting content based on a large model according to claim 1, characterized in that, The domain article library supports dynamic updates. When new domain-related articles are added, the degree of subordination between any two words in the domain thesaurus is recalculated, and the domain semantic association data is updated.

6. A real-time meeting content processing system based on a large model, characterized in that: include: A processor and a memory, wherein the memory stores computer program instructions that, when executed by the processor, implement the real-time processing method for conference content based on a large model according to any one of claims 1-5.

7. Conference equipment, characterized in that, The system integrates the real-time meeting content processing system as described in claim 6, and includes a domain corpus acquisition and processing module: used to determine the meeting discussion domain and acquire relevant articles in the corresponding domain to build a domain article library; The speech processing and main word sequence extraction module is used to collect the speaker's audio information, transcribe the audio into text information through a semantic recognition model, calculate the main theme degree of each word, sort all words by their main theme degree, use an ordered sample clustering algorithm to obtain multiple subsequences, and select the subsequence with the largest mean of main theme degree to form the main word sequence. The real-time dialogue monitoring and deviation index calculation module is used to collect the audio of the participants' dialogue in real time, transcribe it into text information, use the main line degree calculation method to obtain the main line degree of each word in the dialogue, and calculate the dialogue deviation index. For each single dialogue deviation index, the least squares method is used to fit the deviation index of multiple consecutive dialogues to calculate the deviation index slope. Deviation detection and alarm module: Used to determine whether the conversation has deviated from the main topic, ensuring that the meeting proceeds normally.

Images