Meeting minutes generation method and device, electronic equipment and storage medium

Abstract

A conference summary generation method and device, an electronic device, and a storage medium. The conference summary generation method includes: obtaining a text conference record to be segmented; performing first segmentation processing on the text conference record to be segmented according to text length to obtain a plurality of first segments; performing second segmentation processing on each first segment in the plurality of first segments to obtain a plurality of to-be-extracted segments, wherein each to-be-extracted segment includes at least part of text content corresponding to a topic; and generating a conference summary based on the to-be-extracted segments. The conference summary generation method can implement long text segmentation and generate a conference summary according to the plurality of to-be-extracted segments, thereby saving computing resources and being suitable for long conferences. In addition, each to-be-extracted segment includes at least part of text content corresponding to a topic, thereby facilitating the guarantee of topic integrity and the accuracy of conference information extraction.

Description

Method and device for generating meeting minutes, electronic device and storage medium TECHNICAL FIELD

[0001] The present disclosure relates to a method and device for generating meeting minutes, an electronic device and a storage medium. BACKGROUND

[0002] Generating meeting minutes automatically using a large language model is an efficient and innovative method, especially suitable for scenarios that involve processing large amounts of text data and quickly extracting key information. For example, the large language model can include a generative pre-trained transformer (GPT) and a bidirectional encoder representation from transformers (BERT).

[0003] SUMMARY

[0004] The present disclosure provides a method and device for generating meeting minutes, an electronic device and a storage medium.

[0005] At least one embodiment of the present disclosure provides a method for generating meeting minutes, comprising: obtaining a text conference record to be segmented; performing first segmentation processing on the text conference record to be segmented according to the text length to obtain a plurality of first segments; performing second segmentation processing on each of the plurality of first segments to obtain a plurality of extraction segments, wherein the extraction segment includes at least part of the text content corresponding to a topic; and generating the meeting minutes based on the plurality of extraction segments.

[0006] For example, according to the conference summary generation method provided by at least one embodiment of the present disclosure, the second segmentation processing is performed on each of the first segments to obtain a plurality of to-be-extracted segments, which includes: in response to the last part of the last first segment belonging to the same topic, the part of text is merged into the current first segment; identifying the end identifier corresponding to the end sentence of the topic of the segment after merging the part of text; in response to the current first segment not being the last first segment, and identifying a plurality of end identifiers, the text content after the second last end identifier is segmented, and the remaining part is taken as a second segment to obtain the to-be-extracted segment according to the second segment; or, in response to the current first segment not being the last first segment, and only one end identifier being identified, the segment after merging the part of text is segmented within the topic to obtain a plurality of topic segments, the last topic segment is merged into the next first segment, and the to-be-extracted segment is obtained according to the remaining topic segments; or, in response to the current first segment being the last first segment, the current first segment is taken as the second segment to obtain the to-be-extracted segment according to the second segment.

[0007] For example, according to the conference summary generation method provided by at least one embodiment of the present disclosure, the to-be-extracted segment is obtained according to the second segment, which includes: performing the topic segmentation processing on the second segment to obtain at least one topic segment corresponding to the second segment, wherein the topic segment includes at least part of the text content of a topic; and obtaining the to-be-extracted segment according to the topic segment.

[0008] For example, according to the conference summary generation method provided by at least one embodiment of the present disclosure, the to-be-extracted segment is obtained according to the topic segment, which includes: obtaining the text length of the topic segment; in response to the text length of the topic segment being not greater than a first threshold, the topic segment is determined as the to-be-extracted segment; and in response to the text length of the topic segment being greater than the first threshold, the topic segmentation processing is performed on the topic segment to obtain a plurality of to-be-extracted segments.

[0009] For example, according to the meeting minutes generation method provided by at least one embodiment of the present disclosure, the in-topic segmentation processing on the topic segment is performed to obtain a plurality of the to-be-extracted segments, which includes: dividing the text in the topic segment into a plurality of to-be-processed units in units of speakers, wherein the plurality of to-be-processed units include adjacent first to-be-processed unit and second to-be-processed unit, and the first to-be-processed unit is located before the second to-be-processed unit; obtaining an embedding vector of each to-be-processed unit through a model; calculating the similarity between the embedding vector corresponding to the first to-be-processed unit and the embedding vector corresponding to the second to-be-processed unit; in response to the similarity being less than a second threshold, setting the text corresponding to the first to-be-processed unit and the second to-be-processed unit in different topic segments respectively; in response to the similarity being not less than the second threshold, setting the text corresponding to the first to-be-processed unit and the second to-be-processed unit in the same topic segment; and obtaining the to-be-extracted segment according to the topic segment.

[0010] For example, according to the meeting minutes generation method provided by at least one embodiment of the present disclosure, the obtaining of the to-be-extracted segment according to the topic segment includes: judging whether the text length of the topic segment is greater than a first threshold; in response to the text length of the topic segment being not greater than the first threshold, taking the topic segment as the to-be-extracted segment; and in response to the text length of the topic segment being greater than the first threshold, segmenting the topic segment into a plurality of sub-topic segments according to the text length, wherein the text length of the sub-topic segment is not greater than the first threshold, and the sub-topic segment is taken as the to-be-extracted segment.

[0011] For example, according to the meeting minutes generation method provided by at least one embodiment of the present disclosure, the obtaining of the to-be-segmented text conference record includes: obtaining an original text conference record; performing textological regularization on the original text conference record to obtain a textological regularization evaluation value; in response to the textological regularization evaluation value being not less than a third threshold, taking the text after the textological regularization as the to-be-segmented text conference record; and in response to the textological regularization evaluation value being less than the third threshold, taking the original text conference record as the to-be-segmented text conference record.

[0012] For example, according to the meeting minutes generation method provided by at least one embodiment of the present disclosure, the original literal conference record is subjected to text modification to obtain a text modification evaluation value, which includes: subjecting the original literal conference record to model modification to obtain a first modified text; subjecting the first modified text to manual modification to obtain a second modified text; based on the first modified text and the second modified text, obtaining a first word number, a second word number and a third word number, wherein the first word number represents a word number that is modified in both the first modified text and the second modified text relative to the original literal conference record, the second word number represents a word number that is modified in the first modified text and not modified in the second modified text relative to the original literal conference record, and the third word number represents a word number that is not modified in the first modified text and modified in the second modified text relative to the original literal conference record; and based on the first word number, the second word number and the third word number, obtaining the text modification evaluation value.

[0013] For example, according to the meeting minutes generation method provided by at least one embodiment of the present disclosure, the text modification evaluation value is obtained by the following formula:

[0014] In the formula, TP represents the first word number, FP represents the second word number, FN represents the third word number, P represents the accuracy rate, R represents the recall rate, and F represents the text modification evaluation value.

[0015] For example, according to the meeting minutes generation method provided by at least one embodiment of the present disclosure, the first word number, the second word number and the third word number are obtained by: obtaining the minimum edit distance dp[i,j] required for the modified text and the original literal conference record from s1[0:i+1] to s2[0:j+1], wherein the modified text is the first modified text or the second modified text, s1[0:i+1] represents the 0th to ith characters in the original literal conference record, and s2[0:j+1] represents the 0th to jth characters in the modified text; and obtaining the first word number, the second word number and the third word number according to the minimum edit distance dp[i,j].

[0016] For example, according to the meeting minutes generation method provided by at least one embodiment of the present disclosure, the minimum edit distance dp[i,j] is obtained by the following formula:

[0017] In the formula, s1[i] represents the i-th character in the original literal conference record, s2[j] represents the j-th character in the regularized text, i and j are integers, edit = keep represents an edit mode of keeping, edit = insert represents an edit mode of inserting, edit = delete represents an edit mode of deleting, and edit = replace represents an edit mode of replacing.

[0018] For example, according to the conference summary generation method provided in at least one embodiment of the present disclosure, the first word number, the second word number, and the third word number are obtained according to the minimum edit distance dp[i, j], including: taking the number of characters whose edit mode is neither keeping in the first regularized text nor keeping in the second regularized text as the first word number; taking the number of characters whose edit mode is not keeping in the first regularized text and whose edit mode is keeping in the second regularized text as the second word number; and taking the number of characters whose edit mode is keeping in the first regularized text and whose edit mode is not keeping in the second regularized text as the third word number.

[0019] For example, according to the conference summary generation method provided in at least one embodiment of the present disclosure, the conference summary is generated based on the plurality of to-be-extracted segments, including: extracting a paragraph summary for each of the plurality of to-be-extracted segments; and generating the conference summary according to the paragraph summaries.

[0020] For example, according to the conference summary generation method provided in at least one embodiment of the present disclosure, the original literal conference record is obtained, including: obtaining a voice transcription result, a speaker, and a speaking time of a conference at different time points through a voice recognition algorithm and a voiceprint recognition algorithm.

[0021] At least one embodiment of the present disclosure also provides a conference summary generation device, including: an obtaining unit, a first segmenting unit, a second segmenting unit, and a generating unit. The obtaining unit is configured to obtain a literal conference record to be segmented. The first segmenting unit is configured to perform first segmentation processing on the literal conference record to be segmented according to a text length, to obtain a plurality of first segments. The second segmenting unit is configured to perform second segmentation processing on each of the plurality of first segments, to obtain a plurality of to-be-extracted segments. The to-be-extracted segments include at least part of text content corresponding to a theme. The generating unit is configured to generate a conference summary based on the plurality of to-be-extracted segments.

[0022] The electronic device includes a processor, a memory, and one or more computer program modules, wherein the one or more computer program modules are stored in the memory and configured to be executed by the processor, and the one or more computer program modules include a program module for implementing the method for generating a conference summary according to any one of the above embodiments.

[0023] The electronic device includes a processor, a memory, and one or more computer program modules, wherein the one or more computer program modules are stored in the memory and configured to be executed by the processor, and the one or more computer program modules include a program module for implementing the method for generating a conference summary according to any one of the above embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0024] In order to more clearly illustrate the technical solutions of the embodiments of the present disclosure, the drawings of the embodiments will be briefly introduced below. Obviously, the drawings described below only relate to some of the embodiments of the present disclosure, but not limit the present disclosure.

[0025] FIG. 1 is a schematic diagram of a method for generating a conference summary according to at least one embodiment of the present disclosure.

[0026] FIG. 2 is a schematic diagram of one example of step S300 shown in FIG. 1.

[0027] FIG. 3 is a schematic diagram of a first segment according to at least one embodiment of the present disclosure.

[0028] FIG. 4 is a schematic diagram of one example of step S321 shown in FIG. 2.

[0029] FIG. 5 is a schematic diagram of a second segment obtained according to the first segment shown in FIG. 3.

[0030] FIG. 6 is a schematic diagram of another second segment according to at least one embodiment of the present disclosure.

[0031] FIG. 7 is a schematic diagram of one example of step S3212 shown in FIG. 4.

[0032] FIG. 8 is a schematic diagram of one example of step S3230 shown in FIG. 7.

[0033] FIG. 9 is a schematic diagram of intra-topic segmentation according to at least one embodiment of the present disclosure.

[0034] FIG. 10 is a schematic diagram of one example of step S3236 shown in FIG. 8.

[0035] FIG. 11 is a schematic diagram of one example of step S322 shown in FIG. 2.

[0036] FIG. 12 is a schematic diagram of obtaining a segment to be extracted according to at least one embodiment of the present disclosure. FIG. 12 is a schematic diagram of obtaining a segment to be extracted according to at least one embodiment of the present disclosure.

[0037] FIG. 13 is a schematic diagram of an example of step S100 shown in FIG. 1.

[0038] FIG. 14 is a schematic diagram of text normalization of an original literal conference record in at least one embodiment of the present disclosure.

[0039] FIG. 15 is a schematic diagram of a partial text normalization result corresponding to FIG. 14.

[0040] FIG. 16 is a schematic diagram of an editing manner of obtaining characters provided by at least one embodiment of the present disclosure.

[0041] FIG. 17 is a schematic diagram of a text normalization effect provided by at least one embodiment of the present disclosure.

[0042] FIG. 18 is a flowchart of a conference minutes generation provided by at least one embodiment of the present disclosure.

[0043] FIG. 19 is a schematic block diagram of a conference minutes generation apparatus provided by at least one embodiment of the present disclosure.

[0044] FIG. 20 is a schematic block diagram of an electronic device provided by at least one embodiment of the present disclosure.

[0045] FIG. 21 is a schematic diagram of another electronic device provided by at least one embodiment of the present disclosure.

[0046] FIG. 22 is a schematic diagram of a storage medium provided by at least one embodiment of the present disclosure. DETAILED DESCRIPTION

[0047] In order to make the objects, technical solutions and advantages of the embodiments of the present disclosure clearer, the technical solutions of the embodiments of the present disclosure will be described clearly and completely below with reference to the drawings of the embodiments of the present disclosure. Obviously, the described embodiments are a part of the embodiments of the present disclosure, rather than all the embodiments of the present disclosure. Based on the described embodiments of the present disclosure, all other embodiments obtained by a person of ordinary skill in the art without any creative effort belong to the scope of protection of the present disclosure.

[0048] Unless otherwise defined, technical terms or scientific terms used in the present disclosure shall have the ordinary meaning as understood by a person of ordinary skill in the art to which the present disclosure pertains. The terms "first", "second", and similar terms used in the present disclosure do not denote any order, quantity, or importance, but are used to distinguish different components. The terms "include", "contain", and similar terms mean that the elements or objects before the terms encompass the elements or objects listed after the terms and their equivalents, and do not exclude other elements or objects. The features "same" and the like used in the embodiments of the present disclosure include the features "same" and the like in the strict sense, and also include the case of "approximately same" and the like with a certain error, which, taking into account the measurement and the error related to the measurement of a specific value (for example, the limitation of a measurement system), means within an acceptable deviation range for the specific value determined by a person of ordinary skill in the art. For example, "approximately" can mean within one or more standard deviations, or within 10% or 5% of the value. In the embodiments of the present disclosure, when the quantity of a component is not specifically indicated, it means that the component can be one or more, or can be understood as at least one. "At least one" means one or more, and "a plurality of" means at least two.

[0049] Generally, when generating a meeting minutes, different meeting lengths result in a large difference in the number of words in the transcription result, which has a significant impact on the graphics processing unit (GPU) memory required when the complete transcription result is sent to the minutes generation model at one time. For example, for a multi-person meeting lasting one hour or more, the transcription result can reach 15,000 words or even more, so such a data volume is easy to exceed the memory limit of a GPU with an ordinary configuration when processed at one time. Moreover, the cost of purchasing and maintaining a GPU with high memory is high. In addition, a multi-person meeting usually contains multiple discussion topics, and when meeting information extraction is performed on such a meeting, there can be problems such as missing meeting information and difficulty in ensuring the integrity of the discussion topics.

[0050] Therefore, it is particularly crucial to provide a meeting minutes generation method that has low requirements for the configuration of a GPU hardware environment and the length of a meeting, can ensure the integrity of a meeting topic, and has high running performance.

[0051] At least one embodiment of the present disclosure provides a method for generating a meeting minutes, comprising: obtaining a text conference record to be segmented; performing first segmentation processing on the text conference record to be segmented according to the text length to obtain a plurality of first segments; performing second segmentation processing on each of the plurality of first segments to obtain a plurality of to-be-extracted segments, wherein each to-be-extracted segment includes at least part of the text content corresponding to a topic; and generating a meeting minutes based on the plurality of to-be-extracted segments.

[0052] At least one of the embodiments of the present disclosure further provides a conference summary generation apparatus, an electronic device, and a storage medium corresponding to the conference summary generation method.

[0053] The conference summary generation method provided by the embodiments of the present disclosure can implement long text segmentation on the text conference record, and generate a conference summary according to a plurality of to-be-extracted segments, thereby saving computing resources and being applicable to conferences with a long duration. In addition, since the to-be-extracted segments include at least part of the text content corresponding to a topic, the topic integrity can be ensured, thereby facilitating the extraction integrity and accuracy of conference information. Furthermore, the conference summary generation method can save the time of manual writing, effectively improve the defects of time-consuming and labor-consuming of manual recording, and facilitate the reduction of problems such as missing important information caused by human factors. Since different people may have different styles when recording conference summaries, the conference summary generation method can also generate a conference summary according to a preset format and template, thereby facilitating the standardization and consistency of the format and content of the conference summary.

[0054] The embodiments of the present disclosure and some examples thereof will be described in detail below with reference to the accompanying drawings.

[0055] For example, the conference summary generation method provided by at least one of the embodiments of the present disclosure can be implemented in the form of software, hardware, firmware, or any combination thereof, and can be loaded and executed by a processor in a device such as a mobile phone, a digital camera, a tablet computer, a notebook computer, a desktop computer, a network server, and the like.

[0056] For example, the conference summary generation method is applicable to a conference summary generation apparatus, which can be any electronic device having a conference summary generation function, such as a mobile phone, a digital camera, a notebook computer, a tablet computer, a desktop computer, a network server, and the like, which can load and execute the conference summary generation method. For example, the conference summary generation apparatus can include a central processing unit (CPU), a graphics processing unit (GPU), or other forms of processing units having data processing capability and / or instruction execution capability, such as a digital signal processor (DSP), a storage unit, and the like. The conference summary generation apparatus also has an operating system, an application programming interface (such as OpenGL (Open Graphics Library), Metal, and the like), and the like installed thereon, and the conference summary generation method provided by the embodiments of the present disclosure is implemented by running codes or instructions.

[0057] FIG. 1 is a schematic diagram of a method for generating a meeting summary according to at least one embodiment of the present disclosure. As shown in FIG. 1, the method for generating a meeting summary includes steps S100-S400.

[0058] Step S100: Obtain a text conference record to be segmented.

[0059] Step S200: Perform first segmentation processing on the text conference record to be segmented according to text length to obtain a plurality of first segments.

[0060] Step S300: Perform second segmentation processing on each of the plurality of first segments to obtain a plurality of to-be-extracted segments, wherein each to-be-extracted segment includes at least part of the text content corresponding to a topic.

[0061] Step S400: Generate a meeting summary based on the plurality of to-be-extracted segments.

[0062] For example, for step S100, the speech transcription result, speaker, and speaking time of the conference at different times can be obtained by a speech recognition algorithm and a voiceprint recognition algorithm. For example, the above algorithms can identify the speaker and record the start time and end time of the speech. For example, the text conference record can be an original speech transcription text or a text after regular processing (see the detailed description below for regular processing). The text conference record is used to generate a detailed text conference summary (e.g., the conference summary is generated by segmentation processing, extraction, etc., see the detailed description below), so as to facilitate subsequent review and summary of the conference and improve work efficiency.

[0063] Of course, the embodiments of the present disclosure do not limit the manner of obtaining the text conference record to be segmented, for example, it can also be obtained by other manners other than the above-mentioned speech recognition algorithm and voiceprint recognition algorithm.

[0064] For example, for step S200, the text conference record to be segmented obtained by step S100 is first segmented according to text length to obtain a plurality of first segments.

[0065] For example, considering the limitation of computing resource, the first segmentation processing can be a“coarse segmentation processing”, which aims to obtain a plurality of first segments with small text length. For example, the plurality of first segments have the same text length. The length of the first segment is less than the length upper limit supported by the computing resource. For example, the length of the first segment can be about 4000 (such as 4096), and the specific value can be adjusted according to the length upper limit supported by the actual computing resource, and the embodiments of the present disclosure are not limited thereto. Therefore, the length of the text conference record to be segmented in the embodiments of the present disclosure can not be limited, and the plurality of first segments with text length meeting the requirement of the computing resource can be obtained by the first segmentation processing, and then the meeting minutes can be generated by the plurality of first segments, so it can be applied to any length of meeting.

[0066] For example, for step S300, the second segmentation processing can include a plurality of steps, which can identify the topic information contained in the first segment to segment the first segment according to the topic (for example, it can include the topic segmentation processing, intra-topic segmentation processing and the like introduced below), so as to obtain the to-be-extracted segment including at least part of the text content of one topic.

[0067] For example, for step S400, the plurality of to-be-extracted segments can be used to extract the paragraph summary, for example, the to-be-extracted segment can be sent into a large prophecy model (such as ChatGPT) to generate a paragraph summary according to a given prompt.

[0068] In some embodiments, the given prompt can include a prompt for extracting the meeting progress and a prompt for extracting the to-do list. For example, the prompt for extracting the meeting progress can be: “You are a conference secretary, focusing on summarizing the content of the meeting progress. Your main task is to extract the progress of the meeting discussion from a conference article, without adding any subjective speculation-based content, keeping the progress content objective, factual and complete, ensuring that the progress content remains coherent and smooth, and the logical structure is clear. The meeting content is as follows: \n' +'meeting original text'”. For example, the prompt for extracting the to-do list can be: “You are a conference secretary, focusing on summarizing the to-do list of the meeting. Your main task is to identify the segments containing to-do semantics from a conference article and summarize the to-do list, without adding any subjective speculation-based content, keeping the to-do list objective, factual and complete. The meeting content is as follows: \n' +'meeting original text'”.

[0069] For example, each paragraph summary can be used to integrate the full-text meeting summary. For example, the paragraph summary can be generated according to a required template to have a uniform format (for example, each paragraph summary is required to be generated in the format of 'AA: xxx'). Then, the multiple paragraph summaries with uniform format can be integrated by the large speech model to generate the full-text meeting summary according to the preset template requirement, so as to ensure that the format and content of the meeting summary remain standardized and consistent. For example, the prompt for generating the full-text meeting summary can be "This is the meeting summary extracted from a meeting record. Please integrate and output it in the format of 'AA: xxx'".

[0070] The meeting summary generation method provided by the embodiments of the present disclosure can implement long text segmentation on the literal meeting record, and generate a meeting summary according to multiple to-be-extracted segments, thereby saving computing resources and being applicable to meetings of a large duration. In addition, the to-be-extracted segments include at least part of the text content corresponding to a topic, thereby being conducive to ensuring the topic integrity and further being conducive to ensuring the extraction integrity and accuracy of the meeting information. Furthermore, the meeting summary generation method can save the time for manual writing and effectively improve the time-consuming and labor-intensive defects of manual recording, thereby being conducive to reducing the problems of missing important information and the like caused by human factors. Since different people may have style differences in recording meeting summaries, the meeting summary generation method can generate a meeting summary according to a preset format and template, thereby being conducive to ensuring that the format and content of the meeting summary remain standardized and consistent.

[0071] FIG. 2 is a schematic diagram of one example of step S300 shown in FIG. 1; and FIG. 3 is a schematic diagram of a first segment according to at least one embodiment of the present disclosure.

[0072] For example, as shown in FIG. 2, in some examples, step S300 shown in FIG. 1 includes steps S301-S323.

[0073] Step S301: In response to the last first segment dividing the part of the text belonging to the same topic at the end, merging the part of the text to the current first segment.

[0074] Step S310: Identifying the end identifier corresponding to the end sentence of the topic of the segment after merging the part of the text.

[0075] Step S321: In response to the current first segment not being the last first segment, and identifying multiple end identifiers, segmenting the text content after the second-to-last end identifier, and taking the remaining part as a second segment, to obtain the to-be-extracted segment according to the second segment.

[0076] Step S322: in response to the current first segment not being the last first segment and only one end identifier being identified, performing intra-topic segmentation on the segment after merging the partial text to obtain a plurality of topic segments, merging the last topic segment into the next first segment, and obtaining the to-be-extracted segment according to the remaining topic segments.

[0077] Step S323: in response to the current first segment being the last first segment, taking the current first segment as a second segment to obtain the to-be-extracted segment according to the second segment.

[0078] It should be noted that the above steps S321 to S323 are only executed one at a time, that is, when one of steps S321 to S323 is executed, the other two steps are not performed.

[0079] For example, as shown in FIG. 3, for step S301, if the current first segment is not the first first segment, the partial text content received from the last first segment is merged. For example, the partial text content of the last first segment can be the text content corresponding to the last topic of the last first segment. For example, the partial text content of the last first segment can also be the text content corresponding to the last topic of the last first segment.

[0080] For step S310, after merging the partial text content of the last first segment, the segment composed of the partial text content of the last first segment and the current first segment is used to identify the end identifier corresponding to the end sentence of the topic. For example, when multiple end identifiers are identified, it indicates that there are multiple topics. For example, each sentence is punctuated by a punctuation mark (such as a period, a question mark, or an exclamation mark), and each sentence has a corresponding identifier (for example, a corresponding number) before it. The identifier before the end sentence of each topic is the corresponding end identifier. For example, a large language model (for example, qwen1.5_14b) can be used to identify the end identifier corresponding to the end sentence of each topic. For example, for the first segment shown in FIG. 3, when the first to fifth sentences belong to the first topic and the sixth to eleventh sentences belong to the second topic, the large language model can be used to identify the end identifier 5 corresponding to the end sentence (i.e., the fifth sentence) of the first topic and identify the end identifier 11 corresponding to the end sentence (i.e., the eleventh sentence) of the second topic. It should be noted that regardless of whether the last topic of the first segment is complete, the last sentence of the first segment is the end sentence of the last topic. For example, in the case where the first segment shown in FIG. 3 includes 11 sentences as shown, regardless of whether the second topic is complete, the eleventh sentence is the end sentence of the second topic, and the end identifier 11 will be identified.

[0081] For example, for step S321, when multiple end identifiers are identified in the segment composed of the partial text content of the previous first segment and the current first segment, it indicates that the segment includes multiple topics. In the case that the current first segment is not the last first segment, the text content after the second last end identifier, i.e., the text content corresponding to the last topic, is merged into the next first segment, so that the remaining text content in the segment composed of the partial text content of the previous first segment and the current first segment forms a second segment, and then the to-be-extracted segment can be obtained according to the second segment.

[0082] For example, the text lengths of the multiple second segments obtained through step S321 can be different, but the topic integrity can be guaranteed to a great extent. For example, the second segment can include multiple topics, and the content of each topic is complete. For example, the second segment can also include one topic, i.e., the text content in the second segment belongs to the same topic.

[0083] FIG. 4 is a schematic diagram of an example of step S321 shown in FIG. 2.

[0084] For example, as shown in FIG. 4, step S321 includes steps S3211-S3212.

[0085] Step S3211: performing topic segmentation processing on the second segment to obtain at least one topic segment corresponding to the second segment, the topic segment including at least part of the text content of one topic.

[0086] Step S3212: obtaining the to-be-extracted segment according to the topic segment.

[0087] FIG. 5 is a schematic diagram of a second segment obtained according to the first segment shown in FIG. 3; and FIG. 6 is a schematic diagram of another second segment provided by at least one embodiment of the present disclosure.

[0088] For example, for step S3211, the topic segment can be determined according to the end identifier of the topic in the second segment, and the number of end identifiers of the topic in the second segment is equal to the number of topic segments. For example, as shown in FIG. 5, the second segment includes only one end identifier, so the second segment includes only one topic, and thus the second segment can be directly used as the topic segment. For example, as shown in FIG. 6, when the second segment includes multiple end identifiers, the second segment can be segmented into multiple topic segments, and the positions of the arrows represent the segmentation positions of the topic segments.

[0089] For example, for step S3212, the to-be-extracted segment can be determined according to the text length of the topic segment. For example, when the text length of the topic segment meets the requirement of the computing resource (or threshold), the topic segment can be directly taken as the to-be-extracted segment. For example, when the text length of the topic segment exceeds the requirement of the computing resource (or threshold), the topic segment can be subjected to re-segmentation processing (for example, topic-in-topic segmentation and the like, for details, please refer to the description of the embodiments below) to obtain the to-be-extracted segment meeting the requirement.

[0090] The topic segment obtained by subjecting the second segment to topic segmentation only includes at least part of the content of one topic, and then the to-be-extracted segment for extracting the summary of the passage is obtained according to the topic segment, which is beneficial to guarantee the integrity of the topic and in turn the integrity and accuracy of the extraction of the conference information.

[0091] For example, for step S322, if the segment formed by merging the text of the current first segment has only one end identifier, it indicates that the segment only includes one topic, and in the case that the current first segment is not the last first segment, the segment formed by merging can be subjected to topic-in-topic segmentation to obtain a plurality of topic segments, the last topic segment is merged into the next first segment, and the to-be-extracted segment is obtained according to the remaining topic segments. For details of step S322, please refer to the related description of FIG. 11 in the subsequent embodiments.

[0092] For example, for step S323, if the current first segment is the last first segment, the first segment is taken as the second segment. For example, the topic segmentation processing can be performed to obtain at least one topic segment, and the to-be-extracted segment is obtained according to the topic segment. For example, the last first segment can be segmented into a plurality of topic segments according to the identified end identifier in the first segment, and each topic segment includes at least part of the text content corresponding to one topic. For details, please refer to steps S3211-S3212 in the embodiments above, which will not be repeated here.

[0093] FIG. 7 is a schematic diagram of one example of step S3212 shown in FIG. 4.

[0094] For example, as shown in FIG. 7, step S3212 includes steps S3210-S3230.

[0095] Step S3210: Obtain the text length of the topic segment.

[0096] Step S3220: In response to the text length of the topic segment being not greater than the first threshold, determine the topic segment as the to-be-extracted segment.

[0097] Step S3230: In response to the text length of the topic segment being greater than the first threshold value, performing intra-topic segmentation processing on the topic segment to obtain a plurality of to-be-extracted segments.

[0098] For example, for step S3210, as shown in FIG. 5, considering the support capability of the computing resource, the text length of the topic segment can be obtained by a large language model, which is used to determine the to-be-extracted segment.

[0099] For example, for step S3220, the first threshold value can be set for the text length of the to-be-extracted segment, that is, the text length of each to-be-extracted segment needs to be no greater than the first threshold value, so as to perform the extraction of the summary of the paragraph. When the text length of the topic segment is no greater than the first threshold value, it indicates that the text length of the topic segment meets the support range of the computing resource, and therefore the topic segment can be directly used as the to-be-extracted segment.

[0100] For example, for step S3230, when the text length of the topic segment is greater than the first threshold value, it indicates that the topic segment cannot be directly extracted for the summary of the paragraph, and needs to be processed by intra-topic segmentation, thereby obtaining a plurality of to-be-extracted segments that meet the support of the computing resource.

[0101] Therefore, the to-be-extracted segment is determined according to the text length of the topic segment, which is beneficial to obtain the to-be-extracted segment that meets the support of the computing resource while ensuring the integrity of the topic as much as possible.

[0102] FIG. 8 is a schematic diagram of an example of step S3230 shown in FIG. 7; and FIG. 9 is a schematic diagram of intra-topic segmentation in at least one embodiment of the present disclosure.

[0103] For example, as shown in FIG. 8, step S3230 includes steps S3231-S3236.

[0104] Step S3231: dividing the text in the topic segment into a plurality of to-be-processed units in units of speakers, wherein the plurality of to-be-processed units include adjacent first to-be-processed unit and second to-be-processed unit, and the first to-be-processed unit is located before the second to-be-processed unit.

[0105] Step S3232: obtaining an embedding vector of each to-be-processed unit by a model.

[0106] Step S3233: calculating the similarity between the embedding vector corresponding to the first to-be-processed unit and the embedding vector corresponding to the second to-be-processed unit.

[0107] Step S3234: in response to the similarity being less than a second threshold value, setting the text corresponding to the first to-be-processed unit and the second to-be-processed unit in different topic segments, respectively.

[0108] Step S3235: In response to the similarity being not less than the second threshold, the texts corresponding to the first to-be-processed unit and the second to-be-processed unit are arranged in the same topic segment.

[0109] Step S3236: Obtain the to-be-extracted segment according to the topic segment.

[0110] For example, referring to FIG. 9, for step S3231, the texts in the topic segment are divided into a plurality of to-be-processed units in units of speakers, for example, each to-be-processed unit can be marked as “turn”, but is not limited thereto. For example, the to-be-processed unit can be a sentence of one speaker, and the first to-be-processed unit and the second to-be-processed unit can be two adjacent sentences, respectively.

[0111] For example, referring to FIG. 9, for step S3232, the plurality of to-be-processed units obtained by step S3231 can be respectively input into a language model, for example, the language model can be a large language model, for example, an MPNet (Multi-Perspective Sentence Embedding Network) network, but is not limited thereto. For example, the model can convert the to-be-processed unit into an embedding vector, which can represent the semantic content of the to-be-processed unit.

[0112] For example, referring to FIG. 9, for step S3233, the cosine similarity between the embedding vector corresponding to the first to-be-processed unit and the embedding vector corresponding to the second to-be-processed unit can be calculated by the model. For example, the cosine similarity can evaluate the semantic similarity between the first to-be-processed unit and the second to-be-processed unit. For example, the cosine similarity between the to-be-processed units belonging to the same topic needs to be not less than a specific value, i.e., a second threshold, which can be flexibly set according to design requirements (for example, can be 0.2), and when the cosine similarity between different to-be-processed units is not less than the second threshold, it indicates that the different to-be-processed units have a required semantic similarity.

[0113] For example, referring to FIG. 9, for step S3234, when the similarity between the embedding vector corresponding to the first to-be-processed unit and the embedding vector corresponding to the second to-be-processed unit is less than the second threshold, it indicates that the similarity (for example, the semantic similarity) between the first to-be-processed unit and the second to-be-processed unit does not meet the requirements, and the first to-be-processed unit and the second to-be-processed unit do not belong to the same topic. For example, the first to-be-processed unit and the second to-be-processed unit can be taken as a segmentation point, and the first to-be-processed unit and the second to-be-processed unit are arranged in different topic segments. For example, the topic segment can be segmented into a plurality of topic segments, and the embodiments of the present disclosure are not limited in number of topic segments.

[0114] For example, referring to FIG. 9, for step S3235, when the similarity between the embedding vector corresponding to the first to-be-processed unit and the embedding vector corresponding to the second to-be-processed unit is not less than the second threshold value, it indicates that the semantic similarity between the first to-be-processed unit and the second to-be-processed unit can meet the requirements, and the first to-be-processed unit and the second to-be-processed unit belong to the same topic. At this time, the theme segment will not be segmented between the first to-be-processed unit and the second to-be-processed unit. In some embodiments, all to-be-processed units in the theme segment can belong to the same topic, in which case the theme segment can be directly taken as a topic segment. The above steps continue until all to-be-processed units are traversed.

[0115] In this way, by calculating the similarity between different to-be-processed units in the theme segment, it can be evaluated whether different to-be-processed units belong to the same topic, so that the theme segment can be segmented according to the topic to obtain the topic segment, and thus the semantic relevance of the text content in the topic segment can be effectively guaranteed while meeting the support capability of the computing resource, thereby facilitating the improvement of the accuracy of the subsequently extracted summary.

[0116] For example, for step S3236, the to-be-extracted segment can be obtained according to the text length of the topic segment. For example, the text lengths of the plurality of topic segments can be unequal. For example, when the text length of the topic segment exceeds the support capability of the computing resource, for example, is greater than the first threshold value, the topic segment needs to be further processed to obtain a segment that meets the requirements, i.e., the to-be-extracted segment.

[0117] FIG. 10 is a schematic diagram of an example of step S3236 shown in FIG. 8.

[0118] For example, as shown in FIG. 10, step S3236 includes steps S32361-S32363.

[0119] Step S32361: determining whether the text length of the topic segment is greater than the first threshold value.

[0120] Step S32362: in response to the text length of the topic segment being not greater than the first threshold value, taking the topic segment as the to-be-extracted segment.

[0121] Step S32363: in response to the text length of the topic segment being greater than the first threshold value, segmenting the topic segment into a plurality of sub-topic segments according to the text length, wherein the text length of each sub-topic segment is not greater than the first threshold value, and each sub-topic segment is taken as the to-be-extracted segment.

[0122] For example, for step S32361, the text length of the topic segment can be recognized by a large language model, and it is determined whether the text length is greater than the first threshold value.

[0123] For example, for step S32362, when the text length of the topic segment is not greater than the first threshold value, it indicates that the topic segment meets the support capability of the computing resource, and the topic segment can be directly taken as the to-be-extracted segment.

[0124] For example, for step S32363, when the text length of the topic segment is greater than the first threshold value, it indicates that the topic segment exceeds the support capability of the computing resource, and the topic segment needs to be further segmented. For example, when the quotient of the text length of the topic segment and the first threshold value is a positive integer, the topic segment can be evenly segmented into a plurality of sub-topic segments with the text length of the first threshold value. For example, when the quotient of the text length of the topic segment and the first threshold value is not an integer (for example, the quotient is 5.2), the topic segment can be segmented into a plurality of sub-topic segments with the text length of the quotient of the text length of the topic segment and the first threshold value plus 1 (for example, 6). Therefore, the text length of the sub-topic segment is not greater than the first threshold value, and the sub-topic segment can be taken as the to-be-extracted segment.

[0125] Therefore, while meeting the support capability of the computing resource, the semantic relevance of the text content in the sub-topic segment can be effectively ensured, and the accuracy of the subsequent extracted summary can be improved.

[0126] FIG. 11 is a schematic diagram of an example of step S322 shown in FIG. 2.

[0127] For example, as shown in FIG. 11, step S322 includes steps S3221-S3222.

[0128] Step S3221: in response to the text length of the segment formed after the current first segment merges part of the text being greater than the first threshold value, performing intra-topic segmentation processing on the formed segment to obtain a plurality of topic segments.

[0129] Step S3222: in response to obtaining a plurality of topic segments by the first segment, merging the text content in the last topic segment into the next first segment, and obtaining to-be-extracted segments according to the remaining topic segments except the last topic segment in the plurality of topic segments.

[0130] For example, for step S3221, when the segment formed after the current first segment merges part of the text has only one end identifier, it indicates that the segment only includes one topic. If the current first segment is not the last first segment, the text length of the segment formed after the current first segment merges part of the text can be identified by the large language model. Generally, the text length of the segment formed after the current first segment merges part of the text is greater than the first threshold value, that is, the formed segment needs to be further segmented to obtain a segment that can meet the support capability of the computing resource, that is, a to-be-extracted segment.

[0131] For example, if the text length of the segment formed after the current first segment merges the partial text is greater than the first threshold value, it means that the segment exceeds the support capability of the computing resource, and the segment needs to be subject to intra-topic segmentation. For example, the segment can be regarded as a “to-be-processed topic segment”, and the intra-topic segmentation manner of the segment can refer to the description of steps S3231-S3235 in FIG. 8 in the above embodiment, and will not be described herein again. Thus, at least one topic segment can be obtained.

[0132] For example, for step S3222, in a case where the segment formed after the current first segment merges the partial text obtains multiple topic segments, the last topic segment is merged into the next first segment. In some cases, the last topic segment can belong to the same topic as the text content at the beginning of the next first segment. Therefore, by setting in this way, after subsequent topic segmentation and intra-topic segmentation of the next first segment, the content at the beginning of the next first segment is in the same topic segment, which can ensure the integrity of the topic as much as possible.

[0133] For example, the to-be-extracted segment can be obtained according to the remaining topic segments except the last topic segment in the multiple topic segments obtained by the segment formed after the current first segment merges the partial text. For example, the method of obtaining the to-be-extracted segment according to the topic segment can refer to the description of steps S32361-S32363 in FIG. 10 in the above embodiment, and will not be described herein again. Thus, at least one to-be-extracted segment can be obtained.

[0134] FIG. 12 is a schematic diagram of obtaining a to-be-extracted segment in at least one embodiment of the present disclosure.

[0135] For example, as shown in FIG. 12, first, the number n of first segments after the first segmentation processing is calculated according to the text length of the to-be-segmented text conference record. According to the statistical results of more than 1000 conference data, the number of words of a discussion topic is about 1000-2000. Therefore, the text length of the first segment is set to be greater than 2000 in this embodiment, and the upper limit of the length of the first segment is 4096 in this embodiment, but the embodiments of the present disclosure are not limited thereto. Then, the first segmentation processing is performed to obtain n first segments. A large language model, i.e., qwen1.5_14b, is used to perform topic segmentation on each first segment. For example, the end identifier corresponding to the end sentence of each topic in the first segment can be identified.

[0136] For example, as shown in FIG. 12, when multiple end markers are identified in a first segment, and the first segment is not the last first segment, the text content after the second last end marker in the first segment, i.e., the text content corresponding to the last topic, is merged into the next first segment, so that the remaining text content of the first segment constitutes a second segment, and then the to-be-extracted segment can be obtained according to the second segment. For example, the second segment can be subjected to topic segmentation to obtain at least one topic segment, each topic segment including at least part of the text content of a topic. For example, if the current first segment is the last first segment, the first segment is subjected to topic segmentation to obtain multiple topic segments.

[0137] For example, as shown in FIG. 12, the to-be-extracted segment can be determined according to the text length of the topic segment, i.e., when the text length of the topic segment is not greater than the maximum number of words (i.e., the first threshold), the topic segment is determined as the to-be-extracted segment, otherwise, the to-be-extracted segment is obtained by subjecting the topic segment to intra-topic segmentation. For details about the specific process of determining the to-be-extracted segment according to the topic segment, please refer to the specific description of step S3212 in the above embodiment, which will not be repeated here.

[0138] FIG. 13 is a schematic diagram of an example of step S100 shown in FIG. 1.

[0139] For example, as shown in FIG. 13, step S100 includes steps S1010-S1040.

[0140] Step S1010: Obtain an original text conference record.

[0141] Step S1020: Subject the original text conference record to text normalization to obtain a text normalization evaluation value.

[0142] Step S1030: In response to the text normalization evaluation value being not less than a third threshold, take the text after text normalization as the text conference record to be segmented.

[0143] Step S1040: In response to the text normalization evaluation value being less than the third threshold, take the original text conference record as the text conference record to be segmented.

[0144] For example, for step S1010, the speech recognition algorithm (Speech Recognition) and the voiceprint recognition algorithm (Voiceprint Recognition) mentioned in the above embodiment can be used to obtain, i.e., by obtaining the speech transcription result, the speaker, and the speaking time at different times of the conference to obtain the original text conference record. The original text conference record is an unprocessed text.

[0145] For example, for step S1020, in order to remove meaningless modal particles and redundant expressions and the like in the original text conference record, a large language model can be used to perform text processing on the original text conference record, so that the original text conference record is reduced in the number of words under the premise of being more processed and without losing information, so as to improve the accuracy of the conference minutes generated subsequently. However, for the text after language processing, it is necessary to evaluate whether effective information is missing, that is, it is necessary to evaluate the effect of text processing. For example, a text processing evaluation value can be calculated, and it is determined whether the text processing evaluation value meets the set threshold requirement (for example, a third threshold). For example, the third threshold can be 90%, but is not limited thereto, and can be set according to design requirements.

[0146] For example, for step S1030, when the text processing evaluation value is not less than the third threshold, it indicates that the text processing effect meets the requirement, and therefore the text after text processing can be directly used as the text conference record to be segmented.

[0147] For example, for step S1040, when the text processing evaluation value is less than the third threshold, it indicates that the text processing effect is poor and does not meet the requirement, and in this case, the original text conference record is used as the text conference record to be segmented.

[0148] By comparing the text processing evaluation value of the text after text processing with the third threshold, the effect of text processing can be quantitatively evaluated, so as to reduce the risk of missing effective information in the determined text conference record to be segmented, thereby ensuring the accuracy of the conference minutes generated finally.

[0149] FIG. 14 is a schematic diagram of text processing on an original text conference record in at least one embodiment of the present disclosure; and FIG. 15 is a schematic diagram of part of the text processing result corresponding to FIG. 14.

[0150] For example, as shown in FIGS. 14 and 15, for step S1020 described above, the original text conference record can be first processed by a model to obtain a first processed text; then the first processed text is processed by a human to obtain a second processed text; based on the first processed text and the second processed text, a first number of words, a second number of words and a third number of words are obtained; and based on the first number of words, the second number of words and the third number of words, a text processing evaluation value is obtained. The first number of words represents the number of words in the original text conference record that are modified in both the first processed text and the second processed text, the second number of words represents the number of words in the original text conference record that are modified in the first processed text and are not modified in the second processed text, and the third number of words represents the number of words in the original text conference record that are not modified in the first processed text and are modified in the second processed text.

[0151] For example, as shown in FIG. 14 and FIG. 15, a large language model can be employed to perform discourse regularization on the original literal conference record to obtain a first regularized text. The edit mode of each character in the first regularized text relative to the original literal conference record can be keep (e.g., denoted as Keep), insert (e.g., denoted as Insert), delete (e.g., denoted as Delete), or replace (e.g., denoted as Replace). When the edit mode of a character is keep, it means that the character is not modified in the first regularized text. When the edit mode of a character is insert, delete, or replace, it means that the character is modified in the first regularized text.

[0152] For example, as shown in FIG. 14 and FIG. 15, since there can still be characters in the first regularized text that need to be regularized, manual regularization can be directly performed on the first regularized text to further remove unnecessary discourse particles, redundant expressions, etc., thereby obtaining a second regularized text. Similarly, the edit mode of each character in the second regularized text relative to the first regularized text can be keep, insert, delete, or replace. After manual regularization, when the edit mode of a character is keep, it means that the character is not modified in the second regularized text, and when the edit mode of a character is insert, delete, or replace, it means that the character is modified in the second regularized text.

[0153] For example, as shown in FIG. 14 and FIG. 15, based on the first regularized text and the second regularized text, the number of characters that are modified in both the first regularized text and the second regularized text, i.e., the number of characters whose edit mode is not Keep in both the first regularized text and the second regularized text, can be obtained, thereby obtaining a first number. For example, by calculating the number of characters that are modified in the first regularized text and not modified in the second regularized text, i.e., the number of characters whose edit mode is not Keep in the first regularized text and whose edit mode is Keep in the second regularized text, a second number can be obtained. For example, by calculating the number of characters that are not modified in the first regularized text and modified in the second regularized text, i.e., the number of characters whose edit mode is Keep in the first regularized text and whose edit mode is not Keep in the second regularized text, a third number can be obtained. In some embodiments, by calculating the number of characters that are not modified in both the first regularized text and the second regularized text, i.e., the number of characters whose edit mode is Keep in both the first regularized text and the second regularized text, a fourth number can be obtained.

[0154] For example, the discourse regularization evaluation value can be obtained by the following formula:

[0155] In the formula, TP represents the first number, FP represents the second number, FN represents the third number, P represents the accuracy, R represents the recall, and F represents the discourse regularization evaluation value.

[0156] According to the above formula, the first word number and the second word number can be used to calculate the precision, the first word number and the third word number can be used to calculate the recall, and the text normalization evaluation value is obtained based on the precision and the recall. Thus, the calculated text normalization evaluation value can be compared with the third threshold value, and the effect of text normalization is evaluated to determine the to-be-segmented text conference record.

[0157] FIG. 16 is a schematic diagram of an editing mode of obtaining characters according to at least one embodiment of the present disclosure; and FIG. 17 is a schematic diagram of a text normalization effect according to at least one embodiment of the present disclosure.

[0158] For example, the first word number, the second word number and the third word number are obtained by: obtaining a minimum edit distance dp[i,j] required for normalizing text and original text conference record from s1[0:i+1] to s2[0:j+1], the normalizing text being the first normalizing text or the second normalizing text, s1[0:i+1] representing characters from the 0th position to the ith position in the original text conference record, and s2[0:j+1] representing characters from the 0th position to the jth position in the normalizing text; and obtaining the first word number, the second word number and the third word number according to the minimum edit distance dp[i,j]. Hereinafter, the normalizing text is taken as the first normalizing text as an example.

[0159] For example, the minimum edit distance dp[i,j] can be obtained by the following formula:

[0160] In the formula, s1[i] represents the ith character in the original text conference record, s2[j] represents the jth character in the normalizing text, i and j are integers, edit=keep represents the editing mode is keeping, edit=insert represents the editing mode is inserting, edit=delete represents the editing mode is deleting, and edit=replace represents the editing mode is replacing.

[0161] It should be noted that in some cases, when dp[i,j]=min(dp[i,j-1]+1,dp[i-1,j]+1,dp[i-1,j-1]+1), and the number of the minimum value in dp[i,j-1]+1, dp[i-1,j]+1, dp[i-1,j-1]+1 is greater than 1, then one of them can be selected as dp[i,j]. For example, when dp[i,j-1]+1 and dp[i-1,j]+1 are equal, and both are less than dp[i-1,j-1]+1, then any one of dp[i,j-1]+1 and dp[i-1,j]+1 can be taken as dp[i,j]. For example, the editing mode of the minimum edit distance dp[i,j] corresponds to the type of the value of dp[i,j].

[0162] For example, as shown in FIG. 16, the original literal conference record is "play", which includes four characters p, l, a and y, p is the 0th character, l is the 1st character, a is the 3rd character, and y is the 4th character. The first regularized text is "slary", which includes five characters s, l, a, r and y, and similarly, s is the 0th character, and y is the 5th character. For example, the minimum edit distance from the 0th to 3rd characters (i.e., s1[0:3]) in the original literal conference record "play" to the 0th to 4th characters (i.e., s2[0:4]) in the first regularized text "slary" is dp[3,4]. For example, the path for obtaining dp[3,4] can be obtained according to the above-mentioned company, please refer to the part filled with gray in FIG. 16.

[0163] For example, referring to FIG. 17, the editing manner of each character of the first regularized text relative to the original literal conference record can be obtained according to the above-mentioned formula. For example, the character with a deletion line indicates that it is deleted, and the corresponding editing manner is deletion. The character in the black solid line box indicates a newly inserted character, and the corresponding editing manner is insertion. The black character without a deletion line indicates a character that is regularized, and the corresponding editing manner is maintenance. It should be noted that the above-mentioned black solid line box is only a representative manner, and embodiments of the present disclosure are not limited thereto, for example, the color of the newly added character can be different from the color of the original literal, so as to be distinguished. For example, the editing effect of the editing manner of the character being replacement can be equivalent to the editing effect of being deleted and then inserted.

[0164] For example, according to the above-mentioned method, the editing manner of each character of the second regularized text relative to the first regularized text can also be obtained. Thus, by taking the number of characters whose editing manner in the first regularized text and the second regularized text are both not maintenance as the first word number, taking the number of characters whose editing manner in the first regularized text is not maintenance and whose editing manner in the second regularized text is maintenance as the second word number, and taking the number of characters whose editing manner in the first regularized text is maintenance and whose editing manner in the second regularized text is not maintenance as the third word number, the discourse regularized evaluation value can be obtained.

[0165] FIG. 18 is a flowchart of generating a conference minutes according to at least one embodiment of the present disclosure.

[0166] For example, as shown in FIG. 18, the original conference audio is obtained, and then the speech transcription results and corresponding speakers at different times are obtained through a speech recognition algorithm and a voiceprint recognition algorithm. In order to remove adverbs and redundant expressions in the speech transcription results, a large language model can be used for text processing to reduce the number of words under the premise of more processing and without losing information, so as to improve the accuracy of the subsequent generated conference minutes. Then, the text conference record to be segmented is segmented according to the text length to obtain a plurality of to-be-extracted segments. For example, the segmentation processing can include the first segmentation processing and the second segmentation processing described above. In the process of obtaining the plurality of to-be-extracted segments through the second segmentation processing, the subject segmentation process (see FIG. 12) and the intra-subject segmentation process (see FIG. 9) mentioned in the above embodiments can be included, which will not be repeated here. After obtaining the plurality of to-be-extracted segments, the to-be-extracted segments can be input into a large prediction model (such as ChatGPT) to generate a paragraph summary according to a given prompt. For example, the given prompt can include a prompt for extracting the conference process and a prompt for extracting the to-do list. Then, each paragraph summary can be used to integrate the full-text conference minutes. According to the user's preset template, the generated paragraph summary is integrated into the full-text conference minutes that meet the user's preset template requirements through the large language model by designing the prompt.

[0167] In some embodiments, the method of obtaining a plurality of subject segments through subject segmentation processing is not limited to the method provided in the above embodiments, and a small model such as a PoNet (Pooling Network) model can also be used for subject segmentation. The embodiments of the present disclosure do not limit the way of subject segmentation.

[0168] For example, the text conference record to be segmented (for example, the text conference record can be a text after processing) can be input into the PoNet network for subject segmentation. The PoNet model can model, identify and segment the part with consistent subject through its unique network structure. Since the PoNet uses a pooling mechanism to replace the self-attention mechanism in the traditional transformer model, it has higher efficiency when processing long sequence text.

[0169] The PoNet network mainly includes three different granularity pooling networks, i.e., a global pooling module (also referred to as a GA module), a segmented segment max-pooling module (also referred to as an SMP module), and a local max-pooling module (also referred to as an LMP module), which can capture sequence information of different granularities, so as to more accurately understand the topic structure of the text. The role of the global pooling module is to capture global information, i.e., to capture the overall features of the entire text sequence. The implementation manner is to extract global feature representations by performing a pooling operation on the entire sequence. These features can reflect the overall theme or sentiment tendency of the text. The role of the segmented segment max-pooling module is to capture local information of different paragraphs in the text. The implementation manner is to segment the text according to certain rules (such as sentences or paragraphs), and perform a max-pooling operation on each paragraph. In this way, the segment max-pooling module can extract the main features of each paragraph, which helps the model to understand the theme changes in different parts of the text. The role of the local max-pooling module is to focus on local detail information in the text, such as keywords or phrases. The implementation manner is to perform a max-pooling operation in a local range of the sequence (such as a sentence or a few words). This helps the model to capture key information points in the text, further enhancing the model's understanding ability of the text theme.

[0170] FIG. 19 is a schematic block diagram of a meeting summary generation apparatus provided by at least one embodiment of the present disclosure. As shown in FIG. 19, the meeting summary generation apparatus 1000 includes an acquisition unit 1010, a first segmentation unit 1020, a second segmentation unit 1030, and a generation unit 1040. For example, these units can be implemented by hardware (such as a circuit) module or a software module, for example, can be implemented by a central processing unit (CPU), a graphics processing unit (GPU), a tensor processing unit (TPU), a field programmable logic gate array (FPGA), or other forms of processing units with data processing capability and / or instruction execution capability, and corresponding computer instructions, and the embodiments of the present disclosure do not make any limitation in this regard.

[0171] As shown in FIG. 19, the acquisition unit 1010 is configured to acquire a text conference record to be segmented, and the specific implementation manner can refer to the related description of step S100, which will not be described here again.

[0172] As shown in FIG. 19, the first segmentation unit 1020 is configured to perform first segmentation processing on the text conference record to be segmented according to the text length to obtain a plurality of first segments, and the specific implementation manner can refer to the related description of step S200, which will not be described here again.

[0173] As shown in FIG. 19, the second segment unit 1030 is configured to perform second segment processing on each of the plurality of first segments to obtain a plurality of to-be-extracted segments, wherein the to-be-extracted segment includes at least part of the text content corresponding to a topic. For the specific implementation of the second segment unit 1030, please refer to the related description of step S300, which will not be repeated here.

[0174] As shown in FIG. 19, the generation unit 1040 is configured to generate the meeting minutes based on the plurality of to-be-extracted segments. For the specific implementation of the generation unit 1040, please refer to the related description of step S400, which will not be repeated here.

[0175] It should be noted that in the embodiments of the present disclosure, the meeting minutes generation apparatus can include more or fewer circuits or units, and the connection relationship between the circuits or units is not limited, and can be determined according to actual needs. The specific constituting manner of each circuit is not limited, which can be constituted by an analog device according to the circuit principle, or can be constituted by a digital chip, or constituted in other applicable manner.

[0176] FIG. 20 is a schematic block diagram of an electronic device provided by at least one embodiment of the present disclosure. For example, as shown in FIG. 20, the electronic device 2000 includes a processor 2100, a memory 2200, and one or more computer program modules 2210.

[0177] For example, the processor 2100 and the memory 2200 are connected through a bus system 2300. For example, the one or more computer program modules 2210 are stored in the memory 2200. For example, the one or more computer program modules 2210 include instructions for performing the meeting minutes generation method provided by any embodiment of the present disclosure. For example, the instructions in the one or more computer program modules 2210 can be executed by the processor 2100. For example, the bus system 2300 can be a commonly used serial, parallel communication bus, etc., and the embodiments of the present disclosure do not limit this.

[0178] For example, the processor 2100 can be a central processing unit (CPU), a digital signal processor (DSP), a graphics processing unit (GPU), or other forms of processing units with data processing and / or instruction execution capabilities, which can be general-purpose processors or special-purpose processors, and can control other components in the data processing apparatus 2000 to perform desired functions. For example, in the embodiments of the present disclosure, the processor is taken as an example of a graphics processing unit (GPU).

[0179] For example, the memory 2200 can include one or more computer program products that can include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory, for example, can include random access memory (RAM), cache memory, and / or the like. The non-volatile memory, for example, can include read-only memory (ROM), hard disk, flash memory, and / or the like. One or more computer program instructions can be stored on the computer-readable storage media, and the processor 2100 can execute the program instructions to implement the functions (implemented by the processor 2100) in the embodiments of the present disclosure and / or other desired functions, such as the method for generating a meeting minutes and the like. Various application programs and various data, such as various data used and / or generated by the application programs, and the like, can also be stored in the computer-readable storage media.

[0180] It should be noted that, for the sake of clarity and simplicity, the present embodiments do not give all the constituent units of the electronic device 2000. To achieve the necessary functions of the electronic device 2000, those skilled in the art can provide and set other constituent units not shown according to specific needs, and the embodiments of the present disclosure do not limit this.

[0181] FIG. 21 is a schematic diagram of another electronic device according to at least one embodiment of the present disclosure.

[0182] The method and / or apparatus for generating a meeting minutes according to the embodiments of the present disclosure can also be implemented by means of the architecture of an exemplary electronic device 3000 as shown in FIG. 21. As shown in FIG. 21, the electronic device 3000 can include a bus 3010, one or more central processing units (CPUs) or graphics processing units (GPUs) 3020, read-only memory (ROM) 3030, random access memory (RAM) 3040, a communication port connected to a network 3050, input / output components 3060, a hard disk 3070, and the like. The storage devices in the electronic device 3000, such as the ROM 3030 or the hard disk 3070, can store various data or files required for processing and / or communication of the method provided by the present disclosure and program instructions executed by the CPU or GPU. The electronic device 3000 can also include a user interface 3080. Of course, the architecture shown in FIG. 21 is only exemplary, and when implementing different devices, one or more components of the electronic device shown in FIG. 21 can be omitted according to actual needs.

[0183] The at least one embodiment of the present disclosure further provides a storage medium. FIG. 22 is a schematic diagram of a storage medium according to at least one embodiment of the present disclosure. For example, as shown in FIG. 22, the storage medium 400 non-transitorily stores computer readable instructions 401 which, when executed by a computer (including a processor), can perform the method for generating a meeting summary according to any embodiment of the present disclosure.

[0184] For example, the storage medium can be any combination of one or more computer readable storage media, such as one computer readable storage medium containing computer readable program code for obtaining a text conference record to be segmented, another computer readable storage medium containing computer readable program code for performing first segmentation on the text conference record to be segmented according to text length to obtain a plurality of first segments, another computer readable storage medium containing computer readable program code for performing second segmentation on each of the plurality of first segments to obtain a plurality of segments to be extracted, and another computer readable storage medium containing computer readable program code for generating a meeting summary based on the plurality of segments to be extracted. For example, when the program code is read by a computer, the computer can execute the program code stored in the computer storage medium, such as the method for generating a meeting summary according to any embodiment of the present disclosure.

[0185] For example, the storage medium can include a memory card of a smart phone, a memory component of a tablet computer, a hard disk of a personal computer, a random access memory (RAM), a read only memory (ROM), an erasable programmable read only memory (EPROM), a portable compact disc read only memory (CD-ROM), a flash memory, or any combination of the above storage media, and can also be other applicable storage media.

[0186] The following points need to be explained:

[0187] (1) In the drawings of the embodiments of the present disclosure, only the structures related to the embodiments of the present disclosure are involved, and other structures can be referred to the general design.

[0188] (2) In the case of no conflict, the features in the same embodiment and different embodiments of the present disclosure can be combined with each other.

[0189] The above description is only exemplary embodiments of the present disclosure, and is not intended to limit the protection scope of the present disclosure, and the protection scope of the present disclosure is determined by the appended claims.

Claims

1. A method for generating a meeting summary, comprising: obtaining a text conference record to be segmented; performing first segmentation processing on the text conference record to be segmented according to text length to obtain a plurality of first segments; performing second segmentation processing on each of the plurality of first segments to obtain a plurality of to-be-extracted segments, wherein the to-be-extracted segments include at least part of the text content corresponding to a topic; and generating the meeting summary based on the plurality of to-be-extracted segments.

2. The method of claim 1, wherein, The second segmentation processing on each of the plurality of first segments to obtain a plurality of to-be-extracted segments comprises: in response to the last part of the last first segment being divided out of the part of the text belonging to the same topic, merging the part of the text into the current first segment; identifying the end marker corresponding to the topic end sentence of the segment after merging the part of the text; in response to the current first segment not being the last first segment, and identifying a plurality of end markers, segmenting the text content after the second last end marker, and taking the remaining part as a second segment to obtain the to-be-extracted segment according to the second segment; or in response to the current first segment not being the last first segment, and only identifying one end marker, performing intra-topic segmentation on the segment after merging the part of the text to obtain a plurality of topic segments, merging the last topic segment into the next first segment, and obtaining the to-be-extracted segment according to the remaining topic segments; or in response to the current first segment being the last first segment, taking the current first segment as the second segment to obtain the to-be-extracted segment according to the second segment.

3. The method of claim 2, wherein, The obtaining of the to-be-extracted segment according to the second segment comprises: performing the topic segmentation processing on the second segment to obtain at least one topic segment corresponding to the second segment, wherein the topic segment includes at least part of the text content of a topic; and obtaining the to-be-extracted segment according to the topic segment.

4. The method of claim 3, wherein, The obtaining of the to-be-extracted segment according to the topic segment comprises: obtaining the text length of the topic segment; in response to the text length of the topic segment being not greater than a first threshold, determining the topic segment as the to-be-extracted segment; and in response to the text length of the topic segment being greater than the first threshold, performing intra-topic segmentation processing on the topic segment to obtain a plurality of to-be-extracted segments.

5. The method of claim 4, wherein, The intra-topic segmentation processing on the topic segment to obtain a plurality of to-be-extracted segments comprises: dividing the text in the topic segment into a plurality of to-be-processed units in units of speakers, wherein the plurality of to-be-processed units include adjacent first and second to-be-processed units, and the first to-be-processed unit is located before the second to-be-processed unit; obtaining an embedding vector of each to-be-processed unit through a model; calculating the similarity between the embedding vector corresponding to the first to-be-processed unit and the embedding vector corresponding to the second to-be-processed unit; in response to the similarity being less than a second threshold, setting the texts corresponding to the first to-be-processed unit and the second to-be-processed unit in different topic segments respectively; in response to the similarity being not less than the second threshold, setting the texts corresponding to the first to-be-processed unit and the second to-be-processed unit in a same topic segment; and obtaining the to-be-extracted segment according to the topic segment.

6. The method of claim 5, wherein, The obtaining the to-be-extracted segment according to the topic segment comprises: judging whether a text length of the topic segment is greater than the first threshold; in response to the text length of the topic segment being not greater than the first threshold, taking the topic segment as the to-be-extracted segment; and in response to the text length of the topic segment being greater than the first threshold, dividing the topic segment into a plurality of sub-topic segments according to text lengths, wherein a text length of each of the sub-topic segments is not greater than the first threshold, and each of the sub-topic segments is taken as the to-be-extracted segment.

7. The method of claim 1-6, wherein, The obtaining the to-be-segmented text conference record comprises: obtaining an original text conference record; performing text normalization on the original text conference record to obtain a text normalization evaluation value; in response to the text normalization evaluation value being not less than a third threshold, taking the text after the text normalization as the to-be-segmented text conference record; and in response to the text normalization evaluation value being less than the third threshold, taking the original text conference record as the to-be-segmented text conference record.

8. The method of claim 7, wherein, The performing text normalization on the original text conference record to obtain a text normalization evaluation value comprises: performing model normalization on the original text conference record to obtain a first normalized text; performing manual normalization on the first normalized text to obtain a second normalized text; based on the first normalized text and the second normalized text, obtaining a first word number, a second word number and a third word number, wherein the first word number represents a word number that is modified in both the first normalized text and the second normalized text relative to the original text conference record, the second word number represents a word number that is modified in the first normalized text and not modified in the second normalized text relative to the original text conference record, and the third word number represents a word number that is not modified in the first normalized text and modified in the second normalized text relative to the original text conference record; and based on the first word number, the second word number and the third word number, obtaining the text normalization evaluation value.

9. The method of claim 8, wherein, The text regularity evaluation value is obtained by the following formula: In the formula, TP represents the first word number, FP represents the second word number, FN represents the third word number, P represents an accuracy rate, R represents a recall rate, and F represents the text normalization evaluation value.

10. The method of generating a meeting summary according to claim 8 or 9, wherein, The obtaining the first word number, the second word number and the third word number comprises: obtaining a minimum edit distance dp[i, j] required from s1[0:i+1] to s2[0:j+1] in the regularized text and the original literal conference record, wherein the regularized text is the first regularized text or the second regularized text, s1[0:i+1] represents characters from the 0th to the ith in the original literal conference record, and s2[0:j+1] represents characters from the 0th to the jth in the regularized text; and obtaining the first word number, the second word number and the third word number according to the minimum edit distance dp[i, j].

11. The method of claim 10, wherein, The minimum edit distance dp[i,j] is obtained by the following equation: In the formula, s1[i] represents the ith character in the original literal conference record, s2[j] represents the jth character in the regularized text, i and j are integers, edit = keep represents that the edit mode is keeping, edit = insert represents that the edit mode is inserting, edit = delete represents that the edit mode is deleting, and edit = replace represents that the edit mode is replacing.

12. The method of generating a meeting summary according to claim 10 or 11, wherein, The obtaining the first word number, the second word number and the third word number according to the minimum edit distance dp[i, j] comprises: taking the number of characters whose edit mode is neither keeping in the first regularized text nor keeping in the second regularized text as the first word number; taking the number of characters whose edit mode is not keeping in the first regularized text and whose edit mode is keeping in the second regularized text as the second word number; and taking the number of characters whose edit mode is keeping in the first regularized text and whose edit mode is not keeping in the second regularized text as the third word number. The generating the conference summary based on the plurality of to-be-extracted segments comprises:

13. The method of claim 1-12, wherein, extracting a paragraph summary for each of the plurality of to-be-extracted segments; and generating the conference summary according to the paragraph summaries. The obtaining the original literal conference record comprises:

14. The method of claim 7-13, wherein, obtaining a voice transcription result, a speaker and a speaking time of the conference at different time points through a voice recognition algorithm and a voiceprint recognition algorithm.

15. A conference summary generation apparatus, comprising: an obtaining unit configured to obtain a literal conference record to be segmented; a first segmenting unit configured to perform first segmenting processing on the literal conference record to be segmented according to a text length to obtain a plurality of first segments; a second segmenting unit configured to perform second segmenting processing on each of the plurality of first segments to obtain a plurality of to-be-extracted segments, wherein the to-be-extracted segments comprise at least part of text content corresponding to a topic; and a generating unit configured to generate a conference summary based on the plurality of to-be-extracted segments.

16. An electronic device, comprising: a processor; a memory; and one or more computer program modules, wherein the one or more computer program modules are stored in the memory and configured to be executed by the processor, and the one or more computer program modules comprise a program module for implementing a conference summary generation method according to any one of claims 1-14. ​ 17. A storage medium, non-transitorily storing computer readable instructions that, when executed by a computer, perform a method of generating a meeting summary according to any one of claims 1-14.

Images