Text abstract extraction method and device, equipment and storage medium

Abstract

The invention discloses a text abstract extraction method and device, equipment and a storage medium. The method comprises the following steps: determining a target text and a target abstract extraction model; under the condition that the length of the target text is larger than a preset length threshold value and the preset length threshold value is smaller than or equal to the upper limit of the input text length of the target abstract extraction model, determining the target text as the current text, and circularly executing the following steps: splitting the current text into a plurality of current sub-texts, and extracting the current sub-texts based on the target abstract extraction model, extracting sub-text abstracts for different current sub-texts respectively; synthesizing the plurality of extracted sub-text abstracts, and determining a current text abstract; under the condition that the length of the current text abstract is greater than a preset length threshold value, determining the current text abstract as a new current text; and when the length of the current text abstract is less than or equal to a preset length threshold, stopping circulation, and extracting a target text abstract for the current text abstract based on a target abstract extraction model.

Description

Technical Field

[0001] The present invention relates to the field of computer application technology, and in particular to a text summary extraction method, device, equipment and storage medium. Background Art

[0002] With the development of computer technology, it is often necessary to process information in text form. Among them, there are many needs for text summarization, for example, extracting meeting minutes from text generated in a meeting, extracting the main points of an article, and so on.

[0003] However, many current text summarization methods have an upper limit on the length of the input text. For example, 30 words, 50 words, 100 words, etc. When faced with long texts, it is often difficult to extract text summaries. For example, for a meeting that lasts for 5 hours, the length of the text generated is often tens of thousands of words, and current text summarization methods often have difficulty extracting meeting minutes.

[0004] Therefore, the length of the input text has a greater impact on text summary extraction. Summary of the invention

[0005] The present invention provides a text summary extraction method, device, equipment and storage medium to solve the deficiencies in the related art.

[0006] According to a first aspect of an embodiment of the present invention, a text summary extraction method is provided, comprising:

[0007] Determine the target text and target summary extraction model;

[0008] In the case where the length of the target text is greater than a preset length threshold, and the preset length threshold is less than or equal to the upper limit of the input text length of the target summary extraction model, the target text is determined as the current text, and the following steps are executed in a loop:

[0009] The current text is split into a plurality of current subtexts, and based on the target summary extraction model, subtext summaries are extracted for different current subtexts respectively; the length of any current subtext is less than or equal to the preset length threshold; and the current text summary is determined by synthesizing the extracted plurality of subtext summaries;

[0010] When the length of the current text summary is greater than the preset length threshold, determining the current text summary as a new current text;

[0011] When the length of the current text summary is less than or equal to the preset length threshold, the loop is stopped, and a target text summary is extracted for the current text summary based on the target summary extraction model.

[0012] Optionally, the splitting of the current text into a plurality of current sub-texts includes:

[0013] Split the current text into a current sub-text sequence; the current sub-text sequence includes different current sub-texts arranged in the order of the texts in the current text;

[0014] In the current subtext sequence, between any pair of adjacent current subtexts, the last N consecutive characters in the previous current subtext are the same as the first N consecutive characters in the next current subtext.

[0015] Optionally, the current text contains several paragraphs;

[0016] The current text is split into several current sub-texts, including:

[0017] Split the current text into several preliminary sub-texts; any preliminary sub-text contains at least one paragraph;

[0018] When the length of any prepared subtext is less than or equal to the preset length threshold, determining the prepared subtext as the current subtext;

[0019] When the length of any prepared subtext is greater than the preset length threshold, the prepared subtext is split into at least two current subtexts; and the length of any current subtext is less than or equal to the preset length threshold.

[0020] Optionally, the splitting of the current text into a plurality of current sub-texts includes:

[0021] Split the current text into a current sub-text sequence; the current sub-text sequence includes different current sub-texts arranged in the order of the texts in the current text;

[0022] In the current subtext sequence, between any pair of adjacent current subtexts, the latter current subtext includes the subtext summary extracted from the previous current subtext based on the target summary extraction model.

[0023] Optionally, determining the target text includes:

[0024] Determine a number of collected information; any collected information has a corresponding collection time point;

[0025] Convert the determined collected information into text;

[0026] The converted texts are synthesized to determine the target text; in the target text, the sequence of different texts converted from different collected information is arranged according to the sequence of the collected information corresponding to the collection time points.

[0027] Optionally, the method further comprises: determining user information that provides any collected information;

[0028] The step of converting the determined collected information into text includes:

[0029] The determined collected information and user information providing the determined collected information are converted into text.

[0030] Optionally, the preset length threshold comprises: a length obtained by subtracting a length of a preset prompt word from an upper limit of an input text length of the target summary extraction model;

[0031] The preset prompt word is used to prompt the target summary extraction model to extract a summary for the input text;

[0032] The step of extracting a target text summary based on the target summary extraction model for the current text summary includes:

[0033] The preset prompt word and the current text summary are input into the target summary extraction model to obtain the target text summary output by the target summary extraction model.

[0034] Optionally, the structure of the target summary extraction model includes any one of the following: a language model, a natural language processing model, a generative summary model, and an extractive summary model.

[0035] According to a second aspect of an embodiment of the present invention, a text summary extraction method is provided, comprising:

[0036] Determine the target text and target summary extraction method;

[0037] In the case where the length of the target text is greater than a preset length threshold, and the preset length threshold is less than or equal to the upper limit of the input text length of the target summary extraction method, the target text is determined as the current text, and the following steps are executed in a loop:

[0038] Determine a current text summary for the current text;

[0039] When the length of the current text summary is greater than the preset length threshold, determining the current text summary as a new current text;

[0040] When the length of the current text summary is less than or equal to the preset length threshold, the loop is stopped, and a target text summary is extracted for the current text summary based on the target summary extraction method.

[0041] Optionally, determining a current text summary for the current text includes:

[0042] Determine the current summary extraction method;

[0043] When the length of the current text is greater than a specified length threshold, and the specified length threshold is less than or equal to the upper limit of the input text length of the current summary extraction method, the current text is split into several current sub-texts, and based on the current summary extraction method, sub-text summaries are extracted for different current sub-texts respectively; the length of any current sub-text is less than or equal to the specified length threshold; and the current text summary is determined by combining the extracted several sub-text summaries.

[0044] According to a third aspect of an embodiment of the present invention, a method for extracting meeting minutes is provided, comprising:

[0045] Determine the target meeting text and target summary extraction model;

[0046] In the case where the length of the target conference text is greater than a preset length threshold, and the preset length threshold is less than or equal to the upper limit of the input text length of the target summary extraction model, the target conference text is determined as the current text, and the following steps are executed in a loop:

[0047] The current text is split into a plurality of current subtexts, and based on the target summary extraction model, subtext summaries are extracted for different current subtexts respectively; the length of any current subtext is less than or equal to the preset length threshold; and the current text summary is determined by synthesizing the extracted plurality of subtext summaries;

[0048] When the length of the current text summary is greater than the preset length threshold, determining the current text summary as a new current text;

[0049] When the length of the current text summary is less than or equal to the preset length threshold, the loop is stopped, and the target meeting minutes are extracted for the current text summary based on the target summary extraction model.

[0050] According to a fourth aspect of an embodiment of the present invention, a text summary extraction device is provided, comprising:

[0051] A determination unit, used for determining a target text and a target summary extraction model;

[0052] A loop unit is used for determining the target text as the current text and looping the following steps when the length of the target text is greater than a preset length threshold and the preset length threshold is less than or equal to an upper limit of the input text length of the target summary extraction model:

[0053] The current text is split into a plurality of current subtexts, and based on the target summary extraction model, subtext summaries are extracted for different current subtexts respectively; the length of any current subtext is less than or equal to the preset length threshold; and the current text summary is determined by synthesizing the extracted plurality of subtext summaries;

[0054] When the length of the current text summary is greater than the preset length threshold, determining the current text summary as a new current text;

[0055] When the length of the current text summary is less than or equal to the preset length threshold, the loop is stopped, and a target text summary is extracted for the current text summary based on the target summary extraction model.

[0056] Optionally, the loop unit is used to:

[0057] Split the current text into a current sub-text sequence; the current sub-text sequence includes different current sub-texts arranged in the order of the texts in the current text;

[0058] In the current subtext sequence, between any pair of adjacent current subtexts, the last N consecutive characters in the previous current subtext are the same as the first N consecutive characters in the next current subtext.

[0059] Optionally, the current text contains several paragraphs; the loop unit is used to:

[0060] Split the current text into several preliminary sub-texts; any preliminary sub-text contains at least one paragraph;

[0061] When the length of any prepared subtext is less than or equal to the preset length threshold, determining the prepared subtext as the current subtext;

[0062] When the length of any prepared subtext is greater than the preset length threshold, the prepared subtext is split into at least two current subtexts; and the length of any current subtext is less than or equal to the preset length threshold.

[0063] Optionally, the recurrence unit is used to: split the current text into a current sub-text sequence; the current sub-text sequence includes different current sub-texts arranged in the order of texts in the current text;

[0064] In the current subtext sequence, between any pair of adjacent current subtexts, the latter current subtext includes the subtext summary extracted from the previous current subtext based on the target summary extraction model.

[0065] Optionally, the determining unit is used to: determine a number of collected information; any collected information has a corresponding collection time point;

[0066] Convert the determined collected information into text;

[0067] The converted texts are synthesized to determine the target text; in the target text, the sequence of different texts converted from different collected information is arranged according to the sequence of the collected information corresponding to the collection time points.

[0068] Optionally, the determining unit is further used to: determine user information that provides any collected information;

[0069] Determine the unit used for:

[0070] The determined collected information and user information providing the determined collected information are converted into text.

[0071] Optionally, the preset length threshold comprises: a length obtained by subtracting a length of a preset prompt word from an upper limit of an input text length of the target summary extraction model;

[0072] The preset prompt word is used to prompt the target summary extraction model to extract a summary for the input text;

[0073] The loop unit is used to: input the preset prompt word and the current text summary into the target summary extraction model to obtain the target text summary output by the target summary extraction model.

[0074] Optionally, the structure of the target summary extraction model includes any one of the following: a language model, a natural language processing model, a generative summary model, and an extractive summary model.

[0075] According to the above embodiments, by splitting the text once or multiple times to extract text summaries respectively, the length of the text can be reduced until a text summary can be extracted for the entire text after the length is reduced, which can reduce the influence of the length of the input text on the text summary extraction.

[0076] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0078] Figure 1 is a flowchart of a text summary extraction method according to an embodiment of the present invention;

[0079] Figure 2 is a flow chart of another text summary extraction method according to an embodiment of the present invention;

[0080] Figure 3is a flow chart of another text summary extraction method according to an embodiment of the present invention;

[0081] Figure 4 is a flow chart of another text summary extraction method according to an embodiment of the present invention;

[0082] Figure 5 is a flow chart of a method for extracting meeting minutes according to an embodiment of the present invention;

[0083] Figure 6 is a flow chart of another method for extracting meeting minutes according to an embodiment of the present invention;

[0084] Figure 7 is a framework flow chart of generating meeting minutes according to an embodiment of the present invention;

[0085] Figure 8 is a schematic diagram showing a principle of a conference information extraction process according to an embodiment of the present invention;

[0086] Fig. 9 is a framework flow chart of text splitting according to an embodiment of the present invention;

[0087] Fig.10 is a schematic diagram showing a principle of a meeting minutes generation process according to an embodiment of the present invention;

[0088] Fig.11 is a structural schematic diagram of a text summary extraction device according to an embodiment of the present invention;

[0089] Fig.12 is a schematic structural diagram of another text summary extraction device according to an embodiment of the present invention;

[0090] Fig.13 is a structural schematic diagram of a device for extracting meeting minutes according to an embodiment of the present invention;

[0091] Fig.14 The figure is a schematic diagram of the hardware structure of a computer device configured with the method of the embodiment of the present invention according to the embodiment of the present invention. DETAILED DESCRIPTION

[0092] Exemplary embodiments will be described in detail herein, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, the same numbers in different drawings represent the same or similar elements unless otherwise indicated. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Instead, they are merely examples of devices and methods consistent with some aspects of the present invention as detailed in the appended claims.

[0093] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in the embodiments of the present invention are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data need to comply with the relevant laws, regulations and standards of relevant countries and regions, and provide corresponding operation entrances for users to choose to authorize or refuse.

[0094] With the development of computer technology, it is often necessary to process information in text form. Among them, there are many needs for text summarization, for example, extracting meeting minutes from text generated in a meeting, extracting the main points of an article, and so on.

[0095] However, many current text summarization methods have an upper limit on the length of the input text. For example, 30 words, 50 words, 100 words, etc. When faced with long texts, it is often difficult to extract text summaries. For example, for a meeting that lasts for 5 hours, the length of the text generated is often tens of thousands of words, and current text summarization methods often have difficulty extracting meeting minutes.

[0096] Therefore, the length of the input text has a greater impact on text summary extraction.

[0097] The embodiment of the invention discloses a text summary extraction method.

[0098] In this method, for a relatively long input text, the input text can be divided into several sub-texts, so that sub-texts of relatively short lengths can be obtained, which facilitates the extraction of text summaries for each sub-text.

[0099] Afterwards, the text summaries of the various sub-texts may be synthesized, and for the synthesized text, it may be further determined whether the text summary can be directly extracted.

[0100] If the synthesized text still cannot directly extract the text summary, the above steps can be repeated to further divide the synthesized text into several sub-texts, extract the text summaries of each sub-text, and synthesize the text summaries of each sub-text until a text summary can be directly extracted from the synthesized text in a certain cycle.

[0101] Finally, the final text summary can be extracted from the current synthesized text after the loop stops.

[0102] For ease of understanding, in a specific example, the upper limit of the input length of the text summary extraction model may be 500 words, while the input text length required to extract the text summary is currently 50,000 words.

[0103] Therefore, we can first split the 50,000-word input text into several sub-texts of no more than 500 words in length, use the text summary extraction model to extract text summaries from the sub-texts respectively, and then combine the text summaries of each sub-text to determine that the total length is 5,000 words.

[0104] After that, the comprehensive text of 5,000 words can be further split into several sub-texts of no more than 500 words in length. The text summary extraction model is used to extract text summaries from the sub-texts respectively. Then, the text summaries of each sub-text are combined to determine that the total length is 400 words.

[0105] Finally, we can use the text summary extraction model to extract the final text summary for the 400-word comprehensive text.

[0106] In this method, by extracting the text summary after cyclically splitting the input text, it is convenient to extract the text summary for the input text with a longer length, and the influence of the length of the input text on the text summary extraction can be reduced.

[0107] The following is a detailed explanation of a text summary extraction method provided by an embodiment of the present invention.

[0108] like Figure 1 As shown, Figure 1 The figure is a flowchart of a text summary extraction method according to an embodiment of the present invention.

[0109] The embodiment of the present invention does not limit the execution subject of the method process. Optionally, the execution subject can be any computing device or any application. For example, a terminal, a server, an application with a text summary generation function, etc.

[0110] The method may include the following steps.

[0111] S101: Determine a target text and target summary extraction model.

[0112] S102: When the length of the target text is greater than a preset length threshold, the target text is determined as the current text.

[0113] The preset length threshold may be less than or equal to the upper limit of the input text length of the target summary extraction model.

[0114] S103: Split the current text into several current sub-texts, and extract sub-text summaries for different current sub-texts based on the target summary extraction model; the length of any current sub-text is less than or equal to a preset length threshold; and the current text summary is determined by combining the extracted sub-text summaries.

[0115] S104: Determine whether the length of the current text summary is greater than a preset length threshold.

[0116] When the length of the current text summary is greater than the preset length threshold, S105 is executed; when the length of the current text summary is not greater than the preset length threshold, S106 is executed.

[0117] S105: Determine the current text summary as the new current text, and execute S103.

[0118] S106: Extracting a target text summary for the current text summary based on the target summary extraction model.

[0119] Among them, when the length of the target text is not greater than the preset length threshold, that is, when the length of the target text is less than or equal to the preset length threshold, the target text summary can be directly extracted for the target text based on the target summary extraction model.

[0120] For ease of understanding, Figure 2 As shown, Figure 2 FIG. 4 is a flow chart of another text summary extraction method according to an embodiment of the present invention.

[0121] The method may include the following steps.

[0122] S201: Determine a target text and target summary extraction model.

[0123] S202: When the length of the target text is greater than a preset length threshold, and the preset length threshold is less than or equal to the upper limit of the input text length of the target summary extraction model, the target text is determined as the current text, and the following steps are executed in a loop: the current text is split into a number of current sub-texts, and sub-text summaries are extracted for different current sub-texts based on the target summary extraction model; the length of any current sub-text is less than or equal to the preset length threshold; the current text summary is determined by combining the extracted sub-text summaries; when the length of the current text summary is greater than the preset length threshold, the current text summary is determined as the new current text; when the length of the current text summary is less than or equal to the preset length threshold, the loop is stopped, and based on the target summary extraction model, a target text summary is extracted for the current text summary.

[0124] In the above method flow, the text summary can be extracted separately by splitting the text once or multiple times, thereby reducing the length of the text until the text summary can be extracted for the entire text after the length is reduced, which can reduce the impact of the length of the input text on the text summary extraction.

[0125] Furthermore, by reducing the text length by extracting the text summary of the subtext, the text length can be reduced while retaining the information in the target text as much as possible, thereby improving the information integrity of the target text in the target text summary and reducing the information loss rate of the target text in the target text summary.

[0126] The method flow can make the target text adapt to the target summary extraction model, so as to realize the extraction of text summary for the target text based on the target summary extraction model.

[0127] In addition, the above method flow and related explanations are mainly explained for the summary extraction of a single text (ie, the target text).

[0128] It is understandable that for multiple texts, the above method flow can be used to extract summaries respectively, or multiple texts can be combined into one text as the target text and the above method flow can be used to extract summaries. For specific explanations, please refer to the above method flow and related explanations.

[0129] The following is a detailed explanation of each aspect.

[0130] 1. About the target text.

[0131] The process of this method is not limited to the specific content of the target text.

[0132] Optionally, the target text may be any text. For the convenience of description, any text from which a text summary needs to be extracted is referred to as the target text.

[0133] This method does not limit the source of the target text.

[0134] Optionally, the target text may be directly acquired, specifically acquired locally, or received from an external device, or extracted from other forms of media, such as text extracted from voice, image, or video as the target text.

[0135] In a specific example, text may be extracted from information recorded in a meeting as the target text.

[0136] Specifically, it may include text records of the meeting, such as text records of the meeting by the participants, or text content presented at the meeting; it may also include images of the meeting, such as image content presented at the meeting; it may also include audio of the meeting, such as recordings of the speeches of the participants in the meeting; it may also include video of the meeting, such as video recordings of the meeting, or video recordings of the presented content, etc.

[0137] Texts can be extracted from the information recorded in these meetings, specifically by directly obtaining the records in text form, or by extracting text from images, language or video and synthesizing them into target texts.

[0138] Regarding the determination of the target text in S101 and S201, the present method flow does not limit the specific method of determining the target text.

[0139] Specifically, the text to be extracted may be directly obtained and determined as the target text. For example, an article may be determined as the target text, or the text content in a book may be determined as the target text.

[0140] The target text may be determined locally, or text may be acquired from an external device and further determined as the target text.

[0141] Alternatively, text may be extracted from information collected in other media forms and determined as the target text. For example, through voice-to-text conversion, conference recordings or classroom recordings may be converted into text form as the target text to facilitate extracting summaries.

[0142] For the convenience of description, the information collected in the preset scene is referred to as collected information. The preset scene is not limited, for example, a meeting, a classroom, an office, etc. The form of the collected information is not limited, for example, it can be information in the form of video, audio, text, document, and image.

[0143] Therefore, optionally, determining the target text may specifically include: determining the collected information; converting the determined collected information into text; and determining the target text by integrating the converted text.

[0144] 1. About collecting information.

[0145] The method flow does not limit the form of the collected information. Specifically, it can be text, image, voice, video, document, etc. Specifically, the collected information in different forms can be converted into text through corresponding text conversion methods.

[0146] For example, through a multimodal large language model, different forms of collected information can be converted into text form; optical character recognition technology can be used to convert images into text; speech recognition technology can be used to convert speech into text, or text content in a document can be extracted as the converted text.

[0147] Of course, for the collected information in text form, the collected information itself can be directly determined as the converted text. For example, in a meeting, the text content of the meeting written down by the participants can be collected, and the content presented in the meeting in text form can also be collected, so that the collected information in text form can be directly determined as the converted text.

[0148] In addition, the collected information may also be in other forms, such as handwriting information, operation information, or trajectory information, etc. In a specific example, handwriting information may be collected for writing on a blackboard on an electronic screen in a conference scene, and may be further converted into text.

[0149] The collected information may also be information that references other content, such as links, hyperlinks, web page URLs, etc., so that the content referenced by the collected information can be obtained and converted into text. For example, the web page content can be obtained based on the collected web page URLs and converted into text.

[0150] The process of this method does not limit the content of the collected information. Specifically, it can be information collected in various forms from a meeting, or information collected from a classroom, etc.

[0151] The method flow does not limit the number of collected information. Specifically, it can be to determine a number of collected information, or to determine one or more collected information, and then each determined collected information can be converted into text, and the converted text can be further integrated to determine the target text.

[0152] The process of this method does not limit the source and determination method of the collected information. The collected information can be directly obtained locally, or obtained from an external device, or the collected information can be determined based on various collected information.

[0153] For example, from a meeting, you can collect audio, screen recording, video, text, PPT and other information, so that you can determine some collected information.

[0154] In an optional embodiment, in addition to the collected information itself being converted into text, other information related to the collected information can also be converted into text. For example, the attributes of the collected information, the time point of the collected information, the form or format of the collected information, etc. can all be converted into text and added to the target text. The information of the user who provides the collected information can also be converted into text so as to be added to the target text later.

[0155] The user who provides the collected information here, for example, in a meeting, there may be multiple participants speaking, and when collecting the voice information of these speeches, the participant who spoke is also the user who provided the voice information, so that the information of the speaker can be determined, which is convenient for subsequent conversion into text.

[0156] Participants may also display images. When collecting images, the participants who display them are the users who provide the images. Participants may also display documents such as PPT. When collecting documents such as PPT, the participants who display them are the users who provide the documents such as PPT.

[0157] Of course, the above-mentioned users may also be users related to the collected information, or users corresponding to the collected information.

[0158] It is understandable that the information of the user who displays the collected information may be converted into text, the information of the user who generates the collected information may be converted into text, or the information of the user related to the collected information may be converted into text.

[0159] Accordingly, the user information may specifically be a user identification, such as a user name, which can be easily converted into text.

[0160] Therefore, optionally, the step of determining the target text may further include: determining the user information providing any collected information. Converting the determined collected information into text may specifically include: converting the determined collected information and the user information providing the determined collected information into text.

[0161] Specifically, the determined collected information and user information related to the determined collected information may be converted into text.

[0162] This embodiment does not limit the form and content of the user information. The user information may specifically be a user identifier, or a user name, etc. Optionally, the user information may include: user name, user identity, user level, etc.

[0163] This embodiment does not limit the method for determining the user information.

[0164] Optionally, when acquiring the collected information, the user information providing the collected information may also be acquired; or the user information providing the collected information may be acquired from the collected information.

[0165] For example, for information collected in the form of voice, the speaker information of the collected information, that is, the user information providing the collected information, can be determined through voiceprint matching.

[0166] For another example, with respect to information collected in the form of blackboard writing images, information about the writer of the collected information, that is, information about the user who provided the collected information, can be determined through handwriting matching.

[0167] For another example, in an online meeting, each user can display content in a window for communication, thereby facilitating the acquisition of collected information and displaying user information of the collected information.

[0168] Optionally, the corresponding user information may be determined for each piece of collected information, or the user information for providing the targeted collected information may be determined for one or more pieces of collected information.

[0169] This embodiment does not limit the method of converting user information into text. Usually, the user information itself is in text form, and the user information can be directly determined as the converted text. The user information can also be converted into text according to a preset rule. For example, when the user information is a string of coded symbols, the corresponding text can be determined according to the preset rule, and specifically, different user information can be distinguished and different texts can be determined respectively, for example, user 1-N, user A, user B, etc.

[0170] In addition, optionally, the texts converted from the collected information and the user information providing the collected information may be concatenated and combined to facilitate determination of the user providing the collected information.

[0171] This embodiment does not limit the specific splicing combination method. Optionally, the texts converted from the two can be directly spliced, or the texts converted from the two can be adjusted before splicing.

[0172] Optionally, splicing can be performed in combination with the form of the collected information. Specifically, the media form of the collected information can be combined and the media form can also be introduced into the text for splicing. It can be understood that the determined collected information, the determined form of the collected information, and the user information providing the determined collected information can be converted into text. The texts converted from these three can also be spliced ​​and combined.

[0173] In a specific example, the text after specific splicing and combination may include the following examples.

[0174] "User A provides a text: xxxx", "User A displays a text: xxxx", "User A provides an image: xxxx", "User A displays an image: xxxx", "User A provides a video: xxxx", "User A displays a video: xxxx", "User A displays in the video: xxxx", "User A provides a voice: xxxx", "User A says xxxx".

[0175] It is understandable that in the target text, especially in the conference scenario, by clarifying the participants and the text corresponding to the collected information provided, it is convenient to view the user corresponding to the text, and it is also convenient to improve the integrity of the conference information and reduce the information loss rate.

[0176] In an example of target text, the following text may be included: "In a meeting, user A said: xxxx. User B said: xxxxx. User C showed an image containing xxxx information."

[0177] 2. Determine the target text based on the converted text.

[0178] The method flow does not limit the specific way of determining the target text by synthesizing the converted text.

[0179] Optionally, the converted texts may be concatenated to obtain a target text, or the converted texts may be integrated into an overall text, and then the target text may be obtained based on the overall text.

[0180] Specifically, the converted texts can be directly spliced ​​into a whole text and the whole text can be determined as the target text, or all the converted texts can be screened and the screened texts can be spliced ​​into a whole text, and the whole text can be segmented or regularized to obtain the target text, etc.

[0181] This embodiment does not limit the specific method of synthesizing the entire text.

[0182] Specifically, all the converted texts may be directly spliced ​​into a whole text, or some texts may be selected from the whole converted texts and spliced ​​into a whole text, or the converted texts may be spliced ​​in a preset order to obtain a whole text, and so on.

[0183] Optionally, the whole text can be obtained by directly concatenating the converted texts to obtain a whole text. For example, the converted texts "ABCD" and "EFG" can be directly concatenated into a whole text "ABCDEFG".

[0184] Of course, when splicing the converted texts, it can be splicing multiple converted texts, or it can be splicing each converted text.

[0185] Optionally, in the case of filtering the converted part of the text, the present embodiment does not limit the filtering method, and may specifically filter and delete shorter texts, or filter and delete texts that failed to be converted, and so on.

[0186] It should be noted that different collected information can be converted into different texts respectively, and when these different texts are specifically integrated, the converted texts can be spliced ​​together in a preset order to obtain an overall text.

[0187] Therefore, optionally, the overall text can be obtained by concatenating the converted texts in a preset order to obtain an overall text. This embodiment does not limit the specific preset order, and can be arranged in a sequence according to the collection time points of the collection information, or in a pre-specified order, or in a sequence according to the timestamps of the collection information, or in a sequence according to the generation time points of the collection information.

[0188] The collection time point of the collected information may specifically be the time point at which the collected information is collected. For example, when collecting a conference recording, the current time point may be determined as the collection time point of the conference recording. When collecting a classroom blackboard image, the current time point may be determined as the collection time point of the blackboard image.

[0189] In a specific example, the speeches of the participants in the meeting can be recorded in real time during the meeting, so that when collecting the real-time recording of the meeting, the current time point can be determined as the collection time point of the meeting recording.

[0190] The collected information itself may also have a timestamp. For example, a conference screen recording often carries a timestamp of the conference process, which may be 4 to 5 o'clock, so that the timestamp of the collected information can be directly obtained.

[0191] The generation time point of the collected information may specifically be the generation time point of the collected information. For example, when generating a conference recording, the current time point may be determined as the generation time point of the conference recording. When generating a conference record text, specifically, when a participant types and records the conference content during the conference, the current time point may be determined as the generation time point of the conference recording.

[0192] By combining the converted texts in chronological order, the accuracy of the target text in chronological order can be improved, and the accuracy of the extracted text summary can also be conveniently improved.

[0193] Therefore, optionally, determining the target text may specifically include: determining a number of collected information; any collected information has a corresponding collection time point; converting the determined collected information into text; determining the target text by integrating the converted text; in the target text, the order of different texts converted from different collected information is arranged according to the order of the collection time points corresponding to the collected information.

[0194] The determined one or more collected information may be converted into text, or all or part of the determined collected information may be converted into text.

[0195] Furthermore, the target text may be determined by integrating one or more converted texts, or may be determined by integrating all or part of the converted texts.

[0196] This embodiment can improve the accuracy of the time sequence of the target text and also conveniently improve the accuracy of the extracted text summary by splicing and combining the converted text according to the sequence of the collection information corresponding to the collection time points.

[0197] In addition, optionally, determining the target text may also be: determining a number of collected information; any collected information has a corresponding timestamp; converting the determined collected information into text; determining the target text by integrating the converted text; in the target text, the order of different texts converted from different collected information is arranged according to the order of the corresponding timestamps of the collected information.

[0198] Optionally, determining the target text may also be: determining a number of collected information; any collected information has a corresponding generation time point; converting the determined collected information into text; determining the target text by integrating the converted text; in the target text, the order of different texts converted from different collected information is arranged according to the order of the corresponding generation time points of the collected information.

[0199] Of course, the converted texts can also be randomly spliced ​​together to obtain a whole text.

[0200] After obtaining the overall text, the target text can be obtained based on the overall text.

[0201] The process of this method does not limit the specific way of obtaining the target text based on the overall text.

[0202] Optionally, the entire text may be directly determined as the target text, or the entire text may be segmented or text-formatted to obtain the target text, or the entire text may be polished or simplified to obtain the target text, and so on.

[0203] From the perspective of the target text, the target text can be obtained by segmentation, so that it can contain multiple paragraphs; the target text can also be obtained after text regularization, so as to reduce redundant parts or repeated content, etc.

[0204] Regarding segmentation, in an optional embodiment, the determined target text may contain divided paragraphs, so that it can be convenient to further split it into sub-texts later. Accordingly, in the process of determining the target text, a segmentation operation can be performed.

[0205] This embodiment does not limit the timing of performing the segmentation operation. Specifically, the segmentation operation may be performed after the text is obtained by conversion, or in the process of converting the collected information into text, or in the process of obtaining the above-mentioned overall text, or after obtaining the overall text, the segmentation operation may be performed on the overall text.

[0206] It should be noted that in the process of converting the collected information into text, the converted text may contain multiple paragraphs. For example, for collected information in the form of speech, text can be obtained through speech recognition conversion, and in the converted text, a long time interval corresponding to the speech can be recognized as a segmentation mark, thereby achieving segmentation and obtaining multiple paragraphs.

[0207] Furthermore, when synthesizing the converted texts, or specifically splicing the converted different texts, a segmentation operation may be further performed.

[0208] The present embodiment does not limit the specific operation mode of the segmentation operation. Optionally, the segmentation operation can be performed using a large language model, or the segments can be divided according to the length of characters, or the text converted from different collected information can be divided into different paragraphs.

[0209] Furthermore, the specific segmentation operation may be to first perform segmentation operations on each converted text separately, and then obtain the overall text by splicing; or to first splice the converted different texts, and then perform segmentation operations on the splicing result to obtain the overall text.

[0210] Regarding text regularization, in an optional embodiment, the determined target text may be obtained through text regularization. This embodiment does not limit the specific content of text regularization, and specific text regularization may include deleting redundant content, deleting repeated content, deleting colloquial content, text polishing, etc.

[0211] This embodiment does not limit the timing of performing the text regularization operation. Specifically, the text regularization operation may be performed after the text is converted, or in the process of converting the collected information into text, or in the process of obtaining the above-mentioned overall text, or after the overall text is obtained, the text regularization operation may be performed on the overall text.

[0212] This embodiment does not limit the specific implementation method of the text regularization operation. Optionally, the text regularization operation can be performed through a large language model. For example, through the prompt words "Please delete the redundant content in the following text: XXX" or "Please delete the repeated content and colloquial expressions in the following text: XXX", the large language model can perform text regularization operations on the input text. It is also possible to delete the colloquial content by deleting keywords, for example, deleting colloquial keywords such as "eh, ah, oh, huh". It is also possible to delete texts with high repetition or high similarity by comparing different texts.

[0213] In this embodiment, through text regularization, the target text obtained can be more written, concise and regular.

[0214] In the process of determining the target text by synthesizing the converted texts, it is necessary to at least splice the texts together to determine a target text.

[0215] In this process, at least one of the following operations may be included: filtering out part of the text from the converted text, determining the order of text concatenation, segmentation operation, text regularization operation, etc.

[0216] For ease of understanding, several examples of the process of determining the target text are given below for illustrative purposes.

[0217] 1) Performing text regularization operations on all converted texts respectively, determining the splicing order of the texts, and splicing all the texts after text regularization according to the determined order to obtain the target text.

[0218] 2) All converted texts are concatenated, and segmentation and text regularization operations are performed on the concatenated results to obtain the target text.

[0219] 3) Filter out some texts from all the converted texts, determine the splicing order between the filtered texts, and splice the filtered texts according to the determined order to obtain the target text.

[0220] 2. About the summary extraction model.

[0221] The process of this method does not limit the specific form of the target summary extraction model, as long as the target summary extraction model can extract a text summary for the input text.

[0222] The process of this method does not limit the specific structure of the target summary extraction model.

[0223] Optionally, the structure of the target summary extraction model may include any one of the following: a language model, a natural language processing model, a generative summary model, and an extractive summary model.

[0224] In a specific example, the target summary extraction model may be a large language model, so that a summary can be extracted from the input text through prompt words.

[0225] Of course, the target summary extraction model can also be a language model used to directly extract text summaries from input text without the need for prompt words.

[0226] Regarding determining the target summary extraction model in S101 and S201, the present method flow does not limit the specific determination method.

[0227] Optionally, a target summary extraction model may be determined locally, a target summary extraction model may be obtained from an external device, or a target summary extraction model may be determined in an external device, and then the external device may communicate with the external device and perform summary extraction based on the target summary extraction model.

[0228] The process of this method does not limit the specific execution process of the target summary extraction model.

[0229] Optionally, the text may be input into the target summary extraction model, and the target summary extraction model may output a text summary of the input text. The target summary extraction model may have an upper limit on the length of the input text.

[0230] Of course, the input of the target summary extraction model may also include other texts besides the text to be extracted from the summary, such as prompt words.

[0231] In an optional embodiment, the prompt word and the target text may be concatenated into an input text, which is then input into the target summary extraction model. The prompt word may be, for example, "Please extract the summary of the following text" or "Extract summary".

[0232] In an optional embodiment, for the target summary extraction model, inference acceleration and other methods can also be used to improve the efficiency and speed of the target summary extraction model in extracting text summaries, thereby improving the efficiency and speed of the overall text summary extraction.

[0233] This embodiment does not limit the method of reasoning acceleration, and specifically may adopt parallel reasoning, or increase computing power, etc.

[0234] 3. Regarding the preset length threshold.

[0235] The method flow does not limit the specific length of the preset length threshold, as long as the preset length threshold is less than or equal to the upper limit of the input text length of the target summary extraction model.

[0236] By setting a preset length threshold, the length of the target text can be reduced through one or more summary extractions, until the reduced text can be directly input into the target summary extraction model to extract the final target text summary.

[0237] When the length of the current text summary is less than or equal to the preset length threshold, since the preset length threshold itself is less than or equal to the upper limit of the input text length of the target summary extraction model, the length of the current text summary is also less than or equal to the upper limit of the input text length of the target summary extraction model. Therefore, the current text summary can be directly input into the target summary extraction model to extract the final target text summary.

[0238] In this embodiment, by setting the preset length threshold to be less than or equal to the upper limit of the input text length of the target summary extraction model, the flexibility of setting the preset length threshold can be improved.

[0239] The method flow does not limit the method for determining the preset length threshold. Specifically, the upper limit of the input text length of the target summary extraction model can be directly determined as the preset length threshold; or the preset length threshold can be determined by combining the upper limit of the input text length of the target summary extraction model with other information. For example, the length of the prompt word that needs to be input into the target summary extraction model can be considered to determine the preset length threshold.

[0240] Optionally, the preset length threshold may specifically be: a length obtained by subtracting a length of a preset prompt word from an upper limit of an input text length of a target summary extraction model.

[0241] Of course, the preset length threshold may be smaller than the length obtained by subtracting the length of the preset prompt word from the upper limit of the input text length of the target summary extraction model.

[0242] This embodiment does not limit the content of the preset prompt words, wherein the preset prompt words can be used to prompt the target summary extraction model to extract a summary for the input text.

[0243] For ease of understanding, in a specific example, the preset prompt words may be “please extract a summary”, “extract a summary of the following text” or “please extract a summary of the following text reasonably and with as little missing information as possible”.

[0244] Accordingly, based on the target summary extraction model, a target text summary is extracted for the current text summary, which may be specifically: the preset prompt word and the current text summary are input into the target summary extraction model to obtain the target text summary output by the target summary extraction model.

[0245] More specifically, the preset prompt word may be concatenated with the current text summary and input into the target summary extraction model.

[0246] In addition, for the current subtext in S103 and S202, the preset prompt word and the current subtext may be input into the target summary extraction model. Specifically, the preset prompt word and the current subtext may be concatenated and input into the target summary extraction model.

[0247] In addition to the above prompt words, other information can also be introduced. The process of this method is not limited to specific other information, such as the scene description of the target text, or the need to extract a summary.

[0248] For ease of understanding, several specific examples are given below.

[0249] In a meeting scenario, other information may be "In a meeting, the following text was collected", which can be further combined with the input text and input into the target summary extraction model, thereby facilitating the target summary extraction model to determine the scenario of the target text.

[0250] In a conference scenario, other information may be "the length of the final target text summary needs to be less than 500 words." Therefore, the prompt word "the extracted summary has less than 500 words" may be added in combination with this information.

[0251] In the book scenario, you can extract a summary of the text content in a book. Other information can be "extract the summary related to xx in the book", so that you can focus on extracting summaries from a certain aspect. Specifically, you can add the prompt word "extract the summary related to xx".

[0252] Therefore, optionally, the preset prompt word, other information and the current subtext may be concatenated and input into the target summary extraction model, wherein the length of the concatenated text should not be greater than the upper limit of the input text length of the template summary extraction model.

[0253] 4. About text segmentation.

[0254] In the above-mentioned loop process of S103-S105 and the loop process of S202, the current text in any loop can be split.

[0255] The process of the method is not limited to a specific text splitting method, as long as the length of any current sub-text obtained after splitting is less than or equal to a preset length threshold.

[0256] The length of each current subtext obtained after splitting may not be greater than a preset length threshold.

[0257] Optionally, the current text can be split directly according to a preset length threshold, into several current sub-texts with a length of the preset length threshold; it can also be split according to the paragraphs in the current text, and the splitting is performed in units of paragraphs. Accordingly, in the segmentation operation, the length of each divided paragraph can be made not greater than the preset length threshold, so as to facilitate the splitting in units of paragraphs; the splitting can also be performed by comprehensively considering other factors, for example, the integrity of the current sub-text, the relationship between different current sub-texts, the importance of the current sub-text, etc.

[0258] Several embodiments of text splitting are explained in further detail below. It can be understood that different embodiments can be combined with each other.

[0259] Embodiment 1.

[0260] Optionally, considering that after being split into multiple current subtexts, different current subtexts may be associated with each other, so as to improve the continuity of the current subtexts and facilitate more accurate extraction of text summaries.

[0261] Therefore, splitting the current text into a number of current subtexts may specifically include: splitting the current text into a current subtext sequence; the current subtext sequence may include different current subtexts arranged in the order of texts in the current text.

[0262] It is understandable that the text can be split and sorted in order to obtain several current subtexts to form a current subtext sequence. The order of characters in any current subtext can be the same as the order of corresponding characters in the current text. For example, the current text includes "ABCXXX". After this part of characters is split into a current subtext, the current subtext can be specifically "ABCXXX", and the order of characters therein remains unchanged.

[0263] In the current subtext sequence, between any pair of adjacent current subtexts, the last N consecutive characters in the previous current subtext may be the same as the first N consecutive characters in the next current subtext.

[0264] This embodiment does not limit the specific value of N. N may be 1, 2, or 10. Of course, N may be smaller than the preset length threshold.

[0265] It can be understood that, in the current subtext sequence, between each pair of adjacent current subtexts, the last N consecutive characters in the previous current subtext may be the same as the first N consecutive characters in the next current subtext.

[0266] For ease of understanding, in a specific example, a current text may include "...ABCXXX, DEFXXX. OPQXXX, XYZXXX...".

[0267] After the current text is split into a current subtext sequence, for two adjacent current subtexts, the former current subtext may be "ABCXXX, DEFXXX.", and the latter current subtext may be "DEFXXX.OPQXXX, XYZXXX". There are repeated characters "DEFXXX" between the two current subtexts.

[0268] By having a later current subtext repeat the last multiple consecutive characters of a previous current subtext, the relevance and continuity between two adjacent current subtexts can be improved, thereby improving the accuracy of text summary extraction.

[0269] This embodiment does not limit the specific splitting method, as long as the current sub-text sequence obtained by splitting meets the above characteristics. Specifically, it can be split one by one from the current text in order, first splitting the first current sub-text from the 1st character to the Mth character (M is less than the preset length threshold and M is greater than N) of the current text, and further splitting from the (MN)th character to the (2M-N)th character to get the second current sub-text, and so on. The split current sub-texts can be arranged according to the order of the corresponding texts in the current text to obtain the current sub-text sequence.

[0270] Of course, it is also possible to directly split out multiple non-repeating sub-texts first, and then further copy the last several characters of the previous sub-text to the front of the next sub-text for any two adjacent sub-texts according to the order of the corresponding texts in the current text.

[0271] Embodiment 2.

[0272] The current text may contain several paragraphs, which may be obtained by performing segmentation when determining the target text or determining the current text summary in the previous cycle.

[0273] The paragraphs in the current text can be split. Since paragraphs are usually used to express a complete semantics, the paragraphs can be split as division units to improve the integrity of the sub-text.

[0274] Therefore, optionally, the current text may contain several paragraphs; the current text may be split into several current sub-texts, which may be specifically: the current text may be split into several prepared sub-texts; any prepared sub-text may contain at least one paragraph; when the length of any prepared sub-text is less than or equal to a preset length threshold, the prepared sub-text may be determined as the current sub-text; when the length of any prepared sub-text is greater than a preset length threshold, the prepared sub-text may be split into at least 2 current sub-texts; the length of any current sub-text may be less than or equal to the preset length threshold.

[0275] The current text may be split into several preliminary sub-texts according to the paragraphs in the current text. Any preliminary sub-text may contain at least one complete paragraph, and specifically, each preliminary sub-text may contain at least one complete paragraph. Then, in combination with the length limit of the current sub-text, the longer preliminary sub-text may be further split.

[0276] Of course, the length of the paragraphs in the current text may also be limited. Specifically, the length of any paragraph in the current text may not be greater than a preset length threshold, or the length of each paragraph in the current text may be less than or equal to the preset length threshold.

[0277] This embodiment does not limit the specific splitting method to obtain the preliminary subtext. Optionally, the splitting may be started from the beginning of any paragraph, and at least one subsequent paragraph may be split out as the preliminary subtext. Alternatively, the splitting may be performed directly based on the paragraph as the division unit.

[0278] This embodiment does not limit the specific method of splitting the preliminary subtext to obtain the current subtext. Among them, for the preliminary subtext whose length is greater than the preset length threshold, when splitting to obtain the current subtext, it is first necessary to consider that the length of the current subtext should not be greater than the preset length threshold, and further consider retaining the complete paragraph as much as possible.

[0279] For example, for a prepared sub-text containing only a single paragraph, the paragraph may be ignored and the sub-text may be directly divided into at least two current sub-texts whose lengths are not greater than a preset length threshold.

[0280] For a prepared sub-text containing multiple paragraphs whose lengths are not greater than a preset length threshold, the paragraphs may be divided comprehensively to divide each paragraph into one current sub-text.

[0281] In this embodiment, in addition to paragraph division, a current sub-text sequence can also be obtained by combination to further enhance the association between adjacent current sub-texts.

[0282] Optionally, in the process of splitting to obtain several current sub-texts, as the paragraph division is completed, the preliminary sub-texts can be further sorted and processed according to the order of the paragraphs contained in the preliminary sub-texts in the current text to obtain a current sub-text sequence, so that in the current sub-text sequence, between any pair of adjacent current sub-texts, the last N consecutive characters in the previous current sub-text can be the same as the first N consecutive characters in the next current sub-text.

[0283] Of course, when the length of any prepared sub-text is greater than a preset length threshold, the prepared sub-text may be split into at least two current sub-texts; specifically, the prepared sub-text may be split into a current sub-text sequence, so that in the current sub-text sequence, between any pair of adjacent current sub-texts, the last N consecutive characters in the previous current sub-text may be the same as the first N consecutive characters in the next current sub-text.

[0284] For detailed explanation, please refer to the above embodiments.

[0285] Embodiment three.

[0286] In an optional embodiment, the method process needs to extract summaries for sub-texts, and the association between sub-texts is taken into consideration. Moreover, the length of the summaries extracted from the sub-texts can usually be less than a preset length threshold, so that the summary of the previous sub-text can be directly retained in the next sub-text as the beginning, so as to improve the association between the sub-texts and improve the accuracy of summary extraction.

[0287] Therefore, optionally, the current text is split into several current sub-texts, which can be specifically: splitting the current text into a current sub-text sequence; the current sub-text sequence includes different current sub-texts arranged in the order of texts in the current text; in the current sub-text sequence, between any pair of adjacent current sub-texts, the latter current sub-text includes the sub-text summary extracted for the previous current sub-text based on the target summary extraction model.

[0288] Among them, by adding the text summary of the previous current subtext to the next current subtext, the overall information summary of the previous current subtext can be added to the next current subtext, which improves the relevance between the current subtexts and also improves the accuracy of summary extraction.

[0289] This embodiment does not limit the specific splitting method.

[0290] Optionally, the current subtext and the summary of the current subtext may be determined one by one in a loop, and the next current subtext may be determined based on the summary of the current subtext, and so on. In addition, the text may be first split into several subtexts, and the summaries are extracted for each of them, and then several current subtexts are obtained by combining them, and the summaries are further extracted for different current subtexts.

[0291] In an optional embodiment, the current text may be split into a candidate subtext sequence; the candidate subtext sequence includes different candidate subtexts arranged in the order of the texts in the current text. After extracting candidate subtext summaries for the candidate subtexts respectively, the previous candidate subtext summary may be added to the front of the next candidate subtext according to the order in the candidate subtext sequence, or used as the front text of the next candidate subtext, thereby obtaining the current subtext sequence.

[0292] In another optional embodiment, the current text is split into several current sub-texts, and sub-text summaries are extracted for different current sub-texts based on the target summary extraction model, which may be:

[0293] Starting from the first character of the current text, according to the character order in the current text, split the first current subtext whose length is not greater than the preset length threshold, and determine the text summary of the first current subtext based on the target summary extraction model. The text summary of the first current subtext is determined as the current summary to be added.

[0294] The next character of the last character in the first current subtext in the current text is determined as the current character.

[0295] The following steps are executed in a loop until the preset loop stop condition is met:

[0296] Starting from the current character, a part of the text content is split from the current text, so that the current summary to be added and the part of the text content are combined into a current sub-text whose length is not greater than a preset length threshold;

[0297] Based on the target summary extraction model, determine the text summary of the combined current subtext. Determine the determined text summary as the current summary to be added;

[0298] The next character of the last character in the combined current sub-text in the current text is determined as the current character.

[0299] Of course, in the above loop process, the current summary to be added and part of the text content can be combined into the i+1th current subtext whose length is not greater than the preset length threshold. Where i is the number of the current loop. In the first loop process, the second current subtext can be determined, and so on, the order of the current subtexts can be directly determined, so as to combine and obtain the current subtext sequence.

[0300] This embodiment does not limit the preset loop stop condition. It is understandable that the preset loop stop condition can be that the last character in the combined current subtext is the last character of the current text; or it can be that the partial text content split from the current text contains the last character of the current text.

[0301] Through the above cycle process, the current text can be split into several current subtexts in order and combined into a current subtext sequence. Among them, since the subtext summaries are determined one by one, the subtext summaries are used to generate the next current subtext, thereby improving the information integrity of the previous text, improving the correlation and continuity between the current subtexts, and improving the accuracy of summary extraction.

[0302] For example, the second current sub-text contains the summary of the first current sub-text. After a summary is generated based on the second current sub-text, the summary of the second current sub-text can also contain the information in the first current sub-text and be added to the third current sub-text. And so on. In this way, the information completeness of the previous text can be improved for the current sub-text, and the correlation and continuity between the current sub-texts can be improved, thereby improving the accuracy of summary extraction.

[0303] It is understandable that the above three text splitting embodiments are only used for illustrative purposes and cannot limit the scope of the disclosure of this specification. Different embodiments may also be combined with each other.

[0304] 5. Explanation of the cycle process.

[0305] With respect to the loop process in the process of the present method, when the length of the target text is greater than a preset length threshold, the target text can be determined as the current text and the loop process can be started.

[0306] When the length of the target text is not greater than a preset length threshold, it is possible to directly extract a text summary of the target text based on the target summary extraction model without determining the current text.

[0307] In the loop process, the current text in the first loop can be the target text, and the current text in the subsequent loop can be the current text summary determined in the previous loop, so that the text summary can be extracted multiple times with multiple loops to reduce the length of the text.

[0308] The current text is split into several current sub-texts. For a specific explanation, please refer to the above embodiment.

[0309] After splitting to obtain several current sub-texts, sub-text summaries can be further extracted for different current sub-texts based on the target summary extraction model.

[0310] The method flow does not limit the number of current subtexts from which subtext summaries need to be extracted.

[0311] Optionally, part of the current subtext may be screened out, and based on the target summary extraction model, subtext summaries may be extracted for each of the different screened current subtexts. This embodiment does not limit the screening method, and may specifically be based on the length of the current subtext, etc., and current subtexts with shorter lengths may be screened out and subtext summaries may not be extracted. Specifically, subtext summaries may be extracted for current subtexts with lengths greater than a preset length. Of course, it may also be based on the target summary extraction model to extract subtext summaries for each split current subtext.

[0312] The method flow does not limit the order of extracting subtext summaries for different current subtexts. Specifically, subtext summaries can be extracted for different current subtexts in parallel, or in batches.

[0313] The method does not limit the order of the steps of text splitting and extracting subtext summaries. Optionally, the current text can be split into several current subtexts first, and then subtext summaries can be extracted for different current subtexts based on the target summary extraction model; or subtext summaries can be extracted while splitting.

[0314] After extracting the subtext summaries, the extracted subtext summaries can be further integrated to determine the current text summary. The current text summary can be specifically regarded as a text summary extracted for the current text, that is, it can be a text summary indirectly extracted for the current text by splitting the subtexts, extracting the subtext summaries respectively, and then integrating the subtext summaries.

[0315] The method flow does not limit the specific way of synthesizing the sub-text summaries to determine the current text summary.

[0316] Optionally, the sub-text summaries may be directly spliced ​​to obtain the current text summary. Alternatively, some sub-text summaries may be selected from the sub-text summaries and spliced ​​to obtain the current text summary. This embodiment does not limit the screening method, and may specifically be screening based on the importance or length of the sub-text summaries, or screening based on the length of the current sub-text corresponding to the sub-text summaries, and so on.

[0317] Furthermore, the method flow does not limit the number of sub-text summaries to be synthesized, and may be all or part of the extracted sub-text summaries. The specific screening method is not limited.

[0318] In addition, the order of the current sub-texts can be determined, specifically the order of the current sub-texts in the current text. According to the determined order of order, the sub-text summaries corresponding to each current sub-text can be sorted, so that several sub-text summaries can be spliced ​​in the sorted order to obtain the current text summary.

[0319] Of course, segmentation or text regularization operations can also be performed on the splicing results of several sub-text summaries to obtain the current text summary. The specific explanation of the segmentation operation and text regularization operation can be found in the above embodiment.

[0320] Furthermore, it can be determined whether the loop needs to be stopped based on the length of the current text summary.

[0321] If the length of the current text summary is greater than the preset length threshold, the text length still needs to be reduced through a loop, so that the current text summary can be re-determined as the new current text as the current text at the beginning of the next cycle, further reducing the text length.

[0322] If the length of the current text summary is not greater than a preset length threshold, the loop can be stopped, and the final target text summary is extracted for the current text summary directly based on the target summary extraction model.

[0323] The target text summary here can be regarded as a text summary extracted from the target text, that is, it can be a text summary indirectly extracted from the target text by splitting the sub-texts once or multiple times, extracting the sub-text summaries respectively, and then synthesizing the sub-text summaries.

[0324] In addition, the embodiment of the present invention also provides another method embodiment of a text summary extraction method.

[0325] In an embodiment of the present method, in order to make the length of the target text adapt to a target summary extraction method, the length of the target text can be reduced by extracting the text summary once or multiple times to adapt to the upper limit of the input text length of the target summary extraction method, so that the final target text summary can be extracted for the target text based on the target summary extraction method.

[0326] The process of this method does not limit the specific method of the target summary extraction method, and the target summary extraction method can be any summary extraction method, for example, the above-mentioned target summary extraction model.

[0327] Furthermore, the method flow is not limited to the text summary extraction method used when reducing the target text length. The target summary extraction method or other text summary extraction methods may be used.

[0328] like Figure 3 As shown, Figure 3 FIG. 4 is a flow chart of another text summary extraction method according to an embodiment of the present invention.

[0329] The embodiment of the present invention does not limit the execution subject of the method process. Optionally, the execution subject can be any computing device or any application. For example, a terminal, a server, an application with a text summary generation function, etc.

[0330] The method may include the following steps.

[0331] S301: Determine a target text and a target summary extraction method.

[0332] S302: When the length of the target text is greater than a preset length threshold, the target text is determined as the current text.

[0333] The preset length threshold is less than or equal to the upper limit of the input text length of the target summary extraction method.

[0334] S303: Determine a current text summary for the current text.

[0335] S304: Determine whether the length of the current text summary is greater than a preset length threshold.

[0336] When the length of the current text summary is greater than the preset length threshold, S305 is executed; when the length of the current text summary is not greater than the preset length threshold, S306 is executed.

[0337] S305: Determine the current text summary as the new current text, and execute S303.

[0338] S306: Extracting a target text summary for the current text summary based on the target summary extraction method.

[0339] In which, when the length of the target text is not greater than a preset length threshold, that is, when the length of the target text is less than or equal to the preset length threshold, a target text summary can be directly extracted for the target text based on the target summary extraction method.

[0340] For ease of understanding, Figure 4 As shown, Figure 4 FIG. 4 is a flow chart of another text summary extraction method according to an embodiment of the present invention.

[0341] The method may include the following steps.

[0342] S401: Determine a target text and a target summary extraction method.

[0343] S402: When the length of the target text is greater than a preset length threshold, and the preset length threshold is less than or equal to the upper limit of the input text length of the target summary extraction method, the target text is determined as the current text, and the following steps are executed in a loop: a current text summary is determined for the current text; when the length of the current text summary is greater than the preset length threshold, the current text summary is determined as the new current text; when the length of the current text summary is less than or equal to the preset length threshold, the loop is stopped, and based on the target summary extraction method, a target text summary is extracted for the current text summary.

[0344] In the above method flow, the text summary can be extracted separately by splitting the text once or multiple times, thereby reducing the length of the text until the text summary can be extracted for the entire text after the length is reduced, which can reduce the impact of the length of the input text on the text summary extraction.

[0345] Furthermore, by reducing the text length by extracting the text summary of the subtext, the text length can be reduced while retaining the information in the target text as much as possible, thereby improving the information integrity of the target text in the target text summary and reducing the information loss rate of the target text in the target text summary.

[0346] The process of this method does not limit the specific method of the target summary extraction method. The target summary extraction method can be any summary extraction method, for example, the above-mentioned target summary extraction model.

[0347] Furthermore, the method flow is not limited to the text summary extraction method used when reducing the target text length. The target summary extraction method or other text summary extraction methods may be used.

[0348] In the process of this method, it is not limited to using the same or different text summarization methods in different cycles.

[0349] Regarding determining the current text summary for the current text in S303 and S402, the method flow does not limit the specific manner and process of determining the current text summary.

[0350] Optionally, the explanations of S101 - S106 and S201 - S202 above may be referred to to determine the summary of the current text by splitting the sub-texts, or other summary extraction methods may be used to directly determine the summary of the current text for the current text.

[0351] In an optional embodiment, a current text summary is determined for a current text, specifically, the following steps may be performed: determining a current summary extraction method; when the length of the current text is greater than a specified length threshold, the specified length threshold is less than or equal to the upper limit of the input text length of the current summary extraction method, the current text is split into a number of current sub-texts, and based on the current summary extraction method, sub-text summaries are extracted for different current sub-texts; the length of any current sub-text is less than or equal to the specified length threshold; and the current text summary is determined by combining the extracted sub-text summaries.

[0352] Optionally, when the length of the current text is less than or equal to a specified length threshold, a current text summary is extracted for the current text based on the current summary extraction method.

[0353] This embodiment does not limit the specific current summary extraction method. The current summary extraction method can be any summary extraction method, for example, a target summary extraction method, or any summary extraction model.

[0354] It should be noted that the current summary extraction method may be re-determined in each cycle, and the current summary extraction methods determined in different cycles may be the same or different.

[0355] For ease of understanding, in a specific example, the current summary extraction method is determined by directly determining the target summary extraction method as the current summary extraction method, or determining the target summary extraction model as the current summary extraction method. Specifically, the method for extracting a summary based on the target summary extraction model is determined as the current summary extraction method.

[0356] After determining the current summary extraction method, it can be further determined whether the length of the current text is greater than a specified length threshold. The explanation of the specified length threshold here can refer to the explanation of the preset length threshold above.

[0357] Optionally, the current text is split into several current sub-texts, which can be specifically: splitting the current text into a current sub-text sequence; the current sub-text sequence includes different current sub-texts arranged in the order of texts in the current text; in the current sub-text sequence, between any pair of adjacent current sub-texts, the last N consecutive characters in the previous current sub-text are the same as the first N consecutive characters in the next current sub-text.

[0358] Optionally, the current text contains several paragraphs; the current text is split into several current sub-texts, which can be specifically: the current text is split into several prepared sub-texts; any prepared sub-text contains at least one paragraph; when the length of any prepared sub-text is less than or equal to a specified length threshold, the prepared sub-text is determined as the current sub-text; when the length of any prepared sub-text is greater than a specified length threshold, the prepared sub-text is split into at least 2 current sub-texts; the length of any current sub-text is less than or equal to the specified length threshold.

[0359] Optionally, the current text is split into several current sub-texts, including: splitting the current text into a current sub-text sequence; the current sub-text sequence includes different current sub-texts arranged in the order of text in the current text; in the current sub-text sequence, between any pair of adjacent current sub-texts, the latter current sub-text includes a sub-text summary extracted from the previous current sub-text.

[0360] Specifically, it may be a sub-text summary extracted from a previous current sub-text based on the current summary extraction method.

[0361] Optionally, determining the target text includes: determining a number of collected information; any collected information has a corresponding collection time point; converting the determined collected information into text; integrating the converted text to determine the target text; in the target text, the order of different texts converted from different collected information is arranged according to the order of the collection time points corresponding to the collected information.

[0362] Optionally, the above method flow may further include: determining user information that provides any collected information; converting the determined collected information into text, including: converting the determined collected information and user information that provides the determined collected information into text.

[0363] Optionally, the preset length threshold includes: the length obtained by subtracting the length of a preset prompt word from the upper limit of the length of the input text of the target summary extraction method; the preset prompt word is used to prompt the target summary extraction method to extract a summary of the input text; based on the target summary extraction method, extracting a target text summary for the current text summary, including: inputting the preset prompt word and the current text summary into the target summary extraction method, and obtaining the target text summary output by the target summary extraction method.

[0364] Optionally, the specified length threshold includes: a length obtained by subtracting a length of a preset prompt word from an upper limit of a length of an input text of the current summary extraction method; the preset prompt word is used to prompt the current summary extraction method to extract a summary for the input text.

[0365] Optionally, the structure of the target summary extraction model may be any one of the following: a language model, a natural language processing model, a generative summary model, and an extractive summary model.

[0366] For a detailed explanation of the process flow of this method, please refer to the explanation of S101-S106 and S201-S202 above.

[0367] For ease of understanding, the embodiment of the present invention also provides a specific application example.

[0368] This embodiment can be applied in a conference scenario to generate meeting minutes based on the meeting content.

[0369] like Figure 5 As shown, Figure 5 The figure is a flow chart of a method for extracting meeting minutes according to an embodiment of the present invention.

[0370] The embodiment of the present invention does not limit the execution subject of the method process. Optionally, the execution subject can be any computing device or any application. For example, a terminal, a server, an application with a text summary generation function, etc.

[0371] The method may include the following steps.

[0372] S501: Determine a target conference text and a target summary extraction model.

[0373] S502: When the length of the target conference text is greater than a preset length threshold, the target conference text is determined as the current text.

[0374] The preset length threshold may be less than or equal to the upper limit of the input text length of the target summary extraction model.

[0375] S503: Split the current text into several current sub-texts, and extract sub-text summaries for different current sub-texts based on the target summary extraction model; the length of any current sub-text is less than or equal to a preset length threshold; and determine the current text summary by combining the extracted sub-text summaries.

[0376] S504: Determine whether the length of the current text summary is greater than a preset length threshold.

[0377] When the length of the current text summary is greater than the preset length threshold, S505 is executed; when the length of the current text summary is not greater than the preset length threshold, S506 is executed.

[0378] S505: Determine the current text summary as the new current text, and execute S503.

[0379] S506: Extract target meeting minutes based on the target summary extraction model for the current text summary.

[0380] Among them, when the length of the target meeting text is not greater than the preset length threshold, that is, when the length of the target text is less than or equal to the preset length threshold, the target text summary, that is, the target meeting minutes, can be extracted directly based on the target summary extraction model.

[0381] For ease of understanding, Figure 6 As shown, Figure 6 The figure is a flow chart of another method for extracting meeting minutes according to an embodiment of the present invention.

[0382] The method may include the following steps.

[0383] S601: Determine a target conference text and a target summary extraction model.

[0384] S602: When the length of the target conference text is greater than a preset length threshold, and the preset length threshold is less than or equal to the upper limit of the input text length of the target summary extraction model, the target conference text is determined as the current text, and the following steps are executed in a loop: the current text is split into several current sub-texts, and sub-text summaries are extracted for different current sub-texts based on the target summary extraction model; the length of any current sub-text is less than or equal to the preset length threshold; the current text summary is determined by combining the extracted sub-text summaries; when the length of the current text summary is greater than the preset length threshold, the current text summary is determined as the new current text; when the length of the current text summary is less than or equal to the preset length threshold, the loop is stopped, and the target meeting minutes are extracted for the current text summary based on the target summary extraction model.

[0385] In the above method flow, the text summary can be extracted separately by splitting the text once or multiple times, thereby reducing the length of the text until the text summary can be extracted for the entire text after the length is reduced, which can reduce the impact of the length of the input text on the text summary extraction.

[0386] Furthermore, by reducing the text length by extracting the text summary of the subtext, the text length can be reduced while retaining the information in the target text as much as possible, thereby improving the information integrity of the target text in the target text summary and reducing the information loss rate of the target text in the target text summary.

[0387] Optionally, the current text is split into several current sub-texts, including: splitting the current text into a current sub-text sequence; the current sub-text sequence includes different current sub-texts arranged in the order of texts in the current text; in the current sub-text sequence, between any pair of adjacent current sub-texts, the last N consecutive characters in the previous current sub-text are the same as the first N consecutive characters in the next current sub-text.

[0388] Optionally, the current text contains several paragraphs; splitting the current text into several current sub-texts includes: splitting the current text into several prepared sub-texts; any prepared sub-text contains at least one paragraph; when the length of any prepared sub-text is less than or equal to a preset length threshold, the prepared sub-text is determined as the current sub-text; when the length of any prepared sub-text is greater than a preset length threshold, the prepared sub-text is split into at least 2 current sub-texts; the length of any current sub-text is less than or equal to the preset length threshold.

[0389] Optionally, split the current text into several current subtexts, including:

[0390] Split the current text into a current sub-text sequence; the current sub-text sequence includes different current sub-texts arranged in the order of the texts in the current text;

[0391] In the current subtext sequence, between any pair of adjacent current subtexts, the latter current subtext includes the subtext summary extracted for the previous current subtext based on the target summary extraction model.

[0392] Optionally, determining the target conference text includes: determining a number of collected information; any collected information has a corresponding collection time point; converting the determined collected information into text; determining the target conference text by integrating the converted text; in the target conference text, the order of different texts converted from different collected information is arranged according to the order of the collection time points corresponding to the collected information.

[0393] Optionally, the above method flow further includes: determining user information that provides any collected information;

[0394] Convert the determined collected information into text, including:

[0395] The determined collected information and user information providing the determined collected information are converted into text.

[0396] Optionally, the preset length threshold includes: the length obtained by subtracting the length of a preset prompt word from the upper limit of the input text length of the target summary extraction model; the preset prompt word is used to prompt the target summary extraction model to extract a summary of the input text; based on the target summary extraction model, extracting target meeting minutes for the current text summary, including: inputting the preset prompt word and the current text summary into the target summary extraction model to obtain the target meeting minutes output by the target summary extraction model.

[0397] Optionally, the structure of the target summary extraction model includes any one of the following: a language model, a natural language processing model, a generative summary model, and an extractive summary model.

[0398] For a detailed explanation of the process flow of this method, please refer to the explanation of S101-S106 and S201-S202 above.

[0399] For ease of understanding, the embodiment of the present invention further provides an application embodiment.

[0400] This embodiment designs an efficient meeting minutes generation method based on a large model. The overall process framework of the proposed method is:

[0401] First, obtain meeting-related information, such as meeting recordings and meeting blackboards, generate recording text containing speaker information through speech recognition and voiceprint recognition technology, and generate corresponding blackboard text through a multimodal large model.

[0402] Then, the obtained meeting text is regularized through a large model to delete redundant and repeated texts.

[0403] The regularized meeting text is then split into several meeting fragments.

[0404] Finally, the reasoning of the Large Language Model (LLM) is accelerated, and then the meeting fragments are input into the accelerated LLM model to generate the corresponding meeting key points fragments. Finally, these key points fragments are summarized and input into the LLM model to generate the final meeting minutes.

[0405] This embodiment can effectively solve the problem of limited input context length of large language models, and can conduct more rounds of conversations under limited video memory. It also improves the large model's ability to understand ultra-long texts and realizes efficient generation of meeting minutes.

[0406] Among them, Large Language Model (LLM) is an artificial intelligence model designed to understand and generate human language. They are trained on a large amount of text data and can perform a wide range of tasks. Large language models play an important role in scenarios such as finance and education, but they also have disadvantages. The context length that the model can support is limited. In long text scenarios, such as the need to generate meeting minutes in meeting scenarios, meetings may last for several hours, and the generated recording text characters are as high as tens of thousands of words. This large amount of text poses a severe challenge to large language models. If too long text is fed to LLM at one time, the generated meeting minutes will lose a lot of information, and the ability to understand long text is limited. In addition, too long text increases the reasoning time of large language models. Therefore, how to improve the context length and reasoning speed is an urgent problem to be solved.

[0407] This embodiment proposes an efficient meeting minutes generation method based on a large model. Based on a large language model and a multimodal model, corresponding text is generated for meeting recordings and related blackboard photos in meeting scenarios through speech recognition, voiceprint recognition, and a multimodal large model, thereby recording the meeting content in multiple aspects and reducing the loss of important meeting information.

[0408] Secondly, the reasoning of LLM is accelerated through the large model reasoning library FastLLM. Then, the long meeting text is split into several meeting segments with the context length supported by the large model through text splitting technology. The meeting key points are generated through the accelerated large language model, and finally the final meeting minutes are generated based on these meeting key points.

[0409] The overall framework flow chart of this embodiment is as follows Figure 7 As shown, Figure 7 It is a framework flow chart of generating meeting minutes according to an embodiment of the present invention, which is divided into three major modules.

[0410] 1) Meeting information extraction module, responsible for extracting text information in the meeting scene.

[0411] 2) The conference information processing module uses a large model to organize text information and then divides the long conference text into segments.

[0412] 3) The meeting minutes generation module includes two main steps. First, the large model inference library FastLLM is used to accelerate the LLM inference. Then the segmented text is sent to the accelerated large model for key point generation. If the length of the generated key points is within the context length supported by the large model, these key point information are merged and sent to the large model to generate the final meeting summary information. If it exceeds, the text segmentation process is continued until the conditions are met.

[0413] Each model is explained below.

[0414] 1. Meeting information extraction module.

[0415] Meeting recording: Before obtaining the text of the meeting recording, you need to prepare a trained speech recognition model and voiceprint recognition model. The speech recognition model can recognize the speech content and convert it into corresponding text. The voiceprint recognition model can identify the identity of the speaker. You need to prepare a voiceprint library in advance. The voiceprint library stores personal recording files. If you need to identify the identity of the speaker, the model can compare the input recording with the recording in the voiceprint library to derive the speaker's identity. Through this process, you can get the recording text containing the speaker's information.

[0416] Meeting blackboard writing: In a meeting scenario, speakers may write on the blackboard while speaking, which can improve the audience's understanding of the speech. At this time, if only the speaker's recording audio is saved, non-participants who need to understand the meeting may be confused about the speech. Therefore, meeting blackboard writing is also an indispensable part. By obtaining the blackboard text during the meeting and inputting it into the multimodal large model, the blackboard text is generated by the method of image-to-text.

[0417] The specific process is as follows Figure 8 As shown, Figure 8 The figure is a schematic diagram showing a process of extracting conference information according to an embodiment of the present invention.

[0418] 2. Conference information processing module.

[0419] Text regularization: There may be problems with colloquial or repeated expressions in meetings, so it is necessary to regularize inappropriate content in the meeting text before sending it to the big model to generate minutes. The specific method is to input the meeting text into the big model in segments, and use the big model to remove colloquial words in the meeting manuscript, filter redundant words, reduce semantic repetition, and smooth the text, refine and reorganize sentences, and polish the text. The regularized text is more written and the article is more concise and regular.

[0420] Text splitting: The biggest drawback of large language models is the limited length of input context. Training large models with larger context lengths can alleviate this problem to a certain extent, but larger contexts also mean more training data and greater hardware requirements, which is obviously not a good method when resources are limited. At the same time, although some large models have extended the context length, their ability to understand long texts is limited, the generation effect is not ideal, and a lot of information is lost.

[0421] Based on this, this embodiment adopts the form of text segmentation to divide the long meeting text into segments and send it to the large model. The segmentation process is as follows: Fig. 9As shown, Fig. 9 It is a framework flow chart of text splitting according to an embodiment of the present invention.

[0422] For a conference text, we first determine whether the text exceeds the context length supported by the large model. If it does not exceed the context length, the text will not be split and will be directly input into the large model. If it exceeds the context length, the text will be split.

[0423] Splitting method: Input the conference text into the text splitting algorithm, and use the supported input context length as the upper limit of the number of split characters. At the same time, the presence of characters such as "\n\n", "\n", and "space" in the text is also taken into consideration during the splitting process. If these characters are encountered during the splitting process, it is assumed to be a segmentation, which can maintain the integrity of the original text paragraph. At the same time, the splitting of the conference text will affect the contextual semantic information of the segmentation to a certain extent. Therefore, the word overlap is set during the splitting process to retain the continuity between the segmented text blocks.

[0424] 3. Meeting minutes generation module.

[0425] Large model reasoning acceleration: This embodiment uses the FastLLM high-performance large model reasoning library to accelerate the large model, and the accelerated large model is used for the subsequent extraction of meeting minutes.

[0426] Meeting minutes generation: The meeting minutes generation process is as follows Fig.10 As shown, Fig.10 The figure is a schematic diagram showing a process of generating meeting minutes according to an embodiment of the present invention.

[0427] After the above text segmentation process, the segmented text is input into the accelerated large model to generate summary information corresponding to each segment, and then these summaries are merged to determine whether they exceed the context length that the model can support input. If not, they are input into the large model again to generate the final meeting minutes. If they exceed, the above text segmentation process is continued, and then the form of summary first and minutes is adopted again until the conditions are met, and finally the final meeting minutes are generated.

[0428] The meeting minutes generation method proposed in this embodiment is not restricted by the length of the input context that can be supported by the large model, and meeting information of any length can be input. At the same time, the large model is accelerated, and the time used to generate the final meeting minutes is also shortened.

[0429] The beneficial effects of this embodiment are explained below.

[0430] 1. This embodiment proposes an efficient meeting minutes generation method based on a large model, which solves the problem of time-consuming and labor-intensive manual summary writing.

[0431] 2. This embodiment uses speech recognition and voiceprint recognition models to generate conference text containing speaker information, and uses multimodal large model graph-to-text technology to generate conference text corresponding to the conference blackboard, which greatly ensures the integrity of the conference information.

[0432] 3. This embodiment proposes a text splitting technology, which greatly ensures the semantic information of the split paragraphs, and also solves the problem that the input context length that the large model can support is limited. At the same time, it also improves the large model's ability to understand ultra-long texts. Subsequently, the large model is accelerated in reasoning, which greatly shortens the generation time of meeting minutes and realizes the efficient generation of meeting minutes.

[0433] The various technical features in the above embodiments can be arbitrarily combined as long as there is no conflict or contradiction between the combinations of features. However, due to space limitations, they are not described one by one. Therefore, any combination of the various technical features in the above embodiments also falls within the scope of the present invention.

[0434] Corresponding to the above method embodiments, the embodiments of the present invention also provide corresponding device embodiments.

[0435] like Fig.11 As shown, Fig.11 Detailed description of the invention is given in detail.

[0436] The apparatus may include the following units.

[0437] A determination unit 701, used to determine a target text and a target summary extraction model;

[0438] The loop unit 702 is used to determine the target text as the current text and loop through the following steps when the length of the target text is greater than a preset length threshold and the preset length threshold is less than or equal to an upper limit of the input text length of the target summary extraction model:

[0439] The current text is split into several current sub-texts, and based on the target summary extraction model, sub-text summaries are extracted for different current sub-texts respectively; the length of any current sub-text is less than or equal to a preset length threshold; the current text summary is determined by combining the extracted several sub-text summaries;

[0440] When the length of the current text summary is greater than a preset length threshold, determining the current text summary as a new current text;

[0441] When the length of the current text summary is less than or equal to the preset length threshold, the loop is stopped, and a target text summary is extracted for the current text summary based on the target summary extraction model.

[0442] Optionally, the circulation unit 702 is used to:

[0443] Split the current text into a current sub-text sequence; the current sub-text sequence includes different current sub-texts arranged in the order of the texts in the current text;

[0444] In the current subtext sequence, between any pair of adjacent current subtexts, the last N consecutive characters in the previous current subtext are the same as the first N consecutive characters in the next current subtext.

[0445] Optionally, the current text includes several paragraphs; the loop unit 702 is used to:

[0446] Split the current text into several preliminary sub-texts; any preliminary sub-text contains at least one paragraph;

[0447] When the length of any prepared subtext is less than or equal to a preset length threshold, the prepared subtext is determined as the current subtext;

[0448] When the length of any prepared subtext is greater than a preset length threshold, the prepared subtext is split into at least two current subtexts; and the length of any current subtext is less than or equal to the preset length threshold.

[0449] Optionally, the cyclic unit 702 is used to: split the current text into a current subtext sequence; the current subtext sequence includes different current subtexts arranged in the order of text in the current text;

[0450] In the current subtext sequence, between any pair of adjacent current subtexts, the latter current subtext includes the subtext summary extracted for the previous current subtext based on the target summary extraction model.

[0451] Optionally, the determining unit 701 is used to: determine a number of collected information; any collected information has a corresponding collection time point;

[0452] Convert the determined collected information into text;

[0453] The converted texts are comprehensively considered to determine the target text; in the target text, the order of different texts converted from different collected information is arranged according to the order of the collected information corresponding to the collection time points.

[0454] Optionally, the determining unit 701 is further used to: determine user information that provides any collected information;

[0455] The determining unit 701 is used for:

[0456] The determined collected information and user information providing the determined collected information are converted into text.

[0457] Optionally, the preset length threshold includes: the length obtained by subtracting the length of the preset prompt word from the upper limit of the length of the input text of the target summary extraction model;

[0458] The preset prompt word is used to prompt the target summary extraction model to extract a summary of the input text;

[0459] The loop unit 702 is used to input the preset prompt word and the current text summary into the target summary extraction model to obtain the target text summary output by the target summary extraction model.

[0460] Optionally, the structure of the target summary extraction model includes any one of the following: a language model, a natural language processing model, a generative summary model, and an extractive summary model.

[0461] For specific explanations, please refer to the above method embodiments.

[0462] like Fig.12 As shown, Fig.12 FIG. 4 is a schematic diagram of the structure of another text summary extraction device according to an embodiment of the present invention.

[0463] The apparatus may include the following units.

[0464] The first determination unit 801 is used to: determine the target text and the target summary extraction method;

[0465] The first loop unit 802 is used to: when the length of the target text is greater than a preset length threshold, and the preset length threshold is less than or equal to the upper limit of the input text length of the target summary extraction method, determine the target text as the current text, and loop through the following steps:

[0466] Determine a current text summary for the current text;

[0467] When the length of the current text summary is greater than a preset length threshold, determining the current text summary as a new current text;

[0468] When the length of the current text summary is less than or equal to the preset length threshold, the loop is stopped, and a target text summary is extracted for the current text summary based on the target summary extraction method.

[0469] Optionally, the first loop unit 802 is used to: split the current text into a current sub-text sequence; the current sub-text sequence includes different current sub-texts arranged in the order of text in the current text; in the current sub-text sequence, between any pair of adjacent current sub-texts, the last N consecutive characters in the previous current sub-text are the same as the first N consecutive characters in the next current sub-text.

[0470] Optionally, the first loop unit 802 is used for: the current text contains several paragraphs; the current text is split into several prepared sub-texts; any prepared sub-text contains at least one paragraph; when the length of any prepared sub-text is less than or equal to a specified length threshold, the prepared sub-text is determined as the current sub-text; when the length of any prepared sub-text is greater than a specified length threshold, the prepared sub-text is split into at least 2 current sub-texts; the length of any current sub-text is less than or equal to the specified length threshold.

[0471] Optionally, the first loop unit 802 is used to: split the current text into a current sub-text sequence; the current sub-text sequence includes different current sub-texts arranged in the order of text in the current text; in the current sub-text sequence, between any pair of adjacent current sub-texts, the latter current sub-text includes the sub-text summary extracted from the previous current sub-text.

[0472] Specifically, it may be a sub-text summary extracted from a previous current sub-text based on the current summary extraction method.

[0473] Optionally, the first determination unit 801 is used to: determine a number of collected information; any collected information has a corresponding collection time point; convert the determined collected information into text; integrate the converted text to determine the target text; in the target text, the order of different texts converted from different collected information is arranged according to the order of the collection time points corresponding to the collected information.

[0474] Optionally, the first determining unit 801 is further used to: determine user information that provides any collected information; the first determining unit 801 is used to: convert the determined collected information and the user information that provides the determined collected information into text.

[0475] Optionally, the preset length threshold includes: the length obtained by subtracting the length of a preset prompt word from the upper limit of the length of the input text of the target summary extraction method; the preset prompt word is used to prompt the target summary extraction method to extract a summary of the input text; the first loop unit 802: inputs the preset prompt word and the current text summary into the target summary extraction method to obtain a target text summary output by the target summary extraction method.

[0476] Optionally, the specified length threshold includes: a length obtained by subtracting a length of a preset prompt word from an upper limit of a length of an input text of the current summary extraction method; the preset prompt word is used to prompt the current summary extraction method to extract a summary for the input text.

[0477] For specific explanations, please refer to the above method embodiments.

[0478] like Fig.13 As shown, Fig.13The figure is a schematic diagram of the structure of a device for extracting meeting minutes according to an embodiment of the present invention.

[0479] The second determination unit 901 is used to: determine the target conference text and the target summary extraction model;

[0480] The second loop unit 902 is used to: when the length of the target conference text is greater than a preset length threshold, and the preset length threshold is less than or equal to the upper limit of the input text length of the target summary extraction model, determine the target conference text as the current text, and loop through the following steps:

[0481] The current text is split into several current sub-texts, and based on the target summary extraction model, sub-text summaries are extracted for different current sub-texts respectively; the length of any current sub-text is less than or equal to a preset length threshold; the current text summary is determined by combining the extracted several sub-text summaries;

[0482] When the length of the current text summary is greater than a preset length threshold, determining the current text summary as a new current text;

[0483] When the length of the current text summary is less than or equal to the preset length threshold, the loop is stopped, and the target meeting minutes are extracted for the current text summary based on the target summary extraction model.

[0484] Optionally, the second determination unit 901 is used to: split the current text into a current sub-text sequence; the current sub-text sequence includes different current sub-texts arranged in the order of text in the current text; in the current sub-text sequence, between any pair of adjacent current sub-texts, the last N consecutive characters in the previous current sub-text are the same as the first N consecutive characters in the next current sub-text.

[0485] Optionally, the current text contains several paragraphs; the second determination unit 901 is used to: split the current text into several prepared sub-texts; any prepared sub-text contains at least one paragraph; when the length of any prepared sub-text is less than or equal to a preset length threshold, the prepared sub-text is determined as the current sub-text; when the length of any prepared sub-text is greater than a preset length threshold, the prepared sub-text is split into at least 2 current sub-texts; the length of any current sub-text is less than or equal to the preset length threshold.

[0486] Optionally, the second determining unit 901 is used to: split the current text into a current subtext sequence; the current subtext sequence includes different current subtexts arranged in the order of text in the current text;

[0487] In the current subtext sequence, between any pair of adjacent current subtexts, the latter current subtext includes the subtext summary extracted for the previous current subtext based on the target summary extraction model.

[0488] Optionally, the second determination unit 901 is used to: determine a number of collected information; any collected information has a corresponding collection time point; convert the determined collected information into text; integrate the converted text to determine the target conference text; in the target conference text, the order of different texts converted from different collected information is arranged according to the order of the collection time points corresponding to the collected information.

[0489] Optionally, the second determining unit 901 is further used to: determine user information that provides any collected information;

[0490] The second determining unit 901 is used to convert the determined collected information and the user information providing the determined collected information into text.

[0491] Optionally, the preset length threshold includes: the length obtained by subtracting the length of a preset prompt word from the upper limit of the input text length of the target summary extraction model; the preset prompt word is used to prompt the target summary extraction model to extract a summary of the input text; the second loop unit 902 is used to: input the preset prompt word and the current text summary into the target summary extraction model to obtain the target meeting minutes output by the target summary extraction model.

[0492] Optionally, the structure of the target summary extraction model includes any one of the following: a language model, a natural language processing model, a generative summary model, and an extractive summary model.

[0493] For specific explanations, please refer to the above method embodiments.

[0494] An embodiment of the present invention further provides a computer device, which at least includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein any of the above method embodiments is implemented when the processor executes the program.

[0495] An embodiment of the present invention also provides an electronic device, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute any of the above-mentioned method embodiments.

[0496] Fig.141 is a schematic diagram of the hardware structure of a computer device configured with a method according to an embodiment of the present invention, and the device may include: a processor 1010, a memory 1020, an input / output interface 1030, a communication interface 1040, and a bus 1050. The processor 1010, the memory 1020, the input / output interface 1030, and the communication interface 1040 are connected to each other in communication within the device through the bus 1050.

[0497] The processor 1010 can be implemented by a general-purpose CPU (Central Processing Unit), a microprocessor, an application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided by the embodiments of the present invention.

[0498] The memory 1020 may be implemented in the form of ROM (Read On ly Memory), RAM (Random Access Memory), static storage device, dynamic storage device, etc. The memory 1020 may store an operating system and other application programs. When the technical solution provided by the embodiment of the present invention is implemented by software or firmware, the relevant program code is stored in the memory 1020 and is called and executed by the processor 1010.

[0499] The input / output interface 1030 is used to connect the input / output module to realize information input and output. The input / output module can be configured in the device as a component (not shown in the figure), or it can be externally connected to the device to provide corresponding functions. The input device may include a keyboard, a mouse, a touch screen, a microphone, various sensors, etc., and the output device may include a display, a speaker, a vibrator, an indicator light, etc.

[0500] The communication interface 1040 is used to connect a communication module (not shown) to realize communication interaction between the device and other devices. The communication module can realize communication through a wired mode (such as USB, network cable, etc.) or a wireless mode (such as mobile network, WI FI, Bluetooth, etc.).

[0501] The bus 1050 includes a path that transmits information between the various components of the device (eg, the processor 1010, the memory 1020, the input / output interface 1030, and the communication interface 1040).

[0502] It should be noted that, although the above device only shows the processor 1010, the memory 1020, the input / output interface 1030, the communication interface 1040 and the bus 1050, in the specific implementation process, the device may also include other components necessary for normal operation. In addition, it can be understood by those skilled in the art that the above device may also only include the components necessary for implementing the embodiments of the present invention, and does not necessarily include all the components shown in the figure.

[0503] An embodiment of the present invention further provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, any of the above method embodiments is implemented.

[0504] An embodiment of the present invention further provides a computer-readable storage medium storing a computer program, wherein the computer program implements any of the above method embodiments when executed by a processor.

[0505] Computer readable media include permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. Information can be computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer readable media does not include transitory media such as modulated data signals and carrier waves.

[0506] An embodiment of the present invention further provides a computer program product, including a computer program / instruction, wherein the computer program / instruction implements any of the above method embodiments when executed by a processor.

[0507] It can be known from the description of the above implementation mode that a person skilled in the art can clearly understand that the embodiment of the present invention can be implemented by means of software plus a necessary general hardware platform. Based on such an understanding, the technical solution of the embodiment of the present invention can be essentially or in other words, the part that makes the contribution can be embodied in the form of a software product, and the computer software product can be stored in a storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiment of the present invention.

[0508] The systems, devices, modules or units described in the above embodiments may be implemented by computer chips or entities, or by products with certain functions. A typical implementation device is a computer, which may be in the form of a personal computer, a laptop computer, a cellular phone, a camera phone, a smart phone, a personal digital assistant, a media player, a navigation device, an email transceiver, a game console, a tablet computer, a wearable device or a combination of any of these devices.

[0509] Each embodiment in this specification is described in a progressive manner, and the same and similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment. The device embodiment described above is only schematic, wherein the modules described as separate components may or may not be physically separated, and the functions of each module can be implemented in the same one or more software and / or hardware when implementing the embodiment of the present invention. It is also possible to select some or all of the modules according to actual needs to achieve the purpose of the embodiment. Ordinary technicians in this field can understand and implement it without paying creative work.

[0510] The above is only a specific implementation of the embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the embodiment of the present invention. These improvements and modifications should also be regarded as protection for the embodiment of the present invention.

[0511] In the present invention, the terms "first" and "second" are used for descriptive purposes only and cannot be understood as indicating or implying relative importance. The term "plurality" refers to two or more than two, unless otherwise clearly defined.

[0512] Those skilled in the art will readily appreciate other embodiments of the present invention after considering the specification and practicing the disclosure disclosed herein. The present invention is intended to cover any variations, uses or adaptations of the present invention that follow the general principles of the present invention and include common knowledge or customary techniques in the art that are not disclosed by the present invention. The description and examples are to be considered exemplary only, and the true scope and spirit of the present invention are indicated by the following claims.

[0513] It should be understood that the present invention is not limited to the exact construction that has been described above and shown in the drawings and that various modifications and changes may be made without departing from the scope thereof. The scope of the present invention is limited only by the appended claims.

Claims

1. A text summary extraction method, characterized in that: include: Determine the target text and target summary extraction model; In the case where the length of the target text is greater than a preset length threshold, and the preset length threshold is less than or equal to the upper limit of the input text length of the target summary extraction model, the target text is determined as the current text, and the following steps are executed in a loop: The current text is split into a plurality of current subtexts, and based on the target summary extraction model, subtext summaries are extracted for different current subtexts respectively; the length of any current subtext is less than or equal to the preset length threshold; and the current text summary is determined by synthesizing the extracted plurality of subtext summaries; When the length of the current text summary is greater than the preset length threshold, determining the current text summary as a new current text; When the length of the current text summary is less than or equal to the preset length threshold, the loop is stopped, and a target text summary is extracted for the current text summary based on the target summary extraction model.

2. The method according to claim 1, characterized in that The current text is split into several current sub-texts, including: Split the current text into a current sub-text sequence; the current sub-text sequence includes different current sub-texts arranged in the order of the texts in the current text; In the current subtext sequence, between any pair of adjacent current subtexts, the last N consecutive characters in the previous current subtext are the same as the first N consecutive characters in the next current subtext.

3. The method according to claim 1, characterized in that The current text contains several paragraphs; The current text is split into several current sub-texts, including: Split the current text into several preliminary sub-texts; any preliminary sub-text contains at least one paragraph; When the length of any prepared subtext is less than or equal to the preset length threshold, determining the prepared subtext as the current subtext; When the length of any prepared subtext is greater than the preset length threshold, the prepared subtext is split into at least two current subtexts; and the length of any current subtext is less than or equal to the preset length threshold.

4. The method according to claim 1, characterized in that The current text is split into several current sub-texts, including: Split the current text into a current sub-text sequence; the current sub-text sequence includes different current sub-texts arranged in the order of the texts in the current text; In the current subtext sequence, between any pair of adjacent current subtexts, the latter current subtext includes the subtext summary extracted from the previous current subtext based on the target summary extraction model.

5. The method according to claim 1, characterized in that: The determining of the target text comprises: Determine a number of collected information; any collected information has a corresponding collection time point; Convert the determined collected information into text; The converted texts are synthesized to determine the target text; in the target text, the sequence of different texts converted from different collected information is arranged according to the sequence of the collected information corresponding to the collection time points.

6. The method according to claim 5, characterized in that Also includes: Determine the user information that provides any collected information; The step of converting the determined collected information into text includes: The determined collected information and user information providing the determined collected information are converted into text.

7. The method according to claim 1, characterized in that The preset length threshold includes: the length obtained by subtracting the length of the preset prompt word from the upper limit of the input text length of the target summary extraction model; The preset prompt word is used to prompt the target summary extraction model to extract a summary for the input text; The step of extracting a target text summary based on the target summary extraction model for the current text summary includes: The preset prompt word and the current text summary are input into the target summary extraction model to obtain the target text summary output by the target summary extraction model.

8. The method according to claim 1, characterized in that The structure of the target summary extraction model includes any one of the following: a language model, a natural language processing model, a generative summary model, and an extractive summary model.

9. A text summary extraction method, characterized in that: include: Determine the target text and target summary extraction method; In the case where the length of the target text is greater than a preset length threshold, and the preset length threshold is less than or equal to the upper limit of the input text length of the target summary extraction method, the target text is determined as the current text, and the following steps are executed in a loop: Determine a current text summary for the current text; When the length of the current text summary is greater than the preset length threshold, determining the current text summary as a new current text; When the length of the current text summary is less than or equal to the preset length threshold, the loop is stopped, and a target text summary is extracted for the current text summary based on the target summary extraction method.

10. The method according to claim 9, characterized in that Determining a current text summary for the current text includes: Determine the current summary extraction method; When the length of the current text is greater than a specified length threshold, and the specified length threshold is less than or equal to the upper limit of the input text length of the current summary extraction method, the current text is split into several current sub-texts, and based on the current summary extraction method, sub-text summaries are extracted for different current sub-texts respectively; the length of any current sub-text is less than or equal to the specified length threshold; and the current text summary is determined by combining the extracted several sub-text summaries.

11. A method for extracting meeting minutes, characterized in that: include: Determine the target meeting text and target summary extraction model; In the case where the length of the target conference text is greater than a preset length threshold, and the preset length threshold is less than or equal to the upper limit of the input text length of the target summary extraction model, the target conference text is determined as the current text, and the following steps are executed in a loop: The current text is split into a plurality of current subtexts, and based on the target summary extraction model, subtext summaries are extracted for different current subtexts respectively; the length of any current subtext is less than or equal to the preset length threshold; and the current text summary is determined by synthesizing the extracted plurality of subtext summaries; When the length of the current text summary is greater than the preset length threshold, determining the current text summary as a new current text; When the length of the current text summary is less than or equal to the preset length threshold, the loop is stopped, and the target meeting minutes are extracted for the current text summary based on the target summary extraction model.

12. A text summary extraction device, characterized in that: include: A determination unit, used for determining a target text and a target summary extraction model; A loop unit is used for determining the target text as the current text and looping the following steps when the length of the target text is greater than a preset length threshold and the preset length threshold is less than or equal to an upper limit of the input text length of the target summary extraction model: The current text is split into a plurality of current subtexts, and based on the target summary extraction model, subtext summaries are extracted for different current subtexts respectively; the length of any current subtext is less than or equal to the preset length threshold; and the current text summary is determined by synthesizing the extracted plurality of subtext summaries; When the length of the current text summary is greater than the preset length threshold, determining the current text summary as a new current text; When the length of the current text summary is less than or equal to the preset length threshold, the loop is stopped, and a target text summary is extracted for the current text summary based on the target summary extraction model.

13. An electronic device, characterized in that: include: at least one processor; as well as, a memory communicatively connected to the at least one processor; wherein, The memory stores instructions executable by the one processor, and the instructions are executed by the at least one processor so as to enable the at least one processor to perform the method according to any one of claims 1 to 11.

14. A computer-readable storage medium storing a computer program, characterized in that: The computer program implements the method according to any one of claims 1 to 11 when executed by a processor.

15. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the steps of the method according to any one of claims 1 to 11 are implemented.

Images