A method, device, equipment and storage medium for generating minutes

Abstract

The present application provides a method, device, equipment and storage medium for generating minutes, wherein the method includes: obtaining the text of the minutes to be generated as the target text; obtaining the word-level information, sentence-level information and topic-level information of the target text, wherein the word-level information includes the context information of the words in the target text and the keyword information of the target text, the sentence-level information includes the context information of the sentences in the target text and the key sentence information of the target text, and the topic-level information includes the topic segmentation information of the target text; according to the word-level information, sentence-level information and topic-level information of the target text, generating the minutes corresponding to the target text. The method for generating minutes provided by the present application can automatically generate the minutes corresponding to the target text.

Description

Technical Field

[0001] The present application relates to the field of natural language processing technology, and in particular to a minutes generation method, device, equipment and storage medium. Background Art

[0002] In some application scenarios, it is necessary to obtain the minutes corresponding to the text (the minutes corresponding to the text describe the main points of the text). For example, in scenarios involving meetings, it is usually necessary to obtain the meeting minutes to convey and distribute the content and important spirit of the meeting.

[0003] At present, the main way to obtain the minutes corresponding to the text is manual acquisition, that is, manually sorting the text to obtain the minutes corresponding to the text. However, in some cases, the text is long, for example, the duration of government and corporate daily office meetings, university academic seminars and other meetings is generally long, so the length of the meeting text is usually long. It is understandable that it is very time-consuming to manually sort out long texts, that is, the manual method of obtaining minutes is relatively inefficient. Summary of the invention

[0004] In view of this, the present application provides a minutes generation method, device, equipment and storage medium to solve the problem of low efficiency of existing minutes acquisition methods. The technical solution is as follows:

[0005] A method for generating minutes, comprising:

[0006] Get the text of the minutes to be generated as the target text;

[0007] Acquire word-level information, sentence-level information, and topic-level information of the target text, wherein the word-level information includes context information of words in the target text and keyword information of the target text, the sentence-level information includes context information of sentences in the target text and key sentence information of the target text, and the topic-level information includes topic segmentation information of the target text;

[0008] A summary corresponding to the target text is generated according to the word-level information, sentence-level information and topic-level information of the target text.

[0009] Optionally, the acquiring word-level information, sentence-level information, and topic-level information of the target text, and the generating a summary corresponding to the target text according to the word-level information, sentence-level information, and topic-level information of the target text, include:

[0010] Using the text understanding part of the minutes generation module, word-level information, sentence-level information, and topic-level information of the target text are obtained;

[0011] The minutes generation part of the minutes generation module is used to generate minutes corresponding to the target text based on the word-level information, sentence-level information and topic-level information of the target text.

[0012] Optionally, obtaining word-level information and sentence-level information of the target text includes:

[0013] Obtaining word features corresponding to each word in each sentence contained in the target text and integrating context information, and predicting the probability of each word in each sentence contained in the target text being a keyword based on the obtained word features;

[0014] Acquire sentence features corresponding to each sentence in the target text and integrated with context information, and predict the probability that each sentence in the target text is a key sentence based on the acquired sentence features;

[0015] The word features corresponding to each word in each sentence contained in the target text and the probability that each word in each sentence is a keyword are used as the word-level information of the target text, and the sentence features corresponding to each sentence in the target text and the probability that each sentence is a key sentence are used as the sentence-level information of the target text.

[0016] Optionally, obtaining word-level information and sentence-level information of the target text further includes:

[0017] For each sentence in the target text:

[0018] According to the probability that each word in the sentence is a keyword and the probability that the sentence is a key sentence, after determining the probability that the sentence is a key sentence, the probability that each word in the sentence is a keyword is used as the keyword probability after the key sentence probability corresponding to each word in the sentence;

[0019] According to the probability that each word in the sentence is a keyword and the probability that the sentence is a key sentence, after combining the probability that each word in the sentence is a keyword, the probability that the sentence is a key sentence is determined as the key sentence probability after combining the keyword probabilities corresponding to the sentence;

[0020] The word features corresponding to each word in each sentence contained in the target text and the keyword probability after the combined key sentence probability corresponding to each word in each sentence serve as the word-level information of the target text; the sentence features corresponding to each sentence in the target text and the key sentence probability after the combined key word probability corresponding to each sentence serve as the sentence-level information of the target text.

[0021] Optionally, obtaining topic-level information of the target text includes:

[0022] The target text is topic segmented according to the sentence-level information of the target text to obtain a topic segmentation result as the topic-level information of the target text.

[0023] Optionally, performing topic segmentation on the target text according to the sentence-level information of the target text to obtain a topic segmentation result includes:

[0024] For each sentence in the target text, according to the sentence features corresponding to the sentence and the key sentence probability after the joint keyword probability corresponding to the sentence, determine the sentence features after the joint keyword probability corresponding to the sentence, and based on the sentence features after the joint keyword probability corresponding to the sentence, predict the probability that the sentence is a topic ending sentence;

[0025] The topic segmentation result is determined according to the probability that each sentence in the target text is a topic ending sentence.

[0026] Optionally, generating a summary corresponding to the target text based on word-level information, sentence-level information, and topic-level information of the target text includes:

[0027] At each decoding moment:

[0028] Determine the context vectors corresponding to all sentences of the target text at the current decoding time according to the state vector at the current decoding time, the word-level information and the sentence-level information of the target text, wherein the state vector at the current decoding time is a state vector representing the decoded result information;

[0029] Determine, according to the state vector at the current decoding moment and the topic-level information of the target text, the context vectors corresponding to all topics involved in the target text at the current decoding moment;

[0030] The minutes generation result at the current decoding moment is determined according to the state vector at the current decoding moment, the context vectors corresponding to all sentences of the target text at the current decoding moment, and the context vectors corresponding to all topics involved in the target text at the current decoding moment.

[0031] Optionally, determining the context vectors corresponding to all sentences of the target text at the current decoding time according to the state vector at the current decoding time, the word-level information and the sentence-level information of the target text, includes:

[0032] Determine, according to the state vector at the current decoding moment and the word-level information of the target text, the context vectors corresponding to all the words in each sentence contained in the target text at the current decoding moment;

[0033] The context vectors corresponding to all the sentences of the target text at the current decoding time are determined according to the state vector at the current decoding time, the context vectors corresponding to all the words of each sentence contained in the target text at the current decoding time, and the sentence-level information of the target text.

[0034] Optionally, the sentence-level information of the target text includes sentence features corresponding to each sentence contained in the target text and key sentence probabilities after the combined keyword probabilities corresponding to each sentence;

[0035] The determining, based on the state vector at the current decoding moment, the context vectors corresponding to all words in each sentence contained in the target text at the current decoding moment, and the sentence-level information of the target text, the context vectors corresponding to all sentences of the target text at the current decoding moment comprises:

[0036] For each sentence in the target text, determine the weight corresponding to the sentence at the current decoding time according to the state vector at the current decoding time, the context vectors corresponding to all words of the sentence at the current decoding time, the sentence features corresponding to the sentence, and the key sentence probability after the joint keyword probability corresponding to the sentence;

[0037] The context vectors corresponding to all the sentences of the target text at the current decoding time are determined according to the weight corresponding to each sentence contained in the target text at the current decoding time, the sentence features corresponding to each sentence, and the context vectors corresponding to all the words of each sentence at the current decoding time.

[0038] Optionally, determining the context vectors corresponding to all topics of the target text at the current decoding time according to the state vector at the current decoding time and the topic-level information of the target text includes:

[0039] Determining sentences included in each topic involved in the target text according to the topic-level information of the target text;

[0040] For each topic involved in the target text, determine the representation vector of the topic at the current decoding time according to the sentence features corresponding to each sentence included in the topic and the context vectors corresponding to all words in each sentence included in the topic at the current decoding time;

[0041] According to the state vector at the current decoding moment and the representation vector of each topic involved in the target text at the current decoding moment, the context vector corresponding to all topics involved in the target text at the current decoding moment is determined.

[0042] Optionally, the minutes generation module is a minutes generation model, and the minutes generation model is trained using training texts annotated with keywords, key sentences, and topic ending sentences and minutes corresponding to the training texts. The training process of the minutes generation model includes:

[0043] Using the text understanding part of the minutes generation model, predict the probability of each word in each sentence contained in the training text being a keyword, the probability of each sentence being a key sentence, and the probability of each sentence being a topic ending sentence, so as to obtain keyword prediction results, key sentence prediction results, and topic ending sentence prediction results;

[0044] Using the minutes generation part of the minutes generation model, based on the word-level information, sentence-level information and topic-level information of the training text, predicting the minutes corresponding to the training text to obtain the minutes prediction result, wherein the word-level information, sentence-level information and topic-level information of the training text are obtained based on the text understanding part of the minutes generation model;

[0045] The prediction loss is determined based on the keyword prediction results, the key sentence prediction results, the topic end sentence prediction results, the minutes prediction results, and the keywords, key sentences, topic end sentences annotated in the training text and the minutes corresponding to the training text, and the parameters of the minutes generation model are updated based on the prediction loss.

[0046] A minutes generating device, comprising: a text acquisition module and a text processing module;

[0047] The text acquisition module is used to acquire the text of the minutes to be generated as the target text;

[0048] The text processing module is used to obtain word-level information, sentence-level information and topic-level information of the target text, and generate minutes corresponding to the target text according to the word-level information, sentence-level information and topic-level information of the target text;

[0049] Among them, the word-level information includes the context information of the words in the target text and the keyword information of the target text, the sentence-level information includes the context information of the sentences in the target text and the key sentence information of the target text, and the topic-level information includes the topic segmentation information of the target text.

[0050] Optionally, the text processing module is specifically used to:

[0051] Using the text understanding part of the minutes generation module, word-level information, sentence-level information, and topic-level information of the target text are obtained;

[0052] The minutes generation part of the minutes generation module is used to generate minutes corresponding to the target text based on the word-level information, sentence-level information and topic-level information of the target text.

[0053] A minutes generating device, comprising: a memory and a processor;

[0054] The memory is used to store programs;

[0055] The processor is used to execute the program to implement each step of the minutes generation method described in any one of the above items.

[0056] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, each step of the minutes generation method described in any one of the above items is implemented.

[0057] The minutes generation method, device, equipment and storage medium provided by the present application, after obtaining the target text for the minutes to be generated, first obtains the word-level information (including the context information of the words in the target text and the keyword information of the target text), sentence-level information (including the context information of the sentences in the target text and the key sentence information of the target text) and topic-level information (including the topic segmentation information of the target text) of the target text, and then generates the minutes corresponding to the target text according to the word-level information, sentence-level information and topic-level information of the target text. The minutes generation method provided by the present application can automatically generate the minutes corresponding to the target text, which greatly improves the efficiency compared to the manual acquisition of minutes, and because no manual participation is required, the influence of subjective factors is avoided. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0059] Figure 1 A flowchart of a method for generating minutes provided in an embodiment of the present application;

[0060] Figure 2 An example of a training text provided for an embodiment of the present application;

[0061] Figure 3 A flowchart of an implementation method for obtaining word-level information and sentence-level information of a target text by using the text understanding part of the minutes generation model provided in an embodiment of the present application;

[0062] Figure 4A flowchart of another implementation method for obtaining word-level information and sentence-level information of a target text by using the text understanding part of the minutes generation model provided in an embodiment of the present application;

[0063] Figure 5 A flowchart of topic segmentation of a target text based on sentence-level information of the target text using the text understanding part of the minutes generation model provided in the embodiment of the present application;

[0064] Figure 6 A structural schematic diagram of the text understanding part of the minutes generation model provided in the embodiment of the present application;

[0065] Figure 7 A flowchart of a minutes generation part using a minutes generation model provided in an embodiment of the present application, which generates minutes corresponding to a target text based on word-level information, sentence-level information, and topic-level information of the target text;

[0066] Figure 8 A flowchart of a training minutes generation model using training texts annotated with keywords, key sentences, topic ending sentences, and minutes corresponding to the training texts provided in an embodiment of the present application;

[0067] Fig. 9 A schematic diagram of the structure of a minutes generating device provided in an embodiment of the present application;

[0068] Fig.10 A schematic diagram of the structure of the minutes generation device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0069] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0070] In view of the fact that manual acquisition of minutes is time-consuming and inefficient, the applicant attempted to propose a solution for automatically acquiring minutes. To this end, the applicant conducted research. The initial idea was to obtain the minutes corresponding to the text based on the automatic text summarization method, that is, to extract the summary from the text, and to form a generalized text based on the extracted summary to obtain the minutes. When obtaining the minutes corresponding to the text based on the automatic text summarization method, the applicant thought that sentences could be extracted from the text intact to form a summary, that is, the summary content all comes from the text. When extracting the summary from the text, the summary extraction can be regarded as a sequence annotation task, and each sentence in the text can be classified into key sentences / non-key sentences.

[0071] Through research on automatic text summarization methods, the applicant found that automatic text summarization methods are only used for shorter texts, such as news, emails, etc., and are not suitable for longer texts (such as meeting texts). If they are used for longer texts, there will be a problem of high memory consumption. At the same time, the model used to extract summaries cannot capture long-distance dependency information well. In addition, a major feature of longer texts (such as meeting texts) is topic drift. Taking meeting texts as an example, meeting content has different agendas, such as background introduction, point of view elaboration, decision-making voting, etc. Even meetings with clear agendas often include multiple small agendas. In other words, meeting texts will include multiple different topics, and automatic text summarization methods are difficult to automatically learn these contents, resulting in a failure to focus on the core content.

[0072] In view of the many problems in obtaining minutes based on the automatic text summarization method, the inventor of this case conducted further research, and through continuous research, finally proposed a minutes generation method, which can automatically generate minutes corresponding to the text, and can overcome the problems existing in obtaining minutes based on the automatic text summarization method. The minutes generation method proposed in this application can be applied to electronic devices with data processing capabilities. The electronic device can be a terminal used on the user side, such as a smart phone, PC, notebook, PAD, etc. The electronic device can also be a server on the network side (it can be a single server, or it can be multiple servers or server clusters). The electronic device can generate minutes corresponding to the text according to the minutes generation method provided in this application. Next, the minutes generation method provided in this application is introduced through the following embodiments.

[0073] First embodiment

[0074] See also Figure 1 , shows a schematic diagram of the process of the minutes generation method provided in an embodiment of the present application, the method may include:

[0075] Step S101: Acquire the text of the minutes to be generated as the target text.

[0076] Among them, the target text can be a long text (such as a meeting text) or a short text (such as news, emails, etc.). That is, the minutes generation method provided in this application is applicable to long texts as well as short texts, and has strong versatility.

[0077] Step S102: Obtain word-level information, sentence-level information, and topic-level information of the target text.

[0078] In this embodiment, the word-level information of the target text is information related to the words in the target text, which may include context information of the words in the target text and keyword information of the target text. The keyword information of the target text is information related to the keywords in the target text.

[0079] In this embodiment, the sentence-level information in the target text is information related to the sentences in the target text, which may include context information of the sentences in the target text and key sentence information of the target text. The key sentence information of the target text is information related to the key sentences in the target text.

[0080] In this embodiment, the topic-level information of the target text is information related to the topic involved in the target text, which may include topic segmentation information of the target text. The topic segmentation information of the target text can indicate which sentences are included in each topic involved in the target text.

[0081] Step S103: Generate a summary corresponding to the target text according to the word-level information, sentence-level information and topic-level information of the target text.

[0082] The minutes corresponding to the target text are intended to be highly condensed, covering and summarizing the core key contents of the target text. This application automatically generates the minutes corresponding to the target text based on the valid information at three different levels of the target text.

[0083] Optionally, when obtaining the word-level information, sentence-level information and topic-level information of the target text, and generating the minutes corresponding to the target text based on the word-level information, sentence-level information and topic-level information of the target text, a minutes generation module can be used to obtain the word-level information, sentence-level information and topic-level information of the target text, and generate the minutes corresponding to the target text based on the word-level information, sentence-level information and topic-level information of the target text.

[0084] Specifically, the minutes generation module may include a text understanding part and a minutes generation part. The text understanding part of the minutes generation module may be used to understand the target text to obtain information at three levels of the target text, namely, word-level information, sentence-level information and topic-level information. The minutes generation part of the minutes generation module may be used to generate minutes corresponding to the target text based on the information at the three levels of the target text.

[0085] Optionally, the minutes generation module may be a minutes generation model, and the minutes generation model is trained using training texts annotated with keywords, key sentences, and topic ending sentences, and minutes corresponding to the training texts. Figure 2 An example of a training text is shown ( Figure 2 The bold sentences are the key sentences, the shaded words are the keywords, and the sentences marked with numbers are the topic ending sentences). Figure 2 The minutes corresponding to the training text shown are "The key tasks include the publicity and material preparation of Qi**yun, the rendering design of the shipping image store, and the follow-up of the *Tong Spring Fair, including the plan and the search for a venue."

[0086] It should be noted that the present embodiment does not limit the minutes generation module to a minutes generation model, that is, the minutes generation module can also be a module that implements minutes generation by a non-model method, for example, a module that implements minutes generation based on set rules (for example, based on set information extraction rules, extracts information at three levels of the target text, and based on set minutes generation rules, generates minutes corresponding to the target text based on the information at three levels of the target text). As long as the module includes a text understanding part that can understand the target text to obtain information at three levels of the target text, and includes a part that can generate minutes corresponding to the target text based on the information at three levels of the target text, it belongs to the scope of protection of this application, that is, this application does not make specific limitations on the specific implementation form of obtaining the information at three levels of the target text and generating minutes corresponding to the target text based on the information at three levels of the target text.

[0087] The minutes generation method provided in the embodiment of the present application, after obtaining the target text of the minutes to be generated, first obtains the word-level information, sentence-level information and topic-level information of the target text, and then generates the minutes corresponding to the target text according to the word-level information, sentence-level information and topic-level information of the target text. The minutes generation method provided in the embodiment of the present application can obtain information at three different levels of the target text, and then generate the minutes corresponding to the target text based on the information at three different levels of the target text, that is, the minutes generation method provided in the embodiment of the present application can automatically generate the minutes corresponding to the target text, compared with the method of manually obtaining the minutes, the time consumption is greatly reduced, the efficiency is greatly improved, and because no human participation is required, the influence of subjective factors is avoided.

[0088] Second embodiment

[0089] The above-mentioned embodiments mentioned that the text understanding part of the minutes generation module can be used to obtain the word-level information, sentence-level information and topic-level information of the target text, and the minutes generation part of the minutes generation module can be used to generate minutes corresponding to the target text based on the word-level information, sentence-level information and topic-level information of the target text. The above-mentioned embodiments mentioned that the minutes generation module can be a minutes generation model. This embodiment takes the minutes generation module as a minutes generation model as an example to introduce and explain the specific implementation process of generating minutes corresponding to the target text.

[0090] First, the process of obtaining word-level information and sentence-level information of the target text by using the text understanding part of the minutes generation model is introduced.

[0091] There are many ways to use the text understanding part of the minutes generation model to obtain word-level information and sentence-level information of the target text. Please refer to Figure 3 , which shows a flow chart of one implementation method, may include:

[0092] Step S301: Utilize the text understanding part of the minutes generation model to obtain the word features corresponding to each word in each sentence contained in the target text and integrated with the context information, and based on the word features corresponding to each word in each sentence contained in the target text and integrated with the context information, predict the probability that each word in each sentence contained in the target text is a keyword.

[0093] Among them, the process of obtaining the word features corresponding to each word in each sentence contained in the target text and integrating the context information by using the text understanding part of the minutes generation model includes:

[0094] Step a1: Segment the target text to obtain each sentence in the target text.

[0095] There are many ways to implement sentence processing on the target text. In one possible implementation, the target text can be sentence-separated according to punctuation marks. In another possible implementation, a sliding window of a set length can be used to sentence the target text. This embodiment does not limit the implementation method of sentence processing, that is, the minutes generation method provided in this application has no specific requirements for the sentence segmentation method.

[0096] Step a2: perform word segmentation on each sentence obtained through the sentence segmentation process to obtain a word sequence corresponding to each sentence in the target text.

[0097] Each sentence obtained through the sentence segmentation process can be processed by using an existing word segmentation method. For example, a sentence obtained through the sentence segmentation process is "The next key task is a follow-up of the Spring Fair." After word segmentation, "The next / key / task / is / a / follow-up / of / the / Spring Fair" is obtained.

[0098] Step a3: Obtain the representation vector of each word in the word sequence corresponding to each sentence in the target text to obtain a word representation vector sequence corresponding to each sentence in the target text.

[0099] There are many ways to obtain a representation vector of a word: in one possible implementation, a vector representing the word itself (which may be obtained by but is not limited to the word2vec method) may be obtained as the representation vector of the word; in another possible implementation, a vector representing the word itself and a vector representing the speaker corresponding to the word may be obtained, and the vector representing the word itself and the vector representing the speaker corresponding to the word are concatenated, and the concatenated vector is used as the representation vector of the word. For example, the target text includes the speeches of five speakers A, B, C, D, and E, and a sentence in the target text corresponds to the word sequence "next / of / The key point / work / is / a / follow-up / of / the / Spring / Trade Fair. ", assuming that the speaker corresponding to the sentence is speaker (the speaker corresponding to each word in the word sequence corresponding to the sentence is the speaker corresponding to the sentence) is speaker D, then for each word in the word sequence corresponding to the sentence "The next / key / work / is / a / follow-up / of / the / Spring / Trade Fair.", the vector representing the word itself and the vector representing speaker D can be concatenated, and the concatenated vector is used as the representation vector of the word. For example, for the word "key point", the vector representing the word "key point" itself and the vector representing speaker D are concatenated, and the concatenated vector is used as the representation vector of the word "key point".

[0100] Step a3: Using the text understanding part of the minutes generation model, based on the word representation vector sequence corresponding to each sentence in the target text, determine the word features corresponding to each word in each sentence contained in the target text that are integrated with context information.

[0101] Specifically, the word representation vector sequence corresponding to each sentence in the target text is input into the text understanding part of the minutes generation model. The text understanding part of the minutes generation model determines the word features integrated with context information corresponding to each word in each sentence contained in the target text based on the word representation vector sequence corresponding to each sentence in the target text.

[0102] After obtaining the word features that are integrated with context information corresponding to each word in each sentence contained in the target text, the text understanding part of the minutes generation model can be used to predict the probability of each word in each sentence contained in the target text being a keyword based on the word features that are integrated with context information corresponding to each word in each sentence contained in the target text.

[0103] Optionally, after predicting the probability that each word in each sentence included in the target text is a keyword, the keywords included in the target text may be determined according to the predicted probability.

[0104] In this implementation, the word features corresponding to each word in each sentence of the target text and fused with context information, as well as the probability that each word in each sentence of the target text is a keyword, are used as word-level information of the target text.

[0105] Step S302: Utilize the text understanding part of the minutes generation model to obtain sentence features corresponding to each sentence in the target text that incorporate context information, and based on the sentence features corresponding to each sentence in the target text that incorporate context information, predict the probability that each sentence in the target text is a key sentence.

[0106] Among them, the process of obtaining the sentence features corresponding to each sentence in the target text and integrating the context information by using the text understanding part of the minutes generation model may include:

[0107] Step b1, using the text understanding part of the minutes generation model, based on the word features corresponding to each word in each sentence contained in the target text and integrated with context information, obtain the representation vector of each sentence contained in the target text.

[0108] Step b2: using the text understanding part of the minutes generation model, based on the representation vector of each sentence contained in the target text, determine the sentence features corresponding to each sentence in the target text that incorporate context information.

[0109] After obtaining the sentence features that integrate contextual information corresponding to each sentence in the target text, the text understanding part of the minutes generation model can be used to predict the probability of each sentence in the target text being a key sentence based on the sentence features that integrate contextual information corresponding to each sentence in the target text.

[0110] Optionally, after predicting the probability that each sentence in the target text is a key sentence, the key sentences included in the target text may be determined according to the predicted probability.

[0111] In this implementation, the sentence features corresponding to each sentence in the target text and integrated with the context information, as well as the probability that each sentence in the target text is a key sentence, are used as the sentence-level information of the target text.

[0112] Considering that the above-mentioned keyword prediction and key sentence prediction are performed independently and are not linked to each other, in order to improve the subsequent minutes generation effect, this embodiment provides another preferred implementation method of using the text understanding part of the minutes generation model to obtain word-level information and sentence-level information of the target text, please refer to Figure 4 , showing a schematic diagram of the process of the implementation, which may include:

[0113] Step S401: Utilize the text understanding part of the minutes generation model to obtain the word features corresponding to each word in each sentence contained in the target text and integrated with the context information, and based on the word features corresponding to each word in each sentence contained in the target text and integrated with the context information, predict the probability that each word in each sentence contained in the target text is a keyword.

[0114] Step S402: Utilize the text understanding part of the minutes generation model to obtain the word features corresponding to each word in each sentence contained in the target text and integrated with the context information, and based on the word features corresponding to each word in each sentence contained in the target text and integrated with the context information, predict the probability that each word in each sentence contained in the target text is a keyword.

[0115] The specific implementation process of step S401 and step S402 can refer to the specific implementation process of step S301 and step S302, which will not be described in detail in this embodiment.

[0116] Step S403: For each sentence in the target text, based on the probability that each word in the sentence is a keyword and the probability that the sentence is a key sentence, after determining the probability that the sentence is a key sentence, the probability that each word in the sentence is a keyword is used as the keyword probability after the combined key sentence probability corresponding to each word in the sentence.

[0117] Take the i-th sentence S in the target text i The jth word C in ij For example: The word C can be calculated by the following calculation method: ij The corresponding keyword probability after the joint key sentence probability

[0118]

[0119] in, Based on the word C ij The corresponding word C predicted by the word feature fused with context information ij is the probability of the keyword, According to the sentence S i The corresponding sentence S is predicted by the sentence features that integrate the context information i is the probability of the key sentence, m i It is the sentence S i The physical meaning of the above formula is that the joint word C ij Sentence S i The probability of being a key sentence Word C ij The probability of being a keyword is given by Convert to

[0120] Step S404: For each sentence in the target text, based on the probability that each word in the sentence is a keyword and the probability that the sentence is a key sentence, determine the probability that the sentence is a key sentence after combining the probability that each word in the sentence is a keyword, as the key sentence probability after the combined keyword probability corresponding to the sentence.

[0121] Specifically, for each sentence in the target text, we first calculate the average probability of each word in the sentence being a keyword to obtain the average keyword probability corresponding to the sentence. Based on the average keyword probability corresponding to the sentence and the probability that the sentence is a key sentence, we determine the key sentence probability after the combined keyword probability corresponding to the sentence.

[0122] Take the i-th sentence S in the target text i For example: the sentence S can be calculated as shown in the following formula i The corresponding key sentence probability after the joint keyword probability

[0123]

[0124] in, Represents sentence S i The corresponding keyword probability mean, That is, according to the sentence S i The corresponding sentence S is predicted by the sentence features that integrate the context information i is the probability of the key sentence. The physical meaning of the above formula is that the joint sentence S i The corresponding keyword probability mean, sentence S i The probability of the key sentence is given by Convert to n represents the number of sentences contained in the target text.

[0125] Among them, sentence S i is the corresponding keyword probability mean It can be expressed as:

[0126]

[0127] Among them, m i It is the sentence S i The number of words contained.

[0128] In this implementation, the word features corresponding to each word in each sentence contained in the target text, which are integrated with the context information, and the keyword probability after the joint key sentence probability corresponding to each word in each sentence are used as the word-level information of the target text, and the sentence features corresponding to each sentence in the target text, which are integrated with the context information, and the key sentence probability after the joint keyword probability corresponding to each sentence are used as the sentence-level information of the target text.

[0129] Next, we will introduce the process of obtaining topic-level information of the target text using the text understanding part of the minutes generation model.

[0130] The process of obtaining the topic-level information of the target text by using the text understanding part of the minutes generation model may include: using the text understanding part of the minutes generation model to perform topic segmentation on the target text based on the sentence-level information of the target text, and obtaining the topic segmentation result as the topic-level information of the target text.

[0131] Since the sentence-level information of the target text preferably includes the sentence features corresponding to each sentence in the target text that are integrated with the context information and the key sentence probability after the joint keyword probability corresponding to each sentence, the following example gives the text understanding part of the minutes generation model, and the implementation process of topic segmentation of the target text based on the sentence-level information of the target text, such as Figure 5 As shown, it may include:

[0132] Step S501: For each sentence in the target text, the text understanding part of the minutes generation model is used to perform:

[0133] Step S5011: according to the sentence features corresponding to the sentence and the key sentence probability after the joint keyword probability corresponding to the sentence, the sentence features after the joint keyword probability corresponding to the sentence are determined.

[0134] Take the i-th sentence S in the target text i For example, the sentence S can be determined by the following formula i Corresponding sentence features after joint keyword probability

[0135]

[0136] in, Represents sentence S i The corresponding key sentence probability after the joint keyword probability, Represents sentence S i The corresponding sentence features are integrated with contextual information.

[0137] Step S5012: Based on the sentence features after the joint keyword probability corresponding to the sentence, predict the probability that the sentence is a topic ending sentence.

[0138] Through step S501, the probability of each sentence in the target text being a topic ending sentence can be obtained.

[0139] Step S502: Determine the topic segmentation result according to the probability that each sentence in the target text is a topic ending sentence.

[0140] See also Figure 6, shows a structural diagram of the text understanding part of the minutes generation model, which may include: a word feature acquisition module 6011, a keyword prediction module 6012, a sentence representation acquisition module 6021, a sentence feature acquisition module 6022, a key sentence prediction module 6023, a feature extraction module 603 combining keyword distribution and key sentence distribution, and a topic segmentation module 604.

[0141] Optionally, the word feature acquisition module 6011 can adopt the encoder part of Transformer, or adopt a bidirectional LSTM, the keyword prediction module 6012 can adopt a linear layer connected to a non-linear activation function layer, the sentence representation acquisition module 6021 can adopt an attention layer or a pooling layer, the sentence feature acquisition module 6022 can adopt the encoder part of Transformer, or adopt a bidirectional LSTM, the key sentence prediction module 6023 can adopt a linear layer connected to a non-linear activation function layer, and the topic segmentation module 604 can adopt a linear layer connected to a non-linear activation function layer.

[0142] Combine the following Figure 6 , the process of obtaining word-level information, sentence-level information, and topic-level information of the target text using the text understanding part of the minutes generation model is further introduced:

[0143] First, the word representation vector sequence corresponding to each sentence contained in the target text is obtained, and then the word representation vector sequence corresponding to each sentence contained in the target text is input into the word feature acquisition module 6011. The word feature acquisition module 6011 obtains the word feature corresponding to each word in each sentence with context information according to the word representation vector sequence corresponding to each sentence. For example, the word representation vector sequence W corresponding to the first sentence of the target text is 11 ,W 12 …,W 1m1 (W 1j is the representation vector of the jth word in the first sentence) is input into the word feature acquisition module 6011, and the word feature acquisition module 6011 outputs ( is the word feature with context information corresponding to the jth word in the first sentence), after obtaining the word feature with context information corresponding to each word in each sentence, the word feature with context information corresponding to each word in each sentence is input into the keyword prediction module 6012, and the keyword prediction module 6012 predicts the probability of each word in each sentence being a keyword based on the word feature with context information corresponding to each word in each sentence. For example, the word feature with context information corresponding to each word in the first sentence is input into the keyword prediction module 6012. Input keyword prediction module 6012, keyword prediction module 6012 outputs the probability that each word in the first sentence is a keyword

[0144] After obtaining the word features with context information corresponding to each word in each sentence, they can be input into the sentence representation acquisition module 6021. The sentence representation acquisition module 6021 obtains the representation vector of each sentence according to the word features with context information corresponding to each word in each sentence. For example, the word features with context information corresponding to each word in the first sentence are Input the sentence representation acquisition module 6021, the sentence representation acquisition module 6021 outputs the representation vector s1 of the first sentence, and the other sentences are similar. After obtaining the representation vectors s1, s2…, s n After that, the representation vectors s1, s2…, s n Input sentence feature acquisition module 6022, sentence feature acquisition module 6022 according to the representation vector s1, s2 ..., s n Get the sentence features corresponding to each sentence and fused with context information The sentence features corresponding to each sentence are integrated with context information Input the key sentence prediction module 6023, and the key sentence prediction module 6023 predicts the probability of each sentence being a key sentence according to the input And output.

[0145] For each sentence in the target text, after obtaining the probability that each word in the sentence is a keyword and the probability that the sentence is a key sentence, the probability that each word in the sentence is a keyword and the probability that the sentence is a key sentence are input into the feature extraction module 603 of the combined keyword distribution and key sentence distribution. The feature extraction module 603 of the combined keyword distribution and key sentence distribution calculates the keyword probability after the combined key sentence probability corresponding to each word in the sentence according to the calculation method shown in the above formula (1), and calculates the key sentence probability after the combined keyword probability corresponding to the sentence according to the calculation method shown in the above formula (2).

[0146] For each sentence in the target text, after obtaining the sentence features corresponding to the sentence that are integrated with the context information and the key sentence probability after the joint keyword probability corresponding to the sentence, the sentence features corresponding to the sentence that are integrated with the context information and the key sentence probability after the joint keyword probability corresponding to the sentence are input into the topic segmentation module 604. The topic segmentation module 604 first determines the sentence features after the joint keyword probability corresponding to the sentence based on the input, and then predicts the probability that the sentence is a topic-ending sentence based on the sentence features after the joint keyword probability corresponding to the sentence, and finally determines whether the sentence is a topic-ending sentence based on the probability that the sentence is a topic-ending sentence.

[0147] Through the above process, the word-level information, sentence-level information and topic-level information of the target text can be obtained.

[0148] It should be noted that this embodiment does not limit the structure of the text understanding part of the minutes generation model to Figure 6 The structure shown, Figure 6 The structure shown is only an example. Figure 3 to Figure 5 The structures of the three levels of information of the target text obtained by the process shown all fall within the scope of protection of this application.

[0149] The above content introduces the process of using the text understanding part of the pre-established minutes generation model to obtain the word-level information, sentence-level information, and topic-level information of the target text. Next, the process of using the minutes generation part of the minutes generation model to generate the minutes corresponding to the target text based on the word-level information, sentence-level information, and topic-level information of the target text is introduced.

[0150] In this embodiment, the minutes generation part of the minutes generation model is a decoding module. Optionally, the structure of the decoding module can adopt the decoder structure in the Transformer, or the LSTM structure. When generating the minutes corresponding to the target text, the decoding module is used to predict the words used to compose the minutes one by one based on the word-level information, sentence-level information and topic-level information of the target text.

[0151] See also Figure 7 , showing the process of using the decoding module of the minutes generation model to predict the words used to form the minutes one by one based on the word-level information, sentence-level information and topic-level information of the target text, which may include:

[0152] Step S701: At each decoding moment, the context vectors corresponding to all sentences of the target text at the current decoding moment are determined based on the state vector at the current decoding moment, the word-level information and the sentence-level information of the target text.

[0153] The state vector at the current decoding moment is a state vector representing the decoded result information, which is obtained by performing information fusion encoding on the historical decoding information.

[0154] Specifically, the process of determining the context vectors corresponding to all sentences of the target text at the current decoding time according to the state vector at the current decoding time, the word-level information and the sentence-level information of the target text may include:

[0155] Step S7011: Determine the context vectors corresponding to all words in each sentence of the target text at the current decoding time according to the state vector at the current decoding time and the word-level information of the target text.

[0156] This embodiment takes the word-level information of the target text including the word features corresponding to each word in each sentence contained in the target text and the keyword probability after the combined key sentence probability corresponding to each word in each sentence as an example to provide the specific implementation process of step S7011:

[0157] For each sentence in the target text, execute:

[0158] Step c1, determine the weights corresponding to each word in the sentence according to the state vector at the current decoding moment, the word features corresponding to each word in the sentence that are integrated with context information, and the keyword probability after the combined key sentence probability corresponding to each word in the sentence.

[0159] Take the i-th sentence S in the target text i For example: Assume that the current decoding time is the tth time, and the i-th sentence S in the target text i The jth word C in ij The corresponding weights can be determined as shown below:

[0160]

[0161] Among them, d t represents the state vector at time t, Represents the i-th sentence S i The jth word C in ij The corresponding word features are integrated with context information. Attention() represents the attention mechanism calculation function, which can be but not limited to self-attention and additive attention. The specific implementation process is consistent with the prior art and will not be described in detail in this embodiment. Representation word C ij Attention coefficient corresponding to the tth moment, Representation word C ij The corresponding keyword probability after the joint key sentence probability, Representation word C ij The corresponding weight, Through Optimize and improve.

[0162] It should be noted that this embodiment does not limit As word C ij The corresponding weights, for example, can also be As word C ij However, considering the influence of keywords and key sentences on the generation of minutes, in order to obtain better minutes generation effect later, this embodiment preferably As word C ij The corresponding weight.

[0163] Step c2: Determine the context vectors corresponding to all the words in the sentence at the current decoding moment according to the weights corresponding to the words in the sentence and the word features corresponding to the words in the sentence that are integrated with the context information.

[0164] After determining the weights corresponding to each word in the sentence, the weights corresponding to each word in the sentence and the word features corresponding to the corresponding words that are integrated with context information can be weighted summed up, and the vector obtained after the weighted summation is used as the context vector corresponding to all the words in the sentence at the current decoding moment.

[0165] Take the i-th sentence S in the target text i For example: Assume that the current decoding time is time t, sentence S i The context vectors of all words at the tth moment It can be expressed as:

[0166]

[0167] in, Represents sentence S i The jth word C in ij The corresponding word features fused with context information, m i Represents sentence S i The number of words contained.

[0168] It should be noted that if the word-level information of the target text includes the word features corresponding to each word in each sentence of the target text and the probability of each word in each sentence being a keyword, then the above Replace with (Word C ij The probability of being a keyword) is sufficient.

[0169] Step S7012: Determine the context vectors corresponding to all sentences of the target text at the current decoding time based on the state vector at the current decoding time, the context vectors corresponding to all words in each sentence contained in the target text at the previous decoding time, and the sentence-level information of the target text.

[0170] Taking the sentence-level information of the target text including the sentence features corresponding to each sentence in the target text and the key sentence probability after the joint keyword probability corresponding to each sentence as an example, the specific implementation process of step S7012 is given:

[0171] Step d1: For each sentence contained in the target text, determine the weight corresponding to the sentence based on the state vector at the current decoding moment, the context vectors corresponding to all the words in the sentence at the current decoding moment, the sentence features corresponding to the sentence that are integrated with context information, and the key sentence probability after the joint keyword probability corresponding to the sentence.

[0172] Take the i-th sentence S in the target text i For example: Assume that the current decoding time is the tth time, and the i-th sentence S in the target text i The corresponding weights can be determined as shown below:

[0173]

[0174] Among them, d t represents the state vector at time t, Represents the i-th sentence S i The corresponding sentence features are integrated with context information. Represents sentence S i The context vectors corresponding to all words at the tth moment, Attention() represents the attention mechanism calculation function, which can be but not limited to self-attention and additive attention, etc. The specific implementation process is consistent with the prior art, and this embodiment will not be repeated here. Represents sentence S i The corresponding attention coefficient is Represents sentence S i The corresponding key sentence probability after the joint keyword probability, Represents sentence S i The corresponding weight is Optimize and improve.

[0175] It should be noted that this embodiment does not limit As a sentence S i The corresponding weights, for example, can also be As a sentence S i However, considering the influence of keywords and key sentences on the generation of minutes, in order to obtain better minutes generation effect later, this embodiment preferably As a sentence S i The corresponding weight.

[0176] Step d2: determine the context vectors corresponding to all sentences of the target text at the current decoding time based on the weight corresponding to each sentence in the target text, the sentence features corresponding to each sentence that incorporate context information, and the context vectors corresponding to all words of each sentence at the current decoding time.

[0177] After determining the weights corresponding to each sentence in the target text, for each sentence, the sentence features corresponding to the sentence that incorporate context information can be concatenated with the context vectors corresponding to all the words in the sentence at the current decoding moment to obtain the concatenated vector corresponding to each sentence. After obtaining the concatenated vector corresponding to each sentence, the concatenated vectors corresponding to the corresponding sentences can be weighted and summed according to the weights corresponding to each sentence. The weighted summed vector is used as the context vector corresponding to all the sentences in the target text at the current decoding moment.

[0178] Assuming that the current decoding time is the tth time, the context vectors corresponding to all sentences of the target text at the current decoding time are It can be expressed as:

[0179]

[0180] in, Indicates the sentence features corresponding to the pth sentence in the target text that incorporate contextual information, Represents the context vector corresponding to all words in the p-th sentence in the target text at the t-th time. Indicates that and Splicing, Represents the weight corresponding to the p-th sentence in the target text at the t-th time.

[0181] It should be noted that if the sentence-level information of the target text includes the sentence features corresponding to each sentence in the target text that incorporates context information and the probability that each sentence is a key sentence, then the above Replace with (Sentence S i The probability of being the key sentence) is sufficient.

[0182] Step S702: Determine the context vectors corresponding to all topics involved in the target text at the current decoding time according to the state vector at the current decoding time and the topic-level information of the target text.

[0183] Specifically, the implementation process of step S702 may include:

[0184] Step S7021: Determine the sentences included in each topic involved in the target text according to the topic-level information of the target text.

[0185] As mentioned above, the topic-level information of the target text may be a topic segmentation result of the target text. According to the topic segmentation result of the target text, the sentences included in each topic involved in the target text may be determined.

[0186] Step S7022: For each topic, determine the representation vector of the topic at the current decoding time based on the sentence features corresponding to each sentence included in the topic and the context vectors corresponding to all the words in each sentence included in the topic at the current decoding time.

[0187] Take the zth topic among the Z topics involved in the target text as an example: Assuming that the current decoding time is the tth time, the representation vector of the zth topic at the tth time can be determined as shown in the following formula:

[0188]

[0189] in, represents the range of sentence numbers belonging to the zth topic, Indicates the sentence features corresponding to the qth sentence in the sentences contained in the zth topic, which are integrated with context information. Represents the context vector corresponding to all words in the qth sentence of the sentence contained in the zth topic at the tth time, Represents the weight corresponding to the qth sentence in the sentences contained in the zth topic.

[0190] Step S7023: Determine the context vectors corresponding to all topics involved in the target text at the current decoding time according to the state vector at the current decoding time and the representation vector of each topic involved in the target text at the current decoding time.

[0191] Specifically, first, according to the state vector at the current decoding moment and the representation vector of each topic involved in the target text at the current decoding moment, the weight corresponding to each topic involved in the target text at the current decoding moment is determined, and then the representation vectors of each topic involved in the target text at the current decoding moment are weightedly summed according to the weights corresponding to the topics involved in the target text at the current decoding moment. The weighted summed vector serves as the context vector corresponding to all topics involved in the target text at the current decoding moment.

[0192] Take the zth topic among the Z topics involved in the target text as an example: Assuming that the current decoding time is the tth time, the weight corresponding to the zth topic at the tth time can be determined as shown in the following formula:

[0193]

[0194] Among them, d t represents the state vector at time t, That is, the representation vector of the zth topic at the tth moment.

[0195] After obtaining the weights corresponding to each topic involved in the target text at the current decoding time, the context vector corresponding to all topics of the target text at the current decoding time can be determined as follows:

[0196]

[0197] in, represents the weight corresponding to the zth topic at the tth moment, represents the representation vector of the z-th topic at the t-th time, and Z represents the total number of topics involved in the target text.

[0198] Step S703: Determine the minutes generation result at the current decoding moment according to the state vector at the current decoding moment, the context vectors corresponding to all sentences of the target text at the current decoding moment, and the context vectors corresponding to all topics involved in the target text at the current decoding moment.

[0199] Assuming that the current decoding time is the tth time, the minutes generation result y at the tth time can be determined as follows: t :

[0200]

[0201] Among them, class() represents the classification layer calculation. The classification layer can adopt a linear layer connected to a nonlinear activation function layer. The specific calculation process of the classification layer is the same as that of the prior art, and this embodiment will not be described in detail here.

[0202] Third embodiment

[0203] The above embodiment provides a specific process of using a pre-established minutes generation model to obtain word-level information, sentence-level information and topic-level information of a target text, and generating minutes corresponding to the target text based on the word-level information, sentence-level information and topic-level information of the target text. This embodiment focuses on introducing the process of establishing a minutes generation model.

[0204] The above embodiment mentioned that the minutes generation model can be obtained by training with training texts annotated with keywords, key sentences, topic ending sentences, and minutes corresponding to the training texts. Next, the process of training the minutes generation model by using training texts annotated with keywords, key sentences, topic ending sentences, and minutes corresponding to the training texts is given, such as Figure 8 As shown, it may include:

[0205] Step S801: Acquire training text from a training data set.

[0206] The training data set includes multiple training texts, each training text is annotated with keywords, key sentences and topic ending sentences, and each training text has a corresponding summary.

[0207] It should be noted that the process of marking keywords, key sentences and topic-ending sentences for the training text is the process of constructing a keyword label corresponding to each word in each sentence contained in the training text, a key sentence label corresponding to each sentence, and a topic-ending sentence label corresponding to each sentence. When constructing a keyword label corresponding to a word, if the word is a keyword, the keyword label corresponding to the word is "1", if the word is a non-keyword, the keyword label corresponding to the word is "0", when constructing a key sentence label corresponding to a sentence, if the sentence is a key sentence, the key sentence label corresponding to the sentence is "1", if the sentence is a non-key sentence, the key sentence label corresponding to the sentence is "0", when constructing a topic-ending sentence label corresponding to a sentence, if the sentence is a topic-ending sentence, the topic-ending sentence label corresponding to the sentence is "1", if the sentence is a non-topic-ending sentence, the topic-ending sentence label corresponding to the sentence is "0". For example, a training text contains 5 sentences after sentence segmentation. The 1st and 2nd sentences belong to topic 1, and the 3rd to 5th sentences belong to topic 2. For the 5 sentences contained in the training text, the topic-ending sentence label is "0 1 0 0 1" (the 2nd and 5th sentences are topic-ending sentences), and the key sentence label is "0 0 0 1 0" (the 4th sentence is the key sentence).

[0208] Step S802: Utilize the text understanding part of the minutes generation model to predict the probability that each word in each sentence contained in the training text is a keyword, the probability that each sentence contained in the training text is a key sentence, and the probability that each sentence contained in the training text is a topic-ending sentence, so as to obtain keyword prediction results, key sentence prediction results, and topic-ending sentence prediction results.

[0209] The specific implementation process of using the text understanding part of the minutes generation model to predict the probability of each word in each sentence contained in the training text being a keyword, the probability of each sentence contained in the training text being a key sentence, and the probability of each sentence contained in the training text being a topic-ending sentence is similar to the specific implementation process of using the text understanding part of the minutes generation model in the second embodiment to predict the probability of each word in each sentence contained in the target text being a keyword, the probability of each sentence contained in the target text being a key sentence, and the probability of each sentence contained in the target text being a topic-ending sentence. For details, please refer to the relevant parts of the second embodiment, and this embodiment will not be repeated here.

[0210] Step S803: using the minutes generation part of the minutes generation model, based on the word-level information, sentence-level information and topic-level information of the training text, predicting the minutes corresponding to the training text, and obtaining the minutes prediction results.

[0211] Among them, the word-level information, sentence-level information and topic-level information of the training text are obtained using the text understanding part of the minutes generation model.

[0212] The implementation method of using the text understanding part of the minutes generation model to obtain the word-level information, sentence-level information and topic-level information of the training text is similar to the implementation method of using the text understanding part of the minutes generation model to obtain the word-level information, sentence-level information and topic-level information of the target text in the second embodiment. For details, please refer to the relevant parts of the second embodiment, and this embodiment will not be repeated here.

[0213] It should be noted that the word-level information and sentence-level information obtained in the training phase and the actual application phase of the minutes generation model must be consistent. For example, the word-level information obtained in the training phase includes the word features corresponding to each word in each sentence contained in the training text, which are integrated with context information, and the keyword probability after the joint key sentence corresponding to each word. The word information in the actual application phase also includes the word features corresponding to each word in each sentence contained in the target text, which are integrated with context information, and the keyword probability after the joint key sentence corresponding to each word. The sentence-level information is similar.

[0214] Step S804: Determine the prediction loss based on the keyword prediction results, key sentence prediction results, topic end sentence prediction results, minutes prediction results, and the keywords, key sentences, topic end sentences annotated in the training text and the minutes corresponding to the training text, and update the parameters of the minutes generation model based on the prediction loss.

[0215] Specifically, a first prediction loss (such as cross entropy loss) can be determined based on the keyword prediction results, the key sentence prediction results, and the keywords and key sentences annotated in the training text, a second prediction loss (such as cross entropy loss) can be determined based on the topic end sentence prediction results and the topic end sentences annotated in the training text, a third prediction loss (such as cross entropy loss) can be determined based on the minutes prediction results and the minutes corresponding to the training text, and the parameters of the minutes generation model are updated based on the first prediction loss, the second prediction loss, and the third prediction loss. It should be noted that the first prediction loss is used to update the parameters of the keyword and key sentence prediction part in the text understanding part of the minutes generation model, the second prediction loss is used to update the parameters of the topic end sentence prediction part in the text understanding part of the minutes generation model, and the third prediction loss is used to update the parameters of the minutes generation part of the model to be generated. Optionally, when updating the parameters of the model, the update can be based on the SGD algorithm.

[0216] The minutes generation model is iteratively trained multiple times in the above manner until the training end condition is met, and the model obtained after the training is completed is the established minutes generation model.

[0217] The minutes generation method provided in the present application can automatically generate minutes corresponding to the text without the need for manual labor, thus avoiding the problems existing in manually obtaining minutes; in addition to being able to generate minutes corresponding to the text, the minutes generation method provided in the present application can also extract keywords and key sentences from the text and achieve topic segmentation; when training the minutes generation model, the potential connection between tasks is fully considered, and they complement each other and learn together, and finally a minutes generation model with better performance is trained. The minutes generation model with better performance can generate minutes of better quality for the target text of the minutes to be generated.

[0218] Fourth embodiment

[0219] The embodiment of the present application also provides a minutes generating device. The minutes generating device provided in the embodiment of the present application is described below. The minutes generating device described below and the minutes generating method described above can be referenced to each other.

[0220] See also Fig. 9 , shows a schematic diagram of the structure of the minutes generating device provided in an embodiment of the present application, which may include: a text acquisition module 901 and a text processing module 902.

[0221] The text acquisition module 901 is used to acquire the text of the minutes to be generated as the target text;

[0222] The text processing module 902 is used to obtain word-level information, sentence-level information and topic-level information of the target text, and generate minutes corresponding to the target text according to the word-level information, sentence-level information and topic-level information of the target text.

[0223] Among them, the word-level information includes the context information of the words in the target text and the keyword information of the target text, the sentence-level information includes the context information of the sentences in the target text and the key sentence information of the target text, and the topic-level information includes the topic segmentation information of the target text.

[0224] Optionally, the text processing module 902 is specifically used for:

[0225] The minutes generation module is used to obtain word-level information, sentence-level information and topic-level information of the target text; the minutes generation model is used to generate minutes corresponding to the target text based on the word-level information, sentence-level information and topic-level information of the target text.

[0226] Optionally, the minutes generation module includes a text understanding part and a minutes generation part, wherein the text understanding part is used to obtain word-level information, sentence-level information and topic-level information of the target text, and the minutes generation part is used to generate minutes corresponding to the target text based on the word-level information, sentence-level information and topic-level information of the target text.

[0227] Optionally, the text understanding part of the minutes generation module is specifically used to obtain word-level information and sentence-level information of the target text:

[0228] Obtain word features that are integrated with context information and correspond to each word in each sentence contained in the target text, and use the obtained word features as a basis to predict the probability that each word in each sentence contained in the target text is a keyword; obtain sentence features that are integrated with context information and correspond to each sentence in the target text, and use the obtained sentence features as a basis to predict the probability that each sentence contained in the target text is a key sentence; the word features corresponding to each word in each sentence contained in the target text and the probability that each word in each sentence is a keyword are used as word-level information of the target text, and the sentence features corresponding to each sentence in the target text and the probability that each sentence is a key sentence are used as sentence-level information of the target text.

[0229] Optionally, the text understanding part of the minutes generation module is also used to obtain word-level information and sentence-level information of the target text:

[0230] For each sentence in the target text, after determining the probability of each word in the sentence being a keyword and the probability of the sentence being a key sentence, the probability of each word in the sentence being a keyword is used as the keyword probability after the combined key sentence probability corresponding to each word in the sentence according to the probability that each word in the sentence is a keyword and the probability that the sentence is a key sentence; after determining the probability of each word in the sentence being a keyword and the probability that the sentence is a key sentence, the probability of the sentence being a key sentence is used as the key sentence probability after the combined key sentence probability corresponding to the sentence; the word features corresponding to each word in each sentence contained in the target text and the keyword probability after the combined key sentence probability corresponding to each word in each sentence are used as the word-level information of the target text, and the sentence features corresponding to each sentence in the target text and the key sentence probability after the combined key sentence probability corresponding to each sentence are used as the sentence-level information of the target text.

[0231] Optionally, the text understanding part of the minutes generation module is specifically used to obtain the topic-level information of the target text:

[0232] The target text is topic segmented according to the sentence-level information of the target text to obtain a topic segmentation result as the topic-level information of the target text.

[0233] Optionally, when the text understanding part of the minutes generation module performs topic segmentation on the target text according to the sentence-level information of the target text to obtain the topic segmentation result, it is specifically used to:

[0234] For each sentence in the target text, the sentence features corresponding to the sentence and the key sentence probability after the joint keyword probability corresponding to the sentence are determined, and the sentence features after the joint keyword probability corresponding to the sentence are used as a basis to predict the probability that the sentence is a topic-ending sentence; according to the probability that each sentence in the target text is a topic-ending sentence, the topic segmentation result is determined.

[0235] Optionally, the minutes generating part of the minutes generating module, when generating the minutes corresponding to the target text based on the word-level information, sentence-level information and topic-level information of the target text, is specifically used for:

[0236] At each decoding moment: according to the state vector at the current decoding moment, the word-level information and sentence-level information of the target text, the context vectors corresponding to all the sentences of the target text at the current decoding moment are determined, wherein the state vector at the current decoding moment is a state vector representing the decoded result information; according to the state vector at the current decoding moment and the topic-level information of the target text, the context vectors corresponding to all the topics involved in the target text at the current decoding moment are determined; according to the state vector at the current decoding moment, the context vectors corresponding to all the sentences of the target text at the current decoding moment and the context vectors corresponding to all the topics involved in the target text at the current decoding moment, the minutes generation result at the current decoding moment is determined.

[0237] Optionally, the minutes generation part of the minutes generation module is specifically used to determine the context vectors corresponding to all sentences of the target text at the current decoding time according to the state vector at the current decoding time, the word-level information and the sentence-level information of the target text:

[0238] According to the state vector at the current decoding moment and the word-level information of the target text, the context vector corresponding to all the words in each sentence contained in the target text at the current decoding moment is determined; according to the state vector at the current decoding moment, the context vector corresponding to all the words in each sentence contained in the target text at the current decoding moment and the sentence-level information of the target text, the context vector corresponding to all the sentences of the target text at the current decoding moment is determined.

[0239] Optionally, the sentence-level information of the target text includes sentence features corresponding to each sentence contained in the target text and key sentence probabilities after the joint keyword probabilities corresponding to each sentence. The minutes generation part of the minutes generation module determines the context vectors corresponding to all sentences of the target text at the current decoding time according to the state vector at the current decoding time, the context vectors corresponding to all words of each sentence contained in the target text at the current decoding time, and the sentence-level information of the target text, specifically for:

[0240] For each sentence in the target text, the weight corresponding to the sentence at the current decoding moment is determined according to the state vector at the current decoding moment, the context vector corresponding to all the words of the sentence at the current decoding moment, the sentence features corresponding to the sentence and the key sentence probability after the joint keyword probability corresponding to the sentence; the context vector corresponding to all the sentences of the target text at the current decoding moment is determined according to the weight corresponding to each sentence contained in the target text at the current decoding moment, the sentence features corresponding to each sentence and the context vector corresponding to all the words of each sentence at the current decoding moment.

[0241] Optionally, the minutes generation part of the minutes generation module is specifically used to determine the context vectors corresponding to all topics of the target text at the current decoding time according to the state vector at the current decoding time and the topic-level information of the target text:

[0242] According to the topic-level information of the target text, the sentences included in each topic involved in the target text are determined; for each topic involved in the target text, the representation vector of the topic at the current decoding moment is determined according to the sentence features corresponding to each sentence included in the topic and the context vectors corresponding to all words in each sentence included in the topic at the current decoding moment; according to the state vector at the current decoding moment and the representation vector of each topic involved in the target text at the current decoding moment, the context vectors corresponding to all topics involved in the target text at the current decoding moment are determined.

[0243] Optionally, the minutes generation module is a minutes generation model, and the minutes generation model is trained using training texts annotated with keywords, key sentences, and topic ending sentences and minutes corresponding to the training texts. The minutes generation device provided in this embodiment also includes: a model training module. The model training module is used to:

[0244] The text understanding part of the minutes generation model is used to predict the probability of each word in each sentence contained in the training text being a keyword, the probability of each sentence being a key sentence, and the probability of each sentence being a topic ending sentence, so as to obtain keyword prediction results, key sentence prediction results, and topic ending sentence prediction results; the minutes generation part of the minutes generation model is used to predict the minutes corresponding to the training text based on the word-level information, sentence-level information, and topic-level information of the training text, so as to obtain the minutes prediction results, wherein the word-level information, sentence-level information, and topic-level information of the training text are obtained based on the text understanding part of the minutes generation model; the prediction loss is determined according to the keyword prediction results, the key sentence prediction results, the topic ending sentence prediction results, the minutes prediction results, and the keywords, key sentences, topic ending sentences annotated in the training text, and the minutes corresponding to the training text, and the parameters of the minutes generation model are updated according to the prediction loss.

[0245] The minutes generation device provided in the embodiment of the present application, after obtaining the target text for which the minutes are to be generated, first obtains the word-level information, sentence-level information, and topic-level information of the target text, and then generates the minutes corresponding to the target text according to the word-level information, sentence-level information, and topic-level information of the target text. The minutes generation device provided in the embodiment of the present application can obtain information at three different levels of the target text, and then generate the minutes corresponding to the target text based on the information at three different levels of the target text, that is, the minutes generation device provided in the embodiment of the present application can automatically generate the minutes corresponding to the target text, which greatly reduces the time consumption and greatly improves the efficiency compared to the manual acquisition of minutes, and because no manual participation is required, the influence of subjective factors is avoided.

[0246] Fifth embodiment

[0247] The present application also provides a device for generating minutes. Fig.10 , shows a schematic diagram of the structure of the minutes generating device, the minutes generating device may include: at least one processor 1001, at least one communication interface 1002, at least one memory 1003 and at least one communication bus 1004;

[0248] In the embodiment of the present application, the number of the processor 1001, the communication interface 1002, the memory 1003, and the communication bus 1004 is at least one, and the processor 1001, the communication interface 1002, and the memory 1003 communicate with each other through the communication bus 1004;

[0249] The processor 1001 may be a central processing unit (CPU), or an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of the present invention;

[0250] The memory 1003 may include a high-speed RAM memory, and may also include a non-volatile memory, such as at least one disk memory;

[0251] The memory stores a program, and the processor can call the program stored in the memory, wherein the program is used to:

[0252] Get the text of the minutes to be generated as the target text;

[0253] Acquire word-level information, sentence-level information, and topic-level information of the target text, wherein the word-level information includes context information of words in the target text and keyword information of the target text, the sentence-level information includes context information of sentences in the target text and key sentence information of the target text, and the topic-level information includes topic segmentation information of the target text;

[0254] A summary corresponding to the target text is generated according to the word-level information, sentence-level information and topic-level information of the target text.

[0255] Optionally, the detailed functions and extended functions of the program may refer to the above description.

[0256] Sixth embodiment

[0257] The embodiment of the present application further provides a computer-readable storage medium, which may store a program suitable for execution by a processor, wherein the program is used to:

[0258] Get the text of the minutes to be generated as the target text;

[0259] Acquire word-level information, sentence-level information, and topic-level information of the target text, wherein the word-level information includes context information of words in the target text and keyword information of the target text, the sentence-level information includes context information of sentences in the target text and key sentence information of the target text, and the topic-level information includes topic segmentation information of the target text;

[0260] A summary corresponding to the target text is generated according to the word-level information, sentence-level information and topic-level information of the target text.

[0261] Optionally, the detailed functions and extended functions of the program may refer to the above description.

[0262] Finally, it should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the presence of other identical elements in the process, method, article or device including the elements.

[0263] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0264] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for generating minutes, characterized in that: include: Get the text of the minutes to be generated as the target text; Obtain word-level information, sentence-level information and topic-level information of the target text, wherein the word-level information includes context information of words in the target text and keyword information of the target text, the sentence-level information includes context information of sentences in the target text and key sentence information of the target text, the topic-level information includes topic segmentation information of the target text, the word context information is the word feature corresponding to the word integrated with context information, the keyword information is the probability that the word in the target text is a keyword, the sentence context information is the sentence feature corresponding to the sentence integrated with context information, the key sentence information is the probability that the sentence in the target text is a key sentence, and the topic segmentation information is the topic segmentation result obtained by performing topic segmentation on the target text according to the sentence-level information; Generating a summary corresponding to the target text according to the word-level information, sentence-level information and topic-level information of the target text; The acquiring of word-level information, sentence-level information, and topic-level information of the target text, and the generating of minutes corresponding to the target text according to the word-level information, sentence-level information, and topic-level information of the target text include: Using the text understanding part of the minutes generation module, word-level information, sentence-level information, and topic-level information of the target text are obtained; The minutes generation part of the minutes generation module is used to generate minutes corresponding to the target text based on the word-level information, sentence-level information and topic-level information of the target text.

2. The method for generating minutes according to claim 1, characterized in that: Obtaining word-level information and sentence-level information of the target text includes: Obtaining word features corresponding to each word in each sentence contained in the target text and integrating context information, and predicting the probability of each word in each sentence contained in the target text being a keyword based on the obtained word features; Acquire sentence features corresponding to each sentence in the target text and integrated with context information, and predict the probability that each sentence in the target text is a key sentence based on the acquired sentence features; The word features corresponding to each word in each sentence contained in the target text and the probability that each word in each sentence is a keyword are used as the word-level information of the target text, and the sentence features corresponding to each sentence in the target text and the probability that each sentence is a key sentence are used as the sentence-level information of the target text.

3. The method for generating minutes according to claim 2, characterized in that: Acquiring word-level information and sentence-level information of the target text also includes: For each sentence in the target text: According to the probability that each word in the sentence is a keyword and the probability that the sentence is a key sentence, after determining the probability that the sentence is a key sentence, the probability that each word in the sentence is a keyword is used as the keyword probability after the key sentence probability corresponding to each word in the sentence; According to the probability that each word in the sentence is a keyword and the probability that the sentence is a key sentence, after combining the probability that each word in the sentence is a keyword, the probability that the sentence is a key sentence is determined as the key sentence probability after combining the keyword probabilities corresponding to the sentence; The word features corresponding to each word in each sentence contained in the target text and the keyword probability after the combined key sentence probability corresponding to each word in each sentence serve as the word-level information of the target text; the sentence features corresponding to each sentence in the target text and the key sentence probability after the combined key word probability corresponding to each sentence serve as the sentence-level information of the target text.

4. The method for generating minutes according to claim 3, characterized in that: Performing topic segmentation on the target text according to the sentence-level information of the target text to obtain a topic segmentation result, including: For each sentence in the target text, according to the sentence features corresponding to the sentence and the key sentence probability after the joint keyword probability corresponding to the sentence, determine the sentence features after the joint keyword probability corresponding to the sentence, and based on the sentence features after the joint keyword probability corresponding to the sentence, predict the probability that the sentence is a topic ending sentence; The topic segmentation result is determined according to the probability that each sentence in the target text is a topic ending sentence.

5. The method for generating minutes according to claim 1, characterized in that: The generating of the minutes corresponding to the target text based on the word-level information, sentence-level information and topic-level information of the target text includes: At each decoding moment: Determine the context vectors corresponding to all sentences of the target text at the current decoding time according to the state vector at the current decoding time, the word-level information and the sentence-level information of the target text, wherein the state vector at the current decoding time is a state vector representing the decoded result information; Determine, according to the state vector at the current decoding moment and the topic-level information of the target text, the context vectors corresponding to all topics involved in the target text at the current decoding moment; The minutes generation result at the current decoding moment is determined according to the state vector at the current decoding moment, the context vectors corresponding to all sentences of the target text at the current decoding moment, and the context vectors corresponding to all topics involved in the target text at the current decoding moment.

6. The method for generating minutes according to claim 5, characterized in that: The determining, based on the state vector at the current decoding moment, the word-level information and the sentence-level information of the target text, the context vectors corresponding to all the sentences of the target text at the current decoding moment comprises: Determine, according to the state vector at the current decoding moment and the word-level information of the target text, the context vectors corresponding to all the words in each sentence contained in the target text at the current decoding moment; The context vectors corresponding to all the sentences of the target text at the current decoding time are determined according to the state vector at the current decoding time, the context vectors corresponding to all the words of each sentence contained in the target text at the current decoding time, and the sentence-level information of the target text.

7. The method for generating minutes according to claim 6, characterized in that: The sentence-level information of the target text includes sentence features corresponding to each sentence contained in the target text and key sentence probabilities after the combined keyword probabilities corresponding to each sentence; The determining, based on the state vector at the current decoding moment, the context vectors corresponding to all words in each sentence contained in the target text at the current decoding moment, and the sentence-level information of the target text, the context vectors corresponding to all sentences of the target text at the current decoding moment comprises: For each sentence in the target text, determine the weight corresponding to the sentence at the current decoding time according to the state vector at the current decoding time, the context vectors corresponding to all words of the sentence at the current decoding time, the sentence features corresponding to the sentence, and the key sentence probability after the joint keyword probability corresponding to the sentence; The context vectors corresponding to all the sentences of the target text at the current decoding time are determined according to the weight corresponding to each sentence contained in the target text at the current decoding time, the sentence features corresponding to each sentence, and the context vectors corresponding to all the words of each sentence at the current decoding time.

8. The method for generating minutes according to claim 5, characterized in that: The determining, based on the state vector at the current decoding moment and the topic-level information of the target text, the context vectors corresponding to all topics involved in the target text at the current decoding moment comprises: Determining sentences included in each topic involved in the target text according to the topic-level information of the target text; For each topic involved in the target text, determine the representation vector of the topic at the current decoding time according to the sentence features corresponding to each sentence included in the topic and the context vectors corresponding to all words in each sentence included in the topic at the current decoding time; According to the state vector at the current decoding moment and the representation vector of each topic involved in the target text at the current decoding moment, the context vector corresponding to all topics involved in the target text at the current decoding moment is determined.

9. The method for generating minutes according to any one of claims 1 to 8, characterized in that: The minutes generation module is a minutes generation model, which is obtained by training with training texts marked with keywords, key sentences, and topic ending sentences and minutes corresponding to the training texts. The training process of the minutes generation model includes: Using the text understanding part of the minutes generation model, predict the probability of each word in each sentence contained in the training text being a keyword, the probability of each sentence being a key sentence, and the probability of each sentence being a topic ending sentence, so as to obtain keyword prediction results, key sentence prediction results, and topic ending sentence prediction results; Using the minutes generation part of the minutes generation model, based on the word-level information, sentence-level information and topic-level information of the training text, predicting the minutes corresponding to the training text to obtain the minutes prediction result, wherein the word-level information, sentence-level information and topic-level information of the training text are obtained based on the text understanding part of the minutes generation model; The prediction loss is determined based on the keyword prediction results, the key sentence prediction results, the topic end sentence prediction results, the minutes prediction results, and the keywords, key sentences, topic end sentences annotated in the training text and the minutes corresponding to the training text, and the parameters of the minutes generation model are updated based on the prediction loss.

10. A minutes generating device, characterized in that: include: Text acquisition module and text processing module; The text acquisition module is used to acquire the text of the minutes to be generated as the target text; The text processing module is used to obtain word-level information, sentence-level information and topic-level information of the target text, and generate minutes corresponding to the target text according to the word-level information, sentence-level information and topic-level information of the target text; Among them, the word-level information includes the context information of the words in the target text and the keyword information of the target text, the sentence-level information includes the context information of the sentences in the target text and the key sentence information of the target text, the topic-level information includes the topic segmentation information of the target text, the context information of the words is the word features corresponding to the words integrated with the context information, the keyword information is the probability that the words in the target text are keywords, the context information of the sentences is the sentence features corresponding to the sentences integrated with the context information, the key sentence information is the probability that the sentences in the target text are key sentences, and the topic segmentation information is the topic segmentation result obtained by segmenting the target text according to the sentence-level information; The text processing module is specifically used for: Using the text understanding part of the minutes generation module, word-level information, sentence-level information, and topic-level information of the target text are obtained; The minutes generation part of the minutes generation module is used to generate minutes corresponding to the target text based on the word-level information, sentence-level information and topic-level information of the target text.

11. A minutes generating device, characterized in that: include: Memory and processor; The memory is used to store programs; The processor is used to execute the program to implement each step of the minutes generation method according to any one of claims 1 to 9.

12. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, each step of the minutes generation method according to any one of claims 1 to 9 is implemented.

Images