A method for determining minutes and related equipment

Abstract

The present application discloses a method for determining minutes and related equipment, the method comprising: after obtaining the user record to be used and the record text to be used, the record text to be used can be firstly processed into sentences to obtain at least one sentence to be used; then the user record to be used and the sentences to be used are processed for semantic matching to obtain the semantic matching results to be used; finally, the content of the minutes to be used is determined according to the user record to be used and the semantic matching results to be used. Among them, because the above-mentioned "semantic matching results to be used" can indicate the degree of semantic matching between the user record to be used and at least one sentence in the record text to be used, the user can quickly find the meeting content related to the user record to be used from the record text to be used based on the "semantic matching results to be used", which can effectively improve the efficiency of determining minutes.

Description

Technical Field

[0001] The present application relates to the technical field of natural language processing, and in particular to a minutes determination method and related equipment. Background Art

[0002] In some application scenarios (e.g., meetings, speeches, etc.), some participants (e.g., recorders) usually record some concise information points to summarize the speaker's speech (e.g., a summary of the content, conclusions reached in the meeting, or important matters that need to be followed up, etc.), so that these participants can later write minutes content (e.g., meeting minutes, report summaries, etc.) based on these user records.

[0003] In addition, in order to better write the minutes, these participants usually need to listen to the conference recording audio (or, speech recording audio) again and write the minutes based on these user records.

[0004] However, since these participants need to spend a lot of time to re-listen to the recorded audio of the meeting, the process of determining the content of the minutes takes a long time, resulting in low efficiency in determining the minutes. Summary of the invention

[0005] The main purpose of the embodiments of the present application is to provide a minutes determination method and related equipment, which can improve the efficiency of minutes determination.

[0006] An embodiment of the present application provides a method for determining minutes, the method comprising: obtaining user records to be used and record text to be used; performing sentence division processing on the record text to be used to obtain at least one sentence to be used; performing semantic matching processing on the user record to be used and the at least one sentence to be used to obtain a semantic matching result to be used; and determining the content of the minutes to be used based on the user record to be used and the semantic matching result to be used.

[0007] An embodiment of the present application also provides a minutes determination device, including: an information acquisition unit, used to obtain user records to be used and record texts to be used; a text division unit, used to perform sentence division processing on the record text to be used to obtain at least one sentence to be used; a semantic matching unit, used to perform semantic matching processing on the user record to be used and the at least one sentence to be used to obtain a semantic matching result to be used; and a minutes determination unit, used to determine the content of the minutes to be used based on the user record to be used and the semantic matching result to be used.

[0008] An embodiment of the present application provides a device, which includes: a processor, a memory, and a system bus; the processor and the memory are connected via the system bus; the memory is used to store one or more programs, and the one or more programs include instructions, which, when executed by the processor, enable the processor to execute any implementation of the minutes determination method provided in the embodiment of the present application.

[0009] An embodiment of the present application provides a computer-readable storage medium, in which instructions are stored. When the instructions are executed on a terminal device, the terminal device executes any implementation of the minutes determination method provided in the embodiment of the present application.

[0010] An embodiment of the present application provides a computer program product. When the computer program product is run on a terminal device, the terminal device executes any implementation of the minutes determination method provided in the embodiment of the present application.

[0011] Based on the above technical solution, this application has the following beneficial effects:

[0012] In the technical solution provided by the present application, after obtaining the user records to be used and the record text to be used, the record text to be used can be firstly divided into sentences to obtain at least one sentence to be used; then the user records to be used and the sentences to be used can be semantically matched to obtain the semantic matching results to be used; finally, the content of the minutes to be used is determined based on the user records to be used and the semantic matching results to be used.

[0013] Among them, because the above-mentioned "semantic matching result to be used" can indicate the degree of semantic matching between the user record to be used and at least one sentence in the record text to be used, the user can quickly know the position of the user record to be used in the record text to be used based on the "semantic matching result to be used", so that the user can quickly find the meeting content related to the user record to be used from the record text to be used based on the "semantic matching result to be used", which can effectively avoid the time consumption caused by the user determining the meeting content related to the user record to be used by re-listening to the meeting recording audio, thereby effectively improving the efficiency of minutes determination.

[0014] In addition, the embodiment of the present application can also determine at least one key supplementary sentence based on the above-mentioned "semantic matching results to be used" so that these key supplementary sentences can express semantic supplementary information for the user records to be used, so that when referring to these key supplementary sentences, these target sentences, and the user records to be used to determine the content of the minutes to be used, the user can obtain more comprehensive important semantics of the meeting from this information, thereby achieving the purpose of checking for omissions and filling in the gaps in the user records to be used, which is conducive to improving the effect of minutes determination.

[0015] In addition, the embodiments of the present application can also extract key points from each key supplementary sentence so that these key points can more concisely express the important information omitted from the user record to be used, so that the user can more quickly obtain the key supplementary information of the user record to be used from these key points, which is conducive to improving the efficiency of minutes determination. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the embodiments of the present application 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 some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0017] Figure 1 A flowchart of a method for determining minutes provided in an embodiment of the present application;

[0018] Figure 2 A schematic diagram of a display interface for prompt information provided in an embodiment of the present application;

[0019] Figure 3 A schematic diagram of the working principle of a semantic matching model provided in an embodiment of the present application;

[0020] Figure 4 A schematic diagram of the working principle of another semantic matching model provided in an embodiment of the present application;

[0021] Figure 5 A schematic diagram of a diagram to be used provided in an embodiment of the present application;

[0022] Figure 6 A schematic diagram of the working principle of a key point extraction model provided in an embodiment of the present application;

[0023] Figure 7 A schematic diagram of a prompt information update process provided in an embodiment of the present application;

[0024] Figure 8A schematic diagram of the structure of a minutes determination device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0025] The inventor discovered in his research on the minutes determination process that for the minutes writer, when listening to the meeting, the minutes writer usually records some concise information points to summarize the speaker's speech content (for example, the summary, the conclusions reached in the meeting, or important matters that need to be followed up, etc.); and, after the meeting, the minutes writer can first refer to these user records and listen to the conference recording audio again to determine some meeting content related to these user records (for example, the conference speech sentences and / or key supplementary sentences corresponding to the user records); and then refer to these user records and their related meeting content to write the meeting minutes, so that the meeting minutes can be more accurate. However, because the audio duration of the conference recording audio used to record long meetings is relatively long, the minutes writer needs to spend a lot of time to obtain the meeting content related to these user records, which results in a relatively long time to write the meeting minutes, resulting in relatively low efficiency in minutes determination.

[0026] Based on the above findings, in order to solve the technical problems in the background technology part, an embodiment of the present application provides a method for determining minutes, which includes: after obtaining the user records to be used and the record text to be used, the record text to be used can be firstly divided into sentences to obtain at least one sentence to be used; then the user records to be used are semantically matched with these sentences to be used to obtain the semantic matching results to be used; finally, the content of the minutes to be used is determined according to the user records to be used and the semantic matching results to be used.

[0027] Among them, because the above-mentioned "semantic matching result to be used" can indicate the degree of semantic matching between the user record to be used and at least one sentence in the record text to be used, the user can quickly know the position of the user record to be used in the record text to be used based on the "semantic matching result to be used", so that the user can quickly find the meeting content related to the user record to be used from the record text to be used based on the "semantic matching result to be used", which can effectively avoid the time consumption caused by the user determining the meeting content related to the user record to be used by re-listening to the meeting recording audio, thereby effectively improving the efficiency of minutes determination.

[0028] In addition, the embodiments of the present application do not limit the execution subject of the minutes determination method. For example, the minutes determination method provided in the embodiments of the present application can be applied to data processing devices such as terminal devices or servers. Among them, the terminal device can be a smart phone, a computer, a personal digital assistant (PDA) or a tablet computer. The server can be an independent server, a cluster server or a cloud server. For example, the minutes determination method provided in the embodiments of the present application can be executed in cooperation with a terminal device or a server.

[0029] In order to make the purpose, technical solution and advantages of the embodiments of the present application clearer, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0030] Method Example 1

[0031] See also Figure 1 , which is a flow chart of a method for determining minutes provided in an embodiment of the present application.

[0032] The minutes determination method provided in the embodiment of the present application includes S1-S4:

[0033] S1: Obtain the user record to be used and the record text to be used.

[0034] The above-mentioned "user records to be used" are used to represent concise information (e.g., shorthand content, note content, or meeting key points record, etc.) recorded by relevant personnel (e.g., minutes writers, meeting shorthand writers, etc.) during the meeting. For example, the above-mentioned "user records to be used" may include: supervision loopholes, insufficient authority, low efficiency, obvious loopholes in decentralized management efficiency, etc.

[0035] In addition, the embodiment of the present application does not limit the above-mentioned "user record to be used", for example, it may include a word, a short sentence, or a complete sentence.

[0036] In addition, the embodiments of the present application do not limit the acquisition process of the above-mentioned "user records to be used". For example, it may specifically include: after obtaining the meeting note information uploaded by the user, first segment the meeting note information according to the preset segmentation characters to obtain at least one note segmentation unit; then delete the meaningless symbols (such as ellipsis, dashes, or paragraph numbers at the beginning of a sentence, etc.) in each note segmentation unit; finally, determine each note segmentation unit as a user record to be used, so that the user record to be used includes only one sentence as much as possible.

[0037] It should be noted that the above-mentioned "preset segmentation mark" can be preset (for example, set by the user according to personal preference). For example, it can be a regular sentence ending punctuation mark (such as a period, question mark, or exclamation mark, etc.).

[0038] In order to facilitate understanding of the above-mentioned process of obtaining the "user record to be used", an example is provided below for explanation.

[0039] As an example, if the above-mentioned “meeting notes information” includes: “Regulatory loopholes, insufficient authority, low efficiency, … decentralized management efficiency loopholes are obvious. It is recommended that … consensus be reached to unify futures and various derivatives rules”, the two user records to be used that can be determined from the “meeting notes information” are as follows:

[0040] ->User record to be used 1: "Supervision loopholes, insufficient authority, low efficiency, and obvious loopholes in decentralized management efficiency."

[0041] ->User record 2 to be used: "It is recommended to reach a consensus and unify the rules for futures and various derivatives."

[0042] The above-mentioned "record text to be used" is used to represent the content of the conference speech; and the embodiment of the present application does not limit the acquisition process of the "record text to be used". For example, the acquisition process of the "record text to be used" may specifically include: in the process of listening to the meeting, the real-time conference voice is transcribed to obtain the record text to be used. For another example, the acquisition process of the "record text to be used" may include: after the meeting, the meeting record video (or meeting recording audio) is subjected to voice recognition processing to obtain the record text to be used. For another example, the acquisition process of the "record text to be used" may include: in the translation scenario, the meeting record video can be subjected to voice translation processing to obtain the record text to be used.

[0043] Based on the relevant content of S1 above, it can be known that when a user wants to write minutes for a meeting, the user records to be used can be extracted from the meeting notes information uploaded by the user for the meeting, so that the text position of the user records to be used can be located from the text of the records to be used of the meeting later.

[0044] S2: Segment the record text to be used to obtain at least one sentence to be used.

[0045] In an embodiment of the present application, after obtaining the record text to be used, the record text to be used can be divided into sentences according to preset sentence end markers (for example, a period, a question mark, or an exclamation mark, etc.) to obtain a sentence set of the record text to be used, so that the sentence set includes each sentence in the record text to be used; and then each sentence in the sentence set is determined as a sentence to be used.

[0046] S3: Perform semantic matching processing on the user record to be used and at least one sentence to be used to obtain a semantic matching result to be used.

[0047] The above-mentioned "semantic matching results to be used" are used to record the degree of semantic matching between the user record to be used and each sentence to be used (for example, semantic matching score, etc.), so that the "semantic matching results to be used" can indicate whether the semantic information carried by the user record to be used is the same as the semantic information carried by each sentence to be used.

[0048] In addition, the embodiments of the present application do not limit the implementation of S3. For example, it can be implemented by any existing or future semantic matching method. For example, in order to improve the semantic matching effect, it can be used Method Example 2 Any of the embodiments of S3 shown is implemented.

[0049] S4: Determine the content of the minutes to be used according to the user record to be used and the semantic matching result to be used.

[0050] The above-mentioned “minutes content to be used” refers to the minutes content written by the user based on the user record to be used.

[0051] In addition, the embodiments of the present application do not limit the implementation method of S4. For example, when the above-mentioned "semantic matching results to be used" include semantic matching scores between user records to be used and each sentence to be used, S4 can be specifically as follows: first, according to the semantic matching scores between user records to be used and each sentence to be used, one or more sentences to be used with relatively high semantic matching scores are screened out from all sentences to be used; then, these sentences to be used with relatively high semantic matching scores and user records to be used are displayed to the user (such as Figure 2 As shown), the user can write the minutes content to be used based on these displayed contents, so that the user can quickly find the content required for writing the minutes from these displayed contents. This can effectively avoid the time wasted by the user re-listening to the meeting recording audio to determine the meeting content related to the user record to be used, thereby effectively improving the efficiency of minutes determination.

[0052] For example, in order to further improve the effectiveness of the minutes, S4 can adopt the following Method Example 4 , Method Embodiment five , Method Example 6 ,as well as Method Example 7 Any of the embodiments of S4 shown is implemented.

[0053] Based on the relevant contents of S1 to S4 above, it can be known that for the minutes determination method provided in the embodiment of the present application, after obtaining the user records to be used and the record text to be used, the record text to be used can be firstly subjected to sentence division processing to obtain at least one sentence to be used; then the user records to be used and these sentences to be used are subjected to semantic matching processing to obtain the semantic matching results to be used; finally, the content of the minutes to be used is determined according to the user records to be used and the semantic matching results to be used.

[0054] Among them, because the above-mentioned "semantic matching result to be used" can indicate the degree of semantic matching between the user record to be used and at least one sentence in the record text to be used, the user can quickly know the position of the user record to be used in the record text to be used based on the "semantic matching result to be used", so that the user can quickly find the meeting content related to the user record to be used from the record text to be used based on the "semantic matching result to be used", which can effectively avoid the time consumption caused by the user determining the meeting content related to the user record to be used by re-listening to the meeting recording audio, thereby effectively improving the efficiency of minutes determination.

[0055] Method Example 2

[0056] In order to further improve the determination effect of the minutes, the embodiment of the present application also provides another possible implementation of S3 above, which may specifically include: determining the semantic matching result to be used according to the user record to be used, at least one sentence to be used and a pre-built semantic matching model. For ease of understanding, the following is an example.

[0057] As an example, the process of determining the semantic matching result to be used may specifically include S31-S33:

[0058] S31: Determine a matching text to be used according to at least one sentence to be used and a preset sentence identifier.

[0059] The above-mentioned “preset sentence identifier” is used to identify the starting position of a sentence; and the embodiment of the present application does not limit the “preset sentence identifier”, for example, it may be [CLS].

[0060] The above-mentioned "matching text to be used" is used to record at least one sentence to be used; and the embodiment of the present application does not limit the determination process of the "matching text to be used", for example, it can be specifically: first concatenate a preset sentence identifier at the beginning of each sentence to be used to obtain each concatenated sentence; then all the concatenated sentences are processed collectively to obtain the matching text to be used, so that the matching text to be used includes all the concatenated sentences.

[0061] S32: Inputting the user record to be used and the matching text to be used into a pre-built semantic matching model to obtain a matching prediction result output by the semantic matching model.

[0062] The “matching prediction result” is used to indicate the semantic matching result between the user record to be used and at least one sentence in the matching text to be used.

[0063] The above-mentioned “semantic matching model” is used to perform semantic matching processing on the input data of the semantic matching model; and the “semantic matching model” is a machine learning model.

[0064] In addition, the embodiments of the present application do not limit the network structure of the above-mentioned "semantic matching model". For example, it may specifically include: a word encoding module, a content interaction module and a matching prediction module.

[0065] In order to facilitate understanding of the working principle of the above-mentioned "semantic matching model", the following is an explanation taking the determination process of the above-mentioned "matching prediction result" as an example.

[0066] As an example, the process of determining the above-mentioned “matching prediction result” by using the semantic matching model may specifically include S321-S323:

[0067] S321: Determine a user record encoding feature and at least one sentence encoding feature according to the user record to be used, the matching text to be used, and the word encoding module.

[0068] The above-mentioned "word encoding module" is used to encode the input data of the word encoding module; and the embodiment of the present application does not limit the "word encoding module", for example, Figure 3 or Figure 4 As shown, the "word encoding module" can be implemented using a bidirectional encoding (Bidirectional Encoder Representations from Transformers, BERT) pre-trained model.

[0069] The above-mentioned “user record coding feature” is used to represent the character information carried by the user record to be used; and the “user record coding feature” is obtained by encoding the user record to be used by the word encoding module.

[0070] The j-th sentence encoding feature is used to represent the character information carried by the j-th sentence in the matching text to be used; and the "j-th sentence encoding feature" is obtained by encoding the j-th sentence in the matching text to be used by the word encoding module. Wherein, j is a positive integer, j≤J, J is a positive integer, and J represents the number of sentences in the matching text to be used (that is, the number of "sentences to be used" above).

[0071] It should be noted that the above-mentioned "jth sentence in the matching text to be used" includes the jth sentence to be used; and the embodiment of the present application is not limited to the "jth sentence in the matching text to be used", for example, it can specifically be: the jth concatenated sentence obtained by concatenating the jth sentence to be used and the preset sentence identifier.

[0072] S322: Obtain at least one content interaction feature according to the user record coding feature, at least one sentence coding feature, and a content interaction module, wherein the "content interaction module" is used to perform semantic interaction processing on input data of the content interaction module.

[0073] The embodiment of the present application does not limit the implementation method of S322. For ease of understanding, three possible implementation methods are described below.

[0074] In a first possible implementation, S322 may specifically be: inputting the user record coding feature and at least one sentence coding feature into the content interaction module to obtain at least one content interaction feature output by the content interaction module. It should be noted that the first possible implementation of S322 may be implemented using the working principle of the above-mentioned "content interaction module" shown in steps 31-32 below.

[0075] In a second possible implementation, S322 may specifically include step 11-step 12:

[0076] Step 11: based on the similarity representation data between the user record coding features and the sentence coding features, at least one sentence feature to be processed that meets the second similarity condition is selected from at least one sentence coding feature.

[0077] The “similarity representation data between the user record encoding feature and the j-th sentence encoding feature” is used to represent the similarity between the user record encoding feature and the j-th sentence encoding feature, so that the “similarity representation data between the user record encoding feature and the j-th sentence encoding feature” can represent the semantic similarity between the user record to be used and the j-th sentence to be used. Where j is a positive integer, j≤J, and J is a positive integer.

[0078] In addition, the embodiments of the present application do not limit the determination process of the above-mentioned "similarity representation data between the user record coding feature and the jth sentence coding feature". For example, it can be specifically: the cosine distance between the user record coding feature and the jth sentence coding feature is determined as the "similarity representation data between the user record coding feature and the jth sentence coding feature". For another example, it can be specifically: the Euclidean distance between the user record coding feature and the jth sentence coding feature is determined as the "similarity representation data between the user record coding feature and the jth sentence coding feature".

[0079] The above-mentioned "second similarity condition" can be preset. For example, assuming that the larger the "similarity representation data between the user record coding feature and the j-th sentence coding feature", the more similar the semantic information carried by the user record to be used is to the semantic information carried by the j-th sentence to be used. Based on the above assumption, it can be known that when the J sentence coding features are sorted from large to small according to the above-mentioned "similarity representation data", the "second similarity condition" can be specifically: the arrangement number is lower than the first sequence number threshold (for example, 2). Among them, the "first sequence number threshold" can be preset.

[0080] It should be noted that, for the above-mentioned "first serial number threshold", if the "first serial number threshold" is relatively small (such as 2), then the semantic similarity between the sentence features to be processed selected from the above-mentioned "at least one sentence encoding feature" and the user record to be used is relatively high, so that the sentences to be used represented by these sentence features to be processed can better supplement the meeting information for the user record to be used; in addition, using these sentence features to be processed as prompt information can better guide the semantic matching model to determine the sentences in the text of the record to be used that are semantically similar to the user record to be used, so that it can effectively avoid determining multiple sentences with relatively large differences in semantic relevance from the text of the record to be used for the user record to be used, which is conducive to improving the effect of minutes determination.

[0081] The above-mentioned “sentence features to be processed” refer to sentence coding features that meet the second similarity condition.

[0082] Based on the relevant content of the above step 11, it can be known that after obtaining the user record coding feature and at least one sentence coding feature, multiple sentence coding features that are most similar to the user record coding feature can be selected from these sentence coding features (such as Figure 3 As shown), they are respectively determined as the features of the sentences to be processed, so that the sentences to be used represented by these features of the sentences to be processed can be subsequently determined to have the closest semantics to the user record to be used.

[0083] It should be noted that for Figure 3 For example, "r j-1=1" indicates that the j-1th sentence encoding feature is selected (that is, the j-1th sentence encoding feature satisfies the second similarity condition); "r j =0" indicates that the j-th sentence encoding feature is not selected (that is, the j-th sentence encoding feature does not meet the second similarity condition); s j ={w j,1 ,…,w j,m} represents the jth sentence to be used, and w j,1 represents the first semantic unit (e.g., character and / or word) in the jth sentence to be used, ... (and so on), w j,m represents the mth semantic unit in the jth sentence to be used; q = {w1,…,w n} represents the user record to be used, and w1 represents the first semantic unit in the user record to be used, ... (and so on), w n represents the nth semantic unit in the jth sentence to be used; Represents the j-th sentence encoding feature; Indicates the encoding features of the sentence with the sequence number 1; represents the user record encoding feature; q′ represents the “user record characterization feature” below; s″ j represents the jth content interaction feature; y j represents the j-th matching prediction score (that is, the matching prediction score between the user record to be used and the j-th sentence to be used, etc.).

[0084] Step 12: Input the user record coding feature and at least one sentence feature to be processed into a content interaction module to obtain at least one content interaction feature output by the content interaction module.

[0085] It should be noted that step 12 can be implemented using the working principle of the above-mentioned "content interaction module" shown in steps 31 and 32 below.

[0086] Based on the relevant content of the second possible implementation method of S322 above, it can be known that after obtaining the user record coding features and at least one sentence coding feature, sentence coding features that are relatively similar to the user record coding features can be selected from these sentence coding features; then these selected sentence coding features and the user record coding features are subjected to semantic interaction processing to obtain at least one content interaction feature, so that these content interaction features can represent the semantic interaction results between these selected sentence coding features and the user record coding features, so that the sentences that are closest to the semantics of the user record to be used can be determined from the sentences to be used represented by these selected sentence coding features. This can effectively reduce the number of semantic interaction objects, thereby helping to improve the efficiency of semantic interaction, and then helping to improve the efficiency of minutes determination.

[0087] In fact, for the above-mentioned "at least one sentence to be used", some sentences to be used may contain pronouns, and some previous information may be omitted. Therefore, in order to better extract the semantic information carried by these sentences to be used. The embodiment of the present application also provides another possible implementation of the above-mentioned "semantic matching model". In this implementation, in addition to the above-mentioned word encoding module, content interaction module and matching prediction module, the "semantic matching model" may also include: sentence encoding module (such as Figure 4 shown).

[0088] It should be noted that for Figure 4 For example, j ={w j,1 ,…,w j,m} represents the jth sentence to be used; q={w1,…,w n} indicates the user record to be used; Represents the j-th sentence encoding feature; represents the user record encoding feature; q′ represents the “user record characterization feature” below; s′ j represents the jth representation feature; s″ j represents the jth content interaction feature; y j represents the j-th matching prediction score (for example, the matching prediction score between the user record to be used and the j-th sentence to be used, etc.).

[0089] In order to facilitate understanding of the working principle of the above-mentioned "sentence encoding module", the third possible implementation method of S322 is described below.

[0090] In a third possible implementation, when the “semantic matching model” further includes a sentence encoding module, S322 may specifically include steps 21 and 22:

[0091] Step 21: input at least one sentence encoding feature into a sentence encoding module to obtain at least one sentence representation feature output by the sentence encoding module.

[0092] The above-mentioned "sentence encoding module" is used to encode a sentence with reference to the context information of the sentence; and the embodiment of the present application does not limit the implementation method of the "sentence encoding module", for example, it can be implemented using any encoder. It should be noted that the embodiment of the present application does not limit the above-mentioned "encoder", for example, it can include the Encoder layer in the two-layer Transformer.

[0093] The j-th sentence representation feature is used to represent the semantic information carried by the j-th sentence to be used. Wherein, j is a positive integer, j≤J, and J is a positive integer.

[0094] Step 22: Obtain at least one content interaction feature according to the user record coding feature, at least one sentence representation feature, and the content interaction module.

[0095] The present application embodiment does not limit the implementation of step 22. For example, in the first possible implementation, Figure 4 As shown, step 22 may specifically include: inputting the user record coding feature and at least one sentence representation feature into the content interaction module to obtain at least one content interaction feature output by the content interaction module. It should be noted that the first possible implementation of step 22 can be implemented using the working principle of the above-mentioned "content interaction module" shown in steps 31-32 below.

[0096] In addition, in order to improve the efficiency of semantic interaction, the embodiment of the present application also provides a second possible implementation of step 22, which may specifically include steps 221-222:

[0097] Step 221: based on the similarity representation data between the user record coding feature and each sentence representation feature, at least one sentence feature to be referenced that meets a first similarity condition is selected from at least one sentence representation feature.

[0098] The “similarity representation data between the user record encoding feature and the j-th sentence representation feature” is used to represent the similarity between the user record encoding feature and the j-th sentence representation feature, so that the “similarity representation data between the user record encoding feature and the j-th sentence representation feature” can represent the semantic similarity between the user record to be used and the j-th sentence to be used. Where j is a positive integer, j≤J, and J is a positive integer.

[0099] In addition, the embodiments of the present application do not limit the determination process of the above-mentioned "similarity representation data between the user record coding feature and the j-th sentence representation feature". For example, it can be specifically: the cosine distance between the user record coding feature and the j-th sentence representation feature is determined as the "similarity representation data between the user record coding feature and the j-th sentence representation feature". For another example, it can be specifically: the Euclidean distance between the user record coding feature and the j-th sentence representation feature is determined as the "similarity representation data between the user record coding feature and the j-th sentence representation feature".

[0100] The above-mentioned "first similarity condition" can be preset. For example, assuming that the larger the "similarity representation data between the user record encoding feature and the j-th sentence representation feature", the more similar the semantic information carried by the user record to be used is to the semantic information carried by the j-th sentence to be used. Based on the above assumption, it can be known that when the J sentence representation features are sorted from large to small according to the above-mentioned "similarity representation data", the "first similarity condition" can be specifically: the arrangement sequence number is lower than the second sequence number threshold.

[0101] It should be noted that, for the above-mentioned "second serial number threshold", if the "second serial number threshold" is relatively small (such as 2), then the semantic similarity between the sentence features to be referenced selected from the above-mentioned "at least one sentence representation feature" and the user record to be used is relatively high, so that the sentences to be used represented by these sentence features to be referenced can better supplement the meeting information for the user record to be used; in addition, using these sentence features to be referenced as prompt information can better guide the semantic matching model to determine the sentences in the text of the record to be used that are semantically similar to the user record to be used, so that it can effectively avoid determining multiple sentences with relatively large differences in semantic relevance from the text of the record to be used for the user record to be used, which is conducive to improving the effect of minutes determination.

[0102] The above-mentioned “sentence features to be referenced” refer to sentence representation features that meet the first similarity condition.

[0103] Based on the relevant content of the above step 221, it can be known that after obtaining the user record coding feature and at least one sentence representation feature, multiple sentence representation features that are most similar to the user record coding feature can be selected from these sentence representation features and determined as sentence features to be processed respectively, so that the sentence that is closest to the semantics of the user record to be used can be determined from the sentences to be used represented by these sentence features to be processed.

[0104] Step 222: Input the user record coding feature and at least one sentence feature to be referenced into the content interaction module to obtain at least one content interaction feature output by the content interaction module.

[0105] It should be noted that step 222 can be implemented using the working principle of the above-mentioned "content interaction module" shown in steps 31 and 32 below.

[0106] Based on the relevant contents of the above steps 221 to 222, it can be known that after obtaining the user record coding features and at least one sentence representation feature, the sentence representation features that are relatively similar to the user record coding features can be selected from these sentence representation features; then these selected sentence representation features and the user record coding features are subjected to semantic interaction processing to obtain at least one content interaction feature, so that these content interaction features can represent the semantic interaction results between these selected sentence representation features and the user record coding features, so that the sentences that are closest to the semantics of the user record to be used can be determined from the sentences to be used represented by these selected sentence representation features. This can effectively reduce the number of semantic interaction objects, thereby helping to improve the efficiency of semantic interaction, and further helping to improve the efficiency of minutes determination.

[0107] Based on the relevant content of the third possible implementation method of S322 above, it can be known that after obtaining the user record coding features and at least one sentence coding feature, you can first refer to the context information of each sentence to be used, perform sentence coding processing on each sentence coding feature, and obtain each sentence representation feature; then perform semantic interaction processing on the user record coding features, and all or part of the sentence representation features, and obtain at least one content interaction feature, so that you can subsequently better determine the sentence that is closest to the semantics of the user record to be used based on these content interaction features, which is conducive to improving the effect of minutes determination.

[0108] In addition, the embodiments of the present application do not limit the implementation methods of the above-mentioned "content interaction module". For example, it may specifically include: an attention layer and a fusion layer.

[0109] In order to facilitate understanding of the working principle of the above-mentioned "content interaction module", the following is explained by taking the determination process of the above-mentioned "at least one content interaction feature" as an example.

[0110] As an example, when the input data of the above-mentioned "content interaction module" includes user record coding features and at least one feature of a sentence to be interacted, the process of determining "at least one content interaction feature" using the above-mentioned "content interaction module" includes steps 31 and 32:

[0111] Step 31: Input the user record encoding features and at least one feature of the sentence to be interacted into the attention layer to obtain the user record representation features output by the attention layer.

[0112] The above-mentioned "at least one sentence feature to be interacted" is used to represent the sentence features that need to be referred to when semantically interacting with the user record coding features; and the embodiments of the present application do not limit the "at least one sentence feature to be interacted". For example, when the "content interaction module" having the working principle shown in steps 31-32 is used to implement the "first possible implementation method of S322" above, the "at least one sentence feature to be interacted" may include the above "at least one sentence coding feature". For another example, when the "content interaction module" having the working principle shown in steps 31-32 is used to implement the above step 12, the "at least one sentence feature to be interacted" includes the above "at least one sentence feature to be processed". For another example, when the "content interaction module" having the working principle shown in steps 31-32 is used to implement the above "first possible implementation method of step 22", the "at least one sentence feature to be interacted" includes the above "at least one sentence representation feature". Also, when the "content interaction module" having the working principle shown in step 31-step 32 is used to implement the above step 222, the "at least one sentence feature to be interacted" includes the above "at least one sentence feature to be referenced".

[0113] The above-mentioned "attention layer" is used to perform attention processing on the input data of the attention layer; and the embodiments of the present application do not limit the implementation method of the above-mentioned "attention layer", for example, it can be implemented using a multi-head attention mechanism.

[0114] The above-mentioned “user record characterization feature” is used to represent the semantic information carried by the user record to be used.

[0115] In addition, the embodiments of the present application do not limit the determination process of the above-mentioned "user record characterization features". For example, when the above-mentioned "attention layer" is implemented using a multi-head attention mechanism, the determination process of the "user record characterization features" may specifically include: using the user record encoding features as the query parameters of the multi-head attention mechanism (that is, parameter Q), and using the above-mentioned "at least one sentence feature to be interacted with" as the other two parameters of the multi-head attention mechanism (that is, parameter K and parameter V). In this way, the meeting content information carried by the "at least one sentence feature to be interacted with" can be integrated into the user record encoding features, so that when the user record to be used is relatively short, or the description is vague and other information is incomplete, these meeting content information can better supplement the information omitted in the user record to be used, and at the same time, the "user record characterization features" have a global view of the content of the entire meeting, which is conducive to better determining the sentence corresponding to the user record to be used in the above-mentioned "record text to be used".

[0116] Step 32: Determine at least one content interaction feature according to the user record representation feature, at least one feature of the sentence to be interacted, and the fusion layer.

[0117] The “e-th content interaction feature” refers to the semantic interaction result between the user record representation feature and the e-th sentence feature to be interacted. Where e is a positive integer, e≤E, E is a positive integer, and E represents the number of content interaction features (that is, the number of sentence features to be interacted).

[0118] In addition, the embodiments of the present application do not limit the determination process of the above-mentioned "e-th content interaction feature". For example, it may specifically include: inputting the user record characterization features and the e-th sentence features to be interacted into the fusion layer to obtain the e-th content interaction feature output by the fusion layer.

[0119] The above-mentioned "fusion layer" is used to perform feature fusion processing on the input data of the fusion layer; and the embodiment of the present application does not limit the working principle of the "fusion layer", for example, it can be specifically: after the user record representation features and the e-th sentence features to be interacted are input into the fusion layer, the fusion layer can directly splice the user record representation features and the e-th sentence features to be interacted to obtain the e-th content interaction feature.

[0120] In addition, in order to improve the semantic interaction effect, the embodiment of the present application also provides another working principle of the above-mentioned "fusion layer". For ease of understanding, the following is explained by taking the determination process of the above-mentioned "e-th content interaction feature" as an example.

[0121] As an example, the process of determining the “e-th content interaction feature” may specifically include steps 321 to 323:

[0122] Step 321: Determine an e-th similarity characterization feature according to the user record characterization feature and the e-th sentence feature to be interacted with, so that the e-th similarity characterization feature is used to represent the information similarity between the user record characterization feature and the e-th sentence feature to be interacted with.

[0123] The above-mentioned “e-th similarity representation feature” is used to describe the semantic information similarity between the user record representation feature and the e-th sentence feature to be interacted.

[0124] In addition, the embodiments of the present application do not limit the determination process of the above-mentioned "e-th similarity characterization feature". For example, any existing or future feature similarity determination method can be used for implementation. For another example, it can be specifically: according to the dot product between the user record characterization feature and the e-th sentence feature to be interacted, the e-th similarity characterization feature is determined (for example, the dot product between the user record characterization feature and the e-th sentence feature to be interacted is determined as the e-th similarity characterization feature).

[0125] Step 322: Determine the e-th information difference characterization feature according to the user record characterization feature and the e-th sentence feature to be interacted with, so that the e-th information difference characterization feature is used to represent the information difference between the user record characterization feature and the e-th sentence feature to be interacted with.

[0126] The above-mentioned "e-th information difference representation feature" is used to describe the semantic information difference between the user record representation feature and the e-th sentence feature to be interacted with, so that the "e-th information difference representation feature" can represent the information difference between the user record to be used and the sentence to be used represented by the e-th sentence feature to be interacted with.

[0127] In addition, the embodiments of the present application do not limit the determination process of the above-mentioned "e-th information difference characterization feature". For example, it can be implemented by using any existing or future feature difference determination method. For another example, it can be specifically: according to the difference between the user record characterization feature and the e-th sentence feature to be interacted, determine the e-th information difference characterization feature (for example, the difference between the user record characterization feature and the e-th sentence feature to be interacted is determined as the e-th information difference characterization feature).

[0128] Step 323: concatenate the user record representation feature, the e-th sentence feature to be interacted with, the e-th similarity representation feature, and the e-th information difference representation feature to obtain the e-th content interaction feature (as shown in formula (1)).

[0129]

[0130] In the formula, s e ″ represents the e-th content interaction feature; q′ represents the user record representation feature; represents the feature of the e-th sentence to be interacted; * represents dot product processing; represents the e-th similarity representation feature; represents the e-th information difference representation feature.

[0131] Based on the relevant contents of steps 321 to 323 above, it can be known that for the above-mentioned "fusion layer", after obtaining the user record representation features and the e-th sentence features to be interacted with, the e-th content interaction feature can be determined by referring to the user record representation features, the e-th sentence features to be interacted with, the similarity features between the user record representation features and the e-th sentence features to be interacted with, and the difference features between the user record representation features and the e-th sentence features to be interacted with, so that the e-th content interaction feature can more comprehensively describe the semantic similarity between the user record to be used and the sentence to be used represented by the e-th sentence features to be interacted with.

[0132] Based on the relevant contents of steps 31 to 32 above, it can be known that for the above-mentioned "content interaction module", it can first perform attention processing and then feature fusion processing on the input data of the content interaction module to achieve semantic interaction processing on the input data of the content interaction module.

[0133] It should be noted that when using steps 31-32 to implement the above "first possible implementation method of S322", the "sentence features to be interacted" in steps 31-32 can be directly replaced with "sentence encoding features"; when using steps 31-32 to implement the above step 12, the "sentence features to be interacted" in steps 31-32 can be directly replaced with "sentence features to be processed"; when using steps 31-32 to implement the above "first possible implementation method of step 22", the "sentence features to be interacted" in steps 31-32 can be directly replaced with "sentence representation features"; when using steps 31-32 to implement the above step 222, the "sentence features to be interacted" in steps 31-32 can be directly replaced with "sentence features to be referenced".

[0134] Based on the relevant content of S322 above, it can be known that for the semantic matching model, after using the encoding module in the semantic matching model to determine the user record encoding features and at least one sentence encoding feature, the content interaction module in the semantic matching model can refer to the user record encoding features and these sentence encoding features to perform semantic interaction processing to obtain at least one content interaction feature, so that these content interaction features can better represent the semantic similarity between the user record to be used and all or part of the sentences to be used.

[0135] S323: Input at least one content interaction feature into a matching prediction module to obtain a matching prediction result output by the matching prediction module.

[0136] The above-mentioned "matching prediction module" is used to determine the degree of matching (that is, the degree of similarity) between the semantic information carried by a sentence to be used and the semantic information carried by the user record to be used; and the embodiment of the present application does not limit the implementation method of the "matching prediction module".

[0137] The above-mentioned "matching prediction result" is used to indicate the semantic matching result between the above-mentioned "at least one sentence to be used" and the user record to be used; and the embodiment of the present application does not limit the "matching prediction result". For example, if the e-th content interaction feature is determined based on the user record characterization feature and the e-th sentence to be interacted feature, then the "matching prediction result" may include: the matching prediction score between the user record to be used and the sentence to be used represented by the e-th sentence to be interacted feature, so that the "matching prediction score between the user record to be used and the sentence to be used corresponding to the e-th sentence to be interacted feature" can indicate the semantic similarity between the user record to be used and the sentence to be used represented by the e-th sentence to be interacted feature. Wherein, e is a positive integer, e≤E, and E is a positive integer.

[0138] Based on the relevant contents of S321 to S323 above, it can be known that for the pre-built semantic matching model, it can realize the semantic matching process between the user record to be used and at least one sentence in the matching text to be used with the help of the word encoding module, content interaction module and matching prediction module in the semantic matching model. Among them, because the above "semantic matching model" has a good semantic matching effect, the matching prediction result determined based on the "semantic matching model" can better express the semantic similarity between the above "at least one sentence to be used" and the user record to be used.

[0139] In addition, the above-mentioned "semantic matching model" can be constructed based on the pre-collected recorded semantic matching annotated corpus; and the embodiment of the present application does not limit the construction process of the above-mentioned "semantic matching model", for example, it can be implemented using any existing or future model construction method.

[0140] The above-mentioned "recorded semantic matching annotated corpus" may include sample user records, a first meeting text sentence set corresponding to the sample user records, and semantic matching tags corresponding to the sample user records. Among them, "sample user records" refer to the shorthand information of the sample meeting records by relevant personnel. The "first meeting text sentence set corresponding to the sample user records" is obtained by sentence segmentation of the meeting transcription text of the sample meeting. The "semantic matching tags corresponding to the sample user records" are used to indicate which sentences in the first meeting text sentence set actually have the same semantic information as the semantic information of the sample user records, and the "semantic matching tags corresponding to the sample user records" can be generated by manual annotation.

[0141] In addition, in order to improve the noise resistance of the above-mentioned "semantic matching model", some weak supervision data can be constructed by performing data noise processing on the above-mentioned "record semantic matching annotated corpus", so that the above-mentioned "semantic matching model" can be constructed later using these weak supervision data and the above-mentioned "record semantic matching annotated corpus".

[0142] The construction process of the above-mentioned "weakly supervised data" may specifically include: first, for each original sentence in the above-mentioned "conference transcription text of the sample meeting", by operations such as synonym replacement, word fragment deletion, random word fragment insertion, and word fragment shuffling, generate sentence noise data of each original sentence; then use each original sentence in the "conference transcription text of the sample meeting" and the sentence noise data of each original sentence to form a pair of related sentence pairs, so that the above-mentioned "semantic matching model" can be constructed based on these sentence pairs in the future; and the construction process of the "semantic matching model" includes: first use these sentences to perform the first stage training on the semantic matching model, so that the trained semantic matching model can accurately learn that the semantic information carried by the sentence noise data in the sentence pair is the same as the semantic information carried by the original sentence in the sentence pair; then use the above-mentioned "record semantic matching annotated corpus" to perform the second stage training on the semantic matching model, so that the semantic prediction results output by the semantic matching model for the sample user records and the first meeting text sentence set are almost close to the above-mentioned "semantic matching labels corresponding to the sample user records", which is conducive to improving the semantic matching performance of the semantic matching model.

[0143] Based on the relevant content of S32 above, it can be known that after obtaining the user record to be used and the matching text to be used, the pre-built semantic matching model can be used to perform semantic matching processing on the user record to be used and at least one sentence in the matching text to be used to obtain a matching prediction result, so that the "matching prediction result" can represent the degree of semantic similarity between the above "at least one sentence to be used" and the user record to be used, so that subsequently, based on the "matching prediction result", a sentence with semantic similarity to the user record to be used can be determined from the record text to be used.

[0144] S33: Determine the semantic matching result to be used according to the matching prediction result.

[0145] The embodiment of the present application does not limit the implementation of S33. For example, it may specifically include: determining the matching prediction result as the semantic matching result to be used.

[0146] In addition, in order to further improve the accuracy of semantic matching, the embodiment of the present application also provides another possible implementation of S33, which may specifically include: determining the semantic matching result to be used based on the text retrieval result to be used and the matching prediction result.

[0147] The above-mentioned "text search result to be used" refers to the result obtained by searching and processing the user record to be used from the above-mentioned "record text to be used"; and the embodiment of the present application does not limit the "text search result to be used", for example, it may include: the search score between the user record to be used and the jth sentence to be used; and the "search score between the user record to be used and the jth sentence to be used" can indicate the semantic similarity between the user record to be used and the jth sentence to be used. Wherein, j is a positive integer, j≤J, and J is a positive integer.

[0148] In addition, the above-mentioned "text search result to be used" can be determined according to a preset search algorithm, a user record to be used, and at least one sentence to be used. Among them, the "preset search algorithm" can be preset, for example, it can be the best match (Best Match, BM) 25. It should be noted that the above-mentioned "25" refers to the 25th algorithm iteration.

[0149] To facilitate understanding of the above “another possible implementation of S33”, an example is provided below for explanation.

[0150] As an example, when the above-mentioned "matching prediction result" includes the matching prediction score between the user record to be used and the jth sentence to be used, and the above-mentioned "text retrieval result to be used" includes the retrieval score between the user record to be used and the jth sentence to be used (for example, BM25 retrieval score), S33 can be specifically: weighted sum of the "matching prediction score between the user record to be used and the jth sentence to be used" and the "retrieval score between the user record to be used and the jth sentence to be used" to obtain the semantic matching score between the user record to be used and the jth sentence to be used. Wherein, j is a positive integer, j≤J, and J is a positive integer.

[0151] It should be noted that the weighted weight corresponding to the above-mentioned "matching prediction score between the user record to be used and the jth sentence to be used" and the weighted weight corresponding to the above-mentioned "retrieval score between the user record to be used and the jth sentence to be used" can both be preset.

[0152] Based on the relevant contents of S31 to S33 above, it can be known that after obtaining the user record to be used and at least one sentence to be used, the pre-built semantic matching model can be used to implement the semantic matching process between the user record to be used and the at least one sentence to be used, and obtain the semantic matching result to be used. Among them, because the above-mentioned "semantic matching model" has a good semantic matching effect, the semantic matching result to be used determined based on the "semantic matching model" can better express the degree of semantic matching between the user record to be used and at least one sentence in the record text to be used, which is conducive to improving the effect of determining the minutes.

[0153] Method Example 3

[0154] The inventors also discovered that the position distribution of multiple sentences corresponding to a piece of user record information in the record text to be used is usually relatively concentrated. Therefore, based on this position distribution constraint, the embodiment of the present application also provides another possible implementation of the minutes determination method. In this implementation, when the above "at least one sentence to be used" includes at least one semantic matching sentence, and the above "semantic matching result to be used" includes the semantic matching score between the user record to be used and each semantic matching sentence, the minutes determination method includes S5-S7 in addition to the above S1-S4, and the execution order of S5-S7 is between S3 and S4.

[0155] S5: According to the semantic matching scores between the user record to be used and each semantic matching sentence, at least one benchmark sentence satisfying a preset benchmark condition is screened out from at least one semantic matching sentence.

[0156] The above-mentioned "semantic matching sentences" refer to sentences to be used with semantic matching scores (that is, sentences to be used represented by the "content interaction features" above); and the embodiments of the present application do not limit the "semantic matching sentences". For ease of understanding, the following is an explanation with three examples.

[0157] Example 1: If there is no screening process in the determination process of the above "semantic matching results to be used" (for example, the screening process based on the above "first similarity condition" or "second similarity condition"), then the "semantic matching sentence" refers to the above "sentence to be used", so that the above "at least one semantic matching sentence" can include the above "at least one sentence to be used".

[0158] Example 2: If there is a screening process based on the first similarity condition in the determination process of the above "semantic matching results to be used", then the "semantic matching sentence" refers to the sentence to be used represented by the characteristics of the sentence to be referenced, so that the above "at least one semantic matching sentence" can include at least one sentence to be used represented by the characteristics of the sentence to be referenced.

[0159] Example 3: If there is a screening process based on the second similarity condition in the determination process of the above "semantic matching results to be used", then the "semantic matching sentence" refers to the sentence to be used represented by the characteristics of the sentence to be processed, so that the above "at least one semantic matching sentence" can include at least one sentence to be used represented by the characteristics of the sentence to be processed.

[0160] The above-mentioned "preset benchmark condition" can be preset, for example, it can be specifically: the semantic matching score between the user record to be used and the benchmark sentence is higher than the preset first score threshold. For another example, when all semantic matching sentences are sorted from large to small according to the semantic matching score, the "preset benchmark condition" can be specifically: the arrangement number is lower than the preset third number threshold (for example, 2).

[0161] The above-mentioned “benchmark sentence” refers to a semantically matching sentence that meets preset benchmark conditions.

[0162] Based on the relevant content of S5 above, it can be known that after obtaining the semantic matching scores between the user records to be used and each semantic matching sentence, multiple semantic matching sentences with higher semantic matching scores (for example, Top2) can be screened out from all semantic matching sentences based on these semantic matching scores, so that the text positions of these multiple semantic matching sentences with higher semantic matching scores can be used as a benchmark to determine the text position penalty items of each semantic matching sentence.

[0163] S6: Determine a distance penalty score of each semantically matching sentence according to the text position of at least one reference sentence and the text position of each semantically matching sentence.

[0164] The “distance penalty score of the e-th semantic matching sentence” is used to indicate the influence of the text position of the e-th semantic matching sentence on the “degree of semantic matching between the user record to be used and the e-th semantic matching sentence”. Where e is a positive integer, e≤E.

[0165] In addition, the embodiment of the present application does not limit the determination process of the above-mentioned "distance penalty score of the e-th semantic matching sentence", for example, it may specifically include S61-S63:

[0166] S61: Determine a distance set corresponding to the e-th semantic matching sentence according to the distance between the text position of each benchmark sentence and the text position of the e-th semantic matching sentence.

[0167] For example, when the number of the above-mentioned "benchmark sentences" is Y, S61 may specifically include: collectively processing the distance between the text position of the first benchmark sentence and the text position of the e-th semantic matching sentence, the distance between the text position of the second benchmark sentence and the text position of the e-th semantic matching sentence, ..., and the distance between the text position of the Y-th benchmark sentence and the text position of the e-th semantic matching sentence, to obtain the distance set corresponding to the e-th semantic matching sentence.

[0168] It should be noted that the above “text position” refers to the position of a sentence in the record text to be used. For example, the above “text position of the e-th semantic matching sentence” refers to the position of the e-th semantic matching sentence in the record text to be used.

[0169] S62: Performing preset statistical analysis on the distance set corresponding to the e-th semantic matching sentence to obtain the distance to be used corresponding to the e-th semantic matching sentence.

[0170] The above-mentioned "preset statistical analysis processing" can be pre-set; and the embodiment of the present application is not limited to the "preset statistical analysis processing", for example, it can be taking the minimum value, taking the maximum value, taking the average value, taking the sum value, etc.

[0171] The above-mentioned “distance to be used corresponding to the e-th semantic matching sentence” is used to represent the text position distance between the e-th semantic matching sentence and the above-mentioned “at least one benchmark sentence”.

[0172] S63: Determine a distance penalty score of the e-th semantic matching sentence according to the to-be-used distance corresponding to the e-th semantic matching sentence.

[0173] It should be noted that the embodiment of the present application does not limit the implementation of S63. For example, it may specifically include: determining the distance penalty score of the e-th semantic matching sentence by multiplying the preset penalty coefficient by the distance to be used corresponding to the e-th semantic matching sentence. The “preset penalty coefficient” may be preset.

[0174] Based on the relevant contents of S6 above, after obtaining at least one benchmark sentence, the distance penalty score of each semantic matching sentence can be determined according to the distance between the text position of each semantic matching sentence and the text position of these benchmark sentences.

[0175] S7: Using the distance penalty score of each semantic matching sentence, update the semantic matching score between the user record to be used and each semantic matching sentence (as shown in formula (2)).

[0176] Score′ e =Score e -DP e (2)

[0177] In the formula, Score′ e represents the semantic matching score between the updated user record to be used and the e-th semantic matching sentence; Score e represents the semantic matching score between the user record to be used before the update and the e-th semantic matching sentence; DP e Represents the distance penalty score of the e-th semantic matching sentence.

[0178] Based on the relevant contents of S5 to S7 above, it can be known that after obtaining the semantic matching result to be used, the text position of each sentence to be used involved in the semantic matching result to be used can be used to update the semantic matching result to be used, so that the updated semantic matching result to be used satisfies the position distribution constraint of the relatively concentrated description sentences of the actual user record, so that the updated semantic matching result to be used can more accurately represent the degree of semantic matching between the user record to be used and at least one sentence in the text of the record to be used, so that the content of the minutes to be used can be determined based on the updated semantic matching result to be used, which is conducive to improving the effect of minutes determination.

[0179] Method Example 4

[0180] In addition, in order to further improve the effect of determining the minutes, the embodiment of the present application also provides another possible implementation of the above S4, which is explained below with reference to examples.

[0181] As an example, when the above “at least one sentence to be used” includes at least one semantic matching sentence, and the above “semantic matching result to be used” includes the semantic matching score between the user record to be used and each semantic matching sentence, S4 may specifically include S41-S42:

[0182] S41: Filtering out at least one target sentence satisfying a first matching condition from at least one semantically matching sentence according to the semantic matching scores between the user record to be used and each semantically matching sentence.

[0183] The above-mentioned "first matching condition" can be preset. For example, it can be implemented by using the condition achieved by "at least one target sentence" shown in any implementation of S41 below (for example, the condition shown in formula (6) below).

[0184] The above-mentioned “target sentence” refers to a semantic matching sentence that satisfies the first matching condition.

[0185] The embodiment of the present application does not limit the implementation method of S41. For example, it can be implemented using any implementation method of the process of determining "at least one target sentence" shown in steps 41-42 below. It only needs to replace the "candidate sentence" with the "semantically matching sentence" in any implementation method of the process of determining "at least one target sentence" shown in steps 41-42 below.

[0186] In addition, the embodiment of the present application further provides another possible implementation of S41, which may specifically include S411-S412:

[0187] S411: Filter out at least one preliminarily selected sentence satisfying a second matching condition from at least one semantically matching sentence according to the semantic matching scores between the user record to be used and each semantically matching sentence.

[0188] The above-mentioned "second matching condition" can be preset, for example, it can be specifically: the semantic matching score between the user record to be used and the primary selected sentence is higher than the preset second score threshold. For another example, when all semantic matching sentences are sorted from large to small according to the semantic matching score, the "second matching condition" can be specifically: the arrangement sequence number is lower than the preset fourth sequence number threshold.

[0189] The above-mentioned “preliminary selected sentences” refer to semantically matching sentences that meet the second matching condition.

[0190] Based on the relevant content of S411, it can be known that after obtaining the semantic matching scores between the user record to be used and each semantic matching sentence, multiple semantic matching sentences with higher semantic matching scores (for example, TopK) can be screened out from all the semantic matching sentences based on these semantic matching scores, and they can be determined as preliminary sentences respectively, so that the semantic information carried by these preliminary sentences is closer to the semantic information carried by the user record to be used, so that the sentences used to represent the semantic information carried by the user record to be used can be determined from these preliminary sentences subsequently.

[0191] S412: Filter out at least one target sentence satisfying a third matching condition from at least one of the at least one initially selected sentences according to the semantic matching scores between the user record to be used and each of the initially selected sentences.

[0192] The above-mentioned "third matching condition" can be preset, for example, it can specifically be: the condition achieved by "at least one target sentence" shown in steps 41-42 below (for example, the condition shown in formula (6) below).

[0193] The embodiment of the present application does not limit the implementation method of S412. For example, it can be implemented using any implementation method of the process of determining "at least one target sentence" shown in steps 41-42 below. It only needs to replace the "candidate sentence" with the "preliminary sentence" in any implementation method of the process of determining "at least one target sentence" shown in steps 41-42 below.

[0194] Based on the relevant contents of S411 to S412 above, it can be known that after obtaining the semantic matching results to be used, the semantic matching scores recorded in the semantic matching results to be used can be used to determine the preliminary sentences that meet the second matching condition; then the preliminary sentences that meet the third matching condition can be determined as target sentences respectively, so that the target sentences can meet both the second matching condition and the third matching condition, so that these target sentences can better represent the semantic information carried by the user records to be used.

[0195] In addition, the embodiment of the present application also provides another possible implementation of S41. In this implementation, S41 not only includes the above S411, but also may include S413-S414:

[0196] S413: Filtering out at least one re-selected sentence satisfying a fourth matching condition from at least one of the pre-selected sentences according to the number of non-stop words in each pre-selected sentence.

[0197] The above-mentioned “number of non-stop words” refers to the number of all words in a sentence except stop words.

[0198] The above-mentioned "fourth matching condition" can be preset, for example, it can be specifically: the number of non-stop words exceeds a preset number threshold (for example, 1).

[0199] The above-mentioned “re-selected sentence” refers to the initially selected sentence that meets the fourth matching condition (that is, the semantically matching sentence that meets both the second matching condition and the fourth matching condition).

[0200] Based on the relevant content of S413, it can be known that after obtaining at least one preliminary selected sentence, the number of non-stop words in each preliminary selected sentence can be counted first; then each preliminary selected sentence with a relatively high number of non-stop words can be determined as a re-selected sentence, so that filtering processing can be achieved for low-information sentences, thereby effectively avoiding the adverse effects of these low-information sentences on the minutes determination process.

[0201] S414: Filter out at least one target sentence satisfying a third matching condition from at least one re-selected sentence according to the semantic matching scores between the user record to be used and each re-selected sentence.

[0202] It should be noted that the embodiment of the present application does not limit the implementation method of S414. For example, it can be implemented using any implementation method of the "at least one target sentence" determination process shown in steps 41-42 below. It only needs to replace the "candidate sentence" with the "re-selected sentence" in any implementation method of the "at least one target sentence" determination process shown in steps 41-42 below.

[0203] Based on the relevant contents of S413 to S414 above, it can be known that after obtaining the semantic matching results to be used, the semantic matching scores recorded in the semantic matching results to be used can be used to determine the preliminary sentences that meet the second matching condition; then, the low-information sentences are filtered out from these preliminary sentences to obtain at least one re-selected sentence; finally, the re-selected sentences that meet the third matching condition are respectively determined as target sentences, so that the target sentences can simultaneously meet the second matching condition, the third matching condition and the fourth matching condition, so that these target sentences can better represent the semantic information carried by the user records to be used.

[0204] In addition, in order to better determine the above-mentioned "at least one target sentence", the embodiment of the present application also provides a possible implementation method for determining the above-mentioned "at least one target sentence", and for ease of understanding, it is explained below with examples.

[0205] As an example, when the above-mentioned "at least one target sentence" is screened out from at least one candidate sentence, the determination process of the "at least one target sentence" may specifically include step 41-step 42:

[0206] Step 41: According to the semantic matching scores between the user record to be used and each candidate sentence, a sentence to be deleted that meets a first discarding condition is screened out from at least one candidate sentence to obtain a first screening result.

[0207] The above-mentioned "at least one candidate sentence" includes the above-mentioned "at least one target sentence"; and the embodiments of the present application do not limit the above-mentioned "at least one candidate sentence". For example, when the above S41 is implemented by steps 41-42, the above-mentioned "at least one candidate sentence" refers to the above "at least one semantic matching sentence". For another example, when the above S412 is implemented by steps 41-42, the above-mentioned "at least one candidate sentence" refers to the above "at least one preliminary sentence". For another example, when the above S414 is implemented by steps 41-42, the above-mentioned "at least one candidate sentence" refers to the above "at least one re-selected sentence".

[0208] The above-mentioned "first discarding condition" may be preset, for example, it may specifically be: the semantic matching score between the user record to be used and the sentence to be deleted is lower than a preset third score threshold (for example, 0.6).

[0209] The above-mentioned “sentences to be deleted” refer to candidate sentences that meet the first discarding condition.

[0210] The above-mentioned “first screening result” is used to indicate whether there is a sentence to be deleted in the above-mentioned “at least one candidate sentence”.

[0211] Based on the relevant content of the above step 41, it can be known that after obtaining at least one candidate sentence, candidate sentences with relatively low semantic matching scores can be selected from these candidate sentences and determined as sentences to be deleted, so that the sentences that need to be deleted can be determined from these sentences to be deleted later.

[0212] Step 42: Determine at least one target sentence based on the first screening result and at least one candidate sentence.

[0213] As an example, step 42 may specifically include steps 421 to 424:

[0214] Step 421: According to the first screening result, determine whether there is a sentence to be deleted in the above-mentioned "at least one candidate sentence", if so, execute steps 423-424; if not, execute step 422.

[0215] In an embodiment of the present application, after obtaining the first screening result, if the first screening result indicates that there is no sentence to be deleted in the above-mentioned "at least one candidate sentence", it can be determined that the semantic matching scores of all candidate sentences are relatively high, so that it can be determined that the semantic information carried by all candidate sentences is very close to the semantic information carried by the user records to be used; if the first screening result indicates that there is at least one sentence to be deleted in the above-mentioned "at least one candidate sentence", it can be determined that there are some candidate sentences with relatively low semantic matching scores in the above-mentioned "at least one candidate sentence", so that it can be determined that the semantic information carried by these candidate sentences with relatively low semantic matching scores and the semantic information carried by the user records to be used may be quite different.

[0216] Step 422: Each candidate sentence may be determined as a target sentence.

[0217] In an embodiment of the present application, if the first screening result indicates that there is no sentence to be deleted in the above-mentioned "at least one candidate sentence", it can be determined that the semantic information carried by all candidate sentences is very close to the semantic information carried by the user record to be used, so each candidate sentence can be directly determined as a target sentence, so that these target sentences can represent the sentence corresponding to the user record to be used in the record text to be used.

[0218] Step 423: Determine at least one sentence to be retained based on at least one sentence to be deleted and at least one candidate sentence.

[0219] The above-mentioned “sentences to be retained” refer to sentences that need to be retained (that is, sentences that do not need to be deleted).

[0220] The embodiment of the present application does not limit the implementation method of step 423. For example, it can be specifically as follows: if the first screening result indicates that there is at least one sentence to be deleted in the above-mentioned "at least one candidate sentence", all sentences to be deleted can be deleted from the "at least one candidate sentence" to obtain at least one sentence to be retained.

[0221] In some cases, due to defects in the user record to be used (for example, too brief or too general), the semantic matching scores of all candidate sentences obtained by screening based on the user record to be used are relatively low. Therefore, in order to improve the positioning accuracy of the user record information, the embodiment of the present application also provides another possible implementation of step 423, which may specifically include steps 4231-4235:

[0222] Step 4231: If at least one candidate sentence satisfies a preset correction condition, at least one sentence to be evaluated that satisfies a first retention condition is selected from at least one sentence to be deleted.

[0223] The above-mentioned "preset correction conditions" can be set in advance. For example, they can specifically be: the semantic matching scores between the user records to be used and each candidate sentence are not higher than the preset fourth score threshold (that is, the semantic matching scores between the user records to be used and each candidate sentence are relatively low).

[0224] The above-mentioned "first retention condition" can be preset, for example, it can be specifically: the semantic matching score between the user record to be used and the sentence to be evaluated is higher than the preset fifth score threshold (for example, 0.3), and the fifth score threshold is lower than the above "third score threshold".

[0225] The above “sentences to be evaluated” refer to sentences to be deleted that meet the first retention condition. For example, when the above “third score threshold” is 0.6 and the above “fifth score threshold” is 0.3, the semantic matching score between the user record to be used and the sentence to be evaluated is between 0.6 and 0.3.

[0226] Based on the relevant content of the above step 4231, it can be known that after obtaining at least one sentence to be deleted, if it is determined that the above "at least one candidate sentence" meets the preset correction conditions, it can be determined that there are defects in the user records to be used, so some sentences to be deleted with relatively high semantic matching scores can be determined as sentences to be evaluated, so that some sentences that need to be retained can be selected from these sentences to be evaluated later.

[0227] Step 4232: normalize the semantic matching scores between the user record to be used and each sentence to be evaluated to obtain a normalized score of each sentence to be evaluated.

[0228] It should be noted that the embodiment of the present application does not limit the implementation method of step 4232. For example, any existing or future data normalization method can be used for implementation. For another example, step 4232 can be implemented using the normalization method shown in formulas (3)-(5):

[0229]

[0230]

[0231]

[0232] In the formula, score_norm l Represents the normalized score of the lth sentence to be evaluated; score l represents the semantic matching score between the user record to be used and the lth sentence to be evaluated; l is a positive integer, l≤L, L is a positive integer, and L represents the number of sentences to be evaluated; Score d represents the semantic matching score between the user record to be used and the dth candidate sentence; d is a positive integer, d≤D, D is a positive integer, and D represents the number of candidate sentences; σ represents the variance value of the semantic matching score of the above "at least one candidate sentence"; μ represents the average value of the semantic matching score of the above "at least one candidate sentence".

[0233] Step 4233: According to the normalized score of each sentence to be evaluated, at least one first retained sentence satisfying the second retention condition is selected from at least one sentence to be evaluated.

[0234] The above-mentioned "second retention condition" may be preset, for example, it may specifically include: the semantic matching score between the user record to be used and the first reserved sentence is higher than a preset sixth score threshold (for example, 0.3).

[0235] The above-mentioned “first reserved sentence” refers to the sentence to be evaluated that meets the second reserved condition.

[0236] Based on the relevant content of the above step 4233, it can be known that after obtaining the normalized scores of each sentence to be evaluated, some sentences to be evaluated with relatively high normalized scores can be determined as first retained sentences, so that the semantic information carried by these first retained sentences is closer to the semantic information carried by the user records to be used.

[0237] Step 4234: Delete at least one sentence to be deleted from at least one candidate sentence to obtain at least one second retained sentence.

[0238] It should be noted that the embodiments of the present application do not limit the execution order between step 4234 and the above step 4231. For example, step 4234 and the above step 4231 can be executed in sequence, step 4231 and the above step 4234 can be executed in sequence, or step 4234 and the above step 4231 can be executed simultaneously.

[0239] Step 4235: Determine at least one sentence to be retained based on at least one first retained sentence and at least one second retained sentence.

[0240] In the embodiment of the present application, after obtaining at least one first reserved sentence and at least one second reserved sentence, each first reserved sentence and each second reserved sentence can be determined as a sentence to be retained, so that these sentences to be retained all meet the conditions shown in formula (6), so that the target sentence corresponding to the user record to be used can be selected from these sentences to be retained.

[0241] (sc b ≥t1)or((t1>sc b >t2)&&(sc_norm b >t3)) (6)

[0242] In the formula, sc b Indicates the semantic matching score between the user record to be used and the bth sentence to be retained; b is a positive integer, b≤B, B is a positive integer, and B represents the number of sentences to be retained; sc_norm b represents the normalized score of the bth sentence to be retained; t1 represents the third score threshold; t2 represents the fifth score threshold; t3 represents the sixth score threshold.

[0243] Based on the relevant content of the above step 423, it can be known that if the first screening result indicates that there is at least one sentence to be deleted in the above "at least one candidate sentence", then these sentences to be deleted can be referred to and some sentences to be retained can be selected from these candidate sentences so that the semantic information carried by these sentences to be retained is closer to the semantic information carried by the user records to be used.

[0244] Step 424: Determine at least one target sentence based on at least one sentence to be retained.

[0245] It should be noted that the embodiment of the present application does not limit the implementation method of step 424. For example, it may specifically be: each sentence to be retained is determined as a target sentence.

[0246] Based on the relevant contents of steps 41 to 42 above, it can be known that after obtaining at least one candidate sentence, one or more target sentences can be selected from these candidate sentences according to a preset sentence selection strategy (that is, the sentence selection strategy shown in steps 41-42), so that these target sentences can express the semantic information carried by the user records to be used.

[0247] Based on the relevant content of S41 above, it can be known that after obtaining the semantic matching result to be used, you can first refer to the semantic matching score recorded in the semantic matching result to be used, screen out some semantic matching sentences with relatively high semantic matching scores, and determine them as preliminary sentences respectively; then, remove those low-information sentences with the number of non-stop words less than or equal to 1 from these preliminary sentences to obtain some re-selected sentences with relatively high information content; then, when it is determined that the semantic matching scores corresponding to these re-selected sentences are all relatively low, one or more target sentences can be selected from these re-selected sentences according to the conditions shown in formula (6), so that these target sentences can express the semantic information carried by the user records to be used.

[0248] S42: Determine the content of the minutes to be used according to the user record to be used and at least one target sentence.

[0249] It should be noted that the embodiment of the present application does not limit the implementation method of S42. For example, Figure 2 As shown, S42 can specifically be: these target sentences can be highlighted to the user in the record text to be used, and the user records to be used can be displayed around these target sentences, so that the user can write the minutes content to be used based on these displayed contents, so that the user can know the content required for writing the minutes from these displayed contents, thereby effectively avoiding the time consumption caused by the user determining the meeting content related to the user record to be used by re-listening to the meeting recording audio, thereby effectively improving the minutes determination efficiency.

[0250] It should also be noted that the embodiment of the present application does not limit the display triggering condition of the above-mentioned "at least one target sentence". For example, it can be displayed in response to a user record location display request triggered by the user. Among them, the "user record location display request" is used to request to display the location of the target sentence corresponding to the user record in the record text to be used; and the embodiment of the present application does not limit the triggering method of the "user record location display request". For example, the user can click Figure 2 The telescope button shown triggers the "User Record Position Display Request".

[0251] In addition, in order to further improve the effectiveness of the minutes, S42 can adopt the following Method Example 5 as well as Method Example 6 Any of the embodiments of S42 shown is implemented.

[0252] Based on the above S41 to S42, it can be known that after obtaining the semantic matching results to be used, the target sentences corresponding to the user records to be used can be determined from the record text to be used based on the semantic matching results to be used; and then the content of the minutes to be used can be determined based on these target sentences and the user records to be used. Among them, because these target sentences can accurately express the content of the meeting minutes corresponding to the user records to be used, so that these target sentences can more accurately express the semantic information carried by the user records to be used, so that the content of the minutes to be used determined based on these target sentences is more accurate, which is conducive to improving the effect of determining the minutes.

[0253] Method Example 5

[0254] The inventors also found that in the process of listening to the meeting, the relevant personnel need to multitask, judge whether the meeting content is important in real time, and organize the language to form meeting notes. However, due to the excessive distraction of the relevant personnel, errors may occur in the process of writing and judging the meeting notes (for example, misjudging the importance of some matters), or the relevant personnel may miss records due to participating in the meeting discussion or temporarily leaving the meeting, which may lead to inaccurate meeting minutes determined based on the meeting notes, thus making it difficult to finally form high-quality meeting minutes.

[0255] Based on the above findings, in order to further improve the effect of determining the minutes, the present application embodiment further provides another possible implementation of the above S42, which may specifically include S421-S422:

[0256] S421: Determine a model based on at least one target sentence, at least one sentence to be used, and a pre-constructed supplementary sentence to determine at least one key supplementary sentence.

[0257] The above-mentioned “key supplementary sentence” is used to indicate relatively important semantic information existing in the record text to be used but not recorded in the user record to be used; and the above-mentioned “at least one key supplementary sentence” can be used to indicate important semantic information omitted in the user record to be used.

[0258] The above-mentioned "supplementary sentence determination model" is used to perform key supplementary sentence determination processing on the input data of the supplementary sentence determination model; and the "supplementary sentence determination model" is a machine learning model (for example, a machine learning model based on a graph attention network (GAT) structure, etc.).

[0259] The embodiment of the present application does not limit the implementation of S421. For example, when the above-mentioned "supplementary sentence determination model" is a machine learning model based on a graph attention network (GAT) structure, S421 may specifically include S4211-S4213:

[0260] S4211: Construct a to-be-used graph according to at least one target sentence and at least one to-be-used sentence.

[0261] The above-mentioned “graph to be used” is used to represent the association relationship between different sentences in the record text to be used; and the embodiment of the present application does not limit the construction process of the “graph to be used”, for example, it may specifically include steps 51 to 54:

[0262] Step 51: Determine at least one graph node based on at least one target sentence and at least one sentence to be used, so that the “at least one graph node” includes at least one recorded sentence node and at least one candidate sentence node.

[0263] As an example, step 51 may specifically include steps 511 to 514:

[0264] Step 511: construct each record sentence node so that each record sentence node can represent each target sentence.

[0265] It should be noted that the embodiments of the present application do not limit the implementation method of step 511. For example, it may specifically include: determining the g-th target sentence as the g-th record sentence node. For another example, step 511 may specifically include: first performing semantic feature extraction processing on the target sentence to obtain the semantic features of the g-th target sentence; then determining the semantic features of the g-th target sentence as the g-th record sentence node, so that the g-th record sentence node can represent the semantic information carried by the g-th target sentence, so that the g-th record sentence node can represent the g-th target sentence. Wherein, g is a positive integer, g≤G, G is a positive integer, and G represents the number of target sentences.

[0266] The above-mentioned “semantic features” are used to represent the semantic information carried by a sentence; and the present application embodiment does not limit the determination process of the “semantic features”. For example, a pre-built semantic feature extraction model (for example, Figure 5 BERT shown in the figure).

[0267] Step 512: Delete at least one target sentence from at least one sentence to be used to obtain at least one non-recorded sentence.

[0268] Step 513: Determine each candidate sentence node according to the semantic features of each non-recorded sentence, so that each candidate sentence node can represent each non-recorded sentence respectively (that is, each sentence to be used in the above-mentioned "at least one sentence to be used" except the above-mentioned "at least one target sentence").

[0269] It should be noted that the embodiments of the present application do not limit the implementation methods of step 513. For example, it may specifically include: determining the ath non-recorded statement as the ath candidate statement node. For another example, step 513 may specifically include: first performing semantic feature extraction processing on the ath non-recorded statement to obtain the semantic features of the ath non-recorded statement; and then determining the semantic features of the ath non-recorded statement as the ath candidate statement node, so that the ath candidate statement node can represent the semantic information carried by the ath non-recorded statement, thereby enabling the ath candidate statement node to represent the ath non-recorded statement. Wherein, a is a positive integer, a≤A, A is a positive integer, and A represents the number of non-recorded statements.

[0270] Step 514: All recorded statement nodes and all candidate statement nodes are processed collectively to obtain at least one graph node, so that the "at least one graph node" includes at least one recorded statement node and at least one candidate statement node.

[0271] Based on the relevant content of the above step 51, it can be known that after obtaining at least one target sentence, these target sentences and the above "at least one sentence to be used" can be used to determine each node in the to-be-used graph so that each node in the to-be-used graph can represent each sentence to be used.

[0272] Step 52: Construct a neighbor representation edge corresponding to at least one graph node according to preset neighbor description parameters.

[0273] The above-mentioned “preset neighbor description parameter” is used to represent the neighbor range of an image node; and the embodiment of the present application does not limit the “preset neighbor description parameter”, for example, it may include: context window size.

[0274] The above-mentioned "context window size" is used to describe the context scope of a sentence; and the embodiment of the present application does not limit the "context window size", for example, it can be specifically 10 (that is, the 10 sentences adjacent to the text position of a sentence are all determined as the neighbors of the sentence).

[0275] The above-mentioned “neighbor representation edge” is used to represent the connecting edge between two graph nodes having a neighbor relationship, so that the “neighbor representation edge” is used to represent the neighbor relationship between the sentences represented by the two graph nodes.

[0276] In addition, the embodiments of the present application do not limit the determination process of the above-mentioned "neighbor representation edge". For example, it may specifically include: for a graph node, the context window corresponding to the sentence represented by the graph node can be determined according to preset neighbor description parameters; then the graph nodes corresponding to any two sentences in the context window corresponding to the sentence represented by the graph node are connected to obtain the neighbor representation edges between these graph nodes.

[0277] Based on the relevant contents of the above step 52, it can be known that after obtaining at least one graph node, a graph node pair having a neighbor relationship can be determined according to the preset neighbor description parameters; then the two graph nodes included in each graph node pair are connected to obtain the neighbor representation edges corresponding to these graph nodes (for example, Figure 5 The connecting edges shown in ① in the figure are used to represent the neighbor relationships between different graph nodes.

[0278] Step 53: Connect each candidate sentence node to each record sentence node respectively to obtain the relevant representation edge corresponding to each candidate sentence node.

[0279] The above-mentioned “related representation edge” is used to represent the connection relationship between a candidate sentence node and a record sentence node.

[0280] Based on the relevant content of step 53, it can be known that after obtaining at least one record sentence node and at least one candidate sentence node, each candidate sentence node can be connected to each record sentence node respectively, but different record sentence nodes cannot be connected to each other to obtain the relevant representation edges corresponding to these candidate sentence nodes (for example, Figure 5 The connecting edges are shown in ②).

[0281] Step 54: Determine a graph to be used based on at least one graph node, at least one neighbor representation edge corresponding to the graph node, and at least one related representation edge corresponding to the candidate sentence node.

[0282] In an embodiment of the present application, after obtaining at least one graph node, a neighbor representation edge corresponding to the at least one graph node, and a related representation edge corresponding to at least one candidate statement node, these graph nodes and their corresponding connecting edges can be referred to to construct a graph to be used, so that the graph to be used includes the at least one graph node, the neighbor representation edge corresponding to the at least one graph node, and the related representation edge corresponding to at least one candidate statement node.

[0283] It can be seen that for any graph node in the graph to be used, if the graph node belongs to a record sentence node (that is, the graph node is used to represent a target sentence), then the graph node only corresponds to one type of connection edge (that is, the neighbor representation edge); however, if the graph node belongs to a candidate sentence node (that is, the graph node is used to represent a non-record sentence), then the graph node can correspond to two types of connection edges (that is, the neighbor representation edge and the related representation edge).

[0284] Based on the relevant content of S4211 above, it can be known that after obtaining at least one target sentence, these target sentences and the above "at least one sentence to be used" can be used to construct a graph to be used, so that each node in the graph to be used represents each sentence to be used, and each connecting edge in the graph to be used can represent the connection relationship between different nodes, so that key supplementary sentences can be determined and processed based on the graph to be used later.

[0285] S4212: Input the graph to be used into the supplementary sentence determination model to obtain the key sentence prediction result output by the supplementary sentence determination model.

[0286] The above-mentioned “key sentence prediction result” is used to indicate the possibility that each sentence to be used belongs to a key supplementary sentence (especially, the possibility that each non-recorded sentence belongs to a key supplementary sentence).

[0287] The embodiment of the present application does not limit the determination process of the above-mentioned “key sentence prediction result”. For example, when the above-mentioned “supplementary sentence determination model” includes a graph attention module and a classification prediction layer, the determination process of the “key sentence prediction result” may specifically include steps 61 to 63:

[0288] Step 61: Determine the features to be used of each graph node in the graph to be used.

[0289] The "features to be used of the rth graph node" are used to represent the semantic information carried by the sentence to be used represented by the rth graph node. Where r is a positive integer, r≤R, R is a positive integer, and R represents the number of graph nodes.

[0290] In addition, the embodiment of the present application does not limit the determination process of the above-mentioned "features to be used of the rth graph node". For ease of understanding, the following is an explanation combining two situations.

[0291] Case 1: when the rth graph node is represented by the semantic features of the rth sentence to be used, the "features to be used of the rth graph node" are the rth graph node (that is, the semantic features of the rth sentence to be used).

[0292] Case 2: When the r-th graph node is represented by the r-th sentence to be used, the above-mentioned "supplementary sentence determination model" may also include a coding layer, and the determination process of the "features to be used of the r-th graph node" may specifically be: using the coding layer, encoding the r-th graph node (that is, the r-th sentence to be used) to obtain the features to be used of the r-th graph node.

[0293] The above-mentioned "coding layer" is used to perform encoding processing on the input data of the coding layer; and the embodiment of the present application does not limit the implementation method of the "coding layer", for example, it can be implemented using the BERT pre-training model.

[0294] Step 62: Input the to-be-used features of at least one graph node and the to-be-used graph into the graph attention module to obtain the graph attention features of at least one graph node output by the graph attention module.

[0295] The above-mentioned "graph attention module" is used to perform graph attention processing on the input data of the graph attention module; and the embodiment of the present application does not limit the implementation method of the "graph attention module", for example, it may include M graph attention layers (for example, 2 graph attention layers). Among them, the input data of the first graph attention layer includes the above-mentioned "features to be used of at least one graph node" and the graph to be used; and the input data of the m1th graph attention layer includes the output data of the m1-1th graph attention layer and the graph to be used, m1 is a positive integer, 2≤m1≤M, and M is a positive integer.

[0296] The "m2th graph attention layer" is used to perform graph attention processing on the input layer of the m2th graph attention layer; and the embodiment of the present application does not limit the working principle of the "m2th graph attention layer". For ease of understanding, the following is an example. Wherein, m2 is a positive integer, 1≤m2≤M.

[0297] As an example, when the input data of the m2-th graph attention layer includes the pre-processed features of R graph nodes and the graph to be used, and the output data of the m2-th graph attention layer includes the processed features of R graph nodes, the process of determining the processed features of the r-th graph node (that is, the working principle of the above-mentioned "m2-th graph attention layer") may specifically include steps 71 to 74:

[0298] Step 71: Determine at least one neighbor node corresponding to the rth graph node according to the neighbor representation edge corresponding to the rth graph node, where r is a positive integer, r≤R.

[0299] In the embodiment of the present application, for the rth graph node, the other endpoints of each neighbor representation edge connected to the rth graph node can be respectively determined as each neighbor node corresponding to the rth graph node. It can be seen that the rth graph node and each neighbor node corresponding to the rth graph node are connected by the neighbor representation edge.

[0300] Step 72: Based on the pre-processing feature of the r-th graph node and the d-th graph node corresponding to the r-th graph node r The pre-processing features of the neighboring nodes determine the dth r The neighbor attention weight features of neighbor nodes. Among them, d r is a positive integer, d r ≤D r , D r is a positive integer, D r Represents the number of neighboring nodes corresponding to the rth graph node.

[0301] The above-mentioned “pre-processing features of the r-th graph node” refers to the feature representation used to describe the r-th graph node included in the above-mentioned “input data of the m2-th graph attention layer”.

[0302] In addition, the embodiments of the present application do not limit the above-mentioned “pre-processing features of the r-th graph node”. For example, if m2=1, then the “pre-processing features of the r-th graph node” may be the features to be used of the r-th graph node; if m2≥2, then the “pre-processing features of the r-th graph node” may be the feature representation used to describe the r-th graph node included in the above-mentioned “output data of the m2-1-th graph attention layer” (that is, the feature representation obtained after performing m2-1 graph attention processing on the features to be used of the r-th graph nodes).

[0303] The above “Article d r The neighbor attention weight feature of the rth neighbor node is the weight of the neighbor attention weight feature of the rth neighbor node. r The pre-processing features of the dth neighboring nodes are obtained by semantic interaction; and the embodiment of the present application does not limit the “dth r The process of determining the "neighbor attention weight features of neighboring nodes" can be implemented, for example, using formula (7).

[0304]

[0305] In the formula, Indicates the dth node corresponding to the rth graph node r The neighbor attention weight feature of the neighbor nodes; h r represents the pre-processing features of the r-th graph node; Indicates the dth node corresponding to the rth graph node r The pre-processing features of neighboring nodes; h x Represents the pre-processing feature of the xth neighbor node corresponding to the rth graph node, and x is a positive integer, x≤D r ; W1 and θ1 are both layer parameters of the m2th graph attention layer, and these layer parameters can be determined in the construction process of the above-mentioned “supplementary sentence determination model”; ReLu() represents the activation function.

[0306] It should be noted that the embodiment of the present application does not limit the implementation method of the above formula (7). For example, it can be implemented with the help of an attention layer, an activation function layer, and a normalization processing layer.

[0307] Step 73: According to the D corresponding to the rth graph node r The pre-processing features of the neighboring nodes and the D corresponding to the rth graph node r The neighbor attention weight features of the rth neighbor nodes are used to determine the neighbor fusion features of the rth graph node.

[0308] It should be noted that the embodiment of the present application does not limit the implementation method of step 73. For example, it can be implemented specifically using formula (8).

[0309]

[0310] In the formula, Represents the neighbor fusion feature of the rth graph node; Indicates the dth node corresponding to the rth graph node r Neighbor attention weight features of neighboring nodes; Indicates the dth node corresponding to the rth graph node r The pre-processing features of the neighboring nodes; W2 belongs to the layer parameters of the m2th graph attention layer; δ() represents the sigmoid function.

[0311] Step 74: Determine the processed features of the r-th graph node according to the neighbor fusion features of the r-th graph node.

[0312] The embodiment of the present application does not limit the implementation of step 74. For example, it can be implemented using formula (9).

[0313]

[0314] In the formula, h′ r Represents the processed features of the rth graph node; represents the neighbor fusion feature of the rth graph node; W3 belongs to the layer parameter of the m2th graph attention layer; δ() represents the sigmoid function.

[0315] Based on the relevant contents of steps 71 to 74 above, it can be known that for the m2th graph attention layer, the m2th graph attention layer can perform graph attention processing on the input data of the m2th graph attention layer with the help of formulas (7)-(9) to obtain the output data of the m2th graph attention layer.

[0316] In fact, since each candidate sentence node in the to-be-used graph corresponds to two types of connection edges (for example, neighbor representation edges and related representation edges), in order to better perform graph attention processing, the embodiment of the present application also provides another possible implementation method for determining the above-mentioned "processed features of the r-th graph node". In this implementation method, the determination process of the "processed features of the r-th graph node" includes not only the above-mentioned steps 71-73, but also steps 75-77:

[0317] Step 75: Determine the fused information difference feature of the rth graph node.

[0318] The above-mentioned “fused information difference feature of the r-th graph node” is used to represent the semantic supplementary information of the r-th graph node for the user record to be used.

[0319] In addition, the embodiment of the present application does not limit step 75, for example, it may specifically include steps 751 to 755:

[0320] Step 751: Determine whether the rth graph node is a candidate statement node. If so, execute steps 752 to 754; if not, execute step 755.

[0321] In an embodiment of the present application, for the m2th graph attention layer, after the m2th graph attention layer obtains the pre-processing features of the rth graph node, it can be determined whether the rth graph node belongs to a candidate sentence node; if the rth graph node belongs to a candidate sentence node, it can be determined that not only the neighbor representation edges of the rth graph node exist, but also the related representation edges of the rth graph node exist, so these two types of connection edges can be combined to perform graph attention processing on the rth graph node; if the rth graph node does not belong to a candidate sentence node, it can be determined that the rth graph node belongs to a record sentence node, and thus it can be determined that only the neighbor representation edges of the rth graph node exist, so only the neighbor representation edges of the rth graph node can be referred to, and graph attention processing can be performed on the rth graph node.

[0322] Step 752: Determine the g-th recorded sentence attention weight feature corresponding to the r-th graph node according to the pre-processing feature of the r-th graph node and the pre-processing feature of the g-th recorded sentence node, where g is a positive integer, g≤G, and G is a positive integer.

[0323] The above-mentioned “attention weight feature of the g-th record sentence corresponding to the r-th graph node” is used to represent the semantic interaction result between the r-th graph node and the g-th record sentence node.

[0324] In addition, the embodiment of the present application does not limit the determination process of the above-mentioned "attention weight feature of the g-th recorded sentence corresponding to the r-th graph node". For example, it can be implemented using formula (10).

[0325]

[0326] In the formula, represents the attention weight feature of the g-th record sentence corresponding to the r-th graph node; h r represents the pre-processing features of the r-th graph node; h g represents the pre-processing features of the g-th record statement node; h z It represents the pre-processing features of the z-th record sentence node and is a positive integer z, z≤G; W4 and θ2 are both layer parameters of the m2-th graph attention layer, and these layer parameters can be determined in the construction process of the above-mentioned "supplementary sentence determination model"; ReLu() represents the activation function.

[0327] It should be noted that the embodiment of the present application does not limit the implementation method of the above formula (10). For example, it can be implemented with the help of an attention layer, an activation function layer, and a normalization processing layer.

[0328] Step 753: Determine the record fusion feature of the rth graph node based on the pre-processing features of the G record sentence nodes and the G record sentence attention weight features corresponding to the rth graph node.

[0329] It should be noted that the embodiment of the present application does not limit the implementation method of step 753. For example, it can be implemented specifically using formula (11).

[0330]

[0331] In the formula, Represents the record fusion features of the rth graph node; represents the attention weight feature of the g-th record sentence corresponding to the r-th graph node; h g represents the pre-processing features of the g-th record sentence node; W5 belongs to the layer parameter of the m2-th graph attention layer; δ() represents the sigmoid function.

[0332] Step 754: Determine the fused information difference feature of the rth graph node based on the neighbor fusion feature of the rth graph node and the record fusion feature of the rth graph node.

[0333] It should be noted that the embodiment of the present application does not limit the implementation method of step 754. For example, it can be implemented specifically using formula (12).

[0334]

[0335] In the formula, Differ r Represents the fused information difference feature of the rth graph node; Represents the neighbor fusion feature of the rth graph node; Represents the record fusion features of the r-th graph node.

[0336] Step 755: Determine the preset zero vector as the fused information difference feature of the r-th graph node.

[0337] The above-mentioned “preset zero vector” is a full zero vector; and the embodiment of the present application does not limit the data dimension of the “preset zero vector”. For example, the data dimension of the “preset zero vector” is the same as the above-mentioned The data dimension of is consistent to ensure that the feature representation of each record sentence node and the feature representation of each candidate sentence node always maintain the same data dimension.

[0338] Based on the relevant content of the above step 75, it can be known that for the m2th graph attention layer, after the m2th graph attention layer obtains the pre-processing features of the rth graph node, the m2th graph attention layer can determine the fused information difference features of the rth graph node according to the type of connection edges of the rth graph node.

[0339] Step 76: Concatenate the neighbor fusion feature of the r-th graph node with the fused information difference feature of the r-th graph node to obtain the fused concatenated feature of the r-th graph node.

[0340] Step 77: Determine the processed features of the rth graph node according to the fused and concatenated features of the rth graph node.

[0341] The embodiment of the present application does not limit the implementation method of step 77. For example, it can be implemented using formula (13).

[0342]

[0343] In the formula, h′ r Represents the processed features of the rth graph node; Represents the neighbor fusion feature of the rth graph node; Differ rrepresents the fused information difference feature of the rth graph node; || represents the splicing process; represents the fused concatenated features of the rth graph node; W6 belongs to the layer parameter of the m2th graph attention layer; δ() represents the sigmoid function.

[0344] Based on the relevant contents of steps 75 to 77 above, it can be known that for the m2th graph attention layer, the m2th graph attention layer can use formulas (7)-(8) and (10)-(13) to perform graph attention processing on the input data of the m2th graph attention layer to obtain the output data of the m2th graph attention layer.

[0345] Based on the above-mentioned relevant content of the “graph attention module”, it can be known that the “graph attention module” can include multiple graph attention layers, and each graph attention layer can implement the graph attention processing process with the help of formulas (7)-(8), and (10)-(13).

[0346] Based on the relevant content of the above step 62, it can be known that for the supplementary sentence determination model, after obtaining the features to be used of each graph node in the graph to be used, the graph attention module in the supplementary sentence determination model can perform graph attention processing on the features to be used of these graph nodes to obtain and output the graph attention features of each graph node.

[0347] Step 63: Input the graph attention feature of at least one graph node into the classification prediction layer to obtain the key sentence prediction result output by the classification prediction layer.

[0348] The above-mentioned "classification prediction layer" is used to perform binary classification prediction processing on the input data of the classification prediction layer (that is, to determine whether each sentence to be used is a key sentence); and the embodiment of the present application does not limit the implementation method of the "classification prediction layer", for example, it can be implemented using any existing or future classification network.

[0349] Based on the relevant contents of the above steps 61 to 63, it can be known that for the supplementary sentence determination model including the graph attention module and the classification prediction layer, the supplementary sentence determination model can be based on Figure 5 The key sentence extraction process shown performs key supplementary sentence determination processing on the above "at least one sentence to be used" to obtain and output key sentence prediction results.

[0350] In addition, the above-mentioned "supplementary sentence determination model" can be constructed in advance based on the key sentence annotated corpus; and the embodiment of the present application does not limit the construction process of the above-mentioned "supplementary sentence determination model", for example, it can be implemented using any existing or future model construction method.

[0351] The above-mentioned “key sentence annotated corpus” includes a second conference text sentence set, sample recorded sentences corresponding to the second conference text sentence set, and key supplementary sentence labels corresponding to the second conference text sentence set.

[0352] The above-mentioned "second meeting text sentence set" is obtained by sentence segmentation of the meeting transcription text of the sample meeting; and the embodiment of the present application does not limit the relationship between the "second meeting text sentence set" and the above-mentioned "first meeting text sentence set", for example, the two may be the same or different.

[0353] The above-mentioned "sample recorded sentences corresponding to the second conference text sentence set" is used to indicate that the user records corresponding to the second conference text sentence set carry semantic information; and the "sample recorded sentences corresponding to the second conference text sentence set" belongs to the second conference text sentence set. It should be noted that the embodiment of the present application does not limit the acquisition process of the above-mentioned "sample recorded sentences corresponding to the second conference text sentence set", for example, it can be manually annotated. For another example, it can be determined with the help of the above-mentioned "semantic matching model".

[0354] The above-mentioned “key supplementary sentence labels corresponding to the second conference text sentence set” are used to represent the actual key supplementary sentences in the second conference text sentence set; and the embodiment of the present application does not limit the “key supplementary sentence labels corresponding to the second conference text sentence set”, for example, it can be expressed as {(s1, lab1), (s2, lab2), …, (s U ,lab U )}. Among them, s u Represents the uth sentence in the second conference text sentence set. u It is used to indicate whether the u-th sentence in the second conference text sentence set is actually a key supplementary sentence, and lab u =1, it means that the uth sentence is actually a key supplementary sentence. u =0, it means that the u-th sentence is not actually a key supplementary sentence.

[0355] Based on the relevant content of S4212 above, it can be known that after obtaining the graph to be used, the pre-built supplementary sentence determination model can be used to perform key sentence determination processing on the graph to be used, and the key sentence prediction result corresponding to the graph to be used can be obtained and output, so that the key sentence prediction result can indicate the possibility that each sentence to be used belongs to the key supplementary sentence (especially, the possibility that each non-recorded sentence belongs to the key supplementary sentence).

[0356] S4213: Determine at least one key supplementary sentence based on the key sentence prediction result.

[0357] As an example, when the above-mentioned "key sentence prediction results" include supplementary sentence scores of each sentence to be used, S4213 may specifically be: according to the supplementary sentence scores of each sentence to be used, each key supplementary sentence that meets the preset key supplementary condition is screened out from all sentences to be used.

[0358] The above-mentioned "preset key supplementary condition" can be preset, for example, it can be specifically: the supplementary sentence score is higher than the preset seventh score threshold. For another example, after all the sentences to be used are sorted from large to small according to the supplementary sentence scores, the "preset key supplementary condition" can be specifically: the arrangement sequence number is lower than the preset fifth sequence number threshold.

[0359] Based on the relevant content of S4213 above, it can be known that after obtaining the key sentence prediction result, the supplementary sentence score recorded in the key sentence prediction result can be referred to to determine some sentences to be used with relatively high supplementary sentence scores, all of which are determined as key supplementary sentences, so that the key supplementary sentences can better express the semantic supplementary information for the above "user records to be used".

[0360] Based on the relevant content of S421 above, it can be known that after obtaining at least one target sentence corresponding to the user record to be used, the supplementary sentence determination model can refer to these target sentences to determine some key supplementary sentences from all the sentences to be used, so that these key supplementary sentences can express the semantic supplementary information for the above "user record to be used", so that the minutes content can be determined and processed based on these key supplementary sentences in the future.

[0361] S422: Determine the content of the minutes to be used according to the user record to be used, at least one target sentence, and at least one key supplementary sentence.

[0362] It should be noted that the embodiment of the present application does not limit the implementation method of S422. For example, Figure 2 As shown, S422 can specifically be: these key supplementary sentences, target sentences and user records to be used can be displayed to the user, so that the user can write the minutes content to be used based on these displayed contents, so that the user can obtain the relatively complete meeting record content required for writing the minutes from these displayed contents, thereby effectively avoiding the time consumption caused by the user determining the meeting content related to the user record to be used by re-listening to the meeting recording audio, thereby effectively improving the minutes determination efficiency.

[0363] It should also be noted that the embodiment of the present application does not limit the display triggering condition of the above-mentioned "at least one key supplementary sentence". For example, it can be displayed in response to a key supplementary sentence display request triggered by the user. Among them, the "key supplementary sentence display request" is used to request the display of the key supplementary sentence recorded by the user; and the embodiment of the present application does not limit the triggering method of the "key supplementary sentence display request". For example, the user can click Figure 2 The "check for missing items and fill in the gaps" button shown is used to trigger the "key supplementary sentence display request".

[0364] In addition, in order to further improve the effectiveness of the minutes, S422 can adopt the following Method Example 6 Any implementation of S422 shown is implemented.

[0365] Based on the relevant contents of S421 to S422 above, it can be known that after obtaining at least one target sentence corresponding to the user record to be used, you can first refer to these target sentences and determine some key supplementary sentences from the sentences to be used, so that these key supplementary sentences can express the semantic supplementary information for the above "user records to be used"; then refer to these key supplementary sentences, these target sentences, and the user records to be used to determine the content of the minutes to be used.

[0366] Method Example 6

[0367] In fact, in order to further improve the determination effect of the minutes, a key reminder sentence can be extracted from the key supplementary sentence, so that the key reminder sentence can describe the semantic information carried by the key supplementary sentence in a more concise content. Based on this, the embodiment of the present application also provides another possible implementation of S422, which may specifically include S4221-S4222:

[0368] S4221: Extract key points of each key supplementary sentence to obtain the key points of each key supplementary sentence.

[0369] The “main content of the f-th key supplementary sentence” refers to the core semantic information of the f-th key supplementary sentence. Where f is a positive integer, f≤F, F is a positive integer, and F represents the number of key supplementary sentences.

[0370] In addition, the embodiments of the present application do not limit the above-mentioned determination process of the "main content of the f-th key supplementary sentence". For example, any existing or future sentence core extraction method (for example, summary extraction method, sentence simplification method, etc.) can be adopted for implementation.

[0371] For example, in order to improve the effect of key content extraction, the embodiment of the present application also provides another possible implementation method for determining the above-mentioned "key content of the f-th key supplementary sentence", which may specifically include: determining the key content of the f-th key supplementary sentence based on the f-th key supplementary sentence and a pre-built key extraction model.

[0372] The above-mentioned “key point extraction model” is used to perform key point content extraction processing on the input data of the key point extraction model; and the “key point extraction model” can be a machine learning model.

[0373] In addition, the embodiments of the present application are not limited to the above-mentioned “key point extraction model”. For example, it can be a text generation model based on encoder-decoder (ED).

[0374] In addition, the embodiment of the present application does not limit the determination process of the above-mentioned "key points of the f-th key supplementary sentence", for example, it may specifically include: inputting the f-th key supplementary sentence into a pre-constructed key point extraction model, and obtaining the key points of the f-th key supplementary sentence output by the key point extraction model. The key point extraction model is constructed based on a large number of first sample sentences and the actual key points corresponding to the first sample sentences; and the "actual key points corresponding to the first sample sentences" can be determined by manual annotation.

[0375] In fact, the above-mentioned "key supplementary sentence" may have information missing (for example, incomplete information such as reference omission), so in order to improve the effect of extracting key content, the context information of the "key supplementary sentence" can be referred to and the key content of the "key supplementary sentence" can be extracted. Based on this, the embodiment of the present application also provides another possible implementation method for determining the above-mentioned "key content of the f-th key supplementary sentence", which can specifically include S42211-S42212:

[0376] S42211: Determine context information of the fth key supplementary sentence based on at least one sentence to be used.

[0377] The above-mentioned "context information of the f-th key supplementary sentence" is used to describe the contextual semantic information of the f-th key supplementary sentence; and the embodiment of the present application does not limit the "context information of the f-th key supplementary sentence", for example, it may include one or more sentences to be used.

[0378] In addition, the embodiment of the present application does not limit the determination process of the above-mentioned "context information of the f-th key supplementary sentence". For example, it may specifically be: all sentences to be used are determined as the context information of the f-th key supplementary sentence.

[0379] For another example, the determination process of the above “context information of the fth key supplementary sentence” may specifically be: according to the text position of the fth key supplementary sentence and the preset context window, the context information of the fth key supplementary sentence is extracted from the above “at least one sentence to be used” so that all sentences in the “context information of the fth key supplementary sentence” are located within the preset context window. The “preset context window” may be preset.

[0380] S42212: Input the f-th key supplementary sentence and the context information of the f-th key supplementary sentence into the key point extraction model to obtain the key point content of the f-th key supplementary sentence output by the key point extraction model.

[0381] The above-mentioned “key point extraction model” can be constructed based on a large number of second sample sentences, the context information of the second sample sentences and the actual key point content corresponding to the second sample sentences. Among them, the “actual key point content corresponding to the second sample sentences” can be determined by manual annotation.

[0382] In addition, the embodiments of the present application do not limit the network structure of the above-mentioned "key point extraction model". For example, any existing or future text generation model can be used for implementation.

[0383] In fact, the sentences closer to the fth key supplementary sentence are more likely to carry semantic information that is missing from the fth key supplementary sentence. Therefore, in order to further improve the key content extraction effect, the text position of each sentence in the context information of the fth key supplementary sentence can be referred to to perform key content extraction on the fth key supplementary sentence. Based on this, the embodiment of the present application also provides a possible implementation of the above-mentioned "key point extraction model", which can specifically include: an encoder, a position weighting module, an attention module, a fusion module, and a decoder.

[0384] In order to facilitate understanding of the working principle of the above-mentioned "key point extraction model", the following is an explanation taking the process of determining the above-mentioned "context information of the f-th key supplementary sentence" as an example.

[0385] As an example, when the above “context information of the f-th key supplementary sentence” includes K reference sentences, the determination process of the “context information of the f-th key supplementary sentence” may specifically include steps 81 to 85:

[0386] Step 81: Using an encoder, encode the f-th key supplementary sentence and each reference sentence respectively to obtain the encoding features of the f-th key supplementary sentence and the encoding features of each reference sentence.

[0387] The above-mentioned "encoder" is used to encode the input data of the encoder; and the embodiment of the present application does not limit the "encoder", for example, it can be implemented using any existing or future encoding network (for example, BERT pre-training model, etc.).

[0388] The above-mentioned "encoding features of the f-th key supplementary sentence" is used to represent the semantic information carried by the f-th key supplementary sentence; and the determination process of the "encoding features of the f-th key supplementary sentence" can be specifically: input the f-th key supplementary sentence into the encoder, and obtain the encoding features of the f-th key supplementary sentence output by the encoder.

[0389] The “encoding feature of the kth reference sentence” is used to represent the semantic information carried by the kth reference sentence; and the determination process of the “encoding feature of the kth reference sentence” may specifically be: using an encoder to encode the kth reference sentence to obtain the encoding feature of the kth reference sentence output by the encoder. Wherein, k is a positive integer, k≤K, and K is a positive integer.

[0390] Step 82: Input the encoding features of the K reference sentences into the position weighting module to obtain the position weighting features of the K reference sentences output by the position weighting module.

[0391] The “position weighted feature of the k-th reference sentence” is used to represent the semantic information and text position influence information carried by the k-th reference sentence.

[0392] The above-mentioned "position weighting module" is used to perform position weighting processing on the input data of the position weighting module; and the embodiment of the present application does not limit the working principle of the "position weighting module". For ease of understanding, the following is an explanation taking the determination process of the above-mentioned "position weighted feature of the kth reference sentence" as an example.

[0393] As an example, when the module parameters of the “position weighting module” include K position weighting vectors, the determination process of the “position weighting feature of the kth reference sentence” may specifically include steps 821-822:

[0394] Step 821: Determine the position excitation weight corresponding to the kth reference sentence according to the dot product between the kth position weighted vector and the encoding feature of the kth reference sentence.

[0395] The above-mentioned "kth position weighted vector" is used to indicate the degree of influence of the text position of the kth reference sentence on the above-mentioned "fth key supplementary sentence"; and the "kth position weighted vector" belongs to the module parameters of the above-mentioned "position weighted module", so that the "kth position weighted vector" can be determined during the construction process of the above-mentioned "key point extraction model".

[0396] In addition, the embodiment of the present application does not limit the data dimension of the "kth position weighted vector". For example, the data dimension of the "kth position weighted vector" is consistent with the data dimension of the above "encoding feature of the kth reference sentence".

[0397] The above-mentioned "position incentive weight corresponding to the k-th reference sentence" is used to indicate the degree to which the semantic information carried by the k-th reference sentence supplements the information of the above-mentioned "f-th key supplementary sentence"; moreover, the embodiment of the present application does not limit the determination process of the "position incentive weight corresponding to the k-th reference sentence", for example, it can be implemented using formula (14).

[0398] u k =δ(POSMatrix k T h f,[CLS] ) (14)

[0399] In the formula, u k represents the position excitation weight corresponding to the kth reference sentence; h f,[CLS] represents the encoding features of the string [k-th reference sentence, preset sentence identifier]; POSMatrix represents a position weighted matrix, the POSMatrix includes K position weighted vectors, and the POSMatrix belongs to the module parameters of the above-mentioned "position weighted module", so that the POSMatrix can be determined in the construction process of the above-mentioned "key point extraction model"; POSMatrix k Represents the kth position weighted vector; POSMatrix k T represents the transpose of the k-th position weight vector; δ() represents the sigmoid function.

[0400] Step 822: The product of the position excitation weight corresponding to the kth reference sentence and the encoding feature of the kth reference sentence is determined as the position weighted feature of the kth reference sentence (as shown in formula (15)).

[0401]

[0402] In the formula, represents the position weighted feature of the kth reference sentence; u k represents the position excitation weight corresponding to the kth reference sentence; h f represents the encoding features of the kth reference sentence; h f,[CLS]represents the encoding features of the string [k-th reference sentence, preset sentence identifier]; POSMatrix represents a position weighted matrix, the POSMatrix includes K position weighted vectors, and the POSMatrix belongs to the module parameters of the above-mentioned "position weighted module", so that the POSMatrix can be determined in the construction process of the above-mentioned "key point extraction model"; POSMatrix k Represents the kth position weighted vector; POSMatrix k T represents the transpose of the k-th position weight vector; δ() represents the sigmoid function.

[0403] Based on the relevant content of the above step 82, it can be known that for the key point extraction model, after obtaining the encoding features of the kth reference sentence, the encoding features of the kth reference sentence and the position weighted vector corresponding to the kth reference sentence can be fused with the help of the position weighted module in the key point extraction model to obtain the position weighted features of the kth reference sentence, so that the "position weighted features of the kth reference sentence" can better represent the semantic supplementary information carried by the kth reference sentence.

[0404] Step 83: Input the position weighted feature of the kth reference sentence and the encoding feature of the fth key supplementary sentence into the attention module to obtain the kth fusion feature corresponding to the fth key supplementary sentence output by the attention module. Wherein, k is a positive integer, k≤K, and K is a positive integer.

[0405] The above-mentioned "attention module" is used to perform attention processing on the input data of the attention module; and the embodiment of the present application does not limit the implementation method of the "attention module", for example, it can be implemented using a multi-head attention network.

[0406] The above-mentioned “kth fusion feature corresponding to the fth key supplementary sentence” is used to represent the fusion semantic information of the kth reference sentence and the fth key supplementary sentence.

[0407] In addition, the embodiments of the present application do not limit the acquisition process of the above-mentioned "kth fusion feature corresponding to the f-th key supplementary sentence". For example, when the above-mentioned "attention module" is implemented using a multi-head attention network, the encoding feature of the f-th key supplementary sentence can be used as the query parameter (that is, the Q parameter) required by the multi-head attention network, and the position weighted feature of the k-th reference sentence can be used as the other two parameters (that is, the K parameter and the V parameter) required by the multi-head attention network, so that the multi-head attention network can realize information fusion through the multi-head attention mechanism to obtain the "k-th fusion feature corresponding to the f-th key supplementary sentence".

[0408] Step 84: input the K fusion features corresponding to the f-th key supplementary sentence into the fusion module to obtain the representation features of the f-th key supplementary sentence output by the fusion module.

[0409] The above-mentioned “fusion module” is used to perform fusion processing on the input data of the fusion module; and the embodiment of the present application does not limit the implementation method of the “fusion module”, for example, it can be implemented by using the maximum pooling method.

[0410] The above-mentioned “characteristic features of the f-th key supplementary sentence” are used to represent the semantic fusion information of the f-th key supplementary sentence and the K reference sentences.

[0411] In addition, the embodiments of the present application do not limit the above-mentioned “characteristic features of the fth key supplementary sentence”. For example, Figure 6 As shown, the "characterization feature of the f-th key supplementary sentence" can specifically be: the first fusion feature corresponding to the f-th key supplementary sentence, the second fusion feature corresponding to the f-th key supplementary sentence, ..., and the K-th fusion feature corresponding to the f-th key supplementary sentence are subjected to maximum pooling processing to obtain the "characterization feature of the f-th key supplementary sentence", so that the "characterization feature of the f-th key supplementary sentence" can more accurately represent the semantic information carried by the f-th key supplementary sentence.

[0412] Step 85: Input the representation features of the f-th key supplementary sentence into the decoder, and obtain the main content of the f-th key supplementary sentence output by the decoder.

[0413] The above-mentioned "decoder" is used to decode the input data of the decoder; and the embodiment of the present application does not limit the implementation method of the above-mentioned "decoder". For example, it can be implemented using any existing or future decoding network (that is, Decoder).

[0414] Based on the relevant contents of the above steps 81 to 85, it can be known that for the key point extraction model including the encoder, the position weighting module, the attention module, the fusion module and the decoder, it can be as follows: Figure 6 The key point content generation process shown is implemented so that the key point extraction model can accurately determine the key point content of the fth key supplementary sentence.

[0415] It should be noted that for Figure 6 For example, f represents the fth key supplementary statement; s k∈win,k≠f represents the kth reference sentence; Represents the position weighted feature of the kth reference sentence, k∈{1, 2, 3, ..., K}; POSMatrix krepresents the k-th position weight vector; H f,k represents the kth fusion feature corresponding to the fth key supplementary sentence, k∈{1, 2, 3, …, K}; h′ f represents the representation feature of the fth key supplementary sentence; Y f Indicates the main points of the fth key supplementary sentence.

[0416] Based on the relevant content of S4221 above, it can be known that after obtaining the fth key supplementary sentence, the context information of the fth key supplementary sentence can be determined first; then the pre-built key point extraction model refers to the context information of the fth key supplementary sentence, performs key point content extraction processing on the fth key supplementary sentence, obtains and outputs the key point content of the fth key supplementary sentence, so that the "key point content of the fth key supplementary sentence" can more accurately represent the core semantic information of the fth key supplementary sentence. Wherein, f is a positive integer, f≤F, F is a positive integer, and F represents the number of key supplementary sentences.

[0417] S4222: Determine the content of the minutes to be used according to the user record to be used, at least one target sentence, at least one key supplementary sentence, and the key content of the at least one key supplementary sentence.

[0418] It should be noted that the embodiments of the present application do not limit the implementation of S4222. For example, Figure 2 As shown, S4222 can specifically be: these key supplementary sentences and their key points, target sentences and user records to be used can be displayed to the user, so that the user can write the minutes content to be used based on these displayed contents, so that the user can obtain the relatively complete meeting record content required for writing the minutes from these displayed contents, thereby effectively avoiding the time consumption caused by the user determining the meeting content related to the user record to be used by re-listening to the meeting recording audio, thereby effectively improving the minutes determination efficiency.

[0419] It should also be noted that the embodiment of the present application does not limit the display triggering condition of the above-mentioned "key points of at least one key supplementary sentence". For example, it can be displayed in response to a key point supplementary display request triggered by the user. Among them, the "key point supplementary display request" is used to request the display of the key points of the key supplementary sentence recorded by the user; and the embodiment of the present application does not limit the triggering method of the "key point supplementary display request". For example, the user can click Figure 2 The "check for missing items and fill in the gaps" button shown triggers the "key points supplementary display request".

[0420] Based on the relevant contents of S4221 to S4222 above, it can be known that after obtaining at least one key supplementary sentence for the user record to be used, key content extraction processing can be performed on each key supplementary sentence respectively to obtain the key content of each key supplementary sentence, so that these key contents can more concisely express the important information omitted by the user record to be used, so that the user can more quickly obtain the key supplementary information of the user record to be used from these key contents, which is conducive to improving the efficiency of minutes determination.

[0421] Method Example 7

[0422] In fact, in order to improve the effect of determining the minutes, the minutes content writing process can be realized by means of a human-computer interaction process. Based on this, the present application embodiment provides another possible implementation of the above S4, for example, which can specifically include steps 91-92:

[0423] Step 91: Determine prompt information according to the user record to be used and the semantic matching result to be used.

[0424] The above-mentioned "prompt information" refers to the information provided to the user during the process of writing minutes; and the embodiment of the present application does not limit the "prompt information", for example, it may include the user record to be used, at least one target sentence corresponding to the user record to be used, at least one key supplementary sentence, and at least one of the key points of the at least one key supplementary sentence.

[0425] Based on the relevant content of the above step 91, it can be known that after obtaining the user record to be used and the semantic matching result to be used, prompt information can be determined based on the user record to be used and the semantic matching result to be used, so that subsequent users can write minutes content based on the prompt information.

[0426] Step 92: Display prompt information to the user, so that the user can refer to the prompt information to write the minutes content to be used.

[0427] The present application embodiment does not limit the display process of the above-mentioned "prompt information". For example, it may specifically include (1)-(5):

[0428] (1) When the user enters the user record to be used (for example, Figure 2 After the user record to be used and the positioning trigger button corresponding to the user record to be used are displayed (for example, Figure 2 telescope shown).

[0429] (2) After the user clicks the positioning trigger button corresponding to the user record to be used, at least one target sentence corresponding to the user record to be used can be displayed in the text of the record to be used in a first display mode, so that the user can view each target sentence in the first display mode from the text of the record to be used. The "first display mode" can be preset, for example, Figure 2 As shown, the "first display mode" may include: highlighting with a first color (eg, yellow), and displaying the corresponding relationship between the target sentences and the user records to be used.

[0430] (3) When the user triggers a request to display the key supplementary statement (for example, click Figure 2 After clicking the "check missing" button shown in the figure), the "at least one key supplementary sentence" can be displayed in the record text to be used in the second display mode, so that the user can view the key supplementary sentences in the second display mode from the record text to be used. The "second display mode" can be preset, for example, Figure 2 As shown, the "second display mode" may include: highlighting in a second color (eg, red).

[0431] (4) Display additional requests at user-triggered points (e.g., clicking Figure 2 After clicking the "check missing" button shown in the figure), the key points of each key supplementary sentence can be displayed in the third display mode. Among them, the "third display mode" can be preset, for example, Figure 2 As shown, the "third display mode" may include: displaying the corresponding relationship between each key supplementary sentence and the key content of each key supplementary sentence.

[0432] Based on the relevant contents of the above steps 91 to 92, it can be known that after obtaining the user record to be used and the semantic matching result to be used, the prompt information (for example, Figure 2 The "user record to be used", "target sentence", "key supplementary sentence", and "key content of the key supplementary sentence", etc.) are shown; these prompt information are then displayed to the user so that the user can refer to these prompt information to write the content of the minutes to be used, which is conducive to improving the effect of minutes determination.

[0433] In addition, in order to further improve the user's minutes writing experience, the prompt information can be updated according to the minutes content input by the user in real time (for example, updating the key supplementary sentences and their key points, etc.). Based on this, the embodiment of the present application also provides another possible implementation of the minutes determination method. In this implementation, when the above-mentioned "prompt information" includes at least one key supplementary sentence, the minutes determination method includes all or part of the above steps. After the above-mentioned step 92, the minutes determination method may also include steps 93-97:

[0434] Step 93: Obtain the current minutes writing statement input by the user.

[0435] The above-mentioned “current minutes writing sentence” is used to represent the sentence input in real time when the user writes the minutes content.

[0436] Step 94: According to the semantic matching representation data between the current minutes writing sentence and each key supplementary sentence, search for a to-be-deleted supplementary sentence that meets the fifth matching condition from at least one key supplementary sentence to obtain a search result.

[0437] The "semantic matching representation data between the current minutes writing sentence and the fth key supplementary sentence" is used to represent the similarity between the semantic information carried by the current minutes writing sentence and the semantic information carried by the fth key supplementary sentence. Where f is a positive integer, f≤F, F is a positive integer, and F represents the number of key supplementary sentences.

[0438] In addition, the embodiment of the present application does not limit the determination process of the above-mentioned "semantic matching representation data between the current minutes writing sentence and the fth key supplementary sentence", for example, it can be implemented with the help of any existing or future semantic matching method. For another example, it can be implemented with the help of the above-mentioned "semantic matching model".

[0439] The above-mentioned "fifth matching condition" can be preset, for example, it can be specifically: the semantic matching representation data between the current minutes writing sentence and the to-be-deleted supplementary sentence reaches a preset semantic matching threshold.

[0440] The above-mentioned “supplementary sentence to be deleted” is used to represent the key supplementary sentence that meets the fifth matching condition.

[0441] The above-mentioned “search result” is used to indicate whether there is a supplementary sentence to be deleted in at least one key supplementary sentence.

[0442] Step 95: If the search result indicates that there is at least one supplementary sentence to be deleted in the at least one key supplementary sentence, the at least one supplementary sentence to be deleted is deleted from the at least one key supplementary sentence.

[0443] In an embodiment of the present application, if the search result indicates that there is at least one supplementary statement to be deleted in at least one key supplementary statement, it can be determined that the current minutes writing statement has expressed the semantic information carried by the at least one supplementary statement to be deleted. Therefore, in order to avoid these supplementary statements to be deleted from interfering with the subsequent minutes writing process, the at least one supplementary statement to be deleted can be directly deleted from the at least one key supplementary statement, so that these supplementary statements to be deleted do not exist in the at least one key supplementary statement after deletion, so that these supplementary statements to be deleted will no longer be displayed to the user as prompt information subsequently.

[0444] Step 96: Update the prompt information according to at least one key supplementary sentence.

[0445] It should be noted that the embodiment of the present application does not limit the implementation method of step 96. For ease of understanding, two situations are combined for explanation below.

[0446] In case one, when the above-mentioned "prompt information" includes at least one key supplementary sentence, step 96 may specifically be: after obtaining at least one key supplementary sentence after deletion, the at least one key supplementary sentence after deletion may be used to update the prompt information so that the updated prompt information no longer includes the above-mentioned "at least one supplementary sentence to be deleted", so that these supplementary sentences to be deleted are no longer displayed to the user as prompt information, thereby effectively avoiding interference of these supplementary sentences to be deleted with the subsequent minutes writing process.

[0447] In case 2, when the above-mentioned "prompt information" includes at least one key supplementary sentence and the key content of the at least one key supplementary sentence, step 96 can specifically be: updating the prompt information according to at least one key supplementary sentence and the key content of the at least one key supplementary sentence.

[0448] It can be seen that after obtaining at least one key supplementary sentence after deletion, the at least one key supplementary sentence after deletion and the key content of the at least one key supplementary sentence after deletion can be used to update the prompt information, so that the updated prompt information no longer includes the above-mentioned "at least one supplementary statement to be deleted" and the key content of the "at least one supplementary statement to be deleted", so that these supplementary sentences to be deleted and their key contents are no longer displayed to the user as prompt information, which can effectively avoid these supplementary sentences to be deleted and their key contents from interfering with the subsequent minutes writing process.

[0449] Step 97: Determine whether the preset stop condition is met. If so, end the updating process for the prompt information; if not, return to execute step 92 and subsequent steps.

[0450] The above-mentioned “preset stop condition” can be preset, for example, it can specifically be: receiving a request for completing the minutes writing triggered by a user, or all key supplementary sentences have been deleted.

[0451] Based on the above steps 91 to 97, it can be seen that Figure 7 As shown, in the process of the user continuously writing the minutes content, the prompt information displayed to the user can be continuously reduced (for example, the key supplementary sentences displayed to the user are continuously reduced) to show the user in real time the information that is still needed when the user is writing the minutes content, so that when the user has completed writing the minutes content (or all key supplementary sentences have been deleted), the updating process of the prompt information is stopped, which is conducive to improving the minutes writing effect.

[0452] Based on the minutes determination method provided in the above method embodiment, the embodiment of the present application also provides a minutes determination device, which is explained and illustrated below in conjunction with the accompanying drawings.

[0453] Device Embodiment

[0454] The device embodiment introduces the minutes determination device, and related content can be found in the above method embodiment.

[0455] See also Figure 8 , which is a structural schematic diagram of a minutes determination device provided in an embodiment of the present application.

[0456] The minutes determination device 800 provided in the embodiment of the present application includes:

[0457] An information acquisition unit 801 is used to acquire user records to be used and record texts to be used;

[0458] A text segmentation unit 802 is used to perform sentence segmentation processing on the record text to be used to obtain at least one sentence to be used;

[0459] A semantic matching unit 803, configured to perform semantic matching processing on the user record to be used and the at least one sentence to be used, to obtain a semantic matching result to be used;

[0460] The minutes determining unit 804 is used to determine the contents of the minutes to be used according to the user records to be used and the semantic matching results to be used.

[0461] In a possible implementation, the semantic matching unit 803 is specifically used to: determine the semantic matching result to be used based on the user record to be used, the at least one sentence to be used and a pre-built semantic matching model; wherein the semantic matching model is used to perform semantic matching processing on the input data of the semantic matching model.

[0462] In a possible implementation manner, the semantic matching unit 803 includes:

[0463] A first determining subunit, configured to determine a matching text to be used according to the at least one sentence to be used and a preset sentence identifier;

[0464] A matching prediction subunit, used for inputting the user record to be used and the matching text to be used into the semantic matching model to obtain a matching prediction result output by the semantic matching model;

[0465] The second determining subunit is used to determine the semantic matching result to be used according to the matching prediction result.

[0466] In a possible implementation, the semantic matching model includes a word encoding module, a content interaction module, and a matching prediction module;

[0467] The matching prediction subunit is specifically used to: determine the user record encoding feature and at least one sentence encoding feature according to the user record to be used, the matching text to be used, and the word encoding module; obtain at least one content interaction feature according to the user record encoding feature, the at least one sentence encoding feature, and the content interaction module; input the at least one content interaction feature into the matching prediction module to obtain the matching prediction result output by the matching prediction module.

[0468] In a possible implementation, the semantic matching model further includes a sentence encoding module;

[0469] The process of determining the at least one content interaction feature includes: inputting the at least one sentence encoding feature into the sentence encoding module to obtain at least one sentence representation feature output by the sentence encoding module; and obtaining at least one content interaction feature based on the user record encoding feature, the at least one sentence representation feature, and the content interaction module.

[0470] In a possible implementation, the process of determining the at least one content interaction feature includes: inputting the user record coding feature and the at least one sentence representation feature into the content interaction module to obtain the at least one content interaction feature output by the content interaction module.

[0471] In a possible implementation, the process of determining the at least one content interaction feature includes: screening out at least one sentence feature to be referenced that meets a first similarity condition from the at least one sentence representation feature based on similarity representation data between the user record coding feature and each of the sentence representation features; inputting the user record coding feature and the at least one sentence feature to be referenced into the content interaction module to obtain the at least one content interaction feature output by the content interaction module.

[0472] In a possible implementation, the process of determining the at least one content interaction feature includes: screening out at least one sentence feature to be processed that meets a second similarity condition from the at least one sentence coding feature based on similarity representation data between the user record coding feature and each of the sentence coding features; inputting the user record coding feature and the at least one sentence feature to be processed into the content interaction module to obtain the at least one content interaction feature output by the content interaction module.

[0473] In a possible implementation, the content interaction module includes an attention layer and a fusion layer;

[0474] When the input data of the content interaction module includes the user record coding feature and at least one feature of the sentence to be interacted with, the process of determining the at least one content interaction feature includes: inputting the user record coding feature and the at least one feature of the sentence to be interacted with into the attention layer to obtain the user record representation feature output by the attention layer; and determining the at least one content interaction feature according to the user record representation feature, the at least one feature of the sentence to be interacted with, and the fusion layer.

[0475] In a possible implementation, the number of the sentence features to be interacted is E; the process of determining the e-th content interaction feature includes: inputting the user record representation feature and the e-th sentence feature to be interacted into the fusion layer, and obtaining the e-th content interaction feature output by the fusion layer; wherein e is a positive integer, e≤E, and E is a positive integer.

[0476] In a possible implementation, the process of determining the e-th content interaction feature includes: determining the e-th similarity representation feature and the e-th information difference representation feature based on the user record representation feature and the e-th sentence feature to be interacted; wherein the e-th similarity representation feature is used to represent the information similarity between the user record representation feature and the e-th sentence feature to be interacted; the e-th information difference representation feature is used to represent the information difference between the user record representation feature and the e-th sentence feature to be interacted; the user record representation feature, the e-th sentence feature to be interacted, the e-th similarity representation feature, and the e-th information difference representation feature are concatenated to obtain the e-th content interaction feature.

[0477] In a possible implementation, the e-th similarity representation feature is determined based on the dot product between the user record representation feature and the e-th feature of the sentence to be interacted.

[0478] In a possible implementation, the e-th information difference representation feature is determined according to the difference between the user record representation feature and the e-th feature of the sentence to be interacted.

[0479] In a possible implementation, the second determination subunit is specifically used to determine the semantic matching result to be used based on the text retrieval result to be used and the matching prediction result; wherein the text retrieval result to be used is determined based on a preset retrieval algorithm, the user record to be used and the at least one sentence to be used.

[0480] In a possible implementation, the at least one sentence to be used includes at least one semantic matching sentence, and the semantic matching result to be used includes a semantic matching score between the user record to be used and each of the semantic matching sentences;

[0481] The minutes determination device 800 further includes:

[0482] A matching update unit is used to screen out at least one benchmark sentence that meets a preset benchmark condition from the at least one semantic matching sentence according to the semantic matching score between the user record to be used and each of the semantic matching sentences; determine a distance penalty score of each of the semantic matching sentences according to the text position of the at least one benchmark sentence and the text position of each of the semantic matching sentences; and update the semantic matching score between the user record to be used and each of the semantic matching sentences using the distance penalty score of each of the semantic matching sentences.

[0483] In a possible implementation manner, the number of the semantically matching sentences is E; wherein E is a positive integer;

[0484] The process of determining the distance penalty score of the e-th semantic matching sentence includes: determining the distance set corresponding to the e-th semantic matching sentence according to the distance between the text position of each benchmark sentence and the text position of the e-th semantic matching sentence; performing preset statistical analysis on the distance set corresponding to the e-th semantic matching sentence to obtain the distance to be used corresponding to the e-th semantic matching sentence; and determining the distance penalty score of the e-th semantic matching sentence according to the distance to be used corresponding to the e-th semantic matching sentence.

[0485] In a possible implementation, the at least one sentence to be used includes at least one semantic matching sentence, and the semantic matching result to be used includes a semantic matching score between the user record to be used and each of the semantic matching sentences;

[0486] The minutes determination unit 804 includes:

[0487] A first screening subunit, configured to screen out at least one target sentence satisfying a first matching condition from the at least one semantically matching sentence according to a semantic matching score between the user record to be used and each of the semantically matching sentences;

[0488] The third determining subunit is used to determine the content of the minutes to be used according to the user record to be used and the at least one target sentence.

[0489] In a possible implementation manner, the first screening subunit includes:

[0490] A second screening subunit, configured to screen out at least one preliminary selected sentence satisfying a second matching condition from the at least one semantic matching sentence according to a semantic matching score between the user record to be used and each of the semantic matching sentences;

[0491] The third screening subunit is used to screen out at least one target sentence satisfying a third matching condition from the at least one preliminary selected sentence according to the semantic matching score between the user record to be used and each of the preliminary selected sentences.

[0492] In a possible implementation manner, the first screening subunit further includes:

[0493] a fourth screening subunit, configured to screen out at least one re-selected sentence satisfying a fourth matching condition from the at least one pre-selected sentence according to the number of non-stop words in each of the pre-selected sentences;

[0494] The third screening subunit is specifically configured to screen out at least one target sentence satisfying the third matching condition from the at least one re-selected sentence according to the semantic matching score between the user record to be used and each of the re-selected sentences.

[0495] In one possible implementation, the process of determining the at least one target sentence includes: filtering out sentences to be deleted that meet a first discarding condition from at least one candidate sentence based on a semantic matching score between the user record to be used and each candidate sentence, to obtain a first filtering result; and determining the at least one target sentence based on the first filtering result and the at least one candidate sentence.

[0496] In one possible implementation, the process of determining the at least one target sentence includes: if the first screening result indicates that there is at least one sentence to be deleted in the at least one candidate sentence, then determining at least one sentence to be retained based on the at least one sentence to be deleted and the at least one candidate sentence; and determining the at least one target sentence based on the at least one sentence to be retained.

[0497] In a possible implementation, the process of determining the at least one target sentence includes: if the at least one candidate sentence satisfies a preset correction condition, then selecting at least one sentence to be evaluated that satisfies a first retention condition from the at least one sentence to be deleted; normalizing the semantic matching score between the user record to be used and each of the sentences to be evaluated to obtain a normalized score of each of the sentences to be evaluated; based on the normalized score of each of the sentences to be evaluated, selecting at least one first retained sentence that satisfies a second retention condition from the at least one sentence to be evaluated; deleting the at least one sentence to be deleted from the at least one candidate sentence to obtain at least one second retained sentence; and determining the at least one sentence to be retained based on the at least one first retained sentence and the at least one second retained sentence.

[0498] In a possible implementation manner, the third determining subunit includes:

[0499] a fourth determination subunit, configured to determine at least one key supplementary sentence according to the at least one target sentence, the at least one sentence to be used, and a pre-constructed supplementary sentence determination model; wherein the supplementary sentence determination model is configured to perform key supplementary sentence determination processing on input data of the supplementary sentence determination model;

[0500] The fifth determining subunit is used to determine the content of the minutes to be used according to the user record to be used, the at least one target sentence, and the at least one key supplementary sentence.

[0501] In a possible implementation manner, the fourth determining subunit includes:

[0502] A graph construction subunit, configured to construct a graph to be used according to the at least one target sentence and the at least one sentence to be used;

[0503] A supplementary prediction subunit, used for inputting the to-be-used graph into the supplementary sentence determination model to obtain a key sentence prediction result output by the supplementary sentence determination model;

[0504] The sentence supplement subunit is used to determine the at least one key supplement sentence according to the key sentence prediction result.

[0505] In one possible implementation, the graph construction subunit is specifically used to: determine at least one graph node based on the at least one target sentence and the at least one sentence to be used; wherein the at least one graph node includes at least one recorded sentence node and at least one candidate sentence node; construct a neighbor representation edge corresponding to the at least one graph node according to preset neighbor description parameters; connect each of the candidate sentence nodes with each of the recorded sentence nodes respectively to obtain a related representation edge corresponding to each of the candidate sentence nodes; determine the graph to be used based on the at least one graph node, the neighbor representation edge corresponding to the at least one graph node, and the related representation edge corresponding to the at least one candidate sentence node.

[0506] In a possible implementation, the supplementary sentence determination model includes a graph attention module and a classification prediction layer;

[0507] The supplementary prediction subunit is specifically used to: determine the features to be used of each graph node in the graph to be used; input the features to be used of at least one graph node and the graph to be used into the graph attention module to obtain the graph attention features of at least one graph node output by the graph attention module; input the graph attention features of at least one graph node into the classification prediction layer to obtain the key sentence prediction result output by the classification prediction layer.

[0508] In one possible implementation, the graph attention module includes M graph attention layers; wherein, the input data of the first graph attention layer includes the features to be used of the at least one graph node and the graph to be used; the input data of the m1th graph attention layer includes the output data of the m1-1th graph attention layer and the graph to be used, m1 is a positive integer, 2≤m1≤M, and M is a positive integer.

[0509] In a possible implementation, the graph to be used includes a neighbor representation edge corresponding to the at least one graph node; the at least one graph node includes G record statement nodes, and the number of the graph nodes is R; wherein R is a positive integer, and G is a positive integer;

[0510] When the input data of the m2-th graph attention layer includes the pre-processing features of R graph nodes and the to-be-used graph, the output data of the m2-th graph attention layer includes the post-processing features of R graph nodes, and m2 is a positive integer, 1≤m2≤M, the post-processing feature determination process of the r-th graph node includes: determining at least one neighbor node corresponding to the r-th graph node according to the neighbor representation edge corresponding to the r-th graph node; wherein r is a positive integer, r≤R; determining at least one neighbor node corresponding to the r-th graph node according to the pre-processing features of the r-th graph node and the d-th graph node corresponding to the r-th graph node; r The pre-processing features of the neighboring nodes determine the dth r The neighbor attention weight features of neighbor nodes; among them, d r is a positive integer, d r ≤D r , D r is a positive integer, D r represents the number of neighboring nodes corresponding to the r-th graph node; according to the D corresponding to the r-th graph node r The pre-processing features of neighboring nodes, and the D r The neighbor attention weight features of the rth neighbor nodes are used to determine the neighbor fusion features of the rth graph node; and the processed features of the rth graph node are determined according to the neighbor fusion features of the rth graph node.

[0511] In a possible implementation, the at least one graph node also includes at least one candidate sentence node; the process of determining the processed features of the rth graph node includes: if the rth graph node belongs to the candidate sentence node, then according to the pre-processing features of the rth graph node and the pre-processing features of the gth record sentence node, determine the gth record sentence attention weight feature corresponding to the rth graph node; wherein g is a positive integer, g≤G, and G is a positive integer; according to the pre-processing features of the G record sentence nodes and the G record sentence attention weight features corresponding to the rth graph node, determine the record fusion feature of the rth graph node; according to the neighbor fusion feature of the rth graph node and the record fusion feature of the rth graph node, determine the fusion information difference feature of the rth graph node; splice the neighbor fusion feature of the rth graph node with the fusion information difference feature of the rth graph node to obtain the fusion splicing feature of the rth graph node; according to the fusion splicing feature of the rth graph node, determine the processed features of the rth graph node.

[0512] In a possible implementation manner, the fifth determining subunit includes:

[0513] A key point extraction subunit is used to extract key points from each of the key supplementary sentences to obtain key points of each of the key supplementary sentences;

[0514] The sixth determination subunit is used to determine the content of the minutes to be used according to the user record to be used, the at least one target sentence, the at least one key supplementary sentence, and the main content of the at least one key supplementary sentence.

[0515] In one possible implementation, the number of key supplementary sentences is F, and the key content of the f-th key supplementary sentence is determined based on the f-th key supplementary sentence and a pre-built key extraction model; wherein the key extraction model is used to perform key content extraction processing on the input data of the key extraction model; f is a positive integer, f≤F, and F is a positive integer.

[0516] In a possible implementation manner, the key point extraction subunit includes:

[0517] a seventh determining subunit, configured to determine context information of the f-th key supplementary sentence according to the at least one sentence to be used;

[0518] The key point generation subunit is used to input the f-th key supplementary sentence and the context information of the f-th key supplementary sentence into the key point extraction model to obtain the key point content of the f-th key supplementary sentence output by the key point extraction model.

[0519] In a possible implementation, the context information includes K reference sentences; the key point extraction model includes an encoder, a position weighting module, an attention module, a fusion module and a decoder;

[0520] The process of determining the key points of the f-th key supplementary sentence includes: using the encoder to encode the f-th key supplementary sentence and each of the reference sentences respectively to obtain the encoding features of the f-th key supplementary sentence and the encoding features of each of the reference sentences; inputting the encoding features of the K reference sentences into the position weighting module to obtain the position weighted features of the K reference sentences output by the position weighting module; inputting the position weighted features of the k-th reference sentence and the encoding features of the f-th key supplementary sentence into the attention module to obtain the k-th fusion feature corresponding to the f-th key supplementary sentence output by the attention module; wherein k is a positive integer, k≤K, and K is a positive integer; inputting the K fusion features corresponding to the f-th key supplementary sentence into the fusion module to obtain the representation features of the f-th key supplementary sentence output by the fusion module; inputting the representation features of the f-th key supplementary sentence into the decoder to obtain the key points of the f-th key supplementary sentence output by the decoder.

[0521] In one possible implementation, the module parameters of the position weighting module include K position weighting vectors; the process of acquiring the position weighted feature of the kth reference sentence includes: determining the position excitation weight corresponding to the kth reference sentence based on the dot product between the kth position weighting vector and the encoding feature of the kth reference sentence; and determining the product of the position excitation weight corresponding to the kth reference sentence and the encoding feature of the kth reference sentence as the position weighted feature of the kth reference sentence.

[0522] In a possible implementation, the minutes determining unit 804 includes:

[0523] A prompt determination subunit, used to determine prompt information according to the user record to be used and the semantic matching result to be used;

[0524] The prompt display subunit is used to display the prompt information to the user, so that the user can write the minutes content to be used with reference to the prompt information.

[0525] In a possible implementation manner, the minutes determining unit 804 further includes:

[0526] A prompt updating subunit is used for obtaining the current minutes writing sentence input by the user if the prompt information includes the at least one key supplementary sentence; searching for the supplementary sentence to be deleted that meets the fifth matching condition from the at least one key supplementary sentence based on the semantic matching representation data between the current minutes writing sentence and each of the key supplementary sentences to obtain a search result; if the search result indicates that there is at least one supplementary sentence to be deleted in the at least one key supplementary sentence, deleting the at least one supplementary sentence to be deleted from the at least one key supplementary sentence; updating the prompt information based on the at least one key supplementary sentence, and returning to the prompt display subunit to continue to execute the step of displaying the prompt information to the user until a preset stop condition is reached.

[0527] In a possible implementation, the prompt information also includes the key content of the at least one key supplementary sentence; the updating process of the prompt information includes: updating the prompt information according to the at least one key supplementary sentence and the key content of the at least one key supplementary sentence.

[0528] Furthermore, an embodiment of the present application also provides a device, including: a processor, a memory, and a system bus;

[0529] The processor and the memory are connected via the system bus;

[0530] The memory is used to store one or more programs, and the one or more programs include instructions. When the instructions are executed by the processor, the processor executes any one of the implementation methods of the above-mentioned minutes determination method.

[0531] Furthermore, an embodiment of the present application also provides a computer-readable storage medium, in which instructions are stored. When the instructions are executed on a terminal device, the terminal device executes any one of the implementation methods of the above-mentioned minutes determination method.

[0532] Furthermore, an embodiment of the present application also provides a computer program product, which, when executed on a terminal device, enables the terminal device to execute any one of the implementation methods of the above-mentioned minutes determination method.

[0533] It can be known from the description of the above implementation mode that those skilled in the art can clearly understand that all or part of the steps in the above-mentioned embodiment method can be implemented by means of software plus a necessary general hardware platform. Based on such an understanding, the technical solution of the present application can be essentially or partly embodied in the form of a software product that contributes to the prior art. The computer software product can be stored in a storage medium such as ROM / RAM, a disk, an optical disk, etc., including several instructions for enabling a computer device (which can be a personal computer, a server, or a network communication device such as a media gateway, etc.) to execute the methods described in the various embodiments of the present application or certain parts of the embodiments.

[0534] It should be noted that the various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments, and the same or similar parts between the various embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part description.

[0535] It should also 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.

[0536] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those 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 application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for determining minutes, characterized in that: The method comprises: Get the user record to be used and the record text to be used; Performing sentence segmentation processing on the record text to be used to obtain at least one sentence to be used; Performing semantic matching processing on the user record to be used and the at least one sentence to be used to obtain a semantic matching result to be used; Determining the content of the minutes to be used according to the user record to be used and the semantic matching result to be used; The at least one sentence to be used includes at least one semantic matching sentence, and the semantic matching result to be used includes a semantic matching score between the user record to be used and each of the semantic matching sentences; Before determining the content of the minutes to be used according to the user record to be used and the semantic matching result to be used, the method further includes: According to the semantic matching scores between the user record to be used and each of the semantic matching sentences, at least one benchmark sentence satisfying a preset benchmark condition is screened out from the at least one semantic matching sentence; Determining a distance penalty score of each of the semantically matching sentences according to a text position of the at least one benchmark sentence and a text position of each of the semantically matching sentences; The semantic matching score between the to-be-used user record and each of the semantic matching sentences is updated by using the distance penalty score of each of the semantic matching sentences.

2. The method according to claim 1, characterized in that The performing semantic matching processing on the user record to be used and the at least one sentence to be used to obtain a semantic matching result to be used includes: Determining a matching text to be used according to the at least one sentence to be used and a preset sentence identifier; Inputting the user record to be used and the matching text to be used into a semantic matching model to obtain a matching prediction result output by the semantic matching model; The semantic matching result to be used is determined according to the matching prediction result.

3. The method according to claim 2, characterized in that The semantic matching model includes a word encoding module, a content interaction module and a matching prediction module; The process of determining the matching prediction result includes: Determine a user record encoding feature and at least one sentence encoding feature according to the user record to be used, the matching text to be used, and a word encoding module; Obtaining at least one content interaction feature according to the user record coding feature, the at least one sentence coding feature, and the content interaction module; The at least one content interaction feature is input into the matching prediction module to obtain the matching prediction result output by the matching prediction module.

4. The method according to claim 3, characterized in that The semantic matching model also includes a sentence encoding module; The obtaining at least one content interaction feature according to the user record coding feature, the at least one sentence coding feature, and the content interaction module includes: Inputting the at least one sentence encoding feature into the sentence encoding module to obtain at least one sentence representation feature output by the sentence encoding module; At least one content interaction feature is obtained according to the user record coding feature, the at least one sentence representation feature, and the content interaction module.

5. The method according to claim 4, characterized in that The obtaining at least one content interaction feature according to the user record coding feature, the at least one sentence representation feature, and the content interaction module includes: The user record coding feature and the at least one sentence representation feature are input into the content interaction module to obtain the at least one content interaction feature output by the content interaction module.

6. The method according to claim 4, characterized in that The obtaining at least one content interaction feature according to the user record coding feature, the at least one sentence representation feature, and the content interaction module includes: According to the similarity representation data between the user record coding feature and each of the sentence representation features, at least one sentence feature to be referenced that meets a first similarity condition is selected from the at least one sentence representation feature; The user record coding feature and the at least one sentence feature to be referenced are input into the content interaction module to obtain the at least one content interaction feature output by the content interaction module.

7. The method according to claim 3, characterized in that The obtaining at least one content interaction feature according to the user record coding feature, the at least one sentence coding feature, and the content interaction module includes: According to the similarity representation data between the user record coding feature and each of the sentence coding features, at least one sentence feature to be processed that meets a second similarity condition is selected from the at least one sentence coding feature; The user record coding feature and the at least one sentence feature to be processed are input into the content interaction module to obtain the at least one content interaction feature output by the content interaction module.

8. The method according to any one of claims 5 to 7, characterized in that: The content interaction module includes an attention layer and a fusion layer; When the input data of the content interaction module includes the user record coding feature and at least one feature of a sentence to be interacted with, the process of determining the at least one content interaction feature includes: Inputting the user record encoding feature and the at least one sentence feature to be interacted into the attention layer, and obtaining the user record representation feature output by the attention layer; The at least one content interaction feature is determined according to the user record representation feature, the at least one feature of the sentence to be interacted, and the fusion layer.

9. The method according to claim 8, characterized in that The number of sentence features to be interacted is E; The process of determining the e-th content interaction feature includes: The user record representation feature and the e-th sentence feature to be interacted are input into the fusion layer to obtain the e-th content interaction feature output by the fusion layer; wherein e is a positive integer, e≤E, and E is a positive integer.

10. The method according to claim 2, characterized in that The determining the semantic matching result to be used according to the matching prediction result includes: The semantic matching result to be used is determined according to the text retrieval result to be used and the matching prediction result; wherein the text retrieval result to be used is determined according to a preset retrieval algorithm, the user record to be used and the at least one sentence to be used.

11. The method according to claim 1, characterized in that: The number of the semantically matching sentences is E; wherein E is a positive integer; The process of determining the distance penalty score of the e-th semantic matching sentence includes: Determine a distance set corresponding to the e-th semantic matching sentence according to the distance between the text position of each benchmark sentence and the text position of the e-th semantic matching sentence; Performing a preset statistical analysis on the distance set corresponding to the e-th semantic matching sentence to obtain a to-be-used distance corresponding to the e-th semantic matching sentence; Determine a distance penalty score of the e-th semantic matching sentence according to the to-be-used distance corresponding to the e-th semantic matching sentence.

12. The method according to claim 1, characterized in that The at least one sentence to be used includes at least one semantic matching sentence, and the semantic matching result to be used includes a semantic matching score between the user record to be used and each of the semantic matching sentences; The determining the content of the minutes to be used according to the user record to be used and the semantic matching result to be used includes: Filtering at least one target sentence satisfying a first matching condition from the at least one semantically matching sentence according to the semantic matching score between the user record to be used and each of the semantically matching sentences; The content of the minutes to be used is determined according to the user record to be used and the at least one target sentence.

13. The method according to claim 12, characterized in that The step of selecting at least one target sentence satisfying a first matching condition from the at least one semantically matching sentence according to the semantic matching score between the user record to be used and each of the semantically matching sentences comprises: According to the semantic matching scores between the user record to be used and each of the semantic matching sentences, at least one preliminary selected sentence satisfying a second matching condition is screened out from the at least one semantic matching sentence; At least one target sentence satisfying a third matching condition is selected from the at least one initially selected sentence according to the semantic matching score between the user record to be used and each of the initially selected sentences.

14. The method according to claim 13, characterized in that The method further comprises: According to the number of non-stop words in each of the initially selected sentences, at least one re-selected sentence satisfying a fourth matching condition is selected from the at least one initially selected sentence; The step of selecting at least one target sentence satisfying a third matching condition from the at least one preliminary selected sentence according to the semantic matching score between the user record to be used and each of the preliminary selected sentences comprises: At least one target sentence satisfying the third matching condition is selected from the at least one re-selected sentence according to the semantic matching score between the user record to be used and each of the re-selected sentences.

15. The method according to any one of claims 12 to 14, characterized in that: The process of determining the at least one target sentence includes: According to the semantic matching score between the user record to be used and each candidate sentence, a sentence to be deleted that meets a first discarding condition is screened out from at least one candidate sentence to obtain a first screening result; The at least one target sentence is determined according to the first screening result and the at least one candidate sentence.

16. The method according to claim 15, characterized in that The step of determining the at least one target sentence according to the first screening result and the at least one candidate sentence includes: If the first screening result indicates that there is at least one sentence to be deleted in the at least one candidate sentence, determining at least one sentence to be retained according to the at least one sentence to be deleted and the at least one candidate sentence; The at least one target sentence is determined according to the at least one sentence to be retained.

17. The method according to claim 16, characterized in that The step of determining at least one sentence to be retained according to the at least one sentence to be deleted and the at least one candidate sentence includes: If the at least one candidate sentence satisfies a preset correction condition, then at least one sentence to be evaluated that satisfies a first retention condition is selected from the at least one sentence to be deleted; Normalizing the semantic matching scores between the user record to be used and each sentence to be evaluated to obtain a normalized score of each sentence to be evaluated; According to the normalized scores of the sentences to be evaluated, at least one first retained sentence satisfying a second retention condition is selected from the at least one sentence to be evaluated; Deleting the at least one sentence to be deleted from the at least one candidate sentence to obtain at least one second retained sentence; The at least one sentence to be retained is determined according to the at least one first retained sentence and the at least one second retained sentence.

18. The method according to claim 12, characterized in that The step of determining the content of the minutes to be used according to the user record to be used and the at least one target sentence includes: Determine at least one key supplementary sentence according to the at least one target sentence, the at least one sentence to be used, and a pre-constructed supplementary sentence determination model; wherein the supplementary sentence determination model is used to perform key supplementary sentence determination processing on input data of the supplementary sentence determination model; The content of the minutes to be used is determined according to the user record to be used, the at least one target sentence, and the at least one key supplementary sentence.

19. The method according to claim 18, characterized in that The determining of at least one key supplementary sentence according to the at least one target sentence, the at least one sentence to be used, and a pre-built supplementary sentence determination model comprises: constructing a to-be-used graph according to the at least one target sentence and the at least one to-be-used sentence; Inputting the to-be-used graph into the supplementary sentence determination model to obtain a key sentence prediction result output by the supplementary sentence determination model; The at least one key supplementary sentence is determined according to the key sentence prediction result.

20. The method according to claim 19, characterized in that The step of constructing a graph to be used according to the at least one target sentence and the at least one sentence to be used comprises: Determine at least one graph node according to the at least one target sentence and the at least one sentence to be used; wherein the at least one graph node includes at least one recorded sentence node and at least one candidate sentence node; According to preset neighbor description parameters, construct a neighbor representation edge corresponding to the at least one graph node; Connecting each of the candidate sentence nodes to each of the recorded sentence nodes respectively to obtain relevant representation edges corresponding to each of the candidate sentence nodes; The graph to be used is determined according to the at least one graph node, the neighbor representation edge corresponding to the at least one graph node, and the related representation edge corresponding to the at least one candidate sentence node.

21. The method according to claim 19, characterized in that The supplementary sentence determination model includes a graph attention module and a classification prediction layer; The process of determining the key sentence prediction result includes: Determine the to-be-used features of each graph node in the to-be-used graph; Inputting the to-be-used feature of the at least one graph node and the to-be-used graph into the graph attention module, and obtaining the graph attention feature of the at least one graph node output by the graph attention module; The graph attention feature of the at least one graph node is input into the classification prediction layer to obtain the key sentence prediction result output by the classification prediction layer.

22. The method according to claim 18, characterized in that The step of determining the content of the minutes to be used according to the user record to be used, the at least one target sentence, and the at least one key supplementary sentence includes: Performing key point extraction processing on each of the key supplementary sentences to obtain the key point content of each of the key supplementary sentences; The content of the minutes to be used is determined according to the user record to be used, the at least one target sentence, the at least one key supplementary sentence, and the main content of the at least one key supplementary sentence.

23. The method according to claim 22, characterized in that The number of the key supplementary sentences is F; The process of determining the key content of the fth key supplementary sentence includes: Determine the context information of the f-th key supplementary sentence according to the at least one sentence to be used; wherein f is a positive integer, f≤F, and F is a positive integer; The f-th key supplementary sentence and the context information of the f-th key supplementary sentence are input into a key point extraction model to obtain the key point content of the f-th key supplementary sentence output by the key point extraction model.

24. The method according to claim 23, characterized in that The context information includes K reference sentences; the key point extraction model includes an encoder, a position weighting module, an attention module, a fusion module and a decoder; The process of determining the key content of the fth key supplementary sentence includes: Using the encoder, respectively encoding the f-th key supplementary sentence and each of the reference sentences to obtain encoding features of the f-th key supplementary sentence and encoding features of each of the reference sentences; Inputting the encoding features of the K reference sentences into the position weighting module to obtain the position weighting features of the K reference sentences output by the position weighting module; Inputting the position weighted feature of the k-th reference sentence and the encoding feature of the f-th key supplementary sentence into the attention module, obtaining the k-th fusion feature corresponding to the f-th key supplementary sentence output by the attention module; wherein k is a positive integer, k≤K, and K is a positive integer; Inputting K fusion features corresponding to the f-th key supplementary sentence into the fusion module to obtain the representation features of the f-th key supplementary sentence output by the fusion module; The characterization features of the f-th key supplementary sentence are input into the decoder, and the main content of the f-th key supplementary sentence output by the decoder is obtained.

25. The method according to claim 1, characterized in that The determining the content of the minutes to be used according to the user record to be used and the semantic matching result to be used includes: Determining prompt information according to the user record to be used and the semantic matching result to be used; The prompt information is displayed to the user, so that the user can write the minutes content to be used with reference to the prompt information.

26. The method according to claim 25, characterized in that After displaying the prompt information to the user, the method further includes: If the prompt information includes the at least one key supplementary sentence, obtaining the current minutes writing sentence input by the user; According to the semantic matching representation data between the current minutes writing sentence and each of the key supplementary sentences, searching for a to-be-deleted supplementary sentence that satisfies a fifth matching condition from the at least one key supplementary sentence to obtain a search result; If the search result indicates that there is at least one supplementary sentence to be deleted in the at least one key supplementary sentence, deleting the at least one supplementary sentence to be deleted from the at least one key supplementary sentence; According to the at least one key supplementary sentence, the prompt information is updated, and the step of displaying the prompt information to the user is continued until a preset stop condition is reached.

27. A device for determining minutes, characterized in that: include: An information acquisition unit, used to acquire user records to be used and record texts to be used; A text segmentation unit, configured to perform sentence segmentation processing on the record text to be used to obtain at least one sentence to be used; A semantic matching unit, configured to perform semantic matching processing on the user record to be used and the at least one sentence to be used, to obtain a semantic matching result to be used; A minutes determination unit, used to determine the content of the minutes to be used according to the user record to be used and the semantic matching result to be used; The at least one sentence to be used includes at least one semantic matching sentence, and the semantic matching result to be used includes a semantic matching score between the user record to be used and each of the semantic matching sentences; The minutes determination device further includes: A matching update unit, configured to screen out at least one benchmark sentence satisfying a preset benchmark condition from the at least one semantic matching sentence according to the semantic matching score between the user record to be used and each of the semantic matching sentences; and determine a distance penalty score of each of the semantic matching sentences according to the text position of the at least one benchmark sentence and the text position of each of the semantic matching sentences; The semantic matching score between the to-be-used user record and each of the semantic matching sentences is updated by using the distance penalty score of each of the semantic matching sentences.

28. A device, characterized in that The device includes: a processor, a memory, and a system bus; The processor and the memory are connected via the system bus; The memory is used to store one or more programs, wherein the one or more programs include instructions, and when the instructions are executed by the processor, the processor performs the method according to any one of claims 1 to 26.

29. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores instructions, and when the instructions are executed on a terminal device, the terminal device executes the method according to any one of claims 1 to 26.

Images