A text minutes generation method, device, equipment and storage medium

Abstract

The present application proposes a method, device, equipment and storage medium for generating text minutes, the method comprising: obtaining a target text and a reference text, wherein the reference text is determined based on the target text content that the user is concerned about; based on locating the associated content of the reference text from the target text, performing minute generation processing on the target text to obtain a target text minute corresponding to the reference text. By adopting this text minute generation method, even for the same target text, when the reference text is different, the target text can be subjected to text minute generation processing with different focuses by locating the text content related to the reference text from the target text, thereby obtaining a target text minute corresponding to the reference text. Therefore, the method can generate text minutes that meet the needs of different users for the same target text.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to a method, device, equipment and storage medium for generating text minutes. Background Art

[0002] Text minutes generation refers to the extraction of content from long texts, thereby refining information that can represent the core content of the text. Text minutes can help people grasp the content of the text more directly and effectively.

[0003] Conventional text summary generation schemes are usually based on text automatic summarization technology, which extracts key points from the text and forms a summary text. Text automatic summarization technology can be divided into extractive summarization and generative summarization according to the way of generating summaries. Extractive summarization extracts words or sentences from the original text intact to form a summary, and the summary content all comes from the original text; while generative summarization allows the generation of new words and phrases that are not in the original text to form a summary. When generating a summary, the text content is first semantically understood, and a paragraph is generated based on the semantics to summarize the given text.

[0004] Usually, the content of the target text for which text minutes need to be generated is multifaceted, and different personnel may be interested in different aspects of the content. Therefore, different personnel have different requirements for text minutes of the same target text.

[0005] However, current text minutes generation schemes, whether extractive summarization or generative summarization, cannot generate text minutes with different contents according to the needs of different personnel, and cannot meet the needs of different personnel for generating text minutes of the same target text. Summary of the invention

[0006] Based on the above technical status, the present application proposes a text minutes generation method, device, equipment and storage medium. By implementing the technical solution of the present application, it is possible to generate text minutes that meet the needs of different users for the same target text.

[0007] In order to achieve the above objectives, this application proposes the following technical solutions:

[0008] A text minutes generation method, comprising:

[0009] Acquire a target text and a reference text, wherein the reference text is determined based on the content of the target text that the user is interested in;

[0010] Based on locating the associated content of the reference text from the target text, the target text is processed for generating a minutes to obtain a target text minutes corresponding to the reference text.

[0011] A text minutes generating device, comprising:

[0012] A data acquisition unit, used to acquire a target text and a reference text, wherein the reference text is determined based on the content of the target text that the user is concerned about;

[0013] The minutes generating unit is used to perform minutes generating processing on the target text based on locating the associated content of the reference text from the target text, so as to obtain the target text minutes corresponding to the reference text.

[0014] A text minutes generating device, comprising:

[0015] Memory and processor;

[0016] The memory is connected to the processor and is used to store programs;

[0017] The processor is used to implement the above-mentioned text minutes generation method by running the program in the memory.

[0018] A storage medium stores a computer program, and when the computer program is executed by a processor, the above-mentioned text minutes generation method is implemented.

[0019] The text minutes generation method proposed in the present application uses a reference text as a reference for generating text minutes when generating text minutes for a target text, locates the associated content of the reference text from the target text, performs minutes generation processing on the target text, and obtains a target text minutes corresponding to the reference text. When generating text minutes for a target text, the method jointly applies the target text content and the reference text associated content in the target text to determine the text minutes of the target text. With the text minutes generation method, even for the same target text, when the reference text is different, the text minutes generation processing with different focuses can be performed on the target text by locating the text content related to the reference text from the target text, thereby obtaining a target text minutes corresponding to the reference text. Therefore, the method can generate text minutes that meet the needs of different users for the same target text. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] 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 only embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0021] Figure 1 It is a flowchart of a method for generating a text minutes provided in an embodiment of the present application;

[0022] Figure 2 It is a structural diagram of the chapter interactive semantic retrieval model provided in the embodiment of the present application;

[0023] Figure 3 It is a structural diagram of a semantic retrieval model based on enhanced retrieval candidate information provided in an embodiment of the present application;

[0024] Figure 4 It is a structural diagram of a text summary generation model based on an attention mechanism provided in an embodiment of the present application;

[0025] Figure 5 It is a structural diagram of another text summary generation model based on the attention mechanism provided in an embodiment of the present application;

[0026] Figure 6 It is a schematic diagram of the structure of the word-sentence-chapter level information encoding model provided in the embodiment of the present application;

[0027] Figure 7 It is a schematic diagram of word segmentation feature extraction provided in an embodiment of the present application;

[0028] Figure 8 It is a structural diagram of the information fusion model provided in the embodiment of the present application;

[0029] Fig. 9 It is a structural schematic diagram of a text minutes generating device provided in an embodiment of the present application;

[0030] Fig.10 It is a structural schematic diagram of a text minutes generating device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0031] The technical solution of the embodiment of the present application is suitable for the application scenario of generating text minutes. By adopting the technical solution of the embodiment of the present application, it is possible to generate text minutes that are consistent with the user's focus, thereby meeting the text minutes needs of different users.

[0032] The above-mentioned application scenarios for generating text minutes specifically refer to scenarios where minutes content needs to be generated, including but not limited to specific application scenarios such as meeting minutes generation, document abstract generation, and news key points extraction.

[0033] Text minutes generation refers to the extraction of content from long texts, thereby refining information that can represent the core content of the text. Text minutes can help people grasp the content of the text more directly and effectively.

[0034] Conventional text summary generation schemes are usually based on text automatic summarization technology, which extracts key points from the text and forms a summary text. Text automatic summarization technology can be divided into extractive summarization and generative summarization according to the way of generating summaries. Extractive summarization extracts words or sentences from the original text intact to form a summary, and the summary content all comes from the original text; while generative summarization allows the generation of new words and phrases that are not in the original text to form a summary. When generating a summary, the text content is first semantically understood, and a paragraph is generated based on the semantics to summarize the given text.

[0035] Usually, the content of the target text for which text minutes need to be generated is multifaceted, and different personnel may be interested in different aspects of the content. Therefore, different personnel have different requirements for text minutes of the same target text.

[0036] For example, in a meeting scenario, the content of the meeting is usually multifaceted, and different participants are usually concerned about different contents. For example, in a seminar on new product planning, the heads of the company's design department, product department, and marketing department who attended the meeting each care about different aspects of the content. For example, the design department is more concerned about the improvement of the product design plan, the product department is more concerned about product definition and R&D planning, and the marketing department is more concerned about the market positioning of new products. Therefore, the content of the meeting minutes required by different departments is different.

[0037] However, conventional text minutes generation schemes, whether extractive summarization or generative summarization, can only perform technical processing on the text to be summarized to determine the main content of the text, and cannot generate text minutes with different focuses for different concerns.

[0038] Based on the above technical status, the embodiment of the present application proposes a text minutes generation scheme, which can refer to the target text content that the user is concerned about and generate a text minutes for the target text, so that different text minutes can be generated for different points of concern, thereby meeting the personalized needs of different users for the content of text minutes.

[0039] In the subsequent embodiments of the present application, the specific processing content of the technical solution of the embodiment of the present application is introduced by taking the generation of meeting minutes as an example. When the technical solution of the embodiment of the present application is applied to other scenarios, its specific execution process can refer to the introduction of the various embodiments of the present application. It should be noted that the technical solution of the embodiment of the present application is not only applicable to the generation of minutes of text, thereby obtaining minutes in text form, but also applicable to the generation of minutes of speech, obtaining minutes in text or speech form, or generating minutes of text to obtain minutes in speech form. When generating minutes for non-text data, or generating non-text minutes, it can be achieved by converting the non-text data into text data, or converting the generated text minutes into non-text form. The main processing of the minutes generation can still refer to the introduction of the embodiment of the present application.

[0040] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only 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.

[0041] This application embodiment proposes a method for generating a text minutes, see Figure 1 As shown, the method includes:

[0042] S101, obtaining a target text and a reference text.

[0043] The target text mentioned above refers to the text for which minutes need to be generated, and the target text can be a text of any content and any language obtained by any means. Specifically, the target text can be a text directly obtained, such as academic documents, news releases, books, etc., or a text obtained by speech recognition, such as the recognition result text obtained by speech recognition of conference recordings, the recognition result text obtained by recognizing the speech of the speaker, etc. In theory, any form of data content can be converted into text form, and thus used as the target text mentioned above, and the minutes of the target text can be generated through subsequent processing.

[0044] The above-mentioned reference text is determined based on the target text content that the user is concerned about. Specifically, the reference text can represent the user's interest or focus on the target text content, and at the same time represent the user's demand for the generated text minutes content, which is used to provide a reference for generating the text minutes of the above-mentioned target text, so as to generate a target text minutes that meets the user's focus or contains the user's interest content.

[0045] The reference text may be input by the user, or may be preset before the text minutes are generated. The specific form of the reference text may be a fixed sentence pattern, a keyword or a phrase, a short text sentence or a text segment, or even a logical combination of multiple texts. For example, the reference text may be a combination of some search conditions, such as A is relevant and B is relevant (denoted as A&B), A is relevant or B is relevant (denoted as A||B), or A is relevant but B is not relevant (denoted as AB), or even a more complex combination of conditions, such as A is relevant and B is relevant but C is not relevant (denoted as {A&B}-C), wherein the specific forms of A, B, and C are not required, and may be phrases, keywords, sentences, or other forms.

[0046] For example, in the scenario of generating meeting minutes, speech recognition processing is performed on the meeting recording to obtain text corresponding to the meeting recording, and this text is used as the target text mentioned above. At the same time, a reference text representing the meeting content that the user is interested in or concerned about is obtained. The reference text can be a phrase or short sentence summarized by the user based on the content of the meeting, or a simple meeting record recorded by the user at the venue, or keywords, phrases, search conditions, etc. determined by the user based on the content of the meeting minutes that the user hopes to obtain. When the target text corresponding to the meeting and the reference text entered by the user are obtained, the subsequent text minutes generation process can be executed, and the minutes generation process is performed on the above target text to obtain a target text minutes that meets the user's concerns or contains the content that the user is interested in.

[0047] For the processing device, when it processes the text, it actually processes the text features. Therefore, the above-mentioned "obtaining the target text and the reference text" can be to obtain the original text of the target text and the reference text, and then perform feature extraction on the obtained target text and the original reference text to obtain the features of the target text and the reference text for subsequent text minutes generation processing; or, it can also be to directly obtain the features of the target text and the reference text for subsequent text minutes generation processing.

[0048] S102: Based on locating the associated content of the reference text in the target text, a minutes generation process is performed on the target text to obtain a target text minutes corresponding to the reference text.

[0049] Among them, the above-mentioned associated content of the reference text refers to text content related to the reference text, such as text content whose similarity with the reference text is greater than a set similarity threshold, or text content that is semantically similar or related to the reference text, all of which can be used as associated content of the reference text.

[0050] Exemplarily, an embodiment of the present application compares the reference text with each text fragment of the target text in turn, determines the text similarity or semantic similarity between the reference text and each text fragment of the target text, thereby determining the correlation between the reference text and the target text fragment, and locating and identifying related content related to the reference text from the target text.

[0051] It can be understood that the reference text-related content in the target text is more valuable for generating the text minutes corresponding to the reference text. Because the reference text-related content contains text information related to the reference text, if the reference text-related content can be considered when generating the text minutes for the target text, the final generated text minutes can contain more reference text-related information, so that the generated target text minutes can match the reference text.

[0052] Based on the above ideas, when generating the text minutes of the target text, the embodiment of the present application mainly uses the content associated with the reference text in the target text and takes other content in the target text as a supplement to generate the text minutes of the target text, so that the content associated with the reference text accounts for a higher proportion in the final generated target text minutes, thereby making the final generated target text minutes have a higher correlation with the reference text, that is, the target text minutes corresponding to the reference text are obtained.

[0053] As an exemplary implementation, the embodiment of the present application locates text segments related to a reference text from a target text, performs minutes generation processing on the full text content of the target text, and thereby obtains a target text minutes corresponding to the reference text.

[0054] The above-mentioned text fragments may be text sentences, text segments, or text phrases, etc. The embodiment of the present application identifies text fragments related to the reference text from the target text through methods such as text comparison and semantic comparison. For example, as long as the text similarity or semantic similarity between the text fragment in the target text and the reference text is not 0, it can be considered that it is valuable for generating the target text minutes corresponding to the reference text, and therefore it is determined as a text fragment related to the reference text.

[0055] Then, combined with the positioning results of the text fragments related to the reference text, the full text content of the target text is processed for minute generation, that is, a text minute is generated for the full text content of the target text. In the process of generating the minute, the contribution of the text fragments related to the reference text to the generation of the target text minute is set to be higher than the contribution of the text fragments unrelated to the reference text to the generation of the target text minute, so that the generated target text minute contains more reference text related text fragment information. Furthermore, different contributions can be set for each text fragment related to the reference text according to its relevance to the reference text, so that the target text minute finally generated has a higher relevance to the reference text.

[0056] For example, suppose the original text of a certain meeting contains 200 conversation texts, as shown in the following original text (since the original text is long, ellipsis is used to represent the omitted part of the original text for simplicity):

[0057] Original meeting text:

[0058] Paragraph 1: Today we will mainly discuss the launch plan of the new scanning pen product.

[0059] Paragraph 2: It is planned to launch a four-month tour in March 2019 and hold a Science and Technology Innovation Day.

[0060] …………

[0061] Paragraph: 16: The following is a comprehensive analysis of product trends to help products stand out from the fierce market. We will select 14 cities and actively seek self-media cooperation. The cooperation content includes...

[0062] …………

[0063] Paragraph 68: Young people prefer designs that look cool. We should take this market seriously and find a way in. We should strengthen the publicity effect...

[0064] …………

[0065] Paragraph 79: Now there is a demo for you to watch. Before the Double Eleven launch, we conducted a relatively systematic test on the product. The accuracy rate on the test set was as high as 98%, which was 30% higher than that of the competitor, enough to create a generation gap.

[0066] …………

[0067] Paragraph 88: Testing with test set data does not measure the effect. Because if we want to compare with competitors completely, we actually have to compare the same things. If it is the same set, different people will have different feelings.

[0068] …………

[0069] Paragraph 88: Product positioning should take young people into consideration and include this part in the design…

[0070] …………

[0071] Paragraph 162: Interface design also needs to be interactive and reasonable, taking into account the connection between button attributes and jumps... By the way, for the test problem mentioned earlier, we should also consider the user's subjective experience and set up some subjective experience comparisons. For some special cases, such as smoothness, multi-line cross selection, etc., feel the effect of product use.

[0072] …………

[0073] Paragraph 200: This is the end of today’s meeting. All departments please cooperate.

[0074] The above-mentioned original meeting text is taken as the target text. Assuming that the search condition input by user A is "subjective experience of product effect", the search condition input by user A is taken as the reference text. For the determined target text and reference text, by executing the processing of the text minutes generation method proposed in the embodiment of the present application, it can be determined that in the above-mentioned original meeting text, the text content related to the search condition "subjective experience of product effect" input by user A is "By the way, for the test problem mentioned above, we should also consider the user's subjective experience, and need to set up some subjective experience comparisons." and "For some special cases, such as smoothness, multi-line cross selection, etc., feel the effect of product use." By locating the text content related to the search condition "subjective experience of product effect" from the original meeting text, the original meeting text is processed for minutes generation, and finally the meeting minutes "The effect of the new scanning pen product is not only based on data results, but also needs to consider some subjective experience solutions" are obtained. It can be seen that the final meeting minutes match the search condition "subjective experience of product effect" input by user A, that is, the meeting minutes are meeting minutes that characterize the relevant information of the subjective experience of product effect. Therefore, by adopting the technical solution of the embodiment of the present application, it is possible to locate the text content related to the search criteria input by user A from the original meeting text, generate meeting minutes corresponding to the search criteria input by user A, thereby meeting the meeting minutes needs of user A.

[0075] For another example, for the above-mentioned meeting text, assuming that the search condition entered by user B is "product effect subjective experience & product positioning", the meeting text is used as the target text, and the search condition entered by user B is used as the reference text. By executing the processing of the text minutes generation method proposed in the embodiment of the present application, it can be determined that in the above-mentioned meeting text, the text content related to the search condition "product effect subjective experience & product positioning" entered by user B is "Product positioning should consider young people more, and add this part of the design.", "By the way, for the test problem mentioned earlier, we should also consider the user's subjective experience, and need to set up some subjective experience comparisons." and "For some special cases, such as smoothness, multi-line selection, etc., feel the effect of product use." By locating the text content related to the search condition "product effect subjective experience & product positioning" from the meeting text, the meeting text is processed for minutes generation, and finally the meeting minutes are obtained: "The new product positioning of the scanning pen should consider that young people prefer cool designs. The use effect is not only determined by data results, but also needs to consider some subjective experience solutions." It can be seen that the final meeting minutes match the search condition "subjective experience of product effect & product positioning" entered by user B, that is, the meeting minutes are meeting minutes that characterize the relevant information of subjective experience of product effect and product positioning. Therefore, by adopting the technical solution of the embodiment of the present application, it is possible to locate the text content related to the search condition entered by user B from the original meeting text, and generate meeting minutes corresponding to the search condition entered by user B, thereby meeting the meeting minutes needs of user B.

[0076] It can be seen that by adopting the text minutes generation method proposed in the embodiment of the present application, even for the same original meeting text, when the search conditions input by the user are different, text minutes matching the user search conditions can be generated respectively for different user search conditions, thereby meeting the needs of different users for meeting content.

[0077] In summary, the text minutes generation method proposed in the embodiment of the present application, when generating text minutes for a target text, uses a reference text as a reference for generating text minutes, locates the associated content of the reference text from the target text, performs minutes generation processing on the target text, and obtains a target text minutes corresponding to the reference text. When generating text minutes for a target text, the method jointly applies the target text content and the reference text associated content in the target text, and jointly determines the text minutes of the target text. With the text minutes generation method, even for the same target text, when the reference text is different, the text minutes generation processing with different emphases can be performed on the target text by locating the text content related to the reference text from the target text, thereby obtaining a target text minutes corresponding to the reference text. Therefore, the method can generate text minutes that meet the needs of different users for the same target text.

[0078] As an exemplary implementation, the embodiment of the present application pre-trains a text minutes generation model based on an attention mechanism, which is used to perform text minutes generation processing on the target text according to the target text and the reference text, and obtain a target text minutes corresponding to the reference text.

[0079] The attention mechanism-based text minutes generation model is trained based on the pre-collected parallel data of target text-reference text-target text minutes. For example, a large amount of parallel data of original meeting text-user search conditions-meeting minutes text is collected in advance and used for training the model after data preprocessing.

[0080] The original text data of the meeting can collect the audio data of the meeting, and obtain the corresponding text data after voice transcription; of course, the text data of the meeting can also be directly collected, such as the original full text of the meeting compiled by the stenographer. The form of the retrieval condition is not limited to the fixed sentence pattern of the fixed template, but also supports the user to enter the keywords or phrases they are concerned about, or a short meeting record, or even a logical combination of multiple sub-conditions, such as A is related and B is related (recorded as A&B), A is related or B is related (recorded as A||B) and A is related but B is not related (recorded as AB), or even a more complex combination of conditions, such as A is related and B is related but C is not related (recorded as {A&B}-C), where the specific form of A, B, and C is not required, and can be a phrase, keyword, sentence or other form. In the embodiment of the present application, the specific source or specific content of the retrieval condition is not limited. The meeting minutes are intended to be highly condensed and highly relevant to the original long-length meeting text of the retrieval condition, covering and summarizing the key content of the meeting related to the retrieval condition.

[0081] The data preprocessing is first to process the original text of the meeting into sentences. The sentence segmentation can be performed by segmenting the sub-clauses or the whole sentence according to punctuation marks, or by segmenting the sentence according to a fixed word count window and a sliding window. This case does not make specific requirements on the specific method of sentence segmentation. This case uses punctuation marks and the whole sentence segmentation method to process the original text of the meeting into sentences; secondly, the original text data of the meeting after the sentence segmentation is processed into the input sequence form of word components. The text segmentation can be performed using existing technology, which will not be described in detail here. For example, the original text sentence of the meeting is "The next key task is a follow-up of the Spring Fair." The result after word segmentation is "The next / key / task / is / a / follow-up / of / the / Spring Fair." For user search conditions, if the search condition is only a single condition, the search condition is directly regarded as a plain text. The preprocessing only needs to process the search condition text data into the input sequence form of word components. If the search condition is a compound of multiple sub-conditions, each sub-condition is processed into a sequence form after the word components. For the corresponding meeting minutes, the preprocessing is to process the text data into a sequence form of word components.

[0082] The text minutes generation model based on the attention mechanism is a time-series output model, that is, each decoding outputs a segmented word, and the segmented words finally obtained can be combined to obtain the minutes text. During training, the features of the segmented word sequence of the original conference text and the features of the segmented word sequence of the retrieval condition are respectively obtained, and the features of the segmented word sequence of the original conference text and the features of the segmented word sequence of the retrieval condition are input into the text minutes generation model based on the attention mechanism, and the model loss is determined by comparing the meeting minutes segmented word sequence output by the model with the meeting minutes segmented word sequence obtained by pre-processing, and the model parameters are corrected based on the model loss, so that the text minutes generation model based on the attention mechanism can take the features of the target text and the reference text as input to generate the target text minutes corresponding to the reference text.

[0083] Below, taking the generation of meeting minutes that meet different user search conditions as an example, the specific processing process of the text minutes generation method proposed in the embodiment of the present application is introduced. In the following embodiment, the original text of the meeting represents the above-mentioned target text, and the search conditions or search sub-conditions input by the user represent the above-mentioned reference text. By introducing the process of generating text minutes from the original text of the meeting, and obtaining text minutes that meet the search conditions input by the user, the specific processing process of the technical solution of the embodiment of the present application based on locating the associated content of the reference text from the target text and generating the target text minutes corresponding to the reference text is demonstrated.

[0084] First, as for "obtaining the target text and the reference text", as shown above, the original text of the target text and the reference text can be obtained, and then feature extraction can be performed for subsequent text minutes generation processing. Alternatively, the features of the target text and the reference text can be directly obtained and used for subsequent text minutes generation processing.

[0085] For the specific process of extracting target text features and reference text features, please refer to the detailed description of the subsequent embodiments.

[0086] Then, regarding the processing step of "based on locating the associated content of the reference text from the target text, performing a minutes generation process on the target text to obtain a target text minutes corresponding to the reference text" in the text minutes generation method proposed in the embodiment of the present application, as an exemplary implementation, it can be implemented through the following steps A1-A2:

[0087] A1. Locate text segments related to the reference text from the target text by determining the relevance between each text segment in the target text and the reference text.

[0088] Specifically, the above-mentioned text fragments can be text contents of any granularity, such as text sentences, text segments or text phrases. In the embodiment of the present application, the target text is divided into text sentences, and the text sentences obtained by the division are used as the above-mentioned text fragments. Among them, the text sentence division of the target text can be performed according to the punctuation marks of the target text, or it can be performed according to a fixed word count window and a sliding window. This case does not make specific requirements on the specific method of sentence division. In the embodiment of the present application, the target text is divided into sentences according to the punctuation marks in the target text.

[0089] Then, the reference text is compared with each target text sentence obtained by the division, or a semantic comparison is performed to determine the relevance between the reference text and each target text sentence. In the embodiment of the present application, the relevance between each text segment in the target text and the reference text is determined by semantic measurement, that is, the relevance between the reference text and each text segment in the target text is determined by comparing the semantic similarity between the reference text and each text segment in the target text.

[0090] Exemplarily, the features of the target text and the features of the reference text are first obtained respectively. Then, the relevance of each text segment in the target text with the reference text is determined respectively according to the features of each text segment in the target text and the features of the reference text. According to the relevance of the reference text with each text segment in the target text, the text segment related to the reference text can be located from the target text. For example, in the target text, the text segment whose relevance to the reference text is not zero is the text segment related to the reference text.

[0091] The characteristics of the target text may be obtained by extracting characteristics from each text segment in the target text, and then combining the characteristics of each text segment to obtain the characteristics of the target text. Thus, based on the characteristics of each text segment in the target text, the overall characteristics of the target text and the characteristics of each text segment of the target text may be determined respectively. Alternatively, the characteristics of the target text may be directly extracted from the entire target text to obtain the characteristics of the target text, and then, according to the positions of each text segment of the target text in the target text, the characteristics of the corresponding positions are intercepted from the characteristics of the target text to obtain the characteristics of each text segment.

[0092] In an embodiment of the present application, the features of the target text are determined by extracting the features of each text segment of the target text, and the features of the reference text are determined by extracting the features of each text segment of the reference text. The specific feature extraction process will be introduced in subsequent embodiments.

[0093] As an optional implementation, for example, in the application scenario of generating conference minutes, assuming that the user inputs multiple search sub-conditions, the semantic similarity between each search sub-condition and each sentence in the conference original text is determined. At this time, for each search sub-condition, the semantic similarity between each search sub-condition and each sentence in the conference original text is determined, which can be achieved through the above-mentioned process of determining the semantic similarity between the text segment of the target text and the reference text.

[0094] Exemplarily, the embodiment of the present application constructs a semantic fuzzy retrieval model to extract the semantic similarity score between each sub-condition of the user's search condition and each sentence in the original conference text, that is, to determine the relevance of each text fragment in the reference text and the target text.

[0095] The semantic fuzzy retrieval model takes the search sub-condition text and the original conference text as input, and outputs the semantic similarity score between each sentence in the original conference text and the text for a certain search sub-condition text. Based on the above semantic fuzzy retrieval model, after the conference original text and the user search sub-condition text sequence are input into the semantic fuzzy retrieval model, the semantic similarity scores between 1 to n (n is the total number of sentences in the original conference text) sentences in the original conference text and the text for a certain user search sub-condition can be obtained. For the search sub-condition A, the semantic similarity scores between it and the n sentences in the original conference text are expressed as Indicates that for the retrieval sub-condition B, the semantic similarity score between it and the n sentences of the original conference text is expressed as express.

[0096] The embodiments of the present application propose two semantic fuzzy retrieval model frameworks: a paragraph interaction semantic retrieval model, and a semantic retrieval model based on retrieval candidate information enhancement, which are used to measure the semantic similarity scores between retrieval sub-conditions (reference texts) and conference original text (target texts) sentences.

[0097] like Figure 2 As shown in Figure 1, the chapter interactive semantic retrieval model includes a word encoder, a sentence encoder, and a retrieval sub-condition text and conference text interaction module. The word encoder uses the BERT pre-training model. First, the conference transcription text sentence S and the retrieval sub-condition Q are input into the word encoder (where q = {w 1 ,w 2 ,…,w n} represents the n words contained in the search condition Q, S j ={w j,1 ,w j,2 ,…w j,m} represents the m word segments contained in the jth conference text sentence), obtain the context encoding vector of each word in the sentence, and extract the word vector representation of [CLS] as well as As the transcription text sentence encoding and the retrieval sub-condition sentence encoding, it is used to represent the information of the entire sentence.

[0098] Then, the transcribed text sentence encoding introduces context information into the current sentence through a two-layer Transformer sentence encoder model, completing the information omitted by the current sentence context, thereby obtaining a more accurate sentence representation.

[0099] Secondly, through the retrieval sub-condition text and conference text interaction module, the sentence encoding of each text sentence of the original conference text and the sentence encoding of the retrieval sub-condition are interactively operated based on the attention mechanism to obtain the retrieval sub-condition sentence encoding after information improvement. The retrieval sub-condition text and conference text interaction module is composed of an attention structure, which encodes the retrieval sub-condition encoding sentence encoding. As the query Q of the attention mechanism, the conference transcription sentence encoding is used as K and V, so as to integrate the conference content information into the retrieval sub-condition sentence encoding q. When the retrieval sub-condition is relatively short or the description is vague and the information is incomplete, the conference text content can better supplement the information omitted in the retrieval sub-condition; at the same time, the encoding q has a global view of the entire conference content, which is more conducive to selecting better retrieval results in the transcribed text.

[0100] Finally, the retrieval sub-condition sentence encoding q is combined with each conference text sentence encoding Do splicing to generate the final interaction vector S j , input to the output layer to predict the semantic similarity score between the sentence and the retrieval sub-condition.

[0101]

[0102] In the above formula, represents the dot product operation between the retrieval condition and the text sentence j, representing the similarity between the two; Represents the information difference between the retrieval sub-condition and the text sentence j. By comparing from multiple perspectives, more comprehensive similarity judgment information can be obtained.

[0103] Compared with the text interaction semantic retrieval model, the semantic retrieval model based on retrieval candidate information enhancement has made some modifications in the interaction between retrieval sub-conditions and text sentences. Specifically, after obtaining the encoding of the retrieval sub-condition sentence, And get the sentence encoding of each text sentence of the original conference text Afterwards, based on the sentence encoding of the retrieval sub-conditions and the sentence encoding of each text sentence, the similarity between each text sentence of the conference original text and the retrieval sub-conditions is calculated. Then, based on the similarity between each text sentence of the conference original text and the retrieval sub-conditions, the N text sentences with the highest similarity to the reference text are selected from the conference original text. Finally, the sentence encodings of the selected N text sentences and the sentence encodings of the retrieval sub-conditions are interactively operated based on the attention mechanism to obtain the retrieval sub-condition sentence encoding after information completion.

[0104] For example, by encoding the sentences in the conference text By calculating the cosine distance, we can obtain the top N preliminary matching search results that are most similar to the search sub-conditions. That is, based on the similarity between each text sentence in the original conference text and the search sub-conditions, we can select the N text sentences with the highest similarity to the search sub-conditions from the original conference text. Figure 3 As shown in , r=1 represents the selected TopN preliminary matching sentences, and r=0 represents other sentences with lower relevance.

[0105] Then, retrieve the sub-condition sentence encoding Matching results with TopN high quality results Through attention interaction, the encoding vector of the search sub-condition is updated to q. When N is set to a small value, such as 2, the confidence of the N search candidates related to the search sub-condition is high, and the search sub-condition information can be supplemented more accurately; at the same time, by using the N high-quality searches as prompt information, the model can be guided to select results similar to the initial high-quality search results, avoiding too large a difference in semantic relevance between the final selected search results, which reduces the user experience.

[0106] Considering the advantages of BM25 in precise matching scenarios, the embodiment of the present application integrates the BM25 solution with the output results of the semantic retrieval model. Specifically, based on the encoding of each text sentence in the original conference text and the encoding of the retrieval sub-conditions, the semantic similarity between each text sentence in the original conference text and the retrieval sub-conditions is determined by calculating the BM25 algorithm. Then, the semantic similarity scores between each text sentence in the original conference text and the retrieval sub-conditions output by the above-mentioned semantic fuzzy retrieval model, and the semantic similarity between each text sentence in the original conference text and the retrieval sub-conditions calculated by the BM25 algorithm are fused, for example, weighted fusion is performed to obtain the semantic similarity scores between each text sentence in the fused original conference text and the retrieval sub-conditions, that is, the correlation between each text sentence in the fused original conference text and the retrieval sub-conditions is obtained.

[0107] In addition, considering that when the span of sentences related to the search sub-conditions in the conference text is too large, it does not conform to the fact that the actual search sub-condition description content is relatively concentrated, the embodiment of the present application constrains the span of the position distribution of the original conference text sentences related to the search sub-conditions, that is, according to the position distribution of each text sentence in the conference text in the conference text, the relevance of each text sentence in the conference text to the search sub-condition is corrected.

[0108] The embodiment of the present application first selects a second number of text segments having the highest relevance to the search sub-condition from each text sentence in the conference text;

[0109] Then, according to the rule that the greater the distance between other text sentences in the conference text and the second number of selected text sentences, the higher the penalty for the relevance of other text sentences to the search sub-conditions, the penalty for the relevance of other text sentences in the conference text to the search sub-conditions is determined; and based on the penalty for the relevance of other text sentences in the conference text to the search sub-conditions, the relevance of other text sentences in the conference text to the search sub-conditions is penalized.

[0110] Specifically, the embodiment of the present application sets that the Top2 of the semantic similarity scores between each text sentence in the fused conference original text and the search sub-condition is relatively accurate, and the semantic similarity scores between other text sentences in the conference original text and the search sub-condition should be penalized accordingly according to the distance from Top-2. Since there are two sentences in Top2, the penalty for the semantic similarity scores between other text sentences and the search sub-condition should only be selected for the smaller distance between the two sentences in Top2; based on the above processing, the semantic similarity score between other text sentences in the conference original text and the search sub-condition is the original score minus the distance penalty score.

[0111] Furthermore, the embodiment of the present application also performs abnormal text sentence filtering processing on the text sentences related to the search sub-conditions identified from the original conference text, that is, further refining the similarity scores between the conference text sentences and the search sub-conditions. First, based on the relevance of each text sentence in the conference text and the search sub-conditions, a third number of text sentences with the highest relevance to the search sub-conditions are selected from the conference text; then, based on the relevance of each text sentence in the selected third number of text sentences and the search sub-conditions, text sentences whose relevance to the search sub-conditions is greater than a first relevance threshold, or whose relevance to the search sub-conditions is greater than a second relevance threshold and whose normalized relevance to the search sub-conditions is greater than a third relevance threshold are selected from the third number of text sentences as text sentences related to the search sub-conditions;

[0112] The first correlation threshold is greater than the second correlation threshold, and the second correlation threshold is greater than the third correlation threshold.

[0113] Specifically, after extracting the text sentences with the TopK relevance to the search sub-condition from the conference text, the text sentences with the number of non-stop words less than or equal to 1 are deleted to filter out the low-information sentences. If the similarity score of the remaining text sentences with the search sub-condition is less than t1 (such as t1 = 0.6), the credibility of the text sentence is considered to be low, and the sentence is considered to be deleted; if the similarity score of the remaining text sentences with the search sub-condition is relatively low, for example, all less than t1, it means that the content of the search sub-condition is difficult, such as too brief or too general, so the threshold t2 will be lowered (such as t2 = 0.3), and the standardized score score_norm of the remaining text sentences and the search sub-condition will be further checked, and the threshold t3 will be set (such as t3 = 0.2). Finally, only the text sentences with a similarity score greater than t1 with the search sub-condition, or the similarity score with the search sub-condition is greater than t2 and the standardized similarity score with the search sub-condition is greater than t3 are retained as text sentences related to the search sub-condition.

[0114] Specifically, the standardized score score_norm is calculated as follows:

[0115]

[0116]

[0117]

[0118] The final strategy for selecting relevant text sentences is that for a text sentence i in the original conference text, its similarity score with the search sub-condition needs to meet (score i >t 1 )||((score i >t 2 )&&(score_norm i >t 3 )).

[0119] Through the above processing, text sentences related to the search condition or search sub-condition can be located from the original conference text. The text sentences related to the search condition or search sub-condition can also be output to the user so that the user can apply or know what text contents related to the search condition or search sub-condition are in the original conference text.

[0120] Based on the above processing, the semantic similarity between the search sub-conditions and each text sentence in the conference text can be determined respectively, and the text sentences related to the search sub-conditions can be identified from the text of the conference original text.

[0121] When there are multiple search sub-conditions, it is necessary to integrate the semantic similarity scores between each search sub-condition and each text sentence in the conference text to determine the semantic similarity scores between the search condition as a whole and each text sentence in the conference text.

[0122] Specifically, when the search condition contains multiple search sub-conditions (i.e., the number of texts equivalent to the reference text is greater than 1, where the number of texts can be determined based on the number of texts of granularity such as text sentences and text segments), the embodiment of the present application determines the relevance of each text sentence in the conference text to the search condition through the following processing, that is, determines the similarity score with the search condition:

[0123] First, according to the characteristics of the text sentence and the characteristics of each search sub-condition, the relevance between the text sentence and each search sub-condition is determined.

[0124] Specifically, referring to the description of the above embodiment of the application, the similarity scores between the text sentence and each search sub-condition in the search condition can be calculated and determined based on the characteristics of the text sentence and the characteristics of each search sub-condition.

[0125] Then, according to the relationship between each search sub-condition, the relevance of the text sentence and each search sub-condition is fused to determine the relevance of the text sentence and the search condition.

[0126] Specifically, when the user sets multiple search sub-conditions to form a complete search condition, the search condition is usually obtained by logically combining the multiple search sub-conditions. Therefore, there is a clear logical relationship between the various search sub-conditions in the search condition.

[0127] Based on the above-mentioned logical relationship, after determining the similarity scores between the text sentence and each search sub-condition respectively, the embodiment of the present application logically combines the similarity scores between the text sentence and each search sub-condition according to the logical relationship between each search sub-condition, thereby determining the similarity score between the text sentence and the search condition as a whole.

[0128] The above-mentioned logical combination of similarity scores between the text sentence and each search sub-condition can be seen in Table 1:

[0129] Table 1

[0130]

[0131] As shown in Table 1, for the text sentence in the original conference text, assuming that its similarity score with the retrieval sub-condition A is Its similarity score with the search condition B is Then, the similarity score between the text sentence and the search conditions A&B is The similarity score between this text sentence and the search condition A||B is The similarity score between this text sentence and the search condition AB is

[0132] Furthermore, the embodiment of the present application also normalizes the similarity scores between each text sentence in the original conference text and the search condition determined by the above method, so that the similarity scores between each text sentence in the original conference text and the search condition are processed between 0 and 1, so as to more intuitively represent the relevance between the text sentence in the original conference text and the search condition, and make the relevance between different text sentences and the search condition comparable. After the above processing, for the 1st to nth text sentences in the original conference text, their relevance to the search condition is expressed by p 1 ,p 2 …,p n express.

[0133] A2. Based on at least the relevance between each text segment in the target text related to the reference text and the reference text, a minutes generation process is performed on the full text content of the target text to obtain a target text minutes corresponding to the reference text.

[0134] Specifically, when each text segment related to the reference text is located in the target text, the full text content of the target text can be processed for minute generation based on each text segment related to the reference text to obtain a target text minute corresponding to the reference text.

[0135] For example, the text minutes are generated mainly based on the contents of each text segment in the target text related to the reference text, and supplemented by other text contents in the target text. The main content of the obtained target text minutes is the content related to the reference text.

[0136] Alternatively, based on the relevance of each text segment related to the reference text in the target text to the reference text, the contribution of each related text segment to generating the target text minutes is set, so that the higher the relevance of the text segment to the reference text, the higher the contribution to generating the target text minutes, so that the proportion of the text segment content related to the reference text contained in the finally generated target text minutes is proportional to the relevance of the reference text.

[0137] As a preferred implementation, the embodiment of the present application comprehensively considers the relevance of each text segment in the target text with the reference text, and performs a summary generation process on the full text content of the target text. That is, by executing the following steps A21-A22, a summary is generated on the full text content of the target text:

[0138] A21. Determine, based on the relevance of each text segment in the target text and the reference text, the contribution of each text segment in the target text to generating a text summary corresponding to the reference text.

[0139] Specifically, the embodiment of the present application determines the contribution of each text segment in the target text to the generation of the text minutes corresponding to the reference text according to the rule that the higher the correlation between the text segment in the target text and the reference text, the greater the contribution of the text segment in the target text to the generation of the text minutes corresponding to the reference text.

[0140] As an exemplary implementation, the embodiment of the present application adopts a text minutes generation model based on an attention mechanism to generate a text minutes of a target text. The text minutes generation model based on an attention mechanism can determine the contribution of each text segment in the target text to the generation of a text minutes corresponding to the reference text according to the relevance of each text segment in the target text to the reference text, and then generate a text minutes of the target text based on the contribution.

[0141] Since the text minutes generation model is a text decoding model based on the attention mechanism, the model can obtain a text minutes decoding result that meets the requirements by adjusting the attention coefficient of each text segment of the input target text. When the decoding process pays different attention to different text segments, the content of the text minutes finally decoded can change. Since the text minutes generation model based on the attention mechanism is a time-series output model, the attention coefficient of the text minutes decoding at the current moment for each text segment in the target text may also be related to the content of the target text minutes that have been output before the current moment.

[0142] Through training, the model can determine the distribution of attention coefficients of the current text minutes decoding on each text fragment of the target file based on the previous decoding results, that is, assign the correct attention coefficients to each text fragment of the input target text. Therefore, the model can determine the attention coefficients of the model for each text fragment in the target text when generating the text minutes of the target text.

[0143] Then, based on the attention coefficient of the text minutes of the target text to each text segment in the target text and the relevance of each text segment in the target text to the reference text, the contribution of each text segment in the target text to the generation of the text minutes corresponding to the reference text is determined.

[0144] Since the ultimate goal of the embodiment of the present application is to generate a text minutes corresponding to a reference text, only determining the attention coefficient of the text minutes generation to each text segment of the target text is not sufficient to make the target text minutes finally generated correspond to the reference text. In order to make the target text minutes finally generated correspond to the reference text, the embodiment of the present application also combines the attention coefficient of the text minutes generating the target text to each text segment in the target text with the correlation between each text segment in the target text and the reference text, and uses them together to determine the contribution of each text segment in the target text to the generation of the text minutes corresponding to the reference text.

[0145] Exemplarily, the attention coefficient of each text fragment in the target text for generating the text minutes of the target text is multiplied by the relevance of each text fragment in the target text to the reference text, and then the product results corresponding to each text fragment are normalized. The normalized values ​​corresponding to each text fragment are finally obtained as the contribution of each text fragment in the target text to generating the text minutes corresponding to the reference text.

[0146] The above contribution determination scheme can also be referred to in the subsequent embodiments for the example description of the specific process of generating the target text minutes.

[0147] A22. Performing a minutes generation process on the full text content of the target text at least according to the contribution of each text segment in the target text to the generation of the text minutes corresponding to the reference text, to obtain the target text minutes corresponding to the reference text.

[0148] Specifically, the above-mentioned text minutes generation model based on the attention mechanism first generates text minutes decoding features according to the characteristics of the target text and the contribution of each text fragment in the target text to the generation of the text minutes corresponding to the reference text; then, based on the text minutes decoding features, the text minutes decoding process is performed to generate the text minutes of the target text.

[0149] Since the embodiment of the present application determines the characteristics of the target text by obtaining the characteristics of each text segment of the target text, the characteristics of each text segment of the target text are predetermined. However, the contribution of each text segment of the target text to the generation of the text minutes is different. Therefore, the text minutes decoding features can be directly generated based on the characteristics of each text segment of the target text and the contribution of each text segment in the target text to the generation of the text minutes corresponding to the reference text.

[0150] Then, the text minutes decoding features are decoded within the preset dictionary range to obtain a decoding result. According to the above method, the model is used to decode the full text of the target text to obtain the target text minutes corresponding to the reference text.

[0151] Still taking the example of generating meeting minutes that meet the user's search criteria from the original meeting text, Figure 4 The structure of the text minutes generation model based on the attention mechanism and its decoding process of generating the target text minutes are shown.

[0152] Assume that for a user's search condition, the meeting minutes text word sequence that the model has solved historically is represented as y 1 ,y 2 ,…,y t-1 , the hidden layer feature of each sentence in the original conference text (assuming there are n sentences in total) is t represents the current decoding time. For the first to n text sentences in the original conference text, their relevance to the search condition is expressed as p 1 ,p 2 …,p n Represents. The hidden features of each sentence in the original conference text And the relevance p of each sentence to the search condition 1 ,p 2 …,p n , input the model, specifically the decoding end of the model and the original text interactive attention module.

[0153] The historical solved meeting minutes text word sequence y 1 ,y 2 ,…,y t-1 After the model’s decoding hidden layer feature expression module, the hidden layer state feature at the current decoding moment is obtained as d t The decoding hidden layer feature expression module can input a certain search condition of the user, use the historical meeting minutes text word sequence that has been decoded by the model, and output the hidden layer state features at the current decoding moment. The network structure of the decoding hidden layer feature expression module can use the decoder partial encoding model under the Transformer scheme or a unidirectional LSTM structure.

[0154] In the interactive attention module between the decoding end and the original text, firstly, based on the attention mechanism, the hidden state feature d is determined at the current decoding moment. t The attention coefficient of the hidden layer features of the jth text sentence in the original conference text Then, according to the hidden layer state feature d t The attention coefficient of the hidden layer features of the jth text sentence in the original conference text And the relevance p between the jth text sentence in the original conference text and the search condition j , calculate and determine the contribution of the jth text sentence in the original meeting text to the generation of meeting minutes Finally, based on the hidden features of each text sentence in the original conference text and the contribution of each text sentence in the original conference text to the generation of conference minutes, the text minutes decoding features are generated. The specific calculation process of the above processing is as follows:

[0155]

[0156]

[0157]

[0158] Wherein, j = 1, 2, ... n, representing n text sentences in the original meeting text. Attention() represents the attention mechanism calculation function, which can adopt self-attention and additive attention. In the embodiment of the present application, the decoding end fully considers the influence of the information related to the retrieval condition content in all the text sentences of the original meeting text on the generation of the meeting minutes, and optimizes the original attention coefficient to be Text Minutes Decoding Features In order to provide the decoding end with semantic fuzzy retrieval features to guide the selection of relevant content, the context vector representation with different degrees of hidden features of the original sentences of the conference is focused on.

[0159] from and The calculation process shows that if a sentence in the original meeting text has a higher relevance to the search condition, that is, the larger the search matching feature value, the larger the optimized attention coefficient value corresponding to the sentence will be, and the final text minutes decoding feature will be more accurate. The greater the contribution, the greater the ability of selecting and retrieving relevant original content.

[0160] Figure 4 The word prediction module in the input text summary decoding feature Calculate the probability of a word distributed in the dictionary size, and output the word corresponding to the current decoding moment. The network structure of the word prediction module can use a linear layer connected to a nonlinear activation function layer, and the decoding algorithm can use the beamsearch algorithm. According to the above algorithm, the model determines the decoding features of the text minutes based on the original meeting text at each moment, and decodes and outputs it, so that the meeting minutes corresponding to the original meeting text and meeting the search conditions can be obtained.

[0161] As a preferred implementation, since the target text minutes ultimately generated by the embodiment of the present application needs to correspond to the reference text, in addition to clarifying the contribution of each text segment in the target text to the generation of the text minutes corresponding to the reference text, so that the target text minutes ultimately generated contains the target text content related to the reference text, the embodiment of the present application also directly uses the features of the reference text to generate the text minutes of the target text, thereby further improving the correlation between the generated target text minutes and the reference text.

[0162] Based on the above ideas, when generating text minutes decoding features, the embodiment of the present application generates text minutes decoding features according to the features of the target text, the features of the reference text, and the contribution of each text fragment in the target text to generating a text minute corresponding to the reference text.

[0163] For example, in Figure 4 In the attention mechanism-based text summary generation model shown in the figure, in addition to the sentence hidden features of each sentence in the original conference text And the relevance p of each sentence to the search condition 1 ,p 2 …,p n In addition to inputting the model, the features of the search conditions are also input into the model, for example, the hidden layer features of each word in the search condition are input into the model. The model is input so that when the model performs decoding processing, it can decode the text content of the original conference text with reference to the search conditions, thereby improving the correlation between the decoding result and the search conditions.

[0164] For example, the latent features of each sentence in the original conference text The relevance of each sentence to the search condition p 1 ,p 2 …,p n , and the hidden layer features of each word in the search condition The decoding end of the model is input into the interactive attention module of the original text, so that the model can use the retrieval conditions as a reference when determining the attention coefficient of each text sentence of the original meeting text, determining the contribution of each text sentence of the original meeting text to the generation of meeting minutes corresponding to the retrieval conditions, and generating text minutes decoding features, thereby making the final decoded meeting minutes more relevant to the retrieval conditions and avoiding the meeting minutes from deviating from the retrieval conditions.

[0165] As an optional implementation, the present application embodiment proposes the following Figure 5 The structure of the text summary generation model based on the attention mechanism is shown in Figure 4The model structure shown in the figure adds a decoding end and retrieval interactive attention module before the decoding end and original text interactive attention module. The decoding end and retrieval interactive attention module mainly realizes the interaction between the model hidden state features and the retrieval condition features to generate reference decoding features. The reference decoding features then interact with the original text features of the conference through the decoding end and original text interactive attention module.

[0166] based on Figure 5 The model structure shown, when generating text minutes decoding features, first generates reference decoding features based on the features of the reference text; then generates text minutes decoding features based on the reference decoding features, the features of the target text, and the contribution of each text fragment in the target text to generating the text minutes corresponding to the reference text.

[0167] Specifically, see Figure 5 As shown, the features of the retrieval conditions and the hidden state features are input into the decoding end of the model and the retrieval interactive attention module, so that the features of the retrieval conditions are integrated with the hidden state features of the model to obtain reference decoding features.

[0168] Furthermore, the embodiment of the present application determines the characteristics of the reference text by the characteristics of each text fragment of the reference text, and the influence of each text fragment of the reference text on the generation of the target text minutes corresponding to the reference text is also different. For example, the key entity words in the reference text can express the semantics of the reference text to a large extent, so it has a greater reference value for generating the target text minutes corresponding to the reference text, while the non-entity words in the reference text, such as modal particles, modifiers, etc., have a relatively smaller reference value for generating the target text minutes corresponding to the reference text. Therefore, the embodiment of the present application refers to the scheme for determining the contribution of each text fragment in the target text to generating the text minutes corresponding to the reference text introduced in the above embodiment, and determines the contribution of each text fragment in the reference text to generating the text minutes corresponding to the reference text. Among them, the text fragments in the reference text can be text content of any granularity such as words, phrases, sentences, and text paragraphs.

[0169] After the contribution of each text segment in the reference text to the generation of the text minutes corresponding to the reference text is clarified, a reference decoding feature is generated according to the features of the reference text and the contribution of each text segment in the reference text to the generation of the text minutes corresponding to the reference text.

[0170] Specifically, the embodiment of the present application determines the characteristics of the reference text by acquiring the characteristics of each text fragment of the reference text, and the contribution of each text fragment of the reference text to the generation of the text minutes corresponding to the reference text is different. Therefore, the embodiment of the present application generates a reference decoding feature based on the characteristics of each text fragment of the reference text and the contribution of each text fragment in the reference text to the generation of the text minutes corresponding to the reference text.

[0171] Continuing with the above example of generating meeting minutes that meet the user's search criteria from the original meeting text, Figure 5 The text summary generation model based on the attention mechanism shown in the figure assumes that the hidden layer feature of each word in the retrieval condition is The hidden state feature d at the current decoding moment t , and the hidden layer features of each word in the retrieval condition The input decoding end and the retrieval interactive attention module use the attention mechanism to first determine the contribution of each word in the retrieval condition to the generation of the text minutes corresponding to the retrieval condition, and then generate reference decoding features based on the features of each word in the retrieval condition and the contribution of each word in the retrieval condition to the generation of the text minutes corresponding to the retrieval condition. The specific calculation process is as follows:

[0172]

[0173]

[0174] Where i=1, 2, ...m1+m2, represents m1+m2 words in the search condition, and m1+m2 represents the sum of the number of words in the two search sub-conditions contained in the search condition. Attention() represents the attention mechanism calculation function, which can use self-attention and additive attention. It represents the attention coefficient of the decoded hidden state feature to the hidden feature of the i-th word in the retrieval condition at the current decoding moment. This coefficient also represents the contribution of the i-th word in the retrieval condition to the generation of the text minutes corresponding to the retrieval condition. The context vector representations of different degrees of hidden layer features of the focus search condition words obtained after the decoding end interacts with the search, are named reference decoding features in the embodiment of the present application. It can be understood that the technical solution of the embodiment of the present application fully considers the contribution of the information in the search condition to the generation of the meeting minutes at the decoding end.

[0175] See also Figure 5 As shown, after obtaining the reference decoding feature After that, the reference decoding features, the features of each text sentence in the original conference text, and the relevance of each text sentence in the original conference text to the search criteria are input into the decoding end and the original text interactive attention module, so that the model can determine the contribution of each text sentence in the original conference text to the generation of the text minutes corresponding to the search criteria, as well as the text minutes decoding features through attention interaction operations. The specific calculation process is as follows:

[0176]

[0177]

[0178]

[0179] in, is the context vector after the decoding end interacts with the retrieval at time t, that is, the reference decoding feature. is the hidden layer feature of the jth sentence in the original conference text, i = 1, 2, ... n. Attention() represents the attention mechanism calculation function, which can adopt self-attention and additive attention. It represents the attention coefficient of the context vector after the interaction between the decoder and the retrieval end to the jth sentence in the original conference text at the current decoding moment, that is, the attention coefficient of the decoder end to the jth text sentence in the original conference text when generating the text minutes corresponding to the retrieval condition. j Represents the similarity between the search condition and the jth sentence in the original conference text.

[0180] From the above-mentioned embodiments, it can be seen that when decoding the text minutes, the embodiments of the present application not only consider the contribution of each text segment of the target text to the generation of the target text minutes corresponding to the reference text, but also consider the contribution of each text segment of the reference text to the generation of the target text minutes corresponding to the reference text. This ensures that the entire target text minutes generation process has the ability to select and retrieve relevant target text content and reference text content, thereby improving the relevance of the target text minutes finally generated with the reference text and the reference text-related content in the target text, that is, the target text minutes finally generated corresponds to the reference text.

[0181] It should be noted that in the above embodiment, the features of each text segment of the target text and the features of each text segment of the reference text are used to illustrate the process of obtaining the reference decoding features and the text minutes decoding features, as well as the processing process of finally decoding to obtain the target text minutes. When the embodiment of the present application is actually implemented, the overall features of the target text and the overall features of the reference text can be directly input into the above-mentioned text minutes generation model based on the attention mechanism to obtain the reference decoding features and the text minutes decoding features. At this time, the training process of the model and the specific processing process of the model can be performed with reference to the introduction of the above embodiment.

[0182] In addition, the above Figure 4 and Figure 5 The names of the various processing modules of the text minutes generation model based on the attention mechanism shown are named in combination with the specific processing objects. When the target text and reference text actually processed are other types of texts rather than the original meeting text and the retrieval conditions, the names of the various processing modules can be adaptively changed according to the actual processing objects. The embodiments of the present application do not limit the names of the various processing modules of the above-mentioned text minutes generation model based on the attention mechanism, but mainly introduce the functions and processing contents of each processing module, thereby specifically introducing the processing process and the functions implemented of the text minutes generation model based on the attention mechanism.

[0183] In each of the above embodiments, the correlation between each text segment of the target text and the reference text is introduced respectively, and the specific implementation method of the target text to obtain the target text minutes corresponding to the reference text is processed by the text minutes generation process. Because for computer equipment, it processes the text, which is essentially the processing of the characteristics of the text, that is to say, the processing content of the text contained in the text minutes generation method proposed in the embodiment of the present application is essentially the processing of the characteristics of the text. Therefore, the accuracy of the text characteristics will directly affect the accuracy of the minutes generation process of the text. Below, the embodiment of the present application will be illustrated by the method for obtaining the characteristics of the target text and the characteristics of the reference text.

[0184] Normally, text features can be obtained by encoding the text with an encoder. For example, by using a word-level encoder structure, the encoding features of the target text and the reference text can be obtained. However, the target text for which the minutes need to be generated is usually a long text. For example, the main feature of a meeting text is that the text is long. A one-hour meeting may contain 10,000 to 20,000 words. If the features of the meeting text are obtained with the help of a conventional word-level encoder, a lot of content will be consumed, and it will not be able to capture long-distance dependency information well, resulting in inaccurate or incomplete extracted text features. This also leads to the conventional minutes generation method, which can only generate summaries for shorter articles such as news, emails, and interpersonal conversations with fewer rounds, and it is not competent for the task of generating minutes for long target texts.

[0185] In order to improve the effect of text feature extraction, as an exemplary implementation, the embodiment of the present application determines the characteristics of the target text by obtaining the characteristics of each text fragment in the target text, and determines the characteristics of the reference text by obtaining the characteristics of each text fragment in the reference text.

[0186] When acquiring the features of each text segment in the target text, the features of each text segment are extracted by executing the following steps B1-B4:

[0187] B1. Divide the target text into text segments to determine the text segments contained in the target text.

[0188] B2. Perform word segmentation processing on each text segment in the target text to determine each word segment contained in each text segment.

[0189] B3. Extract the word segmentation features of each word segmentation contained in each text fragment and the fused context information.

[0190] B4. Determine the features of each text segment based on the segmentation features of each segmented word contained in each text segment and the fused context information.

[0191] Specifically, the embodiment of the present application constructs a word-sentence-chapter hierarchical information encoding model to extract sentence-level latent features and chapter-level latent features of the target text that integrate contextual information, that is, to extract features of each text segment of the target text and the overall features of the target text.

[0192] The embodiment of the present application first divides the target text into text segments, and performs word segmentation on the divided text segments to determine each word segment contained in each text segment. Partitioning the target text into text segments may be dividing the target text into text sentences, such as dividing the text sentences according to punctuation marks, or extracting text segments by sliding a sliding window with a fixed number of words on the target text. Word segmentation of text segments may be implemented using an existing word segmentation algorithm, which will not be described in detail in the embodiment of the present application. After the above-mentioned text segmentation and word segmentation processing, the text segment features and chapter features of the target text may be extracted using the above-mentioned word-sentence-chapter hierarchical information encoding model for the word-sentence-chapter hierarchical information encoding model.

[0193] See also Figure 6 As shown, the word-sentence-chapter hierarchical information encoding model includes a word hidden layer feature expression module, a sentence representation extraction module, a sentence hidden layer feature expression module and a chapter feature extraction module. Below, the original conference text is used to represent the target text, and the text sentences in the original conference text are used to represent the text fragments in the target text. Taking the extraction of the features of the text sentences and the chapter features of the original conference text as an example, the processing process of extracting the features of the text fragments of the target text and the overall features of the target text is introduced.

[0194] The above-mentioned word hidden layer feature expression module refers to inputting the word representation of each word for each text sentence in the original conference text, and outputting the word hidden layer feature that integrates the context information of the current sentence. The network structure of the word hidden layer feature expression module can use the encoder part model under the Transformer scheme or a bidirectional LSTM structure. Assume that after the original conference text is divided into sentences, there are n sentences in total, and each sentence contains a word sequence of W n1 ,W n2 ,…,W nmn , where n represents the nth sentence of the original conference text, m n Represents the total number of words contained in the nth sentence. It represents the hidden layer features of each word in the first sentence of the original conference text after integrating the context information of the current sentence. m1 represents that there are m1 words in the first sentence. It represents the hidden layer features of the m2 words in the second sentence of the original conference text after integrating the context information of the current sentence. Indicates the mth sentence in the nth sentence of the original conference text n The hidden layer features of a word after integrating the context information of the current sentence.

[0195] The sentence representation extraction module mentioned above compresses the word representations of multiple words in the input sequence to obtain the sentence representation vector. The hidden layer features of all the words in the first sentence of the original text of the meeting After the sentence representation extraction module, the sentence representation vector of the first sentence is s 1 By analogy, the sentence representation vectors of the first to nth sentences in the conference text can be represented as the sequence s 1 ,s 2 …,s n The embodiment of the present application does not limit the network structure of the sentence representation extraction module, and may adopt technologies such as attention mechanism or pooling.

[0196] The sentence hidden layer feature expression module mentioned above refers to inputting all the sentence representation vectors of the original conference text and outputting the sentence hidden layer features that integrate the current sentence context information. Similar to the word hidden layer feature expression module mentioned above, the network structure of the sentence hidden layer feature expression module can utilize the encoder part model under the Transformer solution or the bidirectional LSTM structure. Represents the sentence hidden features of n sentences in the conference text after integrating the context information.

[0197] The above-mentioned chapter feature extraction module is similar to the above-mentioned sentence representation extraction module, which compresses the sentence hidden feature representations of multiple sentences in the input sequence to obtain the chapter representation vector. The sentence representation vectors of the first to the nth sentences of the original conference text can be represented as the sequence After the passage feature extraction module, the original conference text passage feature u is obtained. The embodiment of the present application does not limit the network structure of the passage feature extraction module, and can adopt attention mechanism or pooling technology.

[0198] Through the above-mentioned word-sentence-chapter level information encoding model, the features of each text sentence of the conference original text and the chapter features of the conference original text, that is, the overall features of the conference original text, can be obtained respectively. In addition, the features of the text sentences of the conference original text, the features of the segmented words contained in the text sentences, and the chapter features of the conference original text are all features that integrate contextual information. Therefore, the above-mentioned conference original text feature extraction scheme in the embodiment of the present application can better capture the long-distance dependency information in the long-length conference text and obtain more accurate conference text features.

[0199] As an optional implementation, the above-mentioned word-sentence-paragraph level information encoding model can also omit the sentence hidden layer feature expression module, and directly convert the sentence representation vector s of each text sentence output by the sentence representation extraction module into 1 ,s 2 …,s n , as the features of each text sentence of the conference text, and, according to the features of each text sentence s 1 ,s 2 …,s n , determine the discourse characteristics u of the conference text.

[0200] When acquiring the features of each text segment in the reference text, the features of each text segment are extracted by executing the following steps C1-C2:

[0201] C1. Perform word segmentation processing on the reference text to determine each word segmentation contained in the reference text.

[0202] C2. Extract the word segmentation features of each word contained in the reference text and integrate the context information.

[0203] Specifically, the embodiment of the present application determines the overall features of the reference text by extracting the features of each text segment of the reference text. The text segment of the reference text may be a text content of any granularity such as a word, a phrase, a text sentence, a text segment, etc. in the reference text. In the embodiment of the present application, the features of the reference text are determined by extracting the features of each word segment of the reference text.

[0204] Therefore, firstly, the reference text is segmented, for example, by using a segmentation model or a segmentation algorithm to segment the reference text and determine each segmentation contained in the reference text. Then, the segmentation features of each segmentation contained in the reference text that integrate the context information are extracted respectively, and the overall features of the reference text are obtained by combining the segmentation features of each segmentation contained in the reference text.

[0205] Furthermore, if the number of texts contained in the reference text is greater than 1, that is, the reference text contains multiple text sentences, the embodiment of the present application first merges or filters the text sentences contained in the reference text according to the relationship between the reference texts, integrates them into a reference text, and then performs word segmentation, word segmentation feature extraction and reference text feature extraction on the reference text.

[0206] For example, the embodiments of the present application are as follows Figure 7 The word hidden layer feature expression module shown extracts the word segmentation features of each word of the reference text.

[0207] Take the extraction of features of the search conditions input by the user when obtaining the original meeting minutes as an example. If the user's search condition is a single condition, after the user's search condition is segmented, the search word sequence is input into the above-mentioned word hidden layer feature expression module, and the word hidden layer features fused with the context information of the search condition are output, that is, the word segmentation features of the fused context information of each word segmentation contained in the search condition are obtained.

[0208] If the user's search condition is a combination of multiple search sub-conditions, the word sequence of a single or multiple search sub-conditions is input into the above-mentioned word hidden layer feature expression module according to the method shown in Table 2 below. The word hidden layer feature expression module correspondingly outputs the word hidden layer features of the fused context information of the single or multiple search conditions.

[0209] Table 2

[0210] Compound conditions Input Description Enter the number of sequences A&B Concatenate the A and B sub-condition sequences into one sequence input 1 A||B Enter the A and B sub-condition sequences separately 2 AB Enter only the A sub-condition sequence 1

[0211] According to the idea of ​​merging or filtering search conditions shown in Table 2, for more complex search sub-condition combinations, such as {A&B}-C, the above method can be used to process them into a text sequence, that is, the text sequence input after the sub-conditions A and B are concatenated.

[0212] The network structure of the word hidden layer feature expression module mentioned above can use the encoder part model or bidirectional LSTM structure under the Transformer solution. Figure 7 As shown in , suppose an input sequence in a certain search condition has a total of m1 words, which are After the word sequence is input into the above-mentioned word hidden layer feature expression module, the hidden layer feature representation of each word in the retrieval condition is obtained as follows: In particular, if the search condition is decomposed into multiple sequences according to the method described above, the extraction process is similar. For example, another input sequence in the figure has a total of m2 words, which are After the word sequence is input into the above-mentioned word hidden layer feature expression module, the hidden layer feature representation of each word is obtained as

[0213] According to the above method, the word segmentation features of each search sub-condition can be obtained respectively. Finally, the word segmentation features of each word included in the search sub-condition are spliced ​​according to the word segmentation order to obtain the overall features of the search condition.

[0214] After determining the features of each text segment of the target text and the features of each text segment of the reference text respectively through the above-mentioned processing, the embodiment of the present application also performs feature fusion processing on the features of each text segment of the target text and the features of each text segment of the reference text to obtain target text features fused with reference text features, and / or reference text features fused with target text features.

[0215] That is, the embodiments of the present application integrate the features of the reference text into the features of the target text, and / or integrate the features of the target text into the features of the reference text, so that the features of the reference text and / or the target text include not only their own features but also the features of the other party.

[0216] In the embodiment of the present application, the features of the reference text are integrated into the features of the target text, and at the same time, the features of the target text are integrated into the features of the reference text. When the technical solution of the embodiment of the present application is actually implemented, the features of one party can be selected to be integrated into the other party according to the introduction of the embodiment of the present application.

[0217] When integrating the target text features into the reference text features, the chapter features of the target text and / or the features of each text segment of the target text may be integrated into the features of each text segment of the reference text, or integrated into the overall features of the reference text. The chapter features of the target text are determined based on the features of each text segment of the target text.

[0218] In the embodiment of the present application, the chapter features of the target text are first determined based on the features of each text segment of the target text. Then, the chapter features of the target text and the features of each text segment of the target text are respectively integrated into the features of each text segment of the reference text; in addition, the features of each text segment of the reference text are integrated into the features of each text segment of the target text. Finally, the features of each text segment of the reference text obtained incorporate the chapter features of the target text and the features of each text segment, and the features of each text segment of the target text obtained incorporate the features of each text segment of the reference text.

[0219] Exemplarily, the target text is still represented by the above-mentioned conference text, and the reference text is represented by the user search conditions. The embodiment of the present application fully considers the information fusion of the user search conditions and the conference text, that is, when extracting the hidden features of each sentence in the final conference text, the relevant search condition information will be integrated. At the same time, when extracting the hidden features of each word in the final user's search conditions, the conference text information will also be integrated.

[0220] The embodiment of the present application constructs an information fusion model to realize the information fusion of user search conditions and conference original text. Figure 8 As shown, the information fusion model includes a word hidden layer feature expression module, a word feature extraction module and an information mutual fusion module.

[0221] The functions and processing of the above-mentioned word hidden layer feature expression module can be found in Figure 7 The functional introduction of the word hidden layer feature expression module.

[0222] The word feature extraction module inputs the original conference text chapter-level hidden layer feature u and the retrieval condition. Each word hidden layer feature represents Output the hidden features of each word in the search condition after integrating the original text information In the embodiment of the present application, the word feature extraction module adopts a recursive network structure, the conference original text chapter-level hidden layer feature u is represented as the initial state, the hidden layer feature of each word in the search condition is input, and the calculation process of recursively obtaining the hidden layer feature of each word in the search condition after integrating the conference original text chapter information is as follows:

[0223]

[0224]

[0225] …

[0226]

[0227] The above recursive network structure may adopt a structure such as LSTM or GRU.

[0228] In particular, if the search condition is a combination of multiple search sub-conditions, there will be multiple word hidden layer features after the above processing. Figure 8 As shown, for another word hidden feature representation The calculation process of extracting the hidden features of each word in the search condition after integrating the original conference text information in the embodiment of the present application is consistent with the above process, and finally the result is

[0229] The information mutual fusion module obtains the hidden layer feature of each word in the search condition after integrating the original conference text information, and inputs the hidden layer feature and the sentence hidden layer feature of each text sentence of the original conference text. Output the hidden features of each word in the retrieval condition after further integrating the original sentence information And the hidden features of each sentence in the original conference text that integrates relevant search term information The information fusion module can utilize structures such as self-attention mechanism or bidirectional LSTM.

[0230] From the above introduction, it can be seen that when the embodiment of the present application performs feature extraction on the target text and the reference text, it can not only extract the features of each text fragment of the target text and the reference text that integrates the context information, but also realize the fusion of the target text features and the reference text features, so that the feature information of the target text and the reference text is richer, which is more conducive to obtaining the target text minutes corresponding to the reference text by performing text minutes generation processing on the target text.

[0231] For example, based on the features of the target text extracted in the above manner, a text minutes generation process is performed, or the features of the target text extracted in the above manner and the features of the reference text are combined and applied to perform text minutes generation process. Since the features of the reference text are integrated into the features of the target text, the generated target text minutes can be related to the reference text.

[0232] In the embodiment of the present application, the features of each text segment of the target text extracted in the above manner and the features of each text segment of the reference text are input into Figure 5 The text minutes generation model based on the attention mechanism shown generates a target text minutes corresponding to the reference text. For the specific minutes generation process, please refer to the introduction of the above embodiment.

[0233] Corresponding to the above-mentioned text minutes generation method, the present application embodiment also proposes a text minutes generation device, see Fig. 9 As shown, the device comprises:

[0234] The data acquisition unit 100 is used to acquire a target text and a reference text, wherein the reference text is determined based on the content of the target text that the user is concerned about;

[0235] The minutes generating unit 110 is used to perform minutes generating processing on the target text based on locating the associated content of the reference text from the target text, so as to obtain the target text minutes corresponding to the reference text.

[0236] As an optional implementation, based on locating the associated content of the reference text from the target text, performing a minutes generation process on the target text to obtain a target text minutes corresponding to the reference text includes:

[0237] Based on locating text segments related to the reference text from the target text, a minutes generation process is performed on the full text content of the target text to obtain a target text minutes corresponding to the reference text.

[0238] As an optional implementation, based on locating text segments related to the reference text from the target text, a minutes generation process is performed on the full text content of the target text to obtain a target text minutes corresponding to the reference text, including:

[0239] By determining the relevance of each text segment in the target text to the reference text, a text segment related to the reference text is located from the target text;

[0240] At least based on the relevance between each text segment in the target text related to the reference text and the reference text, the full text content of the target text is processed for minute generation to obtain a target text minute corresponding to the reference text.

[0241] As an optional implementation, based at least on the relevance between each text segment in the target text related to the reference text and the reference text, a minutes generation process is performed on the full text content of the target text to obtain a target text minutes corresponding to the reference text, including:

[0242] Determining, based on the relevance of each text segment in the target text and the reference text, the contribution of each text segment in the target text to generating a text summary corresponding to the reference text;

[0243] At least according to the contribution of each text segment in the target text to the generation of the text minutes corresponding to the reference text, the full text content of the target text is processed to generate the minutes to obtain the target text minutes corresponding to the reference text.

[0244] As an optional implementation, determining the relevance of each text segment in the target text with the reference text includes:

[0245] Obtain the features of the target text and the features of the reference text respectively;

[0246] According to the characteristics of each text segment in the target text and the characteristics of the reference text, the relevance of each text segment in the target text and the reference text is determined respectively.

[0247] As an optional implementation, the minutes generating unit 110 is further configured to:

[0248] The features of each text segment in the target text and the features of the reference text are interactively calculated based on the attention mechanism to obtain the reference text features after information improvement.

[0249] As an optional implementation, the features of each text segment in the target text and the features of the reference text are interactively operated based on the attention mechanism to obtain the reference text features after information improvement, including:

[0250] Based on the features of each text segment in the target text and the features of the reference text, calculating and determining the similarity between each text segment in the target text and the reference text;

[0251] According to the similarity between each text segment in the target text and the reference text, selecting a first number of text segments having the highest similarity with the reference text from the target text;

[0252] An attention-based interactive operation is performed on the features of the first number of text segments selected from the target text and the features of the reference text to obtain information-completed reference text features.

[0253] As an optional implementation, when the number of texts in the reference text is greater than 1, the relevance of each text segment in the target text to the reference text is determined according to the features of each text segment in the target text and the features of the reference text, including:

[0254] For each text segment in the target text, its relevance to the reference text is determined by the following processing:

[0255] Determine the relevance of the text segment to each reference text based on the characteristics of the text segment and the characteristics of each reference text;

[0256] According to the relationship between each reference text, the relevance of the text segment and each reference text is fused to determine the relevance of the text segment and the reference text.

[0257] As an optional implementation, the minutes generating unit 110 is further configured to:

[0258] According to the features of each text segment in the target text and the features of the reference text, the semantic similarity between each text segment in the target text and the reference text is determined by calculating the BM25 algorithm;

[0259] The relevance between each text segment in the target text and the reference text, as well as the semantic similarity between each text segment in the target text and the reference text, are fused to obtain the relevance between each text segment in the fused target text and the reference text.

[0260] As an optional implementation, the minutes generating unit 110 is further configured to:

[0261] According to the position distribution of each text segment in the target text, the relevance between each text segment in the target text and the reference text is corrected.

[0262] As an optional implementation, the relevance between each text segment in the target text and the reference text is modified according to the position distribution of each text segment in the target text, including:

[0263] Selecting, from each text segment in the target text, a second number of text segments having the highest relevance to the reference text;

[0264] Determining the penalty degree for the relevance between other text segments in the target text and the reference text according to a rule that the greater the distance between other text segments in the target text and the selected second number of text segments, the higher the penalty degree for the relevance between other text segments and the reference text;

[0265] According to the penalty degree of the relevance of other text segments in the target text to the reference text, the relevance of other text segments in the target text to the reference text is penalized.

[0266] As an optional implementation, the minutes generating unit 110 is further configured to:

[0267] According to the relevance of each text segment in the target text to the reference text, selecting a third number of text segments with the highest relevance to the reference text from the target text;

[0268] According to the relevance between each text segment in the selected third number of text segments and the reference text, a text segment having a relevance to the reference text greater than a first relevance threshold, or a relevance to the reference text greater than a second relevance threshold and a normalized relevance to the reference text greater than a third relevance threshold is selected from the third number of text segments as a text segment related to the reference text;

[0269] The first correlation threshold is greater than the second correlation threshold, and the second correlation threshold is greater than the third correlation threshold.

[0270] As an optional implementation, determining the contribution of each text segment in the target text to generating a text summary corresponding to the reference text according to the relevance between each text segment in the target text and the reference text includes:

[0271] Determining an attention coefficient of a text summary for generating the target text for each text segment in the target text;

[0272] According to the attention coefficient of the text minutes for generating the target text to each text segment in the target text, and the relevance of each text segment in the target text to the reference text, the contribution of each text segment in the target text to generating the text minutes corresponding to the reference text is determined.

[0273] As an optional implementation, at least according to the contribution of each text segment in the target text to generating the text minutes corresponding to the reference text, the full text content of the target text is processed to generate the minutes to obtain the target text minutes corresponding to the reference text, including:

[0274] Generating text minutes decoding features based at least on the features of the target text and the contribution of each text segment in the target text to generating a text minutes corresponding to the reference text;

[0275] A text summary of the target text is generated according to the text summary decoding features.

[0276] As an optional implementation, generating text minutes decoding features based at least on the features of the target text and the contribution of each text segment in the target text to generating a text minutes corresponding to the reference text includes:

[0277] The text minutes decoding features are generated based at least on the features of the target text, the features of the reference text, and the contribution of each text segment in the target text to the generation of the text minutes corresponding to the reference text.

[0278] As an optional implementation, generating text minutes decoding features based on at least the features of the target text, the features of the reference text, and the contribution of each text segment in the target text to generating a text minutes corresponding to the reference text includes:

[0279] generating a reference decoding feature based at least on the feature of the reference text;

[0280] A text minutes decoding feature is generated according to the reference decoding feature, the feature of the target text, and the contribution of each text segment in the target text to the generation of a text minutes corresponding to the reference text.

[0281] As an optional implementation, the minutes generating unit 110 is further configured to:

[0282] Determining the contribution of each text segment in the reference text to generating a text summary corresponding to the reference text;

[0283] Generating reference decoding features based at least on the features of the reference text includes:

[0284] A reference decoding feature is generated according to the feature of the reference text and the contribution of each text segment in the reference text to the generation of a text summary corresponding to the reference text.

[0285] As an optional implementation, generating reference decoding features according to the features of the reference text and the contribution of each text segment in the reference text to generating a text summary corresponding to the reference text includes:

[0286] Generate reference decoding features according to features of each text segment of the reference text and contribution of each text segment in the reference text to generating a text summary corresponding to the reference text;

[0287] Generating text minutes decoding features according to the reference decoding features, the features of the target text, and the contribution of each text segment in the target text to generating a text minutes corresponding to the reference text includes:

[0288] A text minutes decoding feature is generated according to the reference decoding feature, the features of each text segment of the target text, and the contribution of each text segment in the target text to generating a text minutes corresponding to the reference text.

[0289] As an optional implementation, generating text minutes decoding features according to the features of the target text and the contribution of each text segment in the target text to generating a text minutes corresponding to the reference text includes:

[0290] A text minutes decoding feature is generated according to the features of each text segment in the target text and the contribution of each text segment in the target text to the generation of a text minutes corresponding to the reference text.

[0291] As an optional implementation, a target text and a reference text are obtained, and based on locating the associated content of the reference text from the target text, a minutes generation process is performed on the target text to obtain a target text minutes corresponding to the reference text, including:

[0292] Obtaining the characteristics of the target text and the characteristics of the reference text;

[0293] The features of the target text and the features of the reference text are input into a pre-trained text minutes generation model based on the attention mechanism, so that the text minutes generation model based on the attention mechanism locates the related content of the reference text from the target text, performs minutes generation processing on the target text, and obtains the target text minutes corresponding to the reference text.

[0294] As an optional implementation, the features of the target text are obtained by acquiring the features of each text segment in the target text; the features of the reference text are obtained by acquiring the features of each text segment of the reference text.

[0295] As an optional implementation, the features of each text segment in the target text are obtained, including:

[0296] Performing text segmentation processing on the target text to determine the text segments contained in the target text;

[0297] For each text segment in the target text, perform word segmentation processing respectively to determine each word segment contained in each text segment;

[0298] Extract the word segmentation features of each word fused with context information contained in each text segment respectively;

[0299] The features of each text segment are determined according to the segmentation features of each segmentation contained in each text segment and the fused context information.

[0300] As an optional implementation, the step of obtaining the features of each text segment in the target text further includes:

[0301] The features of each text segment are fused and encoded to obtain the text segment features of each text segment that integrate the context information.

[0302] As an optional implementation, the features of each text segment of the reference text are obtained, including:

[0303] Perform word segmentation processing on the reference text to determine each word segmentation contained in the reference text;

[0304] The word segmentation features of each word contained in the reference text are extracted by integrating the context information.

[0305] As an optional implementation, when the number of texts in the reference text is greater than 1, before performing word segmentation processing on the reference text and determining each word segmentation included in the reference text, the method further includes:

[0306] According to the relationship between the reference texts, the reference texts are merged or filtered.

[0307] As an optional implementation, it also includes:

[0308] The features of each text segment of the target text and the features of each text segment of the reference text are subjected to feature fusion processing to obtain target text features fused with reference text features and / or reference text features fused with target text features.

[0309] As an optional implementation, the features of each text segment of the target text and the features of each text segment of the reference text are subjected to feature fusion processing to obtain reference text features fused with the target text features, including:

[0310] Performing feature fusion processing on the chapter features of the target text and / or the features of each text segment of the target text and the features of each text segment of the reference text to obtain reference text features fused with the features of the target text;

[0311] The chapter features of the target text are determined according to the features of each text segment of the target text.

[0312] As an optional implementation, the features of each text segment of the target text and the features of each text segment of the reference text are subjected to feature fusion processing to obtain target text features fused with reference text features and reference text features fused with target text features, including:

[0313] Determine the text characteristics of the target text based on the characteristics of each text segment of the target text;

[0314] The chapter features of the target text are fused with the features of each text segment of the reference text to obtain the text segment features of each text segment of the reference text fused with the chapter features of the target text;

[0315] The text segment features of each text segment of the reference text fused with the target text chapter features are fused with the features of each text segment of the target text to obtain the target text features fused with the reference text features and the reference text features fused with the target text features.

[0316] Specifically, for the specific working contents of each part in each embodiment of the above-mentioned text minutes generating device, please refer to the specific contents of the corresponding processing steps in each embodiment of the above-mentioned text minutes generating method, which will not be repeated here.

[0317] Another embodiment of the present application also provides a text minutes generation device, see Fig.10 As shown, the device includes:

[0318] Memory 200 and processor 210;

[0319] The memory 200 is connected to the processor 210 and is used to store programs;

[0320] The processor 210 is used to implement the text minutes generation method disclosed in any of the above embodiments by running the program stored in the memory 200.

[0321] Specifically, the above-mentioned text minutes generating device may further include: a bus, a communication interface 220 , an input device 230 and an output device 240 .

[0322] The processor 210, the memory 200, the communication interface 220, the input device 230 and the output device 240 are connected to each other via a bus.

[0323] A bus may include a pathway that transfers information between components of a computer system.

[0324] The processor 210 may be a general-purpose processor, such as a general-purpose central processing unit (CPU), a microprocessor, etc., or an application-specific integrated circuit (ASIC), or one or more integrated circuits for controlling the execution of the program of the scheme of the present invention. It may also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components.

[0325] The processor 210 may include a main processor, and may also include a baseband chip, a modem, and the like.

[0326] The memory 200 stores a program for executing the technical solution of the present invention, and may also store an operating system and other key services. Specifically, the program may include a program code, and the program code includes a computer operation instruction. More specifically, the memory 200 may include a read-only memory (ROM), other types of static storage devices that can store static information and instructions, a random access memory (RAM), other types of dynamic storage devices that can store information and instructions, a disk storage, a flash, and the like.

[0327] The input device 230 may include a device for receiving data and information input by a user, such as a keyboard, a mouse, a camera, a scanner, a light pen, a voice input device, a touch screen, a pedometer, or a gravity sensor.

[0328] Output device 240 may include devices that allow information to be output to a user, such as a display screen, printer, speaker, etc.

[0329] The communication interface 220 may include any transceiver or the like to communicate with other devices or communication networks, such as Ethernet, Radio Access Network (RAN), Wireless Local Area Network (WLAN), etc.

[0330] The processor 210 executes the program stored in the memory 200 and calls other devices, which can be used to implement each step of any text minutes generation method provided in the above embodiments of the present application.

[0331] Another embodiment of the present application further provides a storage medium on which a computer program is stored. When the computer program is executed by a processor, each step of any one of the text minutes generation methods provided in the above embodiments of the present application is implemented.

[0332] Specifically, the specific working contents of each part of the above-mentioned text minutes generating device, as well as the specific processing contents of the computer program on the above-mentioned storage medium when being executed by the processor, can all be found in the contents of each embodiment of the above-mentioned text minutes generating method, and will not be repeated here.

[0333] For the aforementioned method embodiments, for the sake of simplicity, they are all described as a series of action combinations, but those skilled in the art should be aware that the present application is not limited by the order of the actions described, because according to the present application, some steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the present application.

[0334] It should be noted that each embodiment in this specification is described in a progressive manner, and each embodiment focuses on the differences from other embodiments, and the same or similar parts between the embodiments can be referred to each other. For the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0335] The steps in the methods of each embodiment of the present application can be adjusted in order, combined and deleted according to actual needs, and the technical features recorded in each embodiment can be replaced or combined.

[0336] The modules and sub-modules in the devices and terminals of the various embodiments of the present application can be combined, divided and deleted according to actual needs.

[0337] In the several embodiments provided in the present application, it should be understood that the disclosed terminals, devices and methods can be implemented in other ways. For example, the terminal embodiments described above are only schematic, for example, the division of modules or submodules is only a logical function division, and there may be other division methods in actual implementation, for example, multiple submodules or modules can be combined or integrated into another module, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, devices or modules, which can be electrical, mechanical or other forms.

[0338] The modules or submodules described as separate components may or may not be physically separated, and the components of the modules or submodules may or may not be physical modules or submodules, that is, they may be located in one place, or they may be distributed on multiple network modules or submodules. Some or all of the modules or submodules may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0339] In addition, each functional module or submodule in each embodiment of the present application may be integrated into one processing module, or each module or submodule may exist physically separately, or two or more modules or submodules may be integrated into one module. The above-mentioned integrated modules or submodules may be implemented in the form of hardware or in the form of software functional modules or submodules.

[0340] Professionals may further appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in the above description according to function. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians may use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0341] The steps of the method or algorithm described in conjunction with the embodiments disclosed herein may be implemented directly by hardware, software units executed by a processor, or a combination of the two. The software units may be placed in a random access memory (RAM), a memory, a read-only memory (ROM), an electrically programmable ROM, an electrically erasable programmable ROM, a register, a hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art.

[0342] Finally, it should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the presence of other identical elements in the process, method, article or device including the elements.

[0343] 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 generating a textual summary, characterized in that: include: Acquire a target text and a reference text, wherein the reference text is determined based on the content of the target text that the user is interested in; Based on locating the associated content of the reference text from the target text, performing a minutes generation process on the target text to obtain a target text minutes corresponding to the reference text; Based on locating the associated content of the reference text from the target text, performing a minutes generation process on the target text to obtain a target text minutes corresponding to the reference text, including: By determining the relevance of each text segment in the target text to the reference text, a text segment related to the reference text is located from the target text; At least based on the relevance between each text segment in the target text related to the reference text and the reference text, the full text content of the target text is processed for minute generation to obtain a target text minute corresponding to the reference text.

2. The method according to claim 1, characterized in that At least based on the relevance between each text segment in the target text related to the reference text and the reference text, a minutes generation process is performed on the full text content of the target text to obtain a target text minutes corresponding to the reference text, including: Determining, based on the relevance of each text segment in the target text and the reference text, the contribution of each text segment in the target text to generating a text summary corresponding to the reference text; At least according to the contribution of each text segment in the target text to the generation of the text minutes corresponding to the reference text, the full text content of the target text is processed to generate the minutes to obtain the target text minutes corresponding to the reference text.

3. The method according to claim 1, characterized in that Determine the relevance of each text segment in the target text to the reference text, including: Obtain the features of the target text and the features of the reference text respectively; According to the characteristics of each text segment in the target text and the characteristics of the reference text, the relevance of each text segment in the target text and the reference text is determined respectively.

4. The method according to claim 3, characterized in that After respectively acquiring the features of the target text and the features of the reference text, the method further includes: The features of each text segment in the target text and the features of the reference text are interactively calculated based on the attention mechanism to obtain the reference text features after information improvement.

5. The method according to claim 4, characterized in that The features of each text segment in the target text and the features of the reference text are interactively operated based on the attention mechanism to obtain the reference text features after information improvement, including: Based on the features of each text segment in the target text and the features of the reference text, calculating and determining the similarity between each text segment in the target text and the reference text; According to the similarity between each text segment in the target text and the reference text, selecting a first number of text segments having the highest similarity with the reference text from the target text; An attention-based interactive operation is performed on the features of the first number of text segments selected from the target text and the features of the reference text to obtain information-completed reference text features.

6. The method according to claim 3, characterized in that When the number of texts in the reference text is greater than 1, the relevance of each text segment in the target text to the reference text is determined according to the features of each text segment in the target text and the features of the reference text, including: For each text segment in the target text, its relevance to the reference text is determined by the following processing: Determine the relevance of the text segment to each reference text based on the characteristics of the text segment and the characteristics of each reference text; According to the relationship between each reference text, the relevance of the text segment and each reference text is fused to determine the relevance of the text segment and the reference text.

7. The method according to claim 3, characterized in that The method further comprises: According to the features of each text segment in the target text and the features of the reference text, the semantic similarity between each text segment in the target text and the reference text is determined by calculating the BM25 algorithm; The relevance between each text segment in the target text and the reference text, as well as the semantic similarity between each text segment in the target text and the reference text, are fused to obtain the relevance between each text segment in the fused target text and the reference text.

8. The method according to claim 7, characterized in that The method further comprises: According to the position distribution of each text segment in the target text, the relevance between each text segment in the target text and the reference text is corrected.

9. The method according to claim 8, characterized in that According to the position distribution of each text segment in the target text, the relevance of each text segment in the target text and the reference text is corrected, including: Selecting, from each text segment in the target text, a second number of text segments having the highest relevance to the reference text; Determining the penalty degree for the relevance between other text segments in the target text and the reference text according to a rule that the greater the distance between other text segments in the target text and the selected second number of text segments, the higher the penalty degree for the relevance between other text segments and the reference text; According to the penalty degree of the relevance of other text segments in the target text to the reference text, the relevance of other text segments in the target text to the reference text is penalized.

10. The method according to claim 9, characterized in that The method further comprises: According to the relevance of each text segment in the target text to the reference text, selecting a third number of text segments with the highest relevance to the reference text from the target text; According to the relevance between each text segment in the selected third number of text segments and the reference text, a text segment having a relevance to the reference text greater than a first relevance threshold, or a relevance to the reference text greater than a second relevance threshold and a normalized relevance to the reference text greater than a third relevance threshold is selected from the third number of text segments as a text segment related to the reference text; The first correlation threshold is greater than the second correlation threshold, and the second correlation threshold is greater than the third correlation threshold.

11. The method according to claim 2, characterized in that Determining, according to the relevance of each text segment in the target text and the reference text, the contribution of each text segment in the target text to generating a text summary corresponding to the reference text, including: Determining an attention coefficient of a text summary for generating the target text for each text segment in the target text; According to the attention coefficient of the text minutes for generating the target text to each text fragment in the target text, and the relevance of each text fragment in the target text to the reference text, the contribution of each text fragment in the target text to generating the text minutes corresponding to the reference text is determined.

12. The method according to claim 2, characterized in that: At least according to the contribution of each text segment in the target text to the generation of the text minutes corresponding to the reference text, the full text content of the target text is processed to generate the minutes to obtain the target text minutes corresponding to the reference text, including: Generating text minutes decoding features based at least on the features of the target text and the contribution of each text segment in the target text to generating a text minutes corresponding to the reference text; A text summary of the target text is generated according to the text summary decoding features.

13. The method according to claim 12, characterized in that Generating text minutes decoding features based at least on the features of the target text and the contribution of each text segment in the target text to generating a text minutes corresponding to the reference text includes: The text minutes decoding features are generated based at least on the features of the target text, the features of the reference text, and the contribution of each text segment in the target text to the generation of the text minutes corresponding to the reference text.

14. The method according to claim 13, characterized in that Generating text minutes decoding features based at least on the features of the target text, the features of the reference text, and the contribution of each text segment in the target text to generating a text minutes corresponding to the reference text, including: generating a reference decoding feature based at least on the feature of the reference text; A text minutes decoding feature is generated according to the reference decoding feature, the feature of the target text, and the contribution of each text segment in the target text to the generation of a text minutes corresponding to the reference text.

15. The method according to claim 14, characterized in that The method further comprises: Determining the contribution of each text segment in the reference text to generating a text summary corresponding to the reference text; Generating reference decoding features based at least on the features of the reference text includes: A reference decoding feature is generated according to the feature of the reference text and the contribution of each text segment in the reference text to generating a text summary corresponding to the reference text.

16. The method according to claim 15, characterized in that Generate reference decoding features according to the features of the reference text and the contribution of each text segment in the reference text to generating a text summary corresponding to the reference text, including: Generate reference decoding features according to features of each text segment of the reference text and contribution of each text segment in the reference text to generating a text summary corresponding to the reference text; Generating text minutes decoding features according to the reference decoding features, the features of the target text, and the contribution of each text segment in the target text to generating a text minutes corresponding to the reference text includes: A text minutes decoding feature is generated according to the reference decoding feature, the features of each text segment of the target text, and the contribution of each text segment in the target text to generating a text minutes corresponding to the reference text.

17. The method according to claim 12, characterized in that Generating text minutes decoding features according to the features of the target text and the contribution of each text segment in the target text to generating a text minutes corresponding to the reference text includes: A text minutes decoding feature is generated according to the features of each text segment in the target text and the contribution of each text segment in the target text to the generation of a text minutes corresponding to the reference text.

18. The method according to claim 1, characterized in that Acquire a target text and a reference text, and perform a minutes generation process on the target text based on locating the associated content of the reference text from the target text to obtain a target text minutes corresponding to the reference text, including: Obtaining the characteristics of the target text and the characteristics of the reference text; The features of the target text and the features of the reference text are input into a pre-trained text minutes generation model based on the attention mechanism, so that the text minutes generation model based on the attention mechanism locates the related content of the reference text from the target text, performs minutes generation processing on the target text, and obtains the target text minutes corresponding to the reference text.

19. The method according to any one of claims 1 to 18, characterized in that The features of the target text are obtained by obtaining the features of each text segment in the target text; the features of the reference text are obtained by obtaining the features of each text segment of the reference text.

20. The method according to claim 19, characterized in that Get the features of each text segment in the target text, including: Performing text segmentation processing on the target text to determine the text segments contained in the target text; For each text segment in the target text, perform word segmentation processing respectively to determine each word segment contained in each text segment; Extract the word segmentation features of each word fused with context information contained in each text segment respectively; The features of each text segment are determined according to the segmentation features of each segmentation contained in each text segment and the fused context information.

21. The method according to claim 20, characterized in that Also includes: The features of each text segment are fused and encoded to obtain the text segment features of each text segment that integrate the context information.

22. The method according to claim 19, characterized in that Get the features of each text segment of the reference text, including: Perform word segmentation processing on the reference text to determine each word segmentation contained in the reference text; The word segmentation features of each word contained in the reference text are extracted by integrating the context information.

23. The method according to claim 22, characterized in that When the number of texts in the reference text is greater than 1, before the reference text is segmented and each segmentation word included in the reference text is determined, the following is also included: According to the relationship between the reference texts, the reference texts are merged or filtered.

24. The method according to claim 19, characterized in that The method further comprises: The features of each text segment of the target text and the features of each text segment of the reference text are subjected to feature fusion processing to obtain target text features fused with reference text features and / or reference text features fused with target text features.

25. The method according to claim 24, characterized in that The features of each text segment of the target text and the features of each text segment of the reference text are subjected to feature fusion processing to obtain reference text features fused with the target text features, including: Performing feature fusion processing on the chapter features of the target text and / or the features of each text segment of the target text and the features of each text segment of the reference text to obtain reference text features fused with the features of the target text; The chapter features of the target text are determined according to the features of each text segment of the target text.

26. The method according to claim 24, characterized in that The features of each text segment of the target text and the features of each text segment of the reference text are subjected to feature fusion processing to obtain target text features fused with reference text features and reference text features fused with target text features, including: Determine the text characteristics of the target text based on the characteristics of each text segment of the target text; The chapter features of the target text are fused with the features of each text segment of the reference text to obtain the text segment features of each text segment of the reference text fused with the chapter features of the target text; The text segment features of each text segment of the reference text fused with the target text chapter features are fused with the features of each text segment of the target text to obtain the target text features fused with the reference text features and the reference text features fused with the target text features.

27. A text minutes generating device, characterized in that: include: A data acquisition unit, used to acquire a target text and a reference text, wherein the reference text is determined based on the content of the target text that the user is concerned about; A minutes generating unit, configured to perform minutes generating processing on the target text based on locating the associated content of the reference text from the target text, so as to obtain a target text minutes corresponding to the reference text; Based on locating the associated content of the reference text from the target text, performing a minutes generation process on the target text to obtain a target text minutes corresponding to the reference text, including: By determining the relevance of each text segment in the target text to the reference text, a text segment related to the reference text is located from the target text; At least based on the relevance between each text segment in the target text related to the reference text and the reference text, the full text content of the target text is processed for minute generation to obtain a target text minute corresponding to the reference text.

28. A text minutes generating device, characterized in that: include: Memory and processor; The memory is connected to the processor and is used to store programs; The processor is used to implement the text minutes generation method as described in any one of claims 1 to 26 by running the program in the memory.

29. A storage medium, characterized in that: The storage medium stores a computer program, and when the computer program is executed by the processor, the text minutes generation method according to any one of claims 1 to 26 is implemented.

Images