Abstract generation method and apparatus

Abstract

The application provides a summary generation method and device, comprising: determining keywords in each dialogue sentence in dialogue text; determining the relevance between different keywords and the appearance order between keywords in the dialogue sentence; determining the influence degree of the keywords on the dialogue text based on the appearance order; determining the importance degree of the keywords by combining the relevance between different keywords and the influence degree of the keywords on the dialogue text; and generating a summary of the dialogue text based on the importance degree of the keywords, wherein the summary comprises dialogue sentences in which the keywords with the importance degree meeting a condition are located. The scheme of the application can more efficiently and accurately generate a summary of dialogue text.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of text processing, and particularly relates to a summary generation method and device. BACKGROUND

[0002] The dialogue text refers to a text composed of dialogue sentences of chatting or communication between at least two dialogue parties.

[0003] In order to more conveniently and efficiently understand the content of the dialogue text, it is often necessary to generate a summary of the dialogue text. For example, in a customer service scenario, a user and a customer service (such as an artificial customer service or a robot customer service, etc.) can realize information interaction such as problem consultation through dialogue, and by extracting a summary of the dialogue text of the dialogue between the user and the customer service, the interactive information between the user and the customer service can be condensed, which is beneficial to the relevant personnel in the customer service to obtain the summarized content of the problems or solutions that have been discussed.

[0004] In order to more accurately reflect the content of the dialogue text, it is necessary to reasonably extract the summary of the dialogue text. Therefore, how to more reasonably extract the summary from the dialogue text so that the extracted summary can more accurately reflect the content of the dialogue text is an urgent technical problem to be solved by those skilled in the art. SUMMARY

[0005] The present application provides a summary generation method and device.

[0006] The summary generation method comprises the following steps.

[0007] Determine the keywords in each dialogue sentence in the dialogue text;

[0008] Determine the relevance between different keywords and the appearance order of the keywords in the dialogue sentence;

[0009] Determine the influence degree of the keywords on the dialogue text based on the appearance order;

[0010] Determine the importance degree of the keywords in combination with the relevance between different keywords and the influence degree of the keywords on the dialogue text;

[0011] Generate a summary of the dialogue text based on the importance degree of the keywords, wherein the summary comprises dialogue sentences in which the keywords with the importance degree meeting a condition are located.

[0012] In a possible implementation manner, before the summary of the dialogue text is generated, the method further comprises the following steps.

[0013] The dialogue text is divided into at least one dialogue partition, and dialogue sentences in different dialogue partitions represent different categories of dialogue intentions.

[0014] The importance of the keyword is determined based on the importance of the keyword.

[0015] Based on the importance of the keyword, the target dialogue sentence in the dialogue partition for composing the abstract is determined, and the abstract composed of the target dialogue sentence in each dialogue partition is obtained.

[0016] In one possible implementation, determining the appearance order between keywords in a dialogue sentence includes:

[0017] For each keyword, determine the co-occurring keywords of the keyword and the co-occurring dialogue sentence in which the keyword and its co-occurring keywords first appear together, and the co-occurring keywords of the keyword are other keywords that appear in a dialogue sentence at the same time as the keyword.

[0018] Determine the appearance order between the keyword and its co-occurring keywords in the co-occurring dialogue sentence.

[0019] An abstract generation device includes:

[0020] A keyword determination unit is configured to determine keywords in dialogue sentences in dialogue text.

[0021] An association determination unit is configured to determine the relevance between different keywords and the appearance order between keywords in the dialogue sentence.

[0022] An influence determination unit is configured to determine the influence of the keyword on the dialogue text based on the appearance order.

[0023] An importance determination unit is configured to determine the importance of the keyword in combination with the relevance between different keywords and the influence of the keyword on the dialogue text.

[0024] An abstract generation unit is configured to generate an abstract of the dialogue text based on the importance of the keyword, and the abstract includes dialogue sentences in which the keywords in the dialogue text meet the importance condition.

[0025] From the above, in the present application, after determining the keywords of each dialogue sentence in the dialogue text, not only the correlation between different keywords will be determined, but also the influence degree of the keywords on the dialogue text will be determined in combination with the appearance order between the keywords in the dialogue sentence. On this basis, the present application will comprehensively determine the importance of the keywords in combination with the correlation between the keywords and the influence degree of the keywords on the dialogue text, so that the importance of the keywords can more reasonably and accurately reflect the importance of the keywords in the dialogue text. Therefore, the dialogue sentence used for generating the summary can be more reasonably determined from the dialogue text in combination with the importance of the keywords, so that the summary can more accurately reflect the key information in the dialogue text. BRIEF DESCRIPTION OF DRAWINGS

[0026] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed to be used in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only embodiments of the present application, and other drawings can be obtained by the drawings provided by the person skilled in the art without creative labor.

[0027] Figure 1 A flowchart of the summary generation method provided by the embodiments of the present application is shown;

[0028] Figure 2 Another flowchart of the summary generation method provided by the embodiments of the present application is shown;

[0029] Figure 3 A schematic diagram of the directed graph provided by the embodiments of the present application is shown;

[0030] Figure 4 A flowchart of determining the correlation between the keywords in the embodiments of the present application is shown;

[0031] Figure 5 A flowchart of determining the importance of the keywords in the embodiments of the present application is shown;

[0032] Figure 6 Another flowchart of the summary generation method provided by the embodiments of the present application is shown;

[0033] Figure 7 A schematic diagram of the dialogue partition divided by the dialogue text in the embodiments of the present application is shown;

[0034] Figure 8 Another flowchart of the summary generation method provided by the embodiments of the present application is shown;

[0035] Figure 9 Another schematic diagram of the directed graph provided by the embodiments of the present application is shown;

[0036] Figure 10 This illustration shows a schematic diagram of the composition structure of an abstract generation apparatus provided in an embodiment of this application;

[0037] Figure 11 A schematic diagram of the component architecture of an electronic device provided in an embodiment of this application is shown. Detailed Implementation

[0038] The solution in this application is applicable to generating summaries for any type of dialogue text, so that the generated summaries can more reasonably and accurately reflect the key information of the dialogue text.

[0039] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0040] like Figure 1 The diagram illustrates a flowchart of an abstract generation method provided in this embodiment. This method can be applied to any type of electronic device, or a cluster or distributed system composed of multiple electronic devices. For example, the electronic devices can be laptops, desktop computers, or servers, etc., without limitation.

[0041] The method in this embodiment may include:

[0042] S101, Identify the keywords in each dialogue statement in the dialogue text.

[0043] The dialogue text includes at least one dialogue statement. Of course, in general, dialogue texts that require summarization will include multiple dialogue statements. For example, dialogue statements in the dialogue text can be text statements used by different parties for information consultation, answering questions, or communication.

[0044] In this application, the dialogue text from which the summary needs to be extracted can be dialogue text from any scenario. For example, the dialogue text can be the dialogue text between a user and a customer service representative in customer service, such as the text of multiple dialogue statements between a buyer and a seller's customer service representative. Alternatively, the dialogue text can be the dialogue text of an organization or enterprise involving work-related communication.

[0045] Of course, the dialogue text may have other possibilities, and this application does not impose any restrictions on the source or acquisition method of the dialogue text.

[0046] The keywords in the dialogue sentence can be determined in various ways, which are not limited.

[0047] The keywords in the dialogue sentence can be determined in various ways, which are not limited.

[0048] For example, a relatively simple way can be to segment the dialogue sentence, and each segmented word in the dialogue sentence is determined as a keyword of the dialogue sentence.

[0049] For example, after segmenting the dialogue sentence, the keywords of the dialogue sentence can be determined by removing some stop words from each segmented word.

[0050] Of course, according to the specific scene involved in the dialogue text, etc., the words with specific semantics in the dialogue sentence can be determined as keywords, and the like, which will not be described here.

[0051] S102, determine the relevance between different keywords and the appearance order of the keywords in the dialogue sentence.

[0052] The relevance between different keywords in the text dialogue can include the relevance between any two keywords in the dialogue text.

[0053] The relevance between the keywords is only to reflect the degree of correlation between the two keywords. It can be understood that there are various ways to calculate the relevance between two keywords, which are not limited by the present application.

[0054] For a dialogue sentence, the appearance order of the keywords in the dialogue sentence can include the appearance order between any two keywords in the dialogue sentence. For example, the dialogue sentence includes keyword A and keyword B, and the appearance order of keyword A and keyword B in the dialogue sentence can be determined.

[0055] It can be understood that the appearance order of two keywords in the same dialogue sentence also reflects an association relationship between the two keywords, and the appearance order of the two keywords in the same dialogue sentence can represent the influence relationship between the two keywords.

[0056] For example, keyword A appears before keyword B in a dialogue sentence, which means that keyword B is likely to be a keyword derived from keyword A, and the appearance of keyword A can trigger the appearance of keyword B.

[0057] S103, determine the influence degree of the keywords on the dialogue text based on the appearance order.

[0058] It can be understood that for a keyword, the keyword can appear in multiple dialogue sentences, and the appearance order between the keyword and other keywords in different dialogue sentences can be obtained, and the influence relationship between the keyword and other keywords can be obtained, and the influence relationship of other keywords to the keyword can be obtained.

[0059] And through one or both of the influence relationship of the keyword to other keywords and the influence relationship of other keywords to the keyword, the influence degree of the keyword in the dialogue text can be reflected.

[0060] For example, considering that other keywords appearing after the keyword in the same dialogue sentence can be considered as other keywords caused by the keyword. For a keyword, the more other keywords caused by the keyword, or the more other keywords causing the keyword, the more important the keyword is in the dialogue text, and the greater the influence degree is.

[0061] Based on this, in a possible implementation, for a keyword, the first number of other keywords located after the keyword and the second number of other keywords located before the keyword are determined according to the appearance order between the keyword and other keywords in different dialogue texts. Correspondingly, the influence degree of the keyword on the dialogue text can be determined based on the first number and the second number corresponding to the keyword.

[0062] For example, the influence degree of the keyword on the dialogue text can be determined based on the sum of the first number and the second number. The greater the sum of the first number and the second number corresponding to the keyword, the greater the influence degree of the keyword on the dialogue text.

[0063] For another example, the influence degree of the keyword on the dialogue text can be determined based on the larger one of the first number and the second number. The greater the value of the larger one of the first number and the second number corresponding to the keyword, the greater the influence degree of the keyword on the dialogue text. For example, the value of the larger one of the first number and the second number can be taken as an influence value representing the influence degree of the keyword on the dialogue text.

[0064] It can be understood that the influence degree of the keyword in the present application can be represented in the form of score or influence level, and the specific mode is not limited.

[0065] S104, the importance of the keyword is determined according to the relevance between different keywords and the influence degree of the keyword on the dialogue text.

[0066] It can be understood that the relevance between the keyword and each of the other keywords in the dialogue text can reflect the importance of the keyword in the dialogue text, and the influence degree of the keyword on the dialogue text obtained according to the appearance order between the keyword and the other keywords can also reflect the importance of the keyword from another dimension. Based on this, the application combines the relevance between the keywords and the influence degree of the keyword on the dialogue text to determine the importance of the keyword, so that the importance of each keyword in the dialogue text can be more accurately and reasonably reflected.

[0067] The manner of determining the importance of the keyword based on the relevance and the influence degree of the keyword on the dialogue text can be various, and the application does not limit this.

[0068] For example, the first importance value of the keyword can be determined based on the relevance between the keyword and the other keywords, and the second importance value can be determined based on the influence degree of the keyword on the dialogue text. Then, the importance of the keyword can be determined based on the weighted sum or product of the first importance value and the second importance value.

[0069] S105, generating the summary of the dialogue text based on the importance of the keyword.

[0070] The summary includes the dialogue sentence in which the keyword with the importance meeting the condition is located.

[0071] For example, in a possible implementation manner, after determining the importance of each keyword in the dialogue text, for each dialogue sentence in the dialogue text, the importance of the dialogue sentence can be determined based on the importance of the keyword in the dialogue sentence. Accordingly, the first at least dialogue sentence with high importance in the dialogue text can be selected as the summary of the dialogue text.

[0072] The manner of determining the importance of the dialogue sentence based on the importance of the keyword in the dialogue sentence can also be various, for example, the importance of the dialogue sentence can be determined by comprehensively considering the importance of all the keywords in the dialogue sentence.

[0073] In an implementation manner, in order to enable the keyword with high importance to appear in the summary, the application can also determine the importance of the dialogue sentence based on the importance of the keyword with the highest importance in the dialogue sentence. For example, assuming that the importance of the keyword is represented by an importance score, the importance score corresponding to the keyword with the highest importance score in the dialogue sentence can be determined as the score of the dialogue sentence, so that the dialogue sentence with high score can be selected to generate the summary of the dialogue text.

[0074] From the above, in the present application, after determining the keywords of each dialogue sentence in the dialogue text, not only the relevance between different keywords will be determined, but also the influence degree of the keywords on the dialogue text will be determined in combination with the appearance order between the keywords in the dialogue sentence. On this basis, the present application will comprehensively determine the importance of the keywords in combination with the relevance between the keywords and the influence degree of the keywords on the dialogue text, so that the importance of the keywords can more reasonably and accurately reflect the importance of the keywords in the dialogue text. Therefore, the dialogue sentence used for generating the summary can be more reasonably determined from the dialogue text in combination with the importance of the keywords, so that the summary can more accurately reflect the key information in the dialogue text.

[0075] It can be understood that for any two keywords, the two keywords can appear in multiple dialogue sentences at the same time, and the appearance order between the two keywords in different dialogue sentences can be different. In order to avoid the appearance order of two keywords in different dialogue sentences being inconsistent, which affects the determination of the influence degree of the keywords on the dialogue text, the present application can only focus on the appearance order of the two keywords in the dialogue sentence in which the two keywords first appear at the same time.

[0076] It can be understood that in actual application, for any two keywords, the appearance order of the two keywords in the dialogue sentence in which the two keywords first appear at the same time can more represent the importance between the two keywords.

[0077] In order to facilitate understanding, the scheme of the embodiment of the present application will be described below in combination with a possible implementation manner of determining the appearance order. As shown in Figure 2 It shows another flowchart of the summary generation method provided by the embodiment of the present application. The method of the present embodiment can include:

[0078] S201, determining the keywords in each dialogue sentence in the dialogue text.

[0079] S202, determining the relevance between different keywords.

[0080] The above steps S201 and S202 can refer to the related introduction of the previous embodiment, which will not be described here.

[0081] S203, for each keyword, determining the co-occurring keywords of the keyword and the co-occurring dialogue sentence in which the keyword and its co-occurring keywords first appear together.

[0082] Among them, the co-occurring keywords of the keyword are other keywords that appear in a dialogue sentence at the same time as the keyword. That is, for each keyword appearing in the same dialogue sentence, any two keywords are co-occurring keywords of each other.​​​​​​​​

[0083] As the dialogue sentence includes the keyword A and the keyword B, the keyword B is the co-occurrence keyword of the keyword A, and the keyword B is also the co-occurrence keyword of the keyword A.

[0084] It can be understood that for any two keywords which are co-occurrence keywords of each other, the dialogue sentence in which the two keywords appear simultaneously is the dialogue sentence which contains the two keywords simultaneously. In order to distinguish, the dialogue sentence in which the two keywords appear simultaneously is referred to as a co-occurrence dialogue sentence. In actual application, the two keywords may appear simultaneously in multiple co-occurrence dialogue sentences. However, in order to more accurately obtain the influence relationship between the two keywords, the embodiment can only focus on the co-occurrence dialogue sentence in which the two keywords appear simultaneously for the first time.

[0085] S204, for any co-occurrence keyword corresponding to a keyword pair, determining the appearance order between the keyword and the co-occurrence keyword in the co-occurrence dialogue sentence corresponding to the keyword and the co-occurrence keyword.

[0086] It should be noted that the co-occurrence dialogue sentence mentioned here also refers to the co-occurrence dialogue sentence in which the keyword and its co-occurrence keyword appear simultaneously for the first time.

[0087] It can be understood that since the two keywords are co-occurrence keywords of each other, for any two keywords, only the appearance order between the two keywords in the dialogue sentence in which the two keywords appear simultaneously for the first time needs to be determined. Based on this, the step S204 can also be regarded as determining the appearance order of any two keywords which are co-occurrence keywords of each other in the co-occurrence dialogue sentence in which the two keywords appear for the first time.

[0088] S205, for each keyword, based on the appearance order between the keyword and the co-occurrence keyword of the keyword in each co-occurrence dialogue sentence corresponding to the keyword, determining the influence degree of the keyword on the dialogue text.

[0089] As in a possible implementation, for each keyword, based on the appearance order between the keyword and the co-occurrence keyword of the keyword in each co-occurrence dialogue sentence corresponding to the keyword, the first number of other keywords located after the keyword and the second number of other keywords located before the keyword are determined. On this basis, the influence degree of the keyword on the dialogue text can be determined based on the first number and the second number corresponding to the keyword.

[0090] In another possible implementation, in order to determine the order of occurrence of keywords and their co-occurring keywords more quickly and efficiently, and to efficiently count the number of co-occurring keywords before and after each keyword, this application can also combine a directed graph to analyze the influence relationship between keywords and other keywords, and determine the degree of influence of keywords on the dialogue text.

[0091] Specifically, after identifying the co-occurring keywords of a keyword and the co-occurring dialogue statements in which the keyword and its co-occurring keywords first appear together, a directed graph representing the relationship between different keywords within the dialogue text can be constructed.

[0092] The directed graph includes multiple nodes and directed edges between them. Each node represents a keyword, and a directed edge between two nodes indicates that the two keywords corresponding to the two nodes are co-occurring keywords. The direction of the directed edge between the two nodes indicates the order in which the two keywords appear in their co-occurring dialogue statements.

[0093] like Figure 3 As shown, it illustrates a schematic diagram of a directed graph provided in an embodiment of this application. Figure 3 Each circle in the diagram represents a node, such as... Figure 3 It can include nodes A, B, C, D, E, F, and G.

[0094] Different nodes represent different keywords in the dialogue text. For example, node A can represent keyword 1, node B can represent keyword 2, and so on.

[0095] Depend on Figure 3 It can be seen that some nodes are connected by directed edges, while others are not. Specifically, if two nodes represent two keywords that co-occur, then there is a directed edge between them. The direction of the arrow on the directed edge indicates the order in which the two keywords appear in the dialogue statement where they first appear together, with the direction of the directed edge pointing from the keyword that appears earlier to the keyword that appears later.

[0096] For example, keywords 1 and 2 corresponding to nodes A and B appear simultaneously for the first time in dialogue statement S, and in dialogue statement S, keyword 1 appears first and keyword 2 appears later. Therefore, in Figure 3 In this process, node A, representing keyword 1, needs to point to node B, representing keyword 2, so that the keyword represented by node A can lead to the keyword represented by node B.

[0097] Combination Figure 3The directed graph can clearly show whether the keywords represented by each node are co-occurring keywords and the order of the two co-occurring keywords based on the directed graph.

[0098] Accordingly, based on the constructed directed graph, the application can determine the influence degree of the keywords represented by the nodes on the dialogue text according to the number and direction of the directed edges between different nodes in the directed graph.

[0099] For example, the number of other nodes pointed to by each node can be counted based on the directed graph, which is essentially the first number of other keywords appearing after the keyword represented by the node. Accordingly, the number of other nodes pointing to the node can also be counted, which is the second number of other keywords appearing before the keyword represented by the node.

[0100] It can be understood that the first number and the second number can be more conveniently counted by using existing link relationship analysis functions in combination with the directed graph.

[0101] For example, in one possible implementation, the Hyperlink-Induced Topic Search (HITS) algorithm can be used to count the first number and the second number corresponding to each node.

[0102] The HITS algorithm can be used to calculate two values of a page, one is Hub Scores, and the other is Authority Scores, which are interdependent and influence each other. The Hub Scores refer to the sum of the authority values of all outgoing links on the page. The Authority Scores refer to the sum of hubs in all pages where the incoming links are located.

[0103] In the present application, the HITS algorithm can be used to calculate the Hub Scores and the Authority Scores for each node in the directed graph, where the Hub Scores of a node refer to the total number of other nodes pointed to by the node, i.e., the first number corresponding to the keyword represented by the node; and the Authority Scores of a node refer to the total number of nodes pointing to the node, i.e., the second number corresponding to the keyword represented by the node.

[0104] It can be understood that the above is an example of using HITS to analyze the directed graph to determine the first number and the second number corresponding to the keywords represented by each node, and other algorithms can also be used to analyze the directed graph to determine the first number and the second number of each keyword, which is not limited.

[0105] S206, determine the importance of the keywords based on the correlation between different keywords and the influence degree of the keywords on the dialogue text.

[0106] S207, generating an abstract of the dialogue text based on the importance of the keywords.

[0107] The abstract includes dialogue sentences in which the keywords with importance meeting the condition are located in the dialogue text.

[0108] The above steps S206 and S207 can refer to the related description of the foregoing embodiments, and will not be described here again.

[0109] In the foregoing embodiments of the present application, the manner of determining the correlation between keywords can have multiple possibilities, which are not limited by the present application.

[0110] In a possible implementation manner, in order to more accurately and reliably reflect the correlation between keywords, in addition to the similarity between the vectors of the keywords, the present application can also combine the word frequency of the keywords in the dialogue text and the number of times of simultaneous appearance of different keywords in the same dialogue sentence to determine the correlation between the keywords.

[0111] As shown in FIG. 7, which shows a flow diagram of determining the correlation between keywords in an embodiment of the present application, the present embodiment can include: Figure 4

[0112] S401, for each keyword, determining the word vector of the keyword based on the dialogue sentence in which the keyword first appears in the dialogue text.

[0113] It can be understood that a keyword can exist in multiple dialogue sentences, but the dialogue sentence in which the keyword first appears has the greatest influence on the keyword and best represents the semantic features of the keyword. Based on this, the present application determines the word vector of the keyword based on the dialogue sentence in which the keyword first appears in the dialogue text, so that the word vector of the keyword can more accurately reflect the semantic features of the keyword.

[0114] It can be understood that the manner of determining the word vector of the keyword based on the dialogue sentence in which the keyword first appears can have multiple possibilities, which are not limited by the present application.

[0115] In a possible implementation manner, in order to enable the word vector of the keyword to reflect the features of the context in which the keyword is located, and further enable the subsequently calculated similarity between keywords to reflect the similarity between the contexts in which the keywords are located, the present application can utilize a Bidirectional Encoder Representations from Transformers (BERT) model and the dialogue sentence in which the keyword first appears to determine the word vector of the keyword. ​

[0116] As mentioned above, the keyword first appears in the dialogue sentence input to the BERT model, and the encoding vector of the last layer of the BERT model is obtained, which is the word vector of the keyword.

[0117] S402, based on the word vector of the keyword, the vector similarity between different keywords is determined.

[0118] Wherein, the vector similarity is the similarity between keywords. In this application, the similarity between any two keywords is mainly concerned, so this step S402 needs to calculate the vector similarity between each two keywords.

[0119] The way to determine the similarity between keywords based on the word vector of the keyword can also have many possibilities, which is not limited.

[0120] As in one possible implementation, the cosine similarity can be used to calculate the similarity between the word vectors of the keywords. As for any two keywords, such as keyword t i and keyword t j , the vector similarity Sim(t i , t j ) between the two keywords can be calculated by the following formula one:

[0121]

[0122] Wherein, Emb(t i ) is the word vector of keyword t i , Emb(t j ) is the word vector of keyword t j .

[0123] The keyword t i and the keyword t j represent any two keywords in the dialogue text, wherein the keyword t i is the i-th keyword in the dialogue text, and the keyword t j is the j-th keyword in the dialogue text, i and j are natural numbers greater than or equal to 1 and less than or equal to N, N is the total number of keywords in the dialogue text.

[0124] S403, combining the vector similarity between different keywords, the word frequency of the keyword in the dialogue text and the number of times different keywords appear in a dialogue sentence at the same time, the correlation between different keywords is determined.

[0125] Wherein, the word frequency of the keyword in the dialogue text refers to the number of times the keyword appears in the dialogue text.

[0126] The number of times two different keywords appear together in a single dialogue statement refers to the total number of times any two keywords appear simultaneously in the same dialogue statement. For example, if keyword A and keyword B appear simultaneously in dialogue statements 1, 3, and 7, then the total number of times these two keywords appear simultaneously in the same dialogue statement is 3.

[0127] In determining the relevance between keywords, this embodiment comprehensively considers the vector similarity between different keywords, the word frequency of keywords in the dialogue text, and the number of times different keywords appear in the same dialogue statement. This allows for a more comprehensive analysis of the relevance between keywords, making the obtained relevance more accurate and comprehensive in reflecting the association between keywords.

[0128] The specific method for calculating the correlation can be set as needed, and there are no restrictions on it.

[0129] In one possible implementation, this application can calculate the keyword t using the following formula: i With keyword t j The correlation m between them ij :

[0130]

[0131] Among them, TF(t) i ) is the keyword t i Word frequency in dialogue text, TF(t) j ) is the keyword t j Word frequency in dialogue text; TS ij For the keyword t i With keyword t j The number of times they appear in the same dialogue statement, TS ji For keywords TS ji With keyword t i The number of times they appear in the same dialogue statement. This is understandable in TypeScript. ij With TS ji The values ​​are the same.

[0132] In the embodiments described above, there are various ways to determine the importance of keywords by combining the relevance between different keywords and the degree of influence of the keywords on the dialogue text. For ease of understanding, one implementation method for determining the importance of keywords is illustrated below.

[0133] like Figure 5 The diagram illustrates a flowchart of an implementation process for determining the importance of keywords according to an embodiment of this application. The method of this embodiment may include:

[0134] S501, for each keyword, determine a first occurrence dialogue sentence in which the keyword first occurs in the dialogue text and a sentence source side of the first occurrence dialogue sentence.

[0135] In this embodiment, it is also considered that the dialogue sentence in which the keyword first occurs has a greater impact on determining the importance of the keyword. Therefore, the application needs to combine the sentence source side to which the dialogue sentence in which the keyword first occurs belongs to determine the importance of the keyword.

[0136] In order to facilitate distinction, the dialogue sentence in which the keyword first occurs is referred to as the first occurrence dialogue sentence of the keyword in this application.

[0137] The sentence source side of the first occurrence dialogue sentence of the keyword can represent which dialogue side in the multiple dialogue sides involved in the dialogue text the first occurrence dialogue sentence belongs to.

[0138] For example, in the customer service scenario, the dialogue text is mainly the dialogue between the user and the customer service, so the dialogue sides involved in the dialogue text include the user and the customer service, and therefore the sentence source side can be one of the user and the customer service.

[0139] For another example, taking the business communication scenario as an example, the dialogue text can involve multiple business sides, and the sentence source side can belong to a certain side among the multiple business sides.

[0140] S502, combine the sentence source side of the first occurrence dialogue sentence corresponding to the keyword and the correlation between different keywords to determine the reference importance of the keyword.

[0141] It can be understood that the dialogue sentences of different sentence source sides involved in the dialogue text have different characteristics, and when the sentence source side to which the first occurrence dialogue sentence corresponding to the keyword belongs is different, the importance of the keyword reflected based on the correlation between the keyword and other keywords will also be different.

[0142] For example, still taking the customer service scenario as an example, the longer the dialogue sentence issued by the user side, the lower the degree of suitability of the dialogue sentence for generating an abstract. Correspondingly, if the first occurrence dialogue sentence of the keyword is from the user side, the longer the dialogue sentence to which the keyword belongs, the higher the correlation between the keyword and other keywords can be, but the keyword is not an important word in the dialogue text, and the importance of the keyword will be relatively low.

[0143] Conversely, for the dialogue sentence appearing on the customer side, considering that the dialogue sentence issued by the customer side is generally more refined and has less irrelevant information irrelevant to the essential content of the dialogue text, the longer the dialogue sentence from the customer side is, the more important the keyword first appearing in the dialogue sentence can reflect the dialogue information of the dialogue text, and the higher the relevance of the keyword to other keywords is, which can indicate that the keyword is more important.

[0144] In a possible implementation, for the case that the sentence source party includes one of the customer and the user, for any keyword, the relevance sum of the relevance between the keyword and other keywords can be determined; on this basis, the reference importance of the keyword is determined in combination with the sentence source party of the first appearing dialogue sentence corresponding to the keyword and the relevance sum corresponding to the keyword.

[0145] If the first appearing dialogue sentence corresponding to the keyword is from the customer, the reference importance of the keyword is negatively correlated with the relevance sum corresponding to the keyword.

[0146] If the first appearing dialogue sentence corresponding to the keyword is from the user, the reference importance of the keyword is positively correlated with the relevance sum corresponding to the keyword.

[0147] For example, for the keyword t i , if the first appearing dialogue sentence corresponding to the keyword t i is from the customer (Agent), the reference importance SScore(t i |Agent) of the keyword t i can be calculated by the following formula three:

[0148]

[0149] For the keyword t i , if the first appearing dialogue sentence corresponding to the keyword t i is from the user (User), the reference importance SScore(t i |User) of the keyword t i can be calculated by the following formula four:

[0150]

[0151] Further, in order to compare the reference importance of different keywords, the reference importance of different keywords can be normalized, and thus the reference importance of the keyword can be the normalized reference importance.

[0152] For example, assuming that the reference importance of the keyword t i is SScore(t i |Agent), the reference importance of the keyword t i can be calculated by the following formula five:For example, SScore(t i |Agent) or SScore(t i |User) can be normalized by the following Formula Five to obtain the normalized reference importance softmax(t i ):

[0153]

[0154] wherein t l is the lth keyword in the dialogue text, l is a natural number greater than or equal to 1 and less than or equal to N, N is the total number of keywords in the dialogue text.

[0155] S503, in combination with the reference importance of the keyword and the influence degree of the keyword on the dialogue text, determines the importance of the keyword.

[0156] Wherein, the higher the reference importance and the higher the influence degree, the higher the importance of the keyword.

[0157] For example, in a possible implementation, the reference importance of the keyword can be corrected by the influence degree of the keyword on the dialogue text to obtain the importance of the keyword.

[0158] For example, assuming that the influence degree of the keyword on the dialogue text is represented by an influence score, the greater the influence score, the higher the influence degree. And the reference importance is also a reference importance score, then the product of the reference importance score and the influence score can be determined as an importance score representing the importance of the keyword.

[0159] It can be understood that the above is an example of a possible case. In actual application, the sum of the reference importance score corresponding to the reference importance of the keyword and the influence score corresponding to the influence degree of the keyword can be obtained, and the result of the sum can be used as an importance score representing the importance of the keyword. Of course, there can be other possible ways to determine the importance of the keyword, which are not limited.

[0160] In Figure 5 the embodiment, in the determination of the importance of the keyword, the influence of the first occurrence dialogue sentence of the keyword is considered to be greater, and the source of the first occurrence dialogue sentence and the relevance of the keyword to other keywords are considered to determine the basic importance of the keyword, so that the relevance between the keyword and other keywords and the importance of the keyword can be determined more comprehensively and accurately.

[0161] It can be understood that, generally, when different dialogue parties interact with dialogue sentences, they usually involve problem consultation and answers and other parts, and the intents expressed in different parts are different. Based on this, the dialogue text can also be divided into at least one dialogue partition, and the dialogue sentences in different dialogue partitions represent different categories of dialogue intents.

[0162] Considering that the dialogue intents of the dialogue sentences in different dialogue partitions are different, therefore, the importance of the dialogue sentences in different dialogue partitions for representing the key information of the dialogue text is also different. Based on this, in order to more reasonably generate the summary of the dialogue text, so that the generated summary can more comprehensively and accurately express the main idea of the dialogue text, the application can also determine the dialogue sentences used to compose the summary from each dialogue partition.

[0163] Specifically, the target dialogue sentences in the dialogue partition for composing the summary can be determined based on the importance of the keywords, and the summary composed of the target dialogue sentences in each dialogue partition is obtained.

[0164] In one possible implementation, considering that the importance of the dialogue sentences in different dialogue partitions is different, the number of target dialogue sentences determined by the application in different dialogue partitions for generating the summary can also be different.

[0165] The scheme of the application will be introduced below in combination with this possible implementation. As shown in Figure 6 The method provided by the embodiment of the application can include:

[0166] S601, determining the keywords in each dialogue sentence in the dialogue text.

[0167] S602, determining the relevance between different keywords and the appearance order of the keywords in the dialogue sentence.

[0168] S603, determining the influence degree of the keywords on the dialogue text based on the appearance order.

[0169] S604, determining the importance of the keywords in combination with the relevance between different keywords and the influence degree of the keywords on the dialogue text.

[0170] The above steps S601 to S604 can refer to the related introduction of the previous embodiments, which will not be described here.

[0171] S605, dividing the dialogue text into at least one dialogue partition.

[0172] Among them, the dialogue sentences in different dialogue partitions represent different categories of dialogue intents.

[0173] As can be combined with the semantics of each dialogue sentence in the dialogue text, the dialogue intent represented by the semantics of the dialogue sentence is determined, and the dialogue sentences adjacent to each other and with the same dialogue intent are divided into at least one dialogue partition.

[0174] It can be understood that the dialogue sentences in the consultation and chat type dialogue text are carried out according to the specific dialogue logic, and the dialogue intents represented by each logical component involved in the dialogue logic of the dialogue text are different, so the dialogue text can be divided into at least one logical partition according to the dialogue logic, and each logical partition is a dialogue partition.

[0175] As, the dialogue of multiple dialogue parties is generally started from greeting, then asking questions, followed by question discussion or detail confirmation, and finally problem solving, so the dialogue text can be divided into four dialogue partitions, the first dialogue partition is for greeting, which includes dialogue sentences related to greeting; The intent of the second dialogue partition is the title of the question, which includes one or more dialogue sentences related to asking questions and determining the existing problem; The intent of the third dialogue partition can be question detail communication, which can include at least one dialogue sentence related to determining the specific detail content of the question and discussing the possible direction of the question; The intent of the fourth dialogue partition can be to determine the cause of the problem, which can include relevant dialogue sentences related to the cause of the problem and the determined solution.

[0176] Of course, considering that the greeting in the dialogue text and the goodbye sentence at the end of the dialogue text for ending the dialogue are not very meaningful for determining the theme meaning expressed by the dialogue text, the greeting and goodbye related parts in the dialogue text can be removed.

[0177] For example:

[0178] Figure 7 The schematic diagram of the multiple dialogue partitions divided for the dialogue text.

[0179] In Figure 7 The sentence is a dialogue sentence, and the length of the frame where the sentence is located represents the length of the dialogue sentence.

[0180] In Figure 7 The dialogue text shown in the beginning is at least one dialogue sentence 701 related to greeting, which can be removed, such as Figure 7 In the above, the cross indicates deletion. In addition, the last dialogue sentence in the dialogue text is an end dialogue sentence 702 indicating goodbye and the like to end the dialogue, which can also be deleted.

[0181] On this basis, Figure 7The dialogue text is divided into three parts, or three dialogue partitions. These three dialogue partitions are the first dialogue partition 703 at the top, the second dialogue partition 704 in the middle, and the third dialogue partition 705 at the bottom.

[0182] Understandably, since different parties in a dialogue follow a logical progression, statements with the same intent are generally adjacent and consecutive. For example, from Figure 7 It is evident that at least one dialogue statement within each dialogue partition is a consecutive sequence of at least one dialogue statement.

[0183] For example, in Figure 7 In the middle, the first dialogue section 703 can be the part that explains the problem. This part includes: the first to sixth dialogue statements after the greeting, such as... Figure 7 Statements 1 through 6 are shown in the diagram. Accordingly, the main interaction between statements 1 and 6 is to identify the issues that need to be discussed or consulted.

[0184] exist Figure 7 The second dialogue section 704 can be used for communication of problem details, and this section can include statements 7 to 8.

[0185] The third dialog section 705 is the problem-solving section, which may include statement 9 and statement 10.

[0186] Of course, the above is just one possible implementation of dividing the dialogue text into dialogue partitions. In practical applications, the number of dialogue partitions that the dialogue text can be divided into and the division method can be set as needed without any restrictions.

[0187] It is understandable that the execution order of step S605 and any of the preceding steps S601 to S604 is not limited to... Figure 6 The sequence number restriction means that in practical applications, step S605 can be executed first, followed by the steps in S601 to S604. Alternatively, step S605 can be executed simultaneously with any of the steps in S601 to S604.

[0188] S606, for each dialogue partition, based on the target number of statements to be extracted for each dialogue partition with different dialogue intents, and combined with the importance of keywords, determine the target number of target dialogue statements corresponding to that dialogue partition.

[0189] The method for extracting the target number of target dialogue statements from this dialogue partition is similar to the process of determining dialogue statements from the dialogue text mentioned earlier.

[0190] For example, for each dialogue sentence, the importance of the dialogue sentence can be determined based on the importance of each keyword in the dialogue sentence (for specific implementation manners, refer to the foregoing description), and based on this, for each dialogue partition, the first target number of dialogue sentences with high importance can be extracted in the dialogue partition in combination with the target number corresponding to the dialogue partition.

[0191] The target number required to be extracted in the dialogue partition of different dialogue intents can be different or the same, and can be set as needed.

[0192] It is considered that the dialogue sentences in the dialogue partition of different intents have different degrees of fit with the theme content required to be expressed by the dialogue text, that is, different importance, and therefore, the application can set different target numbers for the dialogue partitions of different dialogue intents. For example, it is considered that the dialogue sentences in the dialogue partition related to question raising can reflect the theme direction of the dialogue text, and therefore, the target number corresponding to the dialogue partition related to question raising (for example, the first dialogue partition in the example of Figure 7 ) can be set to be relatively large.

[0193] For example, still taking the example of Figure 7 , it is assumed that the target number of sentences that can be extracted in the dialogue partition of the question raising intent is 3, the target number corresponding to the dialogue partition representing the description of question details is 2, and the target number of the dialogue partition related to question solving can be 1. Then, the first dialogue partition of Figure 7 needs to extract 3 target dialogue sentences with high importance, and similarly, 2 target dialogue sentences are extracted from the second dialogue partition, and one target dialogue sentence is extracted from the third dialogue partition.

[0194] S607, based on the target dialogue sentences in each dialogue partition, generating an abstract of the dialogue text.

[0195] For example, the target dialogue sentences in each dialogue partition can be combined into an abstract in the order of the dialogue partitions and the order of the target dialogue sentences in each dialogue partition.

[0196] In this embodiment, the dialogue text is divided into at least one dialogue partition of different dialogue intents, and then dialogue sentences used for generating an abstract are extracted from the dialogue partition, which is beneficial to making the abstract cover sentences representing different intents, and making the abstract more comprehensively reflect the content of the dialogue text.

[0197] In addition, since the application can generate dialogue abstracts from the dialogue sentences extracted from each dialogue partition in turn, the dialogue abstracts can also reflect the dialogue logic of the dialogue text.

[0198] To facilitate understanding of the scheme in this application, the following example of generating dialogue summaries using a directed graph will be used to illustrate the scheme.

[0199] like Figure 8 As shown, it illustrates another flowchart of the abstract generation method provided in this application embodiment. The method of this embodiment may include:

[0200] S801, determine the keywords in each dialogue statement in the dialogue text, and obtain N keywords in the dialogue text.

[0201] S802, divide the dialogue text into at least one dialogue partition.

[0202] Each dialogue partition includes at least one dialogue statement, and different dialogue partitions contain different dialogue statements.

[0203] The dialogue statements within different dialogue partitions represent different categories of dialogue intentions.

[0204] The order of steps S801 and S802 above is not limited to... Figure 8 As shown.

[0205] S803: For any two keywords in the dialogue text, determine the relevance between the two keywords.

[0206] Understandably, by calculating the relevance between pairs of keywords in the dialogue text, an N*N co-occurrence matrix M can be obtained, where the elements of this co-occurrence matrix are the keywords t. i With keyword t j The correlation m between them ij In this context, different elements correspond to different keyword pairs and their relevance. The keywords include two keywords in the dialogue text.

[0207] The calculation of relevance can be found in the relevant introduction of Formula 2 above, and will not be repeated here.

[0208] S804: For any two keywords in the dialogue text, determine the first co-occurring dialogue statement in which the two keywords appear simultaneously, and obtain the order of appearance of the two keywords in that co-occurring dialogue statement.

[0209] S805. Based on the relevance between pairs of keywords in the dialogue text and the order in which any two keywords appear in their co-occurrence dialogue statements, a directed graph representing the association between different keywords in the dialogue text is constructed.

[0210] The directed graph comprises a plurality of nodes and a plurality of directed edges between the nodes, each node representing a keyword, and a directed edge between two nodes indicating that the relevance between the two keywords represented by the two nodes is not zero, and the direction of the directed edge between the two nodes indicating the appearance order of the two keywords in a co-occurrence dialogue sentence.

[0211] It can be understood that, according to the above formula two, if two keywords do not appear in the same dialogue sentence at the same time, the relevance of the two keywords is zero, and therefore, the existence of a directed edge between the nodes corresponding to the two keywords can be determined in combination with the relevance.

[0212] Meanwhile, in the embodiment, the weight value of the directed edge between two nodes in the directed graph can be the relevance between the two keywords represented by the two nodes. The weight value (or the relevance) of the directed edge can be determined in combination with the co-occurrence matrix M.

[0213] It can be understood that, in order to reflect the dialogue partition from which the dialogue sentence in which the keyword first appears comes, the directed graph constructed in the application can also indicate the dialogue partition from which the keyword represented by the node comes. As shown in Figure 9 The directed graph divides the dialogue partition to which the keyword represented by each node belongs, for example, the keyword represented by the node A and the node B comes from the first dialogue partition, and the other nodes are similar, which will not be described herein.

[0214] S806, for each node in the directed graph, the HITS algorithm is used to determine the authority value and the hub value of the node, and the greater one of the authority value and the hub value of the node is determined as the influence degree value of the keyword represented by the node on the dialogue text.

[0215] S807, for each node in the directed graph, the first appearance dialogue sentence of the keyword represented by the node in the dialogue text and the sentence source side of the first appearance dialogue sentence are determined.

[0216] S808, in combination with the sentence source side of the first appearance dialogue sentence and the relevance on the directed edges between the node and other nodes in the directed graph, the reference importance degree of the keyword represented by the node is determined.

[0217] The process of determining the reference importance degree of the keyword can be referred to the related description of the previous embodiment, which will not be described herein.

[0218] It should be noted that the embodiment is to determine the reference importance degree of the keyword in combination with the directed graph, but it can be understood that the directed graph is only more clearly represented the relevance between the keywords and other keywords, but without the directed graph, the reference importance degree of the keyword is also applicable to the embodiment by directly using the relevance between the keywords.

[0219] S809, in combination with the reference importance degree of the keyword represented by the node and the influence degree value of the keyword represented by the node on the dialogue text, determine the importance score of the keyword represented by the node.

[0220] For example, taking the reference importance degree as the reference importance degree value as an example, assuming that the reference importance degree value of the keyword t i is softmax(t i ) through the above formula three to formula five, then the importance score S(t i ) of this keyword t i can be obtained through the following formula six:

[0221] S(t i )=softmax(t i )*(1+max(HITS A (t i ),HITS H (t i ))) (Formula six);

[0222] Wherein, HITS A (t i ) is the corresponding authority value of the keyword t i , and HITS H (t i ) is the corresponding hub value of the keyword t i .

[0223] S810, for any dialogue sentence in the dialogue text, the importance score corresponding to the keyword with the highest importance score in the dialogue sentence is determined as the score of the dialogue sentence.

[0224] S811, for each dialogue partition, according to the target number of extracted sentences required by the dialogue partition, extract the first target number of target dialogue sentences with higher importance score from the dialogue partition.

[0225] S812, according to the order of each dialogue partition and the order of the target dialogue sentences extracted in the dialogue partition, combine the target dialogue sentences extracted by each dialogue partition into the summary of the dialogue text.

[0226] Corresponding to the summary generation method provided by the embodiment of the present application, the embodiment of the present application further provides a summary generation device.

[0227] As shown in Figure 10 , which shows a component structure schematic diagram of an embodiment of the summary generation device provided by the embodiment of the present application. The device of the embodiment can include:

[0228] A keyword determination unit 1001 is configured to determine keywords in each dialogue sentence in the dialogue text.

[0229] An association determination unit 1002 is configured to determine the relevance between different keywords and the appearance order of the keywords in the dialogue sentence.

[0230] An influence determination unit 1003 is configured to determine the influence degree of the keywords on the dialogue text based on the appearance order.

[0231] An importance determination unit 1004 is configured to determine the importance degree of the keywords in combination with the relevance between different keywords and the influence degree of the keywords on the dialogue text.

[0232] A summary generation unit 1005 is configured to generate a summary of the dialogue text based on the importance degree of the keywords, wherein the summary includes dialogue sentences in which the keywords with the importance degree meeting a condition are located.

[0233] In a possible implementation manner, the device further includes:

[0234] A dialogue partition unit is configured to divide the dialogue text into at least one dialogue partition before the summary generation unit generates the summary of the dialogue text, wherein the dialogue sentences in different dialogue partitions represent different categories of dialogue intents.

[0235] The summary generation unit is specifically configured to determine target dialogue sentences in the dialogue partition for composing a summary based on the importance degree of the keywords, to obtain a summary composed of the target dialogue sentences in each dialogue partition.

[0236] In another possible implementation manner, the summary generation unit includes:

[0237] A sentence determination unit is configured to determine the target number of target dialogue sentences in the dialogue partition based on the target number of required extracted sentences of different dialogue intents in the dialogue partition, in combination with the importance degree of the keywords.

[0238] In another possible implementation manner, the association determination unit includes:

[0239] The co-occurrence determination unit is configured to determine, for each keyword, co-occurrence keywords of the keyword and a co-occurrence dialogue sentence in which the keyword and the co-occurrence keywords first co-occur, the co-occurrence keywords being other keywords that co-occur with the keyword in a dialogue sentence;

[0240] The sequence determination unit is configured to determine an appearance sequence between the keyword and the co-occurrence keywords in the co-occurrence dialogue sentence.

[0241] In yet another possible implementation, the influence determination unit comprises:

[0242] The quantity determination sub-unit is configured to determine, based on the appearance sequence, a first quantity of other keywords located after the keyword and a second quantity of other keywords located before the keyword;

[0243] The influence determination sub-unit is configured to determine, based on the first quantity and the second quantity corresponding to the keyword, an influence degree of the keyword on the dialogue text.

[0244] In yet another possible implementation, the sequence determination unit comprises:

[0245] The directed graph construction unit is configured to construct a directed graph representing an association relationship between different keywords in the dialogue text;

[0246] The directed graph comprises a plurality of nodes and a plurality of directed edges between the nodes, each node representing a keyword, and a directed edge between two nodes indicating that the two keywords corresponding to the two nodes are co-occurrence keywords, and a direction of the directed edge between the two nodes representing an appearance sequence of the two keywords in a co-occurrence dialogue sentence.

[0247] The influence determination unit is specifically configured to determine, according to the quantity and the direction of the directed edges between different nodes in the directed graph, an influence degree of the keyword represented by the node on the dialogue text.

[0248] In yet another possible implementation, the importance determination unit comprises:

[0249] The source determination sub-unit is configured to determine, for each keyword, a first appearance dialogue sentence in which the keyword first appears in the dialogue text and a sentence source of the first appearance dialogue sentence.

[0250] The benchmark determination sub-unit is configured to determine, in combination with the sentence source of the first appearance dialogue sentence corresponding to the keyword and the relevance between different keywords, a benchmark importance degree of the keyword.

[0251] The importance determining sub-unit is configured to determine the importance of the keyword by combining the reference importance of the keyword and the influence degree of the keyword on the dialogue text.

[0252] In yet another possible implementation, the statement source determined by the source determining sub-unit includes one of a customer service and a user;

[0253] The reference determining sub-unit includes:

[0254] The determining sub-unit is configured to determine the relevance sum of the relevance between the keyword and other keywords.

[0255] The reference degree determining sub-unit is configured to determine the reference importance of the keyword by combining the statement source of the first appearance dialogue statement corresponding to the keyword and the relevance sum corresponding to the keyword; if the first appearance dialogue statement corresponding to the keyword is from the customer service, the reference importance of the keyword is negatively correlated with the relevance sum corresponding to the keyword; if the first appearance dialogue statement corresponding to the keyword is from the user, the reference importance of the keyword is positively correlated with the relevance sum corresponding to the keyword.

[0256] In yet another possible implementation, the association determining unit includes:

[0257] The vector determining sub-unit is configured to determine the word vector of each keyword based on the dialogue statement in which the keyword first appears in the dialogue text.

[0258] The similarity determining sub-unit is configured to determine the vector similarity between different keywords based on the word vector of the keyword.

[0259] The relevance determining sub-unit is configured to determine the relevance between different keywords by combining the vector similarity between different keywords, the word frequency of the keyword in the dialogue text, and the number of times different keywords appear simultaneously in a dialogue statement.

[0260] In yet another aspect, the present application also provides an electronic device, such as Figure 11 As shown in the figure, it shows a component structure diagram of the electronic device, which can be any type of electronic device, and the electronic device at least includes a processor 1101 and a memory 1102.

[0261] The processor 1101 is configured to execute the summary generation method in any one of the above embodiments.

[0262] The memory 1102 is configured to store the programs required by the processor for execution.

[0263] It can be understood that the electronic device can further include a display unit 1103 and an input unit 1104.

[0264] Of course, the electronic device can also have more or less components than those shown, and such components can be arranged differently than shown. Furthermore, the Figure 11 components, without limitation.

[0265] In another aspect, the present application also provides a computer readable storage medium, wherein at least one instruction, at least one program, a code set or an instruction set is stored, and the at least one instruction, the at least one program, the code set or the instruction set is loaded and executed by a processor to implement the abstract generation method according to any one of the above embodiments.

[0266] The present application also provides a computer program, which includes computer instructions stored in a computer readable storage medium. The computer program is used to execute the abstract generation method according to any one of the above embodiments when running on an electronic device.

[0267] It should be noted that each of the above embodiments in the present specification is described in a progressive manner, and each embodiment focuses on the difference from other embodiments. The same or similar parts of each embodiment can be referred to each other. Meanwhile, the features recorded in each embodiment in the present specification can be replaced or combined with each other, so that the person skilled in the art can implement or use the present application. For the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts are referred to the part of the method embodiment.

[0268] Finally, it should be noted that in this paper, the relationship terms such as first and second 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 the entities or operations. Moreover, the terms "include", "contain" or any other variants thereof are intended to cover non-exclusive inclusion, so that the process, method, article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or device. Without more limitation, the element defined by the statement "including a…" does not exclude the presence of other identical elements in the process, method, article or device including the element.

[0269] The above description of disclosed embodiments enables one of ordinary skill in the art to make or use the application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein can be applied to other embodiments without departing from the spirit or scope of the application. Thus, the present application is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0270] The preferred embodiments of the application described hereinabove are therefore to be considered in all respects as illustrative and not restrictive, and all changes coming within the meaning and equivalency range of the appended claims are therefore intended to be embraced therein.

Claims

1. A method for generating an abstract, comprising: determining keywords in each dialogue sentence in dialogue text; determining relevance between different keywords and appearance order between keywords in the dialogue sentence; wherein the appearance order between keywords in the dialogue sentence is determined by determining, for each keyword, co-occurrence keywords of the keyword and a co-occurrence dialogue sentence in which the keyword and its co-occurrence keywords first appear together, the co-occurrence keywords of the keyword being other keywords that appear in a dialogue sentence together with the keyword, and determining appearance order between the keyword and its co-occurrence keywords in the co-occurrence dialogue sentence; determining a first number of other keywords located after the keyword and a second number of other keywords located before the keyword based on the appearance order; determining an influence degree of the keyword on the dialogue text based on a sum of the first number and the second number corresponding to the keyword, or determining the influence degree of the keyword on the dialogue text based on a larger one of the first number and the second number corresponding to the keyword; determining an importance degree of the keyword by combining the relevance between different keywords and the influence degree of the keyword on the dialogue text; generating an abstract of the dialogue text based on the importance degree of the keyword, the abstract comprising dialogue sentences in which keywords with an importance degree meeting a condition are located in the dialogue text. 2.The method of claim 1, before the generating the abstract of the dialogue text, further comprising: dividing the dialogue text into at least one dialogue partition, dialogue sentences in different dialogue partitions representing different categories of dialogue intents; and the generating the abstract of the dialogue text based on the importance degree of the keyword comprises: determining target dialogue sentences in the dialogue partition for composing an abstract based on the importance degree of the keyword, to obtain an abstract composed of target dialogue sentences in each dialogue partition. 3.The method of claim 2, the determining target dialogue sentences in the dialogue partition for composing an abstract based on the importance degree of the keyword comprises: determining the target number of target dialogue sentences in the dialogue partition based on a target number of dialogue sentences required by each dialogue partition of different dialogue intents, in combination with the importance degree of the keyword. 4.The method of claim 1, the determining appearance order between the keyword and its co-occurrence keywords in the co-occurrence dialogue sentence comprises: constructing a directed graph representing association relationships between different keywords in the dialogue text; wherein the directed graph comprises a plurality of nodes and a plurality of directed edges between the nodes, each node representing a keyword, and a directed edge between two nodes indicating that the two keywords corresponding to the two nodes are co-occurrence keywords, and a direction of the directed edge between the two nodes representing appearance order of the two keywords corresponding to the two nodes in a co-occurrence dialogue sentence; and the determining an influence degree of the keyword on the dialogue text based on the appearance order comprises: ​ ​ ​ ​ ​ ​ ​ ​ According to the number and direction of the directed edges between different nodes in the directed graph, the influence degree of the keywords represented by the nodes on the dialogue text is determined.

5. The method of claim 1, wherein the importance of the keywords is determined according to the relevance between different keywords and the influence degree of the keywords on the dialogue text, comprising: For each keyword, a first-occurrence dialogue sentence in which the keyword first occurs in the dialogue text and a sentence source of the first-occurrence dialogue sentence are determined; The reference importance of the keyword is determined according to the sentence source of the first-occurrence dialogue sentence corresponding to the keyword and the relevance between different keywords; The importance of the keyword is determined according to the reference importance of the keyword and the influence degree of the keyword on the dialogue text.

6. The method of claim 5, wherein the sentence source comprises one of a customer service source and a user source; The reference importance of the keyword is determined according to the sentence source of the first-occurrence dialogue sentence corresponding to the keyword and the relevance between different keywords, comprising: The relevance sum of the relevance between the keyword and other keywords is determined; The reference importance of the keyword is determined according to the sentence source of the first-occurrence dialogue sentence corresponding to the keyword and the relevance sum corresponding to the keyword; If the first-occurrence dialogue sentence corresponding to the keyword is from the customer service, the reference importance of the keyword is negatively correlated with the relevance sum corresponding to the keyword; If the first-occurrence dialogue sentence corresponding to the keyword is from the user, the reference importance of the keyword is positively correlated with the relevance sum corresponding to the keyword.

7. The method of claim 1, wherein the relevance between different keywords is determined, comprising: For each keyword, a word vector of the keyword is determined based on a dialogue sentence in which the keyword first occurs in the dialogue text; A vector similarity between different keywords is determined based on the word vectors of the keywords; The relevance between different keywords is determined according to the vector similarity between different keywords, the word frequency of the keyword in the dialogue text, and the number of times different keywords appear simultaneously in a dialogue sentence.

8. An abstract generation apparatus, comprising: A keyword determination unit configured to determine keywords in dialogue sentences in dialogue text; An association determination unit configured to determine the relevance between different keywords and the appearance order of the keywords in the dialogue sentences; wherein the appearance order of the keywords in the dialogue sentences is determined by determining, for each keyword, co-occurring keywords of the keyword and a co-occurrence dialogue sentence in which the keyword and the co-occurring keywords first occur together, the co-occurring keywords being other keywords that appear simultaneously with the keyword in a dialogue sentence, and determining the appearance order between the keyword and the co-occurring keywords in the co-occurrence dialogue sentence; An influence determination unit configured to determine the influence degree of the keyword on the dialogue text based on the appearance order. The importance determination unit is configured to determine the importance of the keyword by combining the correlation between different keywords and the influence degree of the keyword on the dialogue text. The abstract generation unit is configured to generate an abstract of the dialogue text based on the importance of the keyword, and the abstract includes a dialogue sentence in which the keyword with the importance meeting a condition is located in the dialogue text. The influence determination unit determines the influence degree of the keyword on the dialogue text based on the appearance order, including: determining a first number of other keywords located after the keyword and a second number of other keywords located before the keyword based on the appearance order; determining the influence degree of the keyword on the dialogue text based on the sum of the first number and the second number corresponding to the keyword, or determining the influence degree of the keyword on the dialogue text based on the larger one of the first number and the second number corresponding to the keyword.

Images