Information processing method and apparatus

Abstract

The present specification provides an information processing method and device, wherein the information processing method comprises: collecting a set of to-be-processed conversations associated with a target service; inputting to-be-processed conversation information included in the set of to-be-processed conversations into a language processing model for processing, obtaining conversation features corresponding to the to-be-processed conversation information, and composing a set of conversation features; performing clustering processing on the conversation features included in the set of conversation features, obtaining a target topic cluster according to the clustering processing result; determining service association information of a sub-topic in the target topic cluster, and determining topic association information of the target topic cluster based on the service association information, and processing the set of to-be-processed conversations of the target service, which realizes the analysis of the association between the to-be-processed conversation information and the target service, does not require marking the to-be-processed conversation information in the set of to-be-processed conversations, reduces resource investment, and is more convenient and flexible for information processing.

Description

Technical Field

[0001] This specification relates to the field of artificial intelligence technology, and more particularly to an information processing method. This specification also relates to an information processing device, a computing device, and a computer-readable storage medium. Background Art

[0002] With the development of Internet technology, scenarios of communicating with users around related businesses are becoming more and more common. In actual scenarios, the types of related businesses are also various, such as product purchases, knowledge Q&A, etc. It is becoming increasingly important to guide users to complete related businesses through dialogue content.

[0003] In the existing technology, before conversing with the user, the topic of the user conversation is determined, and the conversation containing specific features is divided into related topics. However, this method cannot identify conversation categories without a determined topic, and a large amount of annotation is required for the specific features contained in the user conversation, which makes it difficult to control the cost. In addition, this method also limits the number of topics of user conversations, which is not conducive to the analysis of user conversations. Summary of the Invention

[0004] In view of this, the embodiments of this specification provide an information processing method. This specification also relates to an information processing apparatus, a computing device, and a computer-readable storage medium to address the technical deficiencies in the prior art.

[0005] According to a first aspect of the embodiments of this specification, there is provided an information processing method, including:

[0006] Collect the pending dialogue set related to the target business;

[0007] Inputting the dialogue information to be processed contained in the dialogue set to be processed into a language processing model for processing, obtaining dialogue features corresponding to the dialogue information to be processed, and forming a dialogue feature set;

[0008] performing clustering processing on the conversation features included in the conversation feature set, and obtaining a target topic cluster based on the clustering processing result;

[0009] Determine business association information of subtopics in the target topic cluster, and determine topic association information of the target topic cluster based on the business association information; wherein the business association information is used to represent the degree of association between the subtopics and the target business.

[0010] According to a second aspect of the embodiments of this specification, there is provided an information processing device, including:

[0011] A collection module configured to collect a set of pending conversations associated with a target business;

[0012] a processing module configured to input the unprocessed dialogue information contained in the unprocessed dialogue set into a language processing model for processing, obtain dialogue features corresponding to the unprocessed dialogue information, and form a dialogue feature set;

[0013] a clustering module configured to perform clustering processing on the conversation features included in the conversation feature set, and obtain a target topic cluster based on the clustering processing results;

[0014] The determination module is configured to determine the business association information of the subtopics in the target topic cluster, and determine the topic association information of the target topic cluster based on the business association information; wherein the business association information is used to represent the association between the subtopics and the target business.

[0015] According to a third aspect of an embodiment of this specification, a computing device is provided, including:

[0016] memory and processor;

[0017] The memory is configured to store computer-executable instructions, and the processor is configured to execute the computer-executable instructions:

[0018] Collect the pending dialogue set related to the target business;

[0019] Inputting the dialogue information to be processed contained in the dialogue set to be processed into a language processing model for processing, obtaining dialogue features corresponding to the dialogue information to be processed, and forming a dialogue feature set;

[0020] Clustering is performed on the conversation features included in the conversation feature set, and a target topic cluster is obtained according to the clustering result; wherein the business association information is used to characterize the degree of association between the subtopic and the target business.

[0021] Determine business association information of subtopics in the target topic cluster, and determine topic association information of the target topic cluster based on the business association information; wherein the business association information is used to represent the degree of association between the subtopics and the target business.

[0022] According to a fourth aspect of the embodiments of this specification, a computer-readable storage medium is provided, which stores computer-executable instructions, and when the instructions are executed by a processor, the steps of the information processing method are implemented.

[0023] The information processing method provided in this specification collects a set of unprocessed conversations related to a target business; inputs the unprocessed conversation information contained in the unprocessed conversation set into a language processing model for processing to obtain conversation features corresponding to the unprocessed conversation information and form a conversation feature set; clusters the conversation features contained in the conversation feature set to obtain a target topic cluster based on the clustering results; determines the business association information of subtopics in the target topic cluster, and determines the topic association information of the target topic cluster based on the business association information; wherein the business association information is used to characterize the degree of association between the subtopic and the target business.

[0024] An embodiment of this specification realizes processing of a pending conversation set of a target business and analysis of the correlation between the pending conversation information and the target business. There is no need to mark the pending conversation information in the pending conversation set, thus reducing resource investment and making information processing more convenient and flexible. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 This is a flow chart of an information processing method provided in one embodiment of this specification;

[0026] Figure 2 This is a processing flow chart of an information processing method applied to a question-and-answer dialogue provided in an embodiment of this specification;

[0027] Figure 3 This is a processing flow chart of an information processing method provided in one embodiment of this specification;

[0028] Figure 4 This is a schematic diagram of the structure of an information processing device provided in one embodiment of this specification;

[0029] Figure 5 This is a structural block diagram of a computing device provided in one embodiment of this specification. DETAILED DESCRIPTION

[0030] The following description sets forth many specific details to facilitate a thorough understanding of this specification. However, this specification can be implemented in many other ways than those described herein, and those skilled in the art can make similar generalizations without violating the scope of this specification. Therefore, this specification is not limited to the specific implementations disclosed below.

[0031] The terms used in one or more embodiments of this specification are for the purpose of describing specific embodiments only and are not intended to limit one or more embodiments of this specification. The singular forms "a," "the," and "the" used in one or more embodiments of this specification and the appended claims are also intended to include plural forms unless the context clearly indicates otherwise. It should also be understood that the term "and / or" used in one or more embodiments of this specification refers to and includes any or all possible combinations of one or more associated listed items.

[0032] It should be understood that although the terms first, second, etc. may be used to describe various information in one or more embodiments of this specification, such information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other. For example, without departing from the scope of one or more embodiments of this specification, the first may also be referred to as the second, and similarly, the second may also be referred to as the first. Depending on the context, the word "if" as used herein may be interpreted as "at the time of" or "when" or "in response to determining".

[0033] First, the terms involved in one or more embodiments of this specification are explained.

[0034] Mean pooling: Taking the average of all activation values ​​within the pooling domain, high positive activation values ​​may cancel each other out with low negative activation values, resulting in a loss of discriminative information. The sequence-based average pooling method solves this problem by taking the average of the first t largest activation values.

[0035] Clustering: The process of dividing a collection of physical or abstract objects into multiple classes consisting of similar objects is called clustering; the cluster generated by clustering is a collection of data objects that are similar to objects in the same cluster and different from objects in other clusters.

[0036] In this specification, an information processing method is provided. This specification also relates to an information processing apparatus, a computing device, and a computer-readable storage medium, which are described in detail one by one in the following embodiments.

[0037] Figure 1 A flowchart of an information processing method provided according to an embodiment of this specification is shown, which specifically includes the following steps:

[0038] Step S102: Collect a set of to-be-processed conversations related to the target business.

[0039] Among them, the target business can be understood as the business activities carried out by users based on their own demands, such as purchasing goods, communicating problems, etc.; the set of conversations to be processed can be understood as the sum of the users' conversations regarding the target business. It should be noted that in the process of users conducting conversations regarding the target business, the object of the conversation can be either a human customer service representative or an intelligent robot, which is not limited in this embodiment.

[0040] Based on this, the information contained in the pending conversation set is the conversation information generated for users of the same business, and the information in the pending conversation set can be from different users. In the process of collecting the pending conversation set associated with the target salesperson, it should be noted that for the pending conversation set, its source can be the user's conversation information about the target business stored locally on the device, such as a vending machine based on voice recognition, in which the conversation with the user is stored in the vending machine, or it can be stored in a cloud storage space, such as a certain product website, in which the communication between the user and the customer service will be stored in the relevant cloud storage space by the website. The specific storage form of the pending conversation set is determined by the actual usage scenario and is not limited in this embodiment. By collecting the pending conversation set associated with the target business, the conversation content of the users in the pending conversation set can be analyzed later to better guide the user to develop the target business.

[0041] Step S104: inputting the dialogue information to be processed contained in the dialogue set to be processed into a language processing model for processing, obtaining dialogue features corresponding to the dialogue information to be processed, and forming a dialogue feature set.

[0042] Specifically, after collecting the conversation set to be processed, it is necessary to analyze the conversation content of the users included in the conversation set to be processed. However, directly processing the text or voice information through a computing device is too complicated and will occupy a large amount of computing resources. Therefore, it is necessary to convert the conversation content of the users included in the conversation set to be processed into a feature form.

[0043] Among them, the dialogue information to be processed can be understood as the information related to the dialogue content generated when the user conducts a dialogue for the target business. In addition, since the user can conduct the dialogue in text form or voice form during the dialogue for the target business, in the latter case, speech recognition technology is required to convert the voice into text; the language processing model can be understood as a method that can vectorize the dialogue information to be processed through pre-training and determine the dialogue features corresponding to the dialogue information to be processed.

[0044] Based on this, the textual information contained in the pending conversation set is input into the language processing model for processing. During processing, the conversation features corresponding to the pending conversation information are determined based on the semantics of the pending conversation information. After determining the conversation features of all the pending conversation information in the pending conversation set, the resulting conversation features are assembled into a set, namely, a conversation feature set. It should be noted that when processing the pending conversation information in the pending conversation set using the language processing model, the pending conversation information can be input into the language processing model sequentially or simultaneously for concurrent processing. The specific processing method is determined by the actual use case and is not limited in this embodiment. By vectorizing the pending conversation information in the pending conversation set in the above manner, the processing difficulty of the computing device is reduced, effectively improving processing efficiency.

[0045] Furthermore, in the process of generating corresponding dialogue features based on the dialogue information to be processed, it is necessary to ensure the accuracy of the language processing model. If the dialogue information to be processed does not match the relevant dialogue features, it is difficult to ensure the accuracy of the subsequent processing results. In this embodiment, the specific implementation method is as follows:

[0046] The method comprises extracting sample conversation information contained in a sample conversation set and determining sample word information contained in the sample conversation information; replacing the sample word information contained in the sample conversation information in sequence based on a preset replacement strategy to obtain replacement text information; inputting the replacement text information into an initial language processing model for processing to generate a predicted word information vector; determining a replacement word information vector for the replacement text information, and calculating a loss value based on the replacement word information vector and the predicted word information vector; and adjusting parameters of the initial language processing model according to the loss value until the initial language processing model meets a training stop condition, thereby obtaining the language processing model.

[0047] Among them, the sample dialogue set can be understood as a collection of sample dialogue information used to pre-train the initial language processing model; the sample word information can be understood as the information of words or characters contained in the text corresponding to the sample dialogue information; the replacement text information can be understood as the dialogue information obtained by masking the sample word information in the sample dialogue information; the predicted word information vector can be understood as the prediction result of the initial language processing model for the feature vector corresponding to the word information in the replacement text information; the replacement word information vector can be understood as the feature vector corresponding to this information in the replacement text information.

[0048] Based on this, the sample word information contained in each sample conversation information is extracted from the sample conversation information in the sample conversation set. The sample conversation set can be obtained from the historical storage space of the device that processes the target business, or the conversation information associated with the target business can be captured from the network to form a sample conversation set. The specific source of the sample conversation set is determined by the actual usage scenario and is not limited in this embodiment.

[0049] Then, the sample word information in the sample dialogue information is replaced. The replacement process can replace the sample word information with a fixed replacement word, such as a mask, according to a preset ratio, or keep the sample word information unchanged, or replace the sample word information with random word information. The initial language processing model then processes the replaced text information. The initial language processing model predicts the vector of the word information contained in the replacement text information, and then compares it with the vector of the word information contained in the replacement text information. The loss value is calculated according to the comparison result, and the model parameters of the initial language processing model are adjusted based on the loss value until the initial language processing model predicts the vector of the word information in the replacement text information with an accuracy that meets expectations. The pre-training process is then stopped to obtain the language processing model.

[0050] For example, in an online course registration system, the conversation data of customers communicating with customer service about whether to register for the online course is stored. At this time, conversation data similar to this business is crawled from the Internet, and the combined set is used as a sample conversation set. The BERT model is selected as the initial language processing model. Then, based on the mask strategy, the sample conversation set is used to train the initial language processing model to obtain the language processing model.

[0051] In summary, through the above steps, the pre-training process of the language processing model is implemented, ensuring that the language processing model can process the pending dialogue information and obtain the corresponding dialogue features accurately. Through the preset replacement strategy, the ability of the initial language processing model to generate corresponding dialogue features based on the dialogue information is further explored.

[0052] Furthermore, since the length of the conversation information to be processed is determined by the actual conversation situation of the user, it is difficult for the length to be consistent. Inputting conversation information of varying lengths into the language processing model will result in low input efficiency. To solve this problem, in this embodiment, the specific implementation method is as follows:

[0053] The i-th pending conversation information in the pending conversation information set is concatenated with the j-th pending conversation information segment to obtain a target conversation information segment; whether the length of the target conversation information segment is greater than a preset concatenation length is determined; if not, i is incremented by 1, the target conversation information segment is used as the j-th pending conversation information segment, and the step of concatenating the i-th pending conversation information in the pending conversation information set with the j-th pending conversation information segment to obtain the target conversation information segment is performed; if so, the target conversation information segment is segmented to obtain a target length information segment and a segmented sub-information segment having a length equal to the preset concatenation length, j is incremented by 1, the segmented sub-information segment is used as the j-th pending conversation information segment, i is incremented by 1, and the step of concatenating the i-th pending conversation information in the pending conversation information set with the j-th pending conversation information segment to obtain the target conversation information segment is performed; until i equals the number n of pending conversation information in the pending conversation information set, at least one target length information segment is determined, where i and j start at 1 and are positive integers.

[0054] The to-be-processed dialogue information segment can be understood as information obtained by splicing the to-be-processed dialogue information; and the target length information segment can be understood as splicing information for subsequent input into a language processing model.

[0055] Based on this, the pending conversation information contained in the pending conversation information set is spliced ​​with the pending conversation information segment. When the first pending conversation information in the pending conversation information set is spliced, the pending conversation information segment is also the first one. This pending conversation information segment is empty and does not contain any conversation information. After splicing the pending conversation information with the pending conversation information segment, it is determined whether the length of the spliced ​​target conversation information segment is greater than a preset length. If it is less than a preset length, the target conversation information segment is used as the pending conversation information segment, and the next pending conversation information is selected and spliced ​​onto the pending conversation information segment until the length of the spliced ​​target conversation information segment is greater than the preset length. At this time, the portion of the target conversation information segment that is longer than the preset length is split to obtain a target length information segment of the preset length and a split sub-information segment. The split sub-information segment is used as the next pending conversation information segment, and the splicing continues until all the pending conversation information in the pending conversation information set are spliced.

[0056] It should be noted that, in the process of splicing the dialogue information to be processed, there is a case where the length of the target dialogue information segment formed after the last dialogue information to be processed is spliced ​​with the dialogue information segment to be processed is not equal to the preset length. If the length of the target dialogue information segment is less than the preset length, the target dialogue information segment is used as a target length information segment; if the length of the target dialogue information segment is greater than the preset length, the target dialogue information segment is split to obtain a target length information segment with a length equal to the preset length and a split sub-information segment. Then, it is determined whether the length of the split sub-information segment is greater than the preset length. If so, the split sub-information segment is further split to obtain a target length information segment with a preset length, and the length of the split sub-information segment is further judged. Based on the judgment result, it is determined whether the split sub-information segment is further split until the length of the split sub-information segment is less than or equal to the preset length, and the sub-information segment is also used as a target length information segment.

[0057] In addition, all pending dialogue information may be spliced ​​together to obtain spliced ​​global pending dialogue information segments, which are then segmented according to a preset length to obtain target length information segments of a preset length.

[0058] In summary, by concatenating the pending conversation messages in the above manner to generate information segments of uniform target length, and then inputting these target length segments into the language processing model for processing, processing efficiency can be effectively increased. It should be noted that a conversation information identifier can be added to the target length segments to distinguish between two adjacent pending conversation messages within the target length segment.

[0059] Furthermore, the language processing model determines the vector of each word or character in the text information based on the semantic relationship of the input text information context. Therefore, for the target-length information segment after splicing, processing it through the language processing model may result in different dialogue information to be processed being included in the same context analysis scope, resulting in processing errors. To solve this problem, in this embodiment, the specific implementation method is as follows:

[0060] At least one target length information segment is sequentially input into the language processing model; each target length information segment is processed by a dialogue information extraction unit in the language processing model to determine the dialogue information to be processed contained in each target length information segment; and each target length information segment is processed by a dialogue feature determination module in the language processing model to obtain dialogue features of each piece of dialogue information to be processed.

[0061] Among them, the conversation information extraction unit is used to identify the conversation information to be processed contained in the target length information segment. The identification method can determine the position of the conversation information to be processed in the target length information segment through the position information contained in the segment header or segment tail in the target length information segment, or identify the conversation information identifier in the target length information segment to determine the connection position of two adjacent conversation information to be processed, and further identify the position of the two adjacent conversation information to be processed. The specific identification method is determined by the actual usage scenario and is not limited in this embodiment.

[0062] Based on this, after the target length information segment is input into the language processing model, the language processing model first extracts the dialogue information to be processed contained in the target length information segment, and then processes the dialogue information to be processed to determine the dialogue features corresponding to each dialogue information to be processed.

[0063] Continuing with the above example, we collect conversations between customers about whether to register for online courses from the storage space corresponding to the online course registration system. The texts of these conversations are then spliced ​​together. During the splicing process, the corresponding texts can be spliced ​​together in the order of speech in the conversations. A determination is made as to whether the length of the spliced ​​text is greater than 512 words. If not, the next text of the conversation is selected and spliced ​​together until the length of the resulting spliced ​​text is greater than 512 words. The spliced ​​text is then segmented so that the length of the spliced ​​text equals 512 words. The segmented text portion is then spliced ​​into the next spliced ​​text. In this manner, all conversations between customers about whether to register for the online course are spliced ​​together to obtain at least one spliced ​​text, i.e., a target length information segment. It should be noted that the last spliced ​​text may not be 512 words long. In this case, the spliced ​​text that does not meet the 512-word length is also considered a target length information segment.

[0064] In addition, each communication dialogue text spliced ​​in the target length information segment can be given a dialogue information identifier, so that when the target length information segment is input into the language processing model, the dialogue information extraction unit of the language processing model can distinguish the communication dialogue texts in the target length information segment to avoid confusion between different communication dialogue texts.

[0065] Afterwards, the target length information segment obtained by splicing is input into the language processing model, and the dialogue information extraction unit of the language processing model determines the dialogue information identifier contained in the target length information segment, determines the position of the dialogue information to be processed contained in the target length information segment, and extracts the dialogue information to be processed. Then, the dialogue feature determination module of the language processing model determines the dialogue features of each dialogue information to be processed.

[0066] In summary, by extracting the unprocessed dialogue information in the target length information segment and performing feature extraction on the unprocessed dialogue information separately, the accuracy of the language processing model in determining the dialogue features of the unprocessed dialogue information is improved.

[0067] Furthermore, the process of inputting any piece of dialogue information to be processed in the dialogue set to be processed into the language processing model for processing is the same. Therefore, the process of processing any piece of dialogue information to be processed by the language processing model is specifically implemented as follows in this embodiment:

[0068] Inputting target dialogue information to be processed into the language processing model; processing the target dialogue information to be processed by a word information extraction unit in the language processing model to obtain at least one word information contained in the target dialogue information to be processed; processing the at least one word information by a vector determination unit in the language processing model to determine a word vector corresponding to each word information as a word vector associated with the target dialogue information to be processed; generating dialogue features of the target dialogue information to be processed based on the word vector associated with the target dialogue information to be processed, and using them as output of the language processing model.

[0069] Among them, the target unprocessed dialogue information can be understood as the unprocessed dialogue information currently being processed by the language processing model; the word information extraction unit can be understood as being used to identify the word information contained in the unprocessed dialogue information and extract the individual word information contained in the unprocessed dialogue information; the vector determination unit can be understood as being used to determine the corresponding word vectors of the individual word information contained in the unprocessed dialogue information; the dialogue feature can be understood as the vector form expression of the unprocessed dialogue information, and different unprocessed dialogue information corresponds to different dialogue features, and the dialogue feature is related to the semantics of the unprocessed dialogue information.

[0070] It should be noted that the word information extraction unit and the vector determination unit in the language processing model constitute a dialogue feature determination module.

[0071] Based on this, the target dialogue information to be processed is input into the language processing model. The word information extraction unit in the language processing model extracts the word information contained in the target dialogue information to be processed. Then, the vector determination unit in the language processing model determines the word vector corresponding to the extracted word information, and generates the dialogue feature of the target dialogue information to be processed through the word vector associated with the target dialogue information to be processed. The dialogue feature is the output of the language processing model processing the target dialogue information to be processed.

[0072] Continuing with the above example, the customer's communication conversation regarding whether to sign up for an online course is input into the language processing model for processing. First, the word information extraction unit in the language processing model is used to extract the various Chinese characters contained in the customer's communication conversation. Then, the vector determination unit in the language processing model is used to determine the text vectors corresponding to the various Chinese characters extracted from the customer's communication conversation. Based on the text vectors corresponding to the Chinese characters in the same communication conversation, the conversation features corresponding to the communication conversation are determined, and the language processing model is output.

[0073] In summary, word vectors are determined based on the word information contained in the target conversation data to be processed, and the conversation features of the target conversation data to be processed are determined through the word vectors.

[0074] Furthermore, in the process of generating conversation features for the target conversation information to be processed based on the word vectors associated with the target conversation information to be processed, if the conversation features are generated based on only part of the word vectors, the comprehensiveness of the obtained conversation features will be affected. To solve this problem, in this embodiment, the specific implementation method is as follows:

[0075] Mean pooling is performed on the word vectors associated with the target dialogue information to be processed to generate mean features; and based on a preset dimensionality reduction strategy, the mean features are reduced in dimensionality to obtain dialogue features of the target dialogue information to be processed.

[0076] Among them, the execution of mean pooling can be performed based on a deep convolutional neural network, and the dimension reduction of the mean feature through a preset dimensionality reduction strategy can be achieved through algorithms and tools such as UMAP, TSNE, and PCA.

[0077] Based on this, after determining the word vectors of the word information in the target dialogue information to be processed, mean pooling processing can be performed based on all word vectors to obtain the mean feature. At this time, the mean feature corresponds to the target dialogue information to be processed, and its dimension is generally high. Subsequently, clustering is performed using the mean feature, and the number of elements in the resulting cluster will be small, increasing the calculation difficulty of the relevant computing equipment. Therefore, a preset dimensionality reduction strategy is used to reduce the dimensionality of the mean feature, and the reduced mean feature is used as the dialogue feature of the target dialogue information to be processed.

[0078] Continuing with the previous example, based on the word vectors corresponding to the Chinese characters in the customer communication conversation, we generate the mean features of the customer communication conversation through mean pooling. Then, we use the TSNE tool to reduce the dimension of the mean features to obtain the conversation features and generate a visual scatter plot.

[0079] In summary, by performing mean pooling on the word vectors associated with the target dialogue information to be processed, we effectively ensure that the characteristics of the target dialogue information to be processed are fully reflected in the mean features. Subsequently, dialogue features are obtained through dimensionality reduction processing, which ensures the comprehensiveness of the dialogue features to a certain extent while reducing the difficulty of subsequent clustering operations.

[0080] Step S106: performing clustering processing on the conversation features included in the conversation feature set, and obtaining a target topic cluster based on the clustering processing result.

[0081] Specifically, since the pending conversation information in the pending conversation set all revolves around the target business, there will be a large amount of conversation content that is semantically consistent, and the corresponding conversation features will also show a convergence state. In order to reduce the difficulty of analyzing and processing the conversation features of a large amount of pending conversation information, it is necessary to cluster the conversation features.

[0082] Among them, after clustering the conversation features, the conversation features in the same target topic cluster are similar, that is, the semantic information of the corresponding conversation information to be processed is similar, and can be considered to be the same topic. For example, after processing the two conversation information "hello" and "hi", the conversation features obtained will be clustered in the same target topic cluster. Otherwise, the two conversation information to be processed whose corresponding conversation features are no longer in the same target topic cluster do not belong to the same topic.

[0083] Based on this, after calculating the various conversation features corresponding to the conversation information to be processed contained in the conversation information set to be processed, the various conversation features are clustered to obtain the target topic cluster. The conversation features in the same target topic cluster and the semantics of the associated conversation information to be processed are similar and can be considered to be the same topic, which effectively determines the topic type of the conversation information to be processed in the conversation information set to be processed. Subsequently, processing and analyzing the target topic cluster as a whole can effectively reduce the computing resource investment of related equipment and reduce the processing pressure of the equipment.

[0084] Furthermore, in the process of clustering the conversation features included in the conversation feature set, the processing process for any conversation feature is consistent. To achieve clustering of all conversation features included in the conversation feature set, in this embodiment, the specific implementation method is as follows:

[0085] The method further comprises the steps of: selecting the kth conversation feature in the conversation feature set as the target conversation feature; determining the conversation feature in the conversation feature set that is associated with the target conversation feature based on a preset clustering strategy; storing the conversation feature associated with the target conversation feature in a topic set associated with the kth conversation feature; incrementing k by 1, and executing the step of selecting the kth conversation feature in the conversation feature set as the target conversation feature, wherein k starts at 1 and is a positive integer, and k is less than or equal to the number m of conversation features contained in the conversation feature set; and merging the topic sets containing the same conversation features in the m topic sets to obtain a target topic cluster.

[0086] The preset clustering strategy may adopt algorithms such as k-means clustering, mean shift clustering, and density-based clustering. The specific clustering algorithm adopted is determined by the actual usage scenario and is not limited in this embodiment.

[0087] Based on this, in the process of clustering the conversation features in the conversation feature set, one of the conversation features is arbitrarily selected as the target conversation feature, and then the conversation features in the conversation feature set are screened based on the clustering strategy of the preset clustering algorithm, and the conversation features associated with the target conversation feature are selected and stored in the topic set corresponding to the target conversation feature; according to this method, the m conversation features in the conversation feature set are processed to obtain m topic sets that store the conversation features associated with each conversation feature, and finally the topic sets corresponding to each conversation feature are merged to obtain the target topic cluster.

[0088] Continuing with the above example, the conversation features in the conversation feature set are selected one by one. Then, the conversation features associated with the selected conversation features are determined based on algorithms such as density-based clustering. The associated conversation features are stored in a topic set associated with the selected conversation features. After the conversation features associated with all the conversation features in the conversation feature set are determined, the topic sets associated with each conversation feature are merged. The merging method is to merge the topic sets containing the same conversation features to obtain the target topic cluster.

[0089] In summary, through the above method, all conversation features in the conversation feature set can be clustered. The target topic cluster obtained by clustering has similar semantics for the conversation information to be processed associated with the conversation features and can be regarded as the same topic.

[0090] Furthermore, when the conversation features are related vector data, in order to ensure the smooth development of the conversation feature clustering process, the conversation features need to be projected into the related vector space for visualization. In this embodiment, the specific implementation method is as follows:

[0091] Based on a preset clustering strategy, a visual scatter plot is created that includes the conversation features in the conversation feature set; in the visual scatter plot, a circle is drawn with the target conversation feature as the center and a preset length as the radius to obtain a target area; and conversation features located within the target area are selected as conversation features associated with the target conversation feature.

[0092] The visual scatter plot can be understood as a visual image of the conversation features projected into the relevant vector space. Each conversation feature has a position coordinate in the visual scatter plot.

[0093] Based on this, by projecting the conversation features into the relevant vector space, a visual scatter plot is generated. Then, in the visual scatter plot, any conversation feature is selected as the target conversation feature, and a circle is drawn with the target conversation feature as the center and a preset length as the radius to obtain a circular target area. The conversation features contained in the target area are used as conversation features associated with the target conversation features. The above steps are performed in sequence on the conversation features in the conversation feature set to determine the conversation features associated with each conversation feature.

[0094] It should be noted that the size of the target area can be changed by adjusting the preset length, thereby adjusting the conversation features with different degrees of correlation. As conversation features associated with the target conversation features, if the preset length is set to a larger value, conversation features that are not highly correlated with the target conversation features will also be included in the conversation feature selection range associated with the target conversation features. Otherwise, only conversation features that are highly correlated with the target conversation features will be selected and included in the conversation feature selection range associated with the target conversation features.

[0095] Continuing with the previous example, we use the TSNE tool to reduce the dimension of the mean feature. In the process of obtaining the conversation features, we generate a visual scatter plot. We select one of the conversation features as the target conversation feature. With the target conversation feature as the center and a radius of 100 units, we create a circular target area related to the target conversation feature. We then determine the conversation features in the target area as the conversation features associated with the target conversation features. We then use the conversation features included in the conversation feature set as the target conversation features to determine the conversation features associated with each conversation feature.

[0096] In summary, through the above method, the conversation features associated with each conversation feature in the conversation feature set are determined, and further clustering based on the association relationship between the conversation features can be achieved.

[0097] Step S108: determining the business association information of the sub-topics in the target topic cluster, and determining the topic association information of the target topic cluster based on the business association information.

[0098] The business association information is used to represent the degree of association between the sub-topic and the target business.

[0099] Specifically, after determining the target topic cluster, in order to subsequently analyze the driving ability of the topics represented by the target topic cluster on the target business, it is necessary to determine the correlation between the target topic cluster and the target business, that is, the topic correlation information.

[0100] Based on this, the sub-topics in the target topic cluster can be understood as the pending conversation information associated with the conversation features in the target topic cluster. Its business association information represents the correlation between the pending conversation information and the target business. For example, in a shopping scenario, if the user purchases a product after the conversation with customer service, it indicates that there is a correlation between the conversation and the shopping business; otherwise, there is no correlation.

[0101] Furthermore, after obtaining the topic association information of the target topic cluster, the association degree between the target topic cluster and the target business can be analyzed. In this embodiment, the specific implementation is as follows:

[0102] When all target topic clusters associated with the conversation set to be processed are determined, the average value of the topic association information of each target topic cluster is calculated; based on each topic association information and the average value, the deviation rate of each topic association information is calculated; based on a preset deviation rate selection strategy, a target deviation rate is selected from each deviation rate, and the target topic cluster associated with the target deviation rate is determined as a reference topic cluster.

[0103] The deviation rate is used to indicate the driving effect of the pending conversation information associated with the target topic cluster on the target business. For example, in a shopping scenario, if a user has a conversation about a certain type of topic and then shows a clear tendency to buy or not buy, the deviation rate of the topic is high, otherwise it is low.

[0104] Based on this, after the clustering of all conversation features corresponding to the conversation information set to be processed is completed and the topic association information of all target topic clusters obtained by clustering is determined, the average value of each topic association information is calculated, and based on the each topic association information and the calculated average value, the preset deviation rate selection strategy is used to select the target deviation rate in each topic association information, and the target topic cluster associated with the target deviation rate is used as the reference topic cluster.

[0105] Continuing with the above example, clustering yielded three target topic clusters, whose respective topic association information was 0.4, 0.5, and 0.6, respectively. The calculated average value was 0.5. The target deviation rate was then selected based on the preset deviation rate selection strategy. The selection strategy could be to select the deviation rate with the largest difference from the average value, or to select the deviation rate with the smallest difference from the average value, etc. The specific selection strategy is determined by implementation requirements and is not limited in this embodiment.

[0106] In summary, through the above methods, we can analyze the target topic cluster corresponding to the conversation information set to be processed and screen out topics that meet the user's research and analysis needs.

[0107] Furthermore, in the process of determining the association between the target topic cluster and the target business, the topics that are most conducive to achieving the business objectives of the target business and the topics that are least conducive to achieving the business objectives of the target business are most worthy of attention. In order to determine the above two topics, in this embodiment, the specific implementation method is as follows:

[0108] Each topic-related information is compared with the average value, and a deviation value of each topic-related information is obtained according to the comparison result; and the absolute value of the deviation value of each topic-related information is calculated to obtain a deviation rate of each topic-related information.

[0109] The deviation value may be the difference between the topic-related information and the average value of each topic-related information, and may be a positive number or a negative number.

[0110] Based on this, the average value of each topic-related information is calculated, and the difference between each topic-related information and the average value is calculated to obtain the deviation value of each topic-related information. The obtained deviation value is then processed as an absolute value to obtain the deviation rate of each topic-related information.

[0111] Continuing with the above example, the average value of the topic-related information corresponding to the three target topic clusters is calculated to be 0.5. The differences between the topic-related information and the average value are further calculated to be -0.1, 0, and 0.1. The average value is then calculated to obtain the deviation rates corresponding to the three target topic clusters to be 0.1, 0, and 0.1.

[0112] In summary, through the above methods, we can determine the driving effect of the topics corresponding to each target topic cluster on the target business and express it in the form of deviation rate.

[0113] Furthermore, in most cases of actual usage scenarios, more than one reference topic cluster is selected. In order to confirm multiple reference topic clusters, it is necessary to screen out multiple target offset rates. In this example, the specific implementation is as follows:

[0114] Based on a preset deviation rate selection strategy, a deviation rate higher than a preset deviation rate threshold is selected as the target deviation rate; or, based on a preset deviation rate selection strategy, a preset proportion of deviation rates is selected as the target deviation rate in descending order of deviation rate values.

[0115] In actual usage scenarios, it is often necessary to analyze topics that are conducive to the target business achieving its business goals and provide guidance on the corresponding topics, or to analyze topics that are not conducive to the target business achieving its business goals and achieve improvements to the target business, or to guide users not to engage in conversations around the topic.

[0116] Based on this, a deviation rate higher than a preset deviation rate threshold may be selected as the target deviation rate, or a deviation rate with a larger value in a certain proportion may be selected as the target deviation rate.

[0117] Continuing with the above example, if the preset deviation rate threshold is 0, two deviation rates with values ​​higher than 0 are selected as target deviation rates. In addition, if the preset ratio is 2 / 3, two deviation rates with values ​​in the first 2 / 3 are selected as target deviation rates.

[0118] In summary, the above methods can be used to filter out deviation rates with large values ​​and effectively identify topics that have a greater impact on achieving business goals for the target business.

[0119] The information processing method provided in this specification collects a set of pending dialogues associated with the target business, inputs the pending dialogue information in the pending dialogue set into a language model for processing, obtains dialogue features, clusters the dialogue features to obtain a target topic cluster, and determines the topic association information of the target topic cluster based on the business association information of the sub-topics in the target topic cluster, thereby realizing the analysis of the association relationship between the pending dialogue information and the target business. There is no need to mark the pending dialogue information in the pending dialogue set, which reduces resource investment and makes information processing more convenient and flexible.

[0120] The following combined Figure 2 , taking the application of the information processing method provided in this specification in question-answer dialogue as an example, the information processing method is further explained. Figure 2 A processing flow chart of an information processing method for question-and-answer dialogue provided in an embodiment of this specification is shown, which specifically includes the following steps:

[0121] Step S202: Collect a set of to-be-processed conversations related to the target business.

[0122] Specifically, when users activate smart home appliances, they need to use voice to ask questions about them. For example, in the case of a smart air conditioner, users may inquire about current environmental conditions, such as temperature, humidity, and air pollution, to decide whether to turn on the air conditioner. First, a request is sent to the maintenance vendor of the smart air conditioner to collect historical conversation data between users and the smart air conditioner from the vendor's database.

[0123] Step S204: input the target dialogue information to be processed into the language processing model.

[0124] Specifically, the collected historical conversation data between users and smart air conditioners is input into the pre-trained DialogBERT model.

[0125] The pre-trained DialogBERT model is pre-trained by masking sample conversation data, allowing for the generation of word vectors for the text contained in the conversation data based on contextual information. Furthermore, historical conversation data between users and smart air conditioners fed into the DialogBERT model can be concatenated before input, resulting in text word lengths of 512. The DialogBERT model then splits this concatenated data into individual user conversation sentences. The vectors for the text words contained in each sentence are then determined based on the context of each sentence.

[0126] Step S206: Processing the target dialogue information to be processed by the word information extraction unit in the language processing model to obtain at least one word information contained in the target dialogue information to be processed.

[0127] Specifically, the input dialogue sentence is recognized to identify each text word contained therein.

[0128] Step S208: Process at least one word information through the vector determination unit in the language processing model to determine the word vector corresponding to each word information as the word vector associated with the target dialogue information to be processed.

[0129] Specifically, the DialogBERT model is used to generate vectors for each text word in the dialogue sentence.

[0130] Step S210: Generate dialogue features of the target dialogue information to be processed based on the word vectors associated with the target dialogue information to be processed, and use them as outputs of the language processing model.

[0131] Specifically, we perform mean pooling on the vectors of each word in the same conversation sentence to obtain the mean feature of each conversation sentence. We then use the Unified Mapping (UMAP) model to reduce the dimensionality of the 768-dimensional mean feature of each conversation sentence to obtain a 50-dimensional conversation feature. The resulting conversation features are then used to form a conversation feature set.

[0132] Step S212: Select the kth dialogue feature in the dialogue feature set as the target dialogue feature.

[0133] Specifically, any one conversation feature is selected as the target conversation feature.

[0134] Step S214: Based on a preset clustering strategy, determine the conversation features in the conversation feature set that are associated with the target conversation features.

[0135] Specifically, a visual scatter plot of the conversation features in the conversation feature set is created, and a circle is drawn with the target conversation feature as the center and a radius of 50 units. The conversation features contained in the circle are determined to be conversation features associated with the target conversation feature.

[0136] Step S216: storing the conversation features associated with the target conversation features into the topic set associated with the kth conversation feature.

[0137] Specifically, the conversation features associated with the target conversation features are stored in a topic set.

[0138] Step S218: k is incremented by 1 to determine whether k is greater than m.

[0139] If not, execute step S212;

[0140] If so, execute step S220.

[0141] Specifically, k starts at 1 and is a positive integer, and k is less than or equal to the number m of conversation features included in the conversation feature set.

[0142] Step S220: merging the topic sets containing the same conversation features in the m topic sets to obtain a target topic cluster.

[0143] Step S222: Determine the business association information of the sub-topics in the target topic cluster, and determine the topic association information of the target topic cluster based on the business association information.

[0144] Specifically, the sub-topic is the conversation feature contained in the target conversation cluster, and the business association information is used to characterize the probability that the sub-topic guides the smart air conditioner to turn on. After a round of conversation between the user and the smart air conditioner, if the smart air conditioner is finally turned on, the business association information of each conversation sentence in the conversation is recorded as 1, otherwise it is recorded as 0. The topic association information of the target topic cluster is determined by calculating the average value of the business association information of the sub-topics in the target topic cluster through the business association information of the sub-topics in the target topic cluster.

[0145] Step S224: After all target topic clusters associated with the conversation set to be processed are determined, the average value of the topic association information of each target topic cluster is calculated.

[0146] Specifically, 12 target topic clusters were obtained, and the average value of the 12 target topic clusters was calculated to be 0.4.

[0147] Step S226: Compare each topic-related information with the average value, and obtain the deviation value of each topic-related information according to the comparison result.

[0148] Specifically, the calculated average value is subtracted from the 12 target topic clusters to obtain the deviation values ​​of the 12 target topic clusters.

[0149] Step S228: Calculate the absolute value of the deviation value of each topic-related information to obtain the deviation rate of each topic-related information.

[0150] Specifically, the absolute values ​​of the deviation values ​​of the 12 target topic clusters are calculated to obtain the deviation rates of the 12 target topic clusters.

[0151] Step S230: Based on the preset deviation rate selection strategy, a preset ratio of deviation rates is selected as the target deviation rate in descending order of deviation rate values.

[0152] Specifically, the preset ratio is 1 / 4, and 3 of the 12 target topic clusters with larger deviation rates are selected as the target deviation rates.

[0153] Step S232: Determine the target topic cluster associated with the target deviation rate as the reference topic cluster.

[0154] Specifically, based on the three target deviation rates, the corresponding target topic cluster is determined as a reference topic cluster for subsequent analysis of which topics are more likely to guide users to turn on the smart air conditioner. It should be noted that the specific method of analysis is determined by the actual usage scenario and is not limited in this embodiment.

[0155] Combine Figure 3 The provided flowchart of an information processing method illustrates the above process. Conversation sentences s1, s2, s3, and s4 are input into the DialogBERT domain pre-trained model for processing. The vectors e1, e2, e3, and e4 for each sentence are determined and then input into the UMAP dimensionality reduction model for processing, resulting in low-dimensional sentence vectors u1, u2, u3, and u4. Subsequently, the low-dimensional sentence vectors are clustered using the DBSCAN clustering algorithm. The resulting clusters are candidate topics 1 through n. Topics are then filtered using the topic selector SELECTOR to obtain the filtered topics. It should be noted that there is no limit on the number of conversation sentences or sentence vectors.

[0156] The information processing method provided in this specification collects a set of pending dialogues associated with the target business, inputs the pending dialogue information in the pending dialogue set into a language model for processing, obtains dialogue features, clusters the dialogue features to obtain a target topic cluster, and determines the topic association information of the target topic cluster based on the business association information of the sub-topics in the target topic cluster, thereby realizing the analysis of the association relationship between the pending dialogue information and the target business. There is no need to mark the pending dialogue information in the pending dialogue set, which reduces resource investment and makes information processing more convenient and flexible.

[0157] Corresponding to the above method embodiment, this specification also provides an information processing device embodiment, Figure 4 FIG1 shows a schematic diagram of the structure of an information processing device provided by an embodiment of this specification. Figure 4 As shown, the device includes:

[0158] A collection module 402 is configured to collect a set of pending conversations associated with a target service;

[0159] The processing module 404 is configured to input the dialogue information to be processed contained in the dialogue set to be processed into a language processing model for processing, obtain dialogue features corresponding to the dialogue information to be processed, and form a dialogue feature set;

[0160] A clustering module 406 is configured to perform clustering processing on the conversation features included in the conversation feature set, and obtain a target topic cluster based on the clustering processing results;

[0161] The determination module 408 is configured to determine the business association information of the subtopics in the target topic cluster, and determine the topic association information of the target topic cluster based on the business association information; wherein the business association information is used to represent the association between the subtopics and the target business.

[0162] In an optional embodiment, the information processing device further includes:

[0163] The division module is configured to splice the i-th to-be-processed dialogue information in the to-be-processed dialogue information set with the j-th to-be-processed dialogue information segment to obtain a target dialogue information segment; determine whether the length of the target dialogue information segment is greater than a preset splicing length; if not, i is incremented by 1, the target dialogue information segment is used as the j-th to-be-processed dialogue information segment, and the step of splicing the i-th to-be-processed dialogue information in the to-be-processed dialogue information set with the j-th to-be-processed dialogue information segment to obtain a target dialogue information segment is performed; if so, the target dialogue information segment is The segment is split to obtain a target length information segment and a segmented sub-information segment with a length equal to a preset splicing length, j is incremented by 1, and the segmented sub-information segment is used as the j-th conversation information segment to be processed, i is incremented by 1, and the step of splicing the i-th conversation information to be processed in the conversation information set to be processed with the j-th conversation information segment to be processed is performed to obtain a target conversation information segment; until i is equal to the number n of conversation information to be processed contained in the conversation information set to be processed, at least one target length information segment is determined, wherein i and j start from 1 and are positive integers.

[0164] In an optional embodiment, the processing module 404 is further configured to:

[0165] At least one target length information segment is sequentially input into the language processing model; each target length information segment is processed by a dialogue information extraction unit in the language processing model to determine the dialogue information to be processed contained in each target length information segment; and each target length information segment is processed by a dialogue feature determination module in the language processing model to obtain dialogue features of each piece of dialogue information to be processed.

[0166] In an optional embodiment, the processing module 404 is further configured to:

[0167] Inputting target dialogue information to be processed into the language processing model; processing the target dialogue information to be processed by a word information extraction unit in the language processing model to obtain at least one word information contained in the target dialogue information to be processed; processing the at least one word information by a vector determination unit in the language processing model to determine a word vector corresponding to each word information as a word vector associated with the target dialogue information to be processed; generating dialogue features of the target dialogue information to be processed based on the word vector associated with the target dialogue information to be processed, and using them as output of the language processing model.

[0168] In an optional embodiment, the processing module 404 is further configured to:

[0169] Mean pooling is performed on the word vectors associated with the target dialogue information to be processed to generate mean features; and based on a preset dimensionality reduction strategy, the mean features are reduced in dimensionality to obtain dialogue features of the target dialogue information to be processed.

[0170] In an optional embodiment, the information processing device further includes:

[0171] The pre-training module is configured to extract sample conversation information contained in the sample conversation set and determine sample word information contained in the sample conversation information; replace the sample word information contained in the sample conversation information in sequence based on a preset replacement strategy to obtain replacement text information; input the replacement text information into the initial language processing model for processing to generate a predicted word information vector; determine a replacement word information vector for the replacement text information, and calculate a loss value based on the replacement word information vector and the predicted word information vector; and adjust parameters of the initial language processing model according to the loss value until the initial language processing model meets the training stop condition, thereby obtaining the language processing model.

[0172] In an optional embodiment, the clustering module 406 is further configured to:

[0173] The method further comprises the steps of: selecting the kth conversation feature in the conversation feature set as the target conversation feature; determining the conversation feature in the conversation feature set that is associated with the target conversation feature based on a preset clustering strategy; storing the conversation feature associated with the target conversation feature in a topic set associated with the kth conversation feature; incrementing k by 1, and executing the step of selecting the kth conversation feature in the conversation feature set as the target conversation feature, wherein k starts at 1 and is a positive integer, and k is less than or equal to the number m of conversation features contained in the conversation feature set; and merging the topic sets containing the same conversation features in the m topic sets to obtain a target topic cluster.

[0174] In an optional embodiment, the clustering module 406 is further configured to:

[0175] Based on a preset clustering strategy, a visual scatter plot is created that includes the conversation features in the conversation feature set; in the visual scatter plot, a circle is drawn with the target conversation feature as the center and a preset length as the radius to obtain a target area; and conversation features located within the target area are selected as conversation features associated with the target conversation feature.

[0176] In an optional embodiment, the information processing device further includes:

[0177] The reference topic cluster determination module is configured to calculate the average value of the topic association information of each target topic cluster when all target topic clusters associated with the to-be-processed conversation set are determined; calculate the deviation rate of each topic association information based on each topic association information and the average value; select the target deviation rate from each deviation rate based on a preset deviation rate selection strategy, and determine the target topic cluster associated with the target deviation rate as the reference topic cluster.

[0178] In an optional embodiment, the reference topic cluster determination module is further configured to:

[0179] Each topic-related information is compared with the average value, and a deviation value of each topic-related information is obtained according to the comparison result; and the absolute value of the deviation value of each topic-related information is calculated to obtain a deviation rate of each topic-related information.

[0180] In an optional embodiment, the reference topic cluster determination module is further configured to:

[0181] Based on a preset deviation rate selection strategy, a deviation rate higher than a preset deviation rate threshold is selected as the target deviation rate; or, based on a preset deviation rate selection strategy, a preset proportion of deviation rates is selected as the target deviation rate in descending order of deviation rate values.

[0182] The information processing device provided in this specification realizes the analysis of the correlation between the pending dialogue information and the target business by executing the steps in the information processing method provided in this specification. There is no need to mark the pending dialogue information in the pending dialogue set, which reduces resource investment and makes information processing more convenient and flexible.

[0183] The above is a schematic diagram of an information processing device according to this embodiment. It should be noted that the technical solution of the information processing device and the technical solution of the above-mentioned information processing method are based on the same concept. For details not described in detail in the technical solution of the information processing device, please refer to the description of the technical solution of the above-mentioned information processing method.

[0184] Figure 5 1 shows a block diagram of a computing device 500 according to an embodiment of the present disclosure. Components of the computing device 500 include, but are not limited to, a memory 510 and a processor 520. The processor 520 is connected to the memory 510 via a bus 530, and a database 550 is used to store data.

[0185] The computing device 500 also includes an access device 540 that enables the computing device 500 to communicate via one or more networks 560. Examples of such networks include a public switched telephone network (PSTN), a local area network (LAN), a wide area network (WAN), a personal area network (PAN), or a combination of communication networks such as the Internet. The access device 540 may include one or more of any type of network interface (e.g., a network interface card (NIC)), whether wired or wireless, such as an IEEE 802.11 wireless local area network (WLAN) wireless interface, a Worldwide Interoperability for Microwave Access (Wi-MAX) interface, an Ethernet interface, a universal serial bus (USB) interface, a cellular network interface, a Bluetooth interface, a near field communication (NFC) interface, and the like.

[0186] In one embodiment of the present specification, the above components of the computing device 500 and Figure 5 Other components not shown in the figure may also be connected to each other, for example, via a bus. Figure 5 The computing device structure block diagram shown is for illustrative purposes only and is not intended to limit the scope of this specification. Those skilled in the art may add or replace other components as needed.

[0187] Computing device 500 can be any type of stationary or mobile computing device, including a mobile computer or mobile computing device (e.g., a tablet computer, a personal digital assistant, a laptop computer, a notebook computer, a netbook computer, etc.), a mobile phone (e.g., a smartphone), a wearable computing device (e.g., a smartwatch, smart glasses, etc.), or other types of mobile devices, or a stationary computing device such as a desktop computer or PC. Computing device 500 can also be a mobile or stationary server.

[0188] The processor 520 is configured to execute the following computer-executable instructions:

[0189] Collect the pending dialogue set related to the target business;

[0190] Inputting the dialogue information to be processed contained in the dialogue set to be processed into a language processing model for processing, obtaining dialogue features corresponding to the dialogue information to be processed, and forming a dialogue feature set;

[0191] performing clustering processing on the conversation features included in the conversation feature set, and obtaining a target topic cluster based on the clustering processing result;

[0192] Determine business association information of subtopics in the target topic cluster, and determine topic association information of the target topic cluster based on the business association information; wherein the business association information is used to represent the degree of association between the subtopics and the target business.

[0193] The above is a schematic scheme of a computing device of this embodiment. It should be noted that the technical scheme of the computing device and the technical scheme of the above-mentioned information processing method are based on the same concept. For details not described in detail in the technical scheme of the computing device, please refer to the description of the technical scheme of the above-mentioned information processing method.

[0194] An embodiment of the present specification further provides a computer-readable storage medium storing computer instructions, which, when executed by a processor, are used to:

[0195] Collect the pending dialogue set related to the target business;

[0196] Inputting the dialogue information to be processed contained in the dialogue set to be processed into a language processing model for processing, obtaining dialogue features corresponding to the dialogue information to be processed, and forming a dialogue feature set;

[0197] performing clustering processing on the conversation features included in the conversation feature set, and obtaining a target topic cluster based on the clustering processing result;

[0198] Determine business association information of subtopics in the target topic cluster, and determine topic association information of the target topic cluster based on the business association information; wherein the business association information is used to represent the degree of association between the subtopics and the target business.

[0199] The above is a schematic diagram of a computer-readable storage medium according to this embodiment. It should be noted that the technical solution of the storage medium and the technical solution of the information processing method described above are based on the same concept. For details not described in detail in the technical solution of the storage medium, please refer to the description of the technical solution of the information processing method described above.

[0200] The foregoing description of this specification describes specific embodiments. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims can be performed in an order different from that described in the embodiments and still achieve the desired results. Furthermore, the processes depicted in the accompanying drawings do not necessarily require the specific order shown or the sequential order to achieve the desired results. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0201] The computer instructions include computer program code, which may be in source code form, object code form, executable file, or some intermediate form. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signal, telecommunication signal, and software distribution medium, etc. It should be noted that the content contained in the computer-readable medium may be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electric carrier signals and telecommunication signals.

[0202] It should be noted that for the aforementioned method embodiments, for ease of description, they are all expressed as a series of action combinations. However, those skilled in the art should be aware that this specification is not limited to the order of the actions described, because according to this specification, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in this specification are all preferred embodiments, and the actions and modules involved are not necessarily required by this specification.

[0203] In the above embodiments, the description of each embodiment has its own focus. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0204] The preferred embodiments disclosed above are intended only to help illustrate this specification. The optional embodiments do not exhaustively describe all details, nor do they limit the invention to the specific embodiments described. Obviously, many modifications and variations are possible based on the content of this specification. These embodiments are selected and described in detail in this specification to better explain the principles and practical applications of this specification, thereby enabling those skilled in the art to better understand and utilize this specification. This specification is limited only by the claims and their full scope and equivalents.

Claims

1. An information processing method, characterized in that: include: Collect the pending dialogue set related to the target business; Inputting the dialogue information to be processed contained in the dialogue set to be processed into a language processing model for processing, obtaining dialogue features corresponding to the dialogue information to be processed, and forming a dialogue feature set; performing clustering processing on the conversation features included in the conversation feature set, and obtaining a target topic cluster based on the clustering processing result; Determining business association information of subtopics in the target topic cluster, and determining topic association information of the target topic cluster based on the business association information; wherein the business association information is used to characterize the degree of association between the subtopics and the target business; When all target topic clusters associated with the to-be-processed conversation set are determined, an average value of topic association information of each target topic cluster is calculated; each topic association information is compared with the average value, and a deviation value of each topic association information is obtained based on the comparison result; and an absolute value of the deviation value of each topic association information is calculated to obtain a deviation rate of each topic association information; Based on a preset deviation rate selection strategy, a deviation rate that is higher than a preset deviation rate threshold is selected as a target deviation rate; or, based on a preset deviation rate selection strategy, a preset proportion of deviation rates is selected as a target deviation rate in descending order of deviation rate values; The target topic cluster associated with the target deviation rate is determined as a reference topic cluster.

2. The method according to claim 1, characterized in that Before inputting the dialogue information to be processed contained in the dialogue set to be processed into the language processing model for processing and obtaining dialogue features corresponding to the dialogue information to be processed, the method further includes: splicing the i-th to-be-processed conversation information in the to-be-processed conversation information set with the j-th to-be-processed conversation information segment to obtain a target conversation information segment; Determining whether the length of the target conversation information segment is greater than a preset splicing length; If not, i is incremented by 1, the target conversation information segment is used as the j-th conversation information segment to be processed, and the step of concatenating the i-th conversation information to be processed and the j-th conversation information segment to be processed in the conversation information set to be processed is performed to obtain the target conversation information segment; If so, the target conversation information segment is segmented to obtain a target length information segment and a segmented sub-information segment whose length is equal to the preset splicing length, j is incremented by 1, the segmented sub-information segment is used as the j-th conversation information segment to be processed, i is incremented by 1, and the step of splicing the i-th conversation information to be processed in the conversation information set to be processed with the j-th conversation information segment to be processed is performed to obtain the target conversation information segment; Until i is equal to the number n of dialogue information to be processed contained in the dialogue information set to be processed, at least one information segment of target length is determined, wherein i and j start from 1 and are positive integers.

3. The method according to claim 2, characterized in that The step of inputting the dialogue information to be processed contained in the dialogue set to be processed into a language processing model for processing to obtain dialogue features corresponding to the dialogue information to be processed includes: inputting at least one target length information segment into the language processing model in sequence; Processing each target length information segment by a dialogue information extraction unit in the language processing model to determine the dialogue information to be processed contained in each target length information segment; The dialogue information to be processed contained in each target length information segment is processed by the dialogue feature determination module in the language processing model to obtain the dialogue features of each dialogue information to be processed.

4. The method according to claim 1, wherein The process of processing any one of the pending dialogue information in the pending dialogue set by the language processing model includes: Inputting the target dialogue information to be processed into the language processing model; Processing the target dialogue information to be processed by a word information extraction unit in the language processing model to obtain at least one word information contained in the target dialogue information to be processed; Processing the at least one word information by a vector determination unit in the language processing model to determine a word vector corresponding to each word information as a word vector associated with the target dialogue information to be processed; According to the word vector associated with the target dialogue information to be processed, a dialogue feature of the target dialogue information to be processed is generated and used as the output of the language processing model.

5. The method according to claim 4, characterized in that Generating the dialogue features of the target dialogue information to be processed based on the word vector associated with the target dialogue information to be processed includes: Performing mean pooling on the word vectors associated with the target dialogue information to be processed to generate mean features; Based on a preset dimensionality reduction strategy, the mean feature is reduced in dimension to obtain the dialogue feature of the target dialogue information to be processed.

6. The method according to claim 1, characterized in that Before inputting the dialogue information to be processed contained in the dialogue set to be processed into the language processing model for processing and obtaining dialogue features corresponding to the dialogue information to be processed, the method further includes: Extracting sample conversation information contained in the sample conversation set to determine sample word information contained in the sample conversation information; Based on a preset replacement strategy, the sample word information contained in the sample conversation information is replaced in sequence to obtain replacement text information; Inputting the replacement text information into an initial language processing model for processing to generate a predicted word information vector; Determining a replacement word information vector for replacing the text information, and calculating a loss value based on the replacement word information vector and the predicted word information vector; Parameters of the initial language processing model are adjusted according to the loss value until the initial language processing model meets a training stop condition, thereby obtaining the language processing model.

7. The method according to claim 1, characterized in that The clustering process is performed on the conversation features included in the conversation feature set, and a target topic cluster is obtained according to the clustering process result, including: Selecting the kth dialogue feature in the dialogue feature set as the target dialogue feature; Determining, based on a preset clustering strategy, a conversation feature in the conversation feature set that is associated with the target conversation feature; Storing the conversation features associated with the target conversation features into a topic set associated with the kth conversation feature; k is incremented by 1, and the step of selecting the kth conversation feature in the conversation feature set as the target conversation feature is performed, wherein k starts at 1 and is a positive integer, and k is less than or equal to the number m of conversation features included in the conversation feature set; When m topic sets are obtained, the topic sets containing the same conversation features in the m topic sets are merged to obtain the target topic cluster.

8. The method according to claim 7, characterized in that The determining, based on a preset clustering strategy, a conversation feature in the conversation feature set that is associated with the target conversation feature includes: Creating a visual scatter plot including the conversation features in the conversation feature set based on a preset clustering strategy; In the visual scatter plot, a circle is drawn with the target conversation feature as the center and a preset length as the radius to obtain a target area; A conversation feature located in the target area is selected as a conversation feature associated with the target conversation feature.

9. An information processing device, characterized in that include: A collection module configured to collect a set of pending conversations associated with a target business; a processing module configured to input the unprocessed dialogue information contained in the unprocessed dialogue set into a language processing model for processing, obtain dialogue features corresponding to the unprocessed dialogue information, and form a dialogue feature set; a clustering module configured to perform clustering processing on the conversation features included in the conversation feature set, and obtain a target topic cluster based on the clustering processing results; A determination module is configured to determine the business association information of the sub-topics in the target topic cluster, and determine the topic association information of the target topic cluster based on the business association information; wherein the business association information is used to characterize the degree of association between the sub-topic and the target business; when all target topic clusters associated with the to-be-processed conversation set are determined, calculate the average value of the topic association information of each target topic cluster; compare each topic association information with the average value, and obtain the deviation value of each topic association information according to the comparison result; calculate the absolute value of the deviation value of each topic association information to obtain the deviation rate of each topic association information; based on a preset deviation rate selection strategy, select a deviation rate higher than a preset deviation rate threshold as the target deviation rate; or, based on a preset deviation rate selection strategy, select a preset proportion of deviation rates as the target deviation rate in descending order of deviation rate values; The target topic cluster associated with the target deviation rate is determined as a reference topic cluster.

10. A computing device, characterized in that It comprises a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the steps of the information processing method according to any one of claims 1 to 8.

11. A computer-readable storage medium storing computer instructions, characterized in that: When the instruction is executed by a processor, the steps of the information processing method according to any one of claims 1 to 8 are implemented.

Images